<![CDATA[Simply Boring AI]]>https://www.simplyboring.aihttps://www.simplyboring.ai/img/substack.pngSimply Boring AIhttps://www.simplyboring.aiSubstackWed, 08 Jul 2026 09:31:33 GMT<![CDATA[Pruning AI]]>https://www.simplyboring.ai/p/pruning-aihttps://www.simplyboring.ai/p/pruning-aiTue, 07 Jul 2026 12:20:49 GMT

When I was young, I watched my uncle work on his bonsai. Some branches he cut off. Others he wrapped in wire and bent, slowly, until they grew the way he wanted.

Two conversations in the past week have made me think about the connection between his craft and Agentic AI governance.

The first. A comment left by Georg Zoeller on my post. On how constraining agentic AI with rule-based guardrails would not work. His observation was sharp, and justified. If we can pin an agent down with rules, why use AI at all? Why not just write the rule?

My reply was that all models (not only AI) are just mapping functions. Inputs to outputs. Rules too.

When we know what we want, we should certainly use a rule to do the mapping. When we do not know, and we want the machine to learn the pattern that maps input to output, we use AI. The role of rule-based guardrails, then, is not to replace the AI with a rule. It is to reduce the space of patterns AI is allowed to traverse. They prune off the patterns we do not want, and constrain the ones we do.

The “hope” is that the envelope of possible behaviour left after such pruning can be trusted.

The second conversation was during a lunch just now with Lukasz Szpruch and some ex-MAS colleagues. The question was about how far rules could go. Can we realistically have rules for everything in an agentic system? What is the actual role of an LLM as a judge? And where does old-fashioned evaluation and testing fit?

It left me thinking. So let me jot down my quick thoughts.

The tree

Picture an agentic system as a tree. Unlimited branches, each a path the agent might travel. If you let it be, it gets quite unruly.

Not all paths are equal.

In terms of impact - some are truly bad - the ones that cause real, irreversible harm. Some are potentially bad, but fine if controlled. Many are simply not great - low stakes, reversible, the kind you can afford to get wrong now and then. And some are exactly what we want.

In terms of practical controls - some actions can truly be constrained with rules, e.g., a recency check, a sources whitelist. Some may require great flexibility, e.g., filters of open-ended advice or answers to questions. And many in-between.

Take this example - an agent that resolves customer billing disputes. It reads the ticket, looks up the account, decides whether a refund is owed, issues it, replies to the customer, and closes the case.

Even that small workflow branches enormously.

Some branches are truly bad: refunding to an account that is not the customer’s, issuing a refund far larger than the charge, exporting the full transaction history, switching off the audit log.

Some are potentially bad but manageable: a large refund is fine, if a human signs off first.

Many are low stakes and reversible: the wording of the reply, which help article to cite, whether to offer a small goodwill credit.

And one narrow set is what we actually want: the dispute resolved within policy, logged, the customer treated fairly.

This is what changes when we move from a normal model to an agent. A normal model gives you an output. One thing to check. An agent chooses an action, then another, then another - and each choice opens more. The paths branch faster than anyone can list them. That explosion is the risk. Not a bad answer, but a bad route to a fine-looking one.

Better to fail loudly, than silently.

Cut, and wire

But you cannot check every path.

So I’m thinking if we can do what my uncle did. You cut, and you wire.

Cut - You cut the paths that must never happen. The truly bad ones. Delete the database. Move money with no sign-off. Send the customer list outside. You do not manage these, or monitor them, or hope. You remove the branch, so the agent can never walk down it. Zero trust.

Everything that is left, there could be three kinds of wire.

Rigid wire - rules. Stiff, certain, holds its shape on its own. Where a path can be constrained by a plain rule - refunds under a set amount only, always to the original card, stop and ask above a threshold - use it. A rule adds no new paths of its own. It is the strongest wire you can use, and the cheapest to trust. Implement as many of them as you can find.

Supple wire - a model as judge. Some paths are too fuzzy for a rule. Whether a complaint is genuine. Whether a reply is rude. There you let an LLM, or a smaller model (ML or SLM), watch and bend the branch. It is far more flexible. But this wire has a mind of its own. The judge is itself a model, with its own paths, its own ways to be wrong. You gain judgement, and you grow a second tree beside the first. So use it deliberately, and govern it as carefully as the thing it guards. Still better than leaving it to a system prompt and a prayer.

Your own hands - evaluation and testing. Some paths that you cannot trust to either wire. Those you bend yourself. You sit with the agent, predict where it will go, probe the branches and edges, and correct by hand. It is the most flexible control there is, and the least scalable - you only have two hands. It is also how you learn where the other two wires need to go. Every test that surprises you is a new rule waiting to be written, or a judge waiting to be posted.

Which branch gets which wire

Three wires. Different trade-offs.

Rules are certain but rigid.

Judges are flexible but grow their own paths.

Your hands working on evaluation and testing are wisest but do not scale.

And the order should perhaps be this.

Push as much as you can onto rules, because they are certain and add nothing. What rules cannot hold, wire with model judges, and accept that you have added paths in exchange for flexibility. What neither can hold, keep in your own hands, and let your evaluation and testing tell you what to promote to a rule next season.

Every wire is a way of narrowing the envelope of patterns the agent can express. A rule narrows it hard. A model judge narrows it softly. Your hands narrow it one case at a time, and you get better at it with practice. None of them turn the AI back into a rule. They only decide how much room it gets.

None of it kills the tree. My uncle’s bonsai lived for decades.

Is it so for Agentic AI? Not sure. But this is how I am starting to think about it.

#AgenticAI #AIRiskManagement #AIGovernance #AISafety

]]><![CDATA[Gradient Descent, Again]]>https://www.simplyboring.ai/p/gradient-descent-againhttps://www.simplyboring.ai/p/gradient-descent-againMon, 06 Jul 2026 06:38:09 GMT

I wrote about gradient descent three months ago on LinkedIn. I was figuring out how to wander the uneven terrain of independent life. The failures. The plans that drifted. Finding the right angle to navigate.

I’ve learned to be comfortable with the uncertainties and disappointments of this new phase. Even relaxed enough to pick up the brush again.

But somehow I met gradient descent again. This time when trying to learn to learn art again.

A light week of meetings. A bunch of regulators from around the world I am mentoring for a course on AI in financial services. A collaborator on the Agentic MRM paper who was in town. A consultant thinking of moving to a senior AI role in a bank, asking for my views. An ex-colleague from my strategic planning days, now in communications. A senior risk manager between jobs, developing scented cleaning supplies as a hustle.

Lots of building otherwise. And art. I’ve built up the discipline of doing watercolors every night before bed. Which started another round of gradient descent. And a bit of angst.

The Gap

A retrospective I wrote last week made the art sound resolved. An interest that could be an uncorrelated asset in this world of exhausting AI hype. (Side-note: I’m starting to think I have a love-hate relationship with AI. I’ve always loved the discipline, but it has increasingly become populated by hypesters and scammers that rile me.)

I thought I had hit the ground running with my watercolors. The practice had returned. A satisfying piece here. A passable piece there. Enough momentum to feel like my old friend was back. I thought the next five pieces would be even better.

The next five pieces across the week made me realize it had not. Almost none of them worked. My old friend had not forgiven the gap.

Gradient descent is how a model learns. It makes a prediction. Measures the gap between the prediction and the right answer. Then adjusts in the direction that reduces the gap. Then repeats.

It’s called gradient descent because of how it finds that direction. Imagine a hilly landscape. You’re standing somewhere on it, blindfolded. You want to get to the lowest point - the valley. You can’t see the whole landscape. But you can feel the slope beneath your feet. Gradient descent says - follow the slope downhill. One small step at a time. Measure the ground. Adjust. Step again. Eventually you reach the bottom. Or at least somewhere lower and better than where you started.

The gradient is the slope. The descent is following it.

Three months ago I wrote about gradient descent in independent life. Misses. Proposals that went nowhere. Plans that drifted. All of it telling me the slope and which way to go next. Lots of uncertainty. But I got reasonably comfortable with that. Not because it got easier. Just because I could detach from the results. And go with the flow.

Gradient descent in learning to learn art seems different.

Somehow, I find it harder to detach from the gap. The gap between what I see in my head and that muddy mess that appeared on paper in front of me.

And there’s the obsession and ambition problem. When you restart a practice you start looking at the masters. Chung Chien Wei and his effortless wet-on-wet scenes. Thomas Schaller’s architectural washes, luminous and controlled. Alvaro Castagnet’s explosive loose strokes that somehow resolve into something complete.

You know what good looks like. You can feel the gap between that and what’s on your desk. And the gap isn’t abstract. It’s a piece you spent an hour on that looks nothing like what you intended.

One piece almost hit the spot. For a moment something was clicking. Then I lost it. The next piece drifted back toward the same problems as before.

Getting close and losing it is worse than consistent misses. When nothing works you can tell yourself you’re still finding the slope. When something almost works you know exactly what you’re reaching for. But then it becomes elusive again.

The Way Forward

The difference between knowing gradient descent and feeling it is considerable. I wrote about it in the context of independent life.

What I didn’t write about is what it actually feels like to sit with a miss. To look at something that isn’t working and try to stay patient. To not force it. To trust that the signal is in there somewhere even when you can’t see it yet. And to let go if it really just needs to go.

Three months ago the gradient descent was in independent life. Now it’s in a practice. More immediate. More personal. More uncomfortable in a different way. But not very different.

And I think the way forward is simple.

Pick up the brush again tomorrow. That’s all gradient descent really requires.

#GradientDescent #Art #IndependentLife #AIRiskManagement #Transitions #Reflections

]]><![CDATA[Convergence & Divergence]]>https://www.simplyboring.ai/p/convergence-and-divergencehttps://www.simplyboring.ai/p/convergence-and-divergenceSun, 28 Jun 2026 08:20:26 GMT

Don’t force it. Let things flow.

I had scribbled these words on a post-it. I was two months from leaving MAS. I think it was a salve for my anxiety.

For most of the six months since I left, I don’t think I listened to my own note. I forced plenty. Endless coffees, lunches, talks, plans.

Recently, I’ve started to get back to these six words. And last week, for the first time in more than a year, I managed to dive deep into watercolours again.

A good friend at work had once remarked that seeing me do art was a good indication of how bored I was at work. I’ve come to the realisation that it’s a bit more than that.

And since this is the six-month mark, it’s a good time for reflections on some personal lessons. And how my view of AI has shifted. Which could perhaps be useful for some others.

My lessons

The first, and the main one. Run experiments and reflect on them.

I left MAS after eighteen years and walked into a void. I did not search for another job before leaving. Even the plan I did write down at the end of 2025 never happened.

In the end, the last 6 months became a series of experiments.

More than two hundred conversations across twenty-five weeks. At least two to three talks or panels every month. Old friends and complete strangers. CROs and PhD students. An anthropologist studying money, and a CEO I mistook for a fellow trainer. Collaborations across the world.

Every week I reflected on what I learned from these conversations. Some of it was connected to AI. Some of it was just about life. People ask how I find the connections between these conversations. I read The Book of Ichigo Ichie recently. It says “notice coincidences” as one of the ten rules of ichigo ichie. I guess this is what it means.

In the age of AI, I think this is how learning looks like.

Two, the filter beats the plan.

My original plan was 80% wrong. My filter - interesting work, people I like, freedom, the work done properly - was almost always right. A plan imposes a conclusion and makes you bend everything to fit it. Even if it feels wrong. A filter just tells you what to walk away from.

Build a filter early, even before you can fully articulate it.

Between a plan and a filter, I’ll take the filter every time. They’ll help you avoid some early missteps.

Three. You have to actually cross.

Not observe. Cross. Watching other worlds from a safe distance is not enough. The value is in gate-crashing them, awkwardly if one must. The insurance world, where I mistook a CEO for a fellow trainer. Legal AI. Digital assets. An MBA classroom on the other side of the world. US AI governance. African bank and insurance boards. Healthcare workers learning deep learning. Wealth managers and family offices. A central-banking podcast. And many more.

I am an introvert who loses energy when I meet too many people. A side-effect of these crossings - I’m feeling at ease now in crowds. A simple rule is all it takes. If I feel like chatting with someone, I do it. If not, I walk away.

Four. Flow is just gradient descent.

Misses are information. The proposals that die, the rooms that don’t want your lines, the conversations that trail off - they feel like failures, but they point you in a better direction.

In gradient descent a wrong step isn’t failure; it’s the signal that tells you the slope, and which way to step next. You cannot know which direction is right without the wrong ones. Expect a lot of them early, and don’t mistake them for verdicts.

Five. Guard the freedom you left for.

If one leaves for some kind of freedom, protect it like the scarce thing it is.

That means cleaner no’s - to the gig that pays but owns your calendar, to the room that drains you, to the prestige that quietly puts you back in a box.

It means letting no single party own you, and having no single point of failure. And it means keeping at least one thing in your life that owes nothing to the work.

If the freedom is the whole point, don’t trade it back the first time saying no feels awkward.

How my view of AI shifted

I started by seeing the unevenness.

AI was in every single conversation, but understanding was an uneven valley - someone shipping AI products daily, someone else whose work was untouched, same city, same week. I stopped seeing one AI race and started seeing dozens, on different tracks, most runners unable to see each other.

Awareness had stopped being the solution and become the problem, setting expectations nobody could square with reality.

While some folks were going from prompt and context engineering to harness and loop engineering, more sane folks were wondering how an LLM that made a silly hallucination one in 20 times could be trusted to perform an action that mattered.

The boring fundamentals matter more than the hype.

When sixteen agents built a working C compiler, what made it work wasn’t the frontier. It was the boring list - tests, documentation, monitoring - the same list from the risk guidelines I wrote.

Real understanding compounds and builds a flywheel. Prompt engineering resets with every model update. AI courses that just teach one to prompt and pray are not building AI literacy but just scams in disguise.

The scaffolding is important but thinning.

The same week a few of us published a paper extending model risk discipline into the agentic world, SR 11-7 - the standard I’d been referring to for more than a decade - was quietly retired.

The real rules seem to be thinning out just as the systems get more complex. We have plenty of publications on responsible AI. But tell me, how many bite and are effective?

So I guess we just have to fall back on first principles, not frameworks. And I learned to distrust fluent reasoning, in models and in myself.

Coherence is not correctness.

And I ended more sceptical than I began.

Not of the technology - of the expectations around it. I think they are overblown. Whether AI becomes as common as electricity or quietly stalls, an identity built entirely on it is fragile either way. That is what sent me back to water and paper.

What I do now

The shape that emerged is not the one I planned.

Training for a few associations and institutes. A few adjunct roles. Some writing gigs. A startup I’m building with a collaborator. A few advisory gigs on the horizon.

No single party owns me. No exclusivity. If one fails, no big deal. I can still survive.

Before I start any gigs these days, I know to zoom in on the non-solicitation, non-competes, indemnity and intellectual property clauses. These are the ones that can matter more than the money.

I’ve stopped saying yes quickly. The no’s have gotten cleaner. I show up to speak when I have something to say, and otherwise I’d rather listen.

And I’ve started hedging the one risk I couldn’t diversify away. Everything I do sits in the same asset class - AI. So the uncorrelated asset is the art. By hand. Slow. On paper. No prompt. No FOMO. Just observation and whatever is left of Singapore before it disappears.

It may not take off. The world of art is brutal. But it’s irreducibly human.

Gradient descent has one catch. You follow the slope down, you correct, you flow, and you arrive somewhere that looks exactly right. But it might only be a local minimum. A valley within a valley. A few more steps might have taken you somewhere better, and you’d never know.

So I won’t pretend six months converged on an answer. It’s just a start.

The twenty-four weeks

Every Monday for six months, I tied that week to an AI concept. I stopped numbering them around week twenty - but here they are, in order. Start anywhere.

(Each links to the original LinkedIn post.)

Week 1 · 12 Jan 2026 — Valleys, Shifting Sands & Pillars in AI
AI everywhere, understanding nowhere near even. The uneven landscape.

Week 2 · 19 Jan 2026 — The Darn Messy Middle of AI
When awareness stops being the solution and becomes the problem.

Week 3 · 26 Jan 2026 — Phase Shifts in AI and Life
Not pivots. Changes of state. Why I’m optimistic about AI.

Week 4 · 2 Feb 2026 — Resonance - AI at Work and in the Classroom
Becoming a LinkedIn Top Voice the same week I rediscovered my voice.

Week 5 · 9 Feb 2026 — The Flywheel
Sixteen agents, one C compiler, and the boring fundamentals that made it work.

Week 6 · 16 Feb 2026 — Many Different Worlds
There isn’t one AI race. There are dozens, and most runners can’t see each other.

Week 7 · 23 Feb 2026 — The Outsider
Transfer learning, and why nothing we learn is wasted.

Week 8 · 2 Mar 2026 — Alignment
The boat-race agent that spun in circles, and freedom as non-negotiable.

Week 9 · 9 Mar 2026 — Context
The context window, finite attention, and learning to filter.

Week 10 · 16 Mar 2026 — The Edge of Action
Three gaps, ten weeks of evidence, and what I finally decided to build.

Week 11 · 23 Mar 2026 — Grounding
The symbol grounding problem, and meeting rooms where they actually are.

Week 12 · 30 Mar 2026 — Diffusion
Graph diffusion, signal versus noise, and the cost of saying yes too much.

Week 13 · 6 Apr 2026 — Gradient Descent
Why the misses weren’t failures. They were the gradient.

Week 14 · 13 Apr 2026 — Unfinished
No prepared deck survived the room. And fallow isn’t drought.

Week 15 · 20 Apr 2026 — Irony
SR 11-7 retired the same week we argued for more discipline, not less.

Week 16 · 27 Apr 2026 — Thinking Out Loud
Chain-of-thought can be confident, fluent bullshit. So can I.

Week 17 · 4 May 2026 — Serendipity and Emergence
Rojak, introversion, and building the conditions for emergence.

Week 18 · 11 May 2026 — Strange Attractors Again
Chaos isn’t random. It’s shaped by topology.

Week 19 · 18 May 2026 — Impermanence and Drift
Concept drift, disappearing spaces, and a plan turned unrecognisable.

Week 20 · 25 May 2026 — The Crossings
The value isn’t in observing the worlds. It’s in actually crossing.

Week 21 · 1 Jun 2026 — Sum of the Parts
Skill-stacking, freedom of choice, and a hard question about privilege.

Week 22 · 8 Jun 2026 — The Dream and the Plan
Breaking from the Plan is the beginning, not the answer.

Week 23 · 15 Jun 2026 — 5 Useless Things
Useless to box, to force, to perform, to accommodate, to follow convention.

Week 24 · 22 Jun 2026 — The Riskiest Play
Concentration risk, an uncorrelated asset, and picking up the brush again.

I write a lot on both LinkedIn at https://www.linkedin.com/in/garyang/ and also Simply Boring AI, my Substack at simplyboring.ai. Sometimes I cross post. Sometimes only on one. I’m still figuring out works best. So follow me on both.

And if you’d like to work together - workshops, teaching, advisory - I’m at quaintitative.com.

#IndependentLife #AIRiskManagement #Transitions #Reflections #GradientDescent

]]><![CDATA[Units of AI Governance]]>https://www.simplyboring.ai/p/units-of-ai-governancehttps://www.simplyboring.ai/p/units-of-ai-governanceFri, 26 Jun 2026 03:23:56 GMTFor the past few months I have been giving talks at the meeting point of two disciplines I know well. Model risk management, from my years supervising models at MAS. And computer science, from an ill-advised side quest to finish a PhD at 42.

The AI risk management guidelines I wrote for Singapore’s financial sector grew out of both, and they rest on three units of governance: the model, the system, and the use case. Lately something seems missing.

These three units - the model, the system, and the use case - are the ones we already know how to govern, but they share a weakness. Each is specific, tied to one particular application.

Above them sits a second set of three that we are beginning to think about for governance: the archetype, the capability, and the workflow. These seem to be better as they generalize.

Why is this important? A unit of governance is something you point a control at. Too granular and it may work, but not scale.

Let’s draw a line between 6 units. The units below the line are where governance lives today. The units above it may be worth considering for governing AI at scale.

The six of them, and the line between, look like this.

Below the line

The three units below the line are not new. We imported each of them from a discipline that already existed - model risk management, technology risk management, and the use-case registers that AI governance frameworks already lean on. That is their strength and their limit. The methods are mature. But they were built for a world where each thing you governed was specific and countable.

The model

I’ll start where I always do, with the model. It is my favourite, and it is the one thing that makes an AI system different from any other piece of software. The first models I met were simple curves - bootstrapping zero rates off swap quotes in a financial engineering classroom, then their more sophisticated cousins, Black-Scholes, SABR, Hull-White, Heston.

For years a single supervisory letter, SR 11-7, was the only guidance that governed any of it. I took a Masters in Financial Engineering but was not a great student, and so I stayed confused about those models in finance for a long time.

What finally made them click was meeting their successors - decision trees, random forests, gradient boosted trees, then the deep learning models, the RNNs and CNNs, and eventually LLMs. They are all more general versions of the same idea. Even an agent is just a system wrapped around an LLM.

Once you can see the parts every model shares - the function you choose, the objective, the optimization method, the way you evaluate it - the questions you need to ask barely change from one to the next. They all come back to those parts, and to the assumptions beneath them.

But something did change on the way from the curve to the LLM, and it matters for governance. A Black-Scholes model is written down. You can read its assumptions off the page and argue with them. A learned model is not written down in that sense - its behaviour lives in weights shaped by data, not in an equation you can inspect. So validation has to change. It stops being a check of the maths and becomes a probe of the behaviour. You can no longer read the model; you can only interrogate it.

That is the real break, and it is why so much traditional model risk practice feels half-applicable to AI. The questions survive. How you answer them does not. And we see this in Agentic AI increasingly due to their ability to take actions and not just generate outputs.

But that shared structure is also where we see commonalities in risks. I think of them as three U’s.

  • Uncertainty - the irreducible randomness you can never remove, and the reducible gaps you can close with more data. Every model has both. The work is knowing which is which, because you waste effort trying to engineer away the part that won’t move.

  • Unexpectedness - the way more complex models behave in ways nobody designed, in emergent capabilities, adversarial weaknesses, and misalignment. The more general the model, the more of its behaviour was never specified by anyone, which means more of it is found rather than built.

  • Unexplainability - the varying degree to which we can actually say why a model decided what it did. Transparency, explainability, interpretability are all really one question asked at different depths: can we understand this enough to stand behind it? Understand. Not quote some figures generated by an explainability method you don’t understand.

Governing the model, then, is mostly documentation, independent validation, and benchmarking - the same spine as classical model risk. The trap is treating the benchmark as the answer. A model can top every public benchmark and still fail in your context, on your data, against your adversaries, because the benchmark measured a general skill and you are deploying a specific one. The benchmark tells you the model is capable. It does not tell you it is safe here.

It is telling that when the Fed, the OCC and the FDIC replaced SR 11-7 this April with SR 26-2, after fifteen years, they carved generative and agentic models out of scope as too novel and fast-moving. I understand the caution. But the effect is that the fastest-moving, least-understood models are the ones left with the least guidance. Given how tightly the quantitative and the modern connect, that exclusion still puzzles me.

The system

A model, though, never runs on its own. Wrap code, tools and memory around it and you have a system - and for most of my career the system was someone else’s problem. I came up in model risk: development documentation, validation reports, stress tests. Technology risk lived down the corridor, a different team with different acronyms, and I treated it as none of my business. AI changed that.

To see why, it helps to name what a system actually is. Around the model sits the prompt scaffolding that frames every request, the tools it is allowed to call, the memory it carries between turns, the retrieval that decides what context it sees, and the guardrails bolted on at the edges. Every one of those is a knob. Turn any of them and the same model produces different behaviour. The same prompt, through two different stacks, lands in two different embedding spaces and comes back with two different answers. So a model card tells you very little on its own. Two deployments of the same weights are two different risk objects, and the thing you actually have to govern is the configuration, not the checkpoint. We call them fancy names like harness or loop engineering these days for Agentic AI. But it’s really just a system with slightly more dynamism and autonomy.

This is where most so-called AI incidents really live. The system around the model is increasingly where the performance, the variance, and most of the failure modes actually live, and it cuts both ways. On the upside, changing nothing but the harness around a model can lift its score on the same benchmark dramatically, sometimes close to double - the model didn’t get smarter, the system around it got better at using it. On the downside, the system is where the incidents are landing - remote code execution in the Model Context Protocol, supply-chain compromises spreading through AI tooling, an over-permissioned tool, an unsanitized input, a secret written to a log. In almost none of them did the model fail. The thing around it did. We spent a great deal of worry on model monoculture. The system deserves more of that attention.

The use case

If the model and the system are how AI gets built, the use case is what it gets built for - and it is the layer almost every governance framework reaches for first. It is also the question I am always asked: what are the use cases? I find it frustrating because no one wants to understand the underlying mechanics and connect it to their tasks, but just want an AI use case served on a platter.

The pull is easy to understand. The use case is the unit the business names, the unit that goes into a register, the unit a board can hold in its head. It is the unit of value. So it becomes the unit of the inventory, and the inventory becomes the comfortable thing we can look at and boast about - “We have 1000 AI use cases!”

That comfort is the problem. A register of use cases gives you the feeling of coverage while the real risk objects sit a layer down, uncounted. It is the unit I trust least, for three reasons.

Not deep enough. A single fraud detection use case can hide transaction monitoring, an anomaly model, a generative assistant and a retrieval system underneath, each with its own failure modes. The use case enters the inventory once. Ten real things go ungoverned.

Not specific enough. Two banks can list the same wealth-advisory use case and mean completely different workflows, models, and risks. The label matches. Nothing beneath it does. So a control written against the label protects neither of them well.

Not scalable enough. Twenty use cases you can govern one by one. At two hundred the patterns underneath them simply repeat - retrieval, extraction, summarization, classification, recommendation, anomaly detection - and governing at the use-case level means redoing the same work every time you meet the same pattern wearing a new name.

The use case is the right unit to decide whether to build. It is the wrong unit to decide how to control. The unit that actually scales is one layer further down, or one layer up.

Above the line

Everything below the line is specific by nature, which is why the work never compounds - each new system is a fresh job. The units above the line are different in kind. They are patterns, not instances. Govern a pattern once and every instance that fits it inherits the work. That is the whole reason to climb.

The archetype

Go up a layer and you reach the archetype - the first unit that scales. I saw it during the generative-AI wave, before agents arrived: an LLM is general-purpose, but people kept getting more out of it by putting it in a box. Retrieval-augmented generation. Summarization. Classification. Information extraction. Code generation. Chat.

A general-purpose tool is almost ungovernable precisely because it can do anything - there is no fixed set of risks to reason about. The archetype trades a little of that generality for something you can actually hold: a known shape, with known failure modes.

Naming the box helps twice over. For building, it brings the recipe with it - the reference architecture, the prompt patterns, the libraries and settings that already work - so a new project starts from a known shape instead of a blank page. For governance, it brings the risk package with it - the failure modes that apply, the evaluation and testing metrics that matter, the controls you need - none of it worked out fresh each time. This is the clearest answer to the worry that risk management blocks innovation: done at the right unit, it is exactly what lets you reuse.

Governance stops being per-project checklist bureaucracy and becomes a library. You write the controls for retrieval once, and every retrieval system in the firm draws from them.

The same idea shows up across the common archetypes:

  • Summarization - the dominant risk is the model inventing detail that was never in the source, so faithfulness is the metric you cannot skip, and high-stakes summaries earn a human check.

  • Code generation - the risk shifts to security, so the check is correctness plus a vulnerability scan, with the small mercy that bad code tends to fail loudly rather than silently.

  • Retrieval - the risk lives in what gets retrieved, wrong chunks or leaked embeddings, so you test whether the right context was pulled, whether the answer is grounded in it, and you watch the vector store closely.

Three archetypes, three different shapes of risk, but within each one the work is done once and inherited by every use case that adopts it. Constraints, it turns out, are what make a thing governable at all.

The capability

Archetypes tamed generative AI. Agents need the same move, one level harder, and it comes down to two questions: what can the agent do, and how far can it go on its own?

The first question is capability. In my mind, a capability is not what the underlying model can do, or how it scores on a benchmark; it is what an agent can actually do in the world, on real tasks - a repeatable building block with three things attached to it. The actions it covers. The authority it is granted to take them. And the evidence it must produce that it did so within bounds. Retrieve a customer’s transaction history is a capability. Move money between accounts is another, with far more authority and a far higher bar of evidence.

Defined this way, an agent stops being a mysterious opaque whole and becomes a set of capabilities you can reason about one at a time - decompose the agent into the capabilities it exercises, and compose your oversight back up from there.

This is why the benchmark misleads. A model topping a coding benchmark tells you nothing about whether an agent built on it should be allowed to merge to the main branch unattended. The benchmark measures skill in the abstract. Capability is about authority in the world. The two come apart, and governance lives in the gap.

Two pieces of work from very different directions arrive at the same place: GovTech’s Agentic Risk and Capability framework (https://arxiv.org/abs/2512.22211), by Shaun Khoo, Roy Ka-Wei Lee and Jessica Foo, and the work on runtime governance in finance (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6567199) that Łukasz Szpruch, Agus Sudjianto, Tanveer Bhatti and I did. Both land on capability as the better unit for governing agents.

And it scales the way the archetype does, but with guards baked in. Four agentic systems at a wealth firm - a service bot, a trade drafter, a compliance checker, a KYC assembler - look like four separate jobs, but they lean on the same handful of capabilities: document retrieval, fact extraction, policy checking. Govern those once - the authority each is granted, the evidence each must leave behind - and the rest is configuration, not construction. The fifth, sixth, seventh and even the hundredth systems cost almost nothing to govern, because they are new arrangements of capabilities you have already controlled.

The workflow

The second question - how far an agent can go - is the workflow. The workflow is what you allow it to do, designed up front: the branches, the approval gates, the points where it must stop and ask, or abstain. The trajectory is what it actually did, visible only afterwards. Most of the time the two match. The risk lives in the times they don’t, and runtime governance is the layer that watches the live trajectory against the allowed workflow and acts when they diverge - pausing for approval, refusing an action, stopping a run before it compounds.

Singapore’s Model AI Governance Framework for agentic AI (https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/mgf-for-agentic-ai.pdf) frames the same thing as action space and autonomy - what the agent is allowed to touch, and how far it goes alone. They are two dials. Widen the action space and the agent can reach more of the world. Loosen the autonomy and it asks permission less often. Turn both up and the set of possible trajectories explodes, far faster than anyone’s intuition tracks. That is how almost every agentic incident happens: the workflow quietly permitted an action it shouldn’t have, the autonomy let the agent take it without asking, and nothing was watching while it ran.

The simplest test I know is whether someone can draw your agent’s workflow on a whiteboard - the branches, the gates, the places it must stop. If they can’t, then no one has actually drawn its limits, which means the limits don’t exist yet. And the trajectories, when they arrive, will not be the pleasant kind of surprise.

The whole picture

So there are six units, not three. Three below the line - model, system, use case - the established disciplines, and the specific. Three above it - archetype, capability, workflow - the ones we are still working out, and the ones that scale. The move up the line is a move from governing things to governing patterns, and it is the only version of AI governance I can see that keeps pace as the number of systems climbs.

I am still thinking this through. The line may move, and new units may yet appear above it.

But the question worth thinking about is a ridiculously simple one. Of the six, which is your AI actually governed at - and is it the one that can keep up as the number of AI systems grows? And it will.

This began as a series of short notes, written over a few weeks while I worked the idea out in public. Here it is in one piece.

]]><![CDATA[Mr Kiasu vs. Mr FOMO]]>https://www.simplyboring.ai/p/mr-kiasu-vs-mr-fomohttps://www.simplyboring.ai/p/mr-kiasu-vs-mr-fomoThu, 04 Jun 2026 10:21:39 GMT
Spot Mr Kiasu in this piece I drew.

This is the third and final part of my conversation with Johnny Lau.

Part 1, “My Shifu“ was how I met him, and his belief that technology forces artistic purity rather than threatening it.

Part 2, “Garfield’s Eyes“ was about separating the concept from the production. This part starts with the question - will AI replace us, and then goes to the quieter question underneath - how do you live alongside something like this?

Two Fears

Watch how people talk about AI and work, and almost everyone is running on one of two fears.

The first is the fear of losing out. Will I be automated? It’s a reasonable fear. One CEO recently described it as replacing “lower-value human capital”. The fear in the air is palpable.

The second fear is the opposite. It doesn’t feel like fear. It feels like ambition. Everyone’s using AI. If I don’t adopt it now, I’ll be left behind. So people automate before they understand, buy tools they don’t need, and call it transformation.

Fear of losing out. Fear of missing out.

In Singapore we already have a word for the first one. Kiasu - Singlish for the anxiety of losing out - is the national neurosis Johnny Lau turned into a character that sold over half a million copies. What I didn’t know, until this conversation, is that he also drew the second one.

I’d asked Johnny whether AI could ever generate a real Mr Kiasu strip.

“An AI produced Mr Kiasu strip will show Mr Kiasu and Ah Kow, his pet dog in different scenarios and it will be funny. But it is unlikely to create an Ang Mo character who calls himself Mr FOMO, and competes with Mr Kiasu for Ai Swee’s love.”

Mr FOMO. A Westerner whose neurosis isn’t losing out but missing out, competing with Singapore’s most kiasu character for the girl. It’s a cultural invention. The joke is in the gap between kiasu and FOMO. I realized Johnny had done something stranger than draw a character. He’d named both of the fears now running the AI conversation and made them chase the same girl. Two anxieties, one love triangle. Both running on fear. Neither of them living.

Co-Exist with the Storm

So I asked him the real question. Is AI just another wave? He’d lived through the comic boom, dot-com, mobile, animation outsourcing. Was this one different?

“AI is a climatic shift that is going to fundamentally alter the way we think and live. If the internet era represents a tsunami, then we are smacked right in the middle of a perfect storm. Almost no one will be spared, and it’s only a matter of how many of us can co-exist with the storm.”

Not survive the storm. Not beat it, and not ride it either. We should be thinking about how to co-exist with it.

That’s a different posture from both fears. Mr Kiasu wants to outlast the storm by building a wall. Mr FOMO wants to outrun it by chasing it. Johnny is proposing you live inside it - that you stop treating it as an enemy or a prize and start treating it as weather you’ll be making things in for the rest of your life.

Here’s how he does it. Three moves.

One: Stay Fluid

“Part of the solution is to stay fluid and go with the flow because machines does not comprehend ‘going with the flow’ as easy as we do. It can fake going with the flow, but it will never get it.”

AI learns patterns. That’s the whole of what it does. So the moment you organise yourself into a recognisable pattern - the same task, the same output, the same way every time - you become trainable. “Lower-value human capital” is just the corporate name for work that had become a pattern.

The kiasu instinct is to defend that pattern harder, to prove your worth by doing the routine thing faster. It’s exactly backwards. Staying fluid means refusing to be the thing the machine can most easily learn.

Two: Orchestrate, Don’t Just Execute

“It will take something for AI to mimic the way I had orchestrated a project. The orchestration of an idea or a series of ideas; the people coming together as collaborators; outcome that will have certain level of social impact.”

This is the answer to the FOMO fear. You don’t beat missing out by adopting everything. You beat it by getting clear on what only you can do - and for Johnny, that was never the drawing. It was the convening. Deciding which ideas were worth pursuing, putting the right people in a room, choosing what impact even looked like.

AI can produce the strip. It can’t run a creative enterprise. The machine executes. The human orchestrates. Living with AI means moving up the stack - from doing the task to deciding which tasks matter.

Three: Don’t Argue with Time

I’d argued that a character rooted in a Singaporean word might be permanently untranslatable - that kiasuism was a moat. Johnny didn’t buy it:

“Don’t argue with time if the matter is a matter of time, especially dealing with AI. To AI, there is no such thing as a Western or an Eastern language to be tackled. It’s only a matter of how fast it can learn the ropes.”

No moat. Not kiasuism, not Singlish, not the routine task you’re clinging to. Given time, the machine learns every rope.

That sounds like defeat until you turn it over. If nothing fixed is safe, then stop betting on AI’s limits - and start betting on your own capacity to keep making what hasn’t been made yet. Don’t protect the output. Protect the thing that generates new outputs. The wall always falls. The river keeps moving.

What This Looks Like for the Rest of Us

I’m not a comic artist. Maybe you’re not either. But the three moves translate cleanly, because they were never really about drawing.

  • Stay fluid - be hard to reduce to a single repeatable function.

  • Orchestrate - own the judgment, not just the task.

  • Don’t argue with time - invest in the part of you that learns, not the skill that’s about to be commoditised.

None of that is use more AI, and none of it is hide from AI. It’s a third posture. Neither Mr Kiasu nor Mr FOMO. The person who has made peace with the weather and gone back to work.

There’s a coda that makes this literal. The Mr Kiasu robot - Singlish-speaking, built with Dex-Lab and the National Library - goes on more permanent display this year. You can go and ask the character himself what he thinks about AI. He didn’t fight the technology and he didn’t chase it. He absorbed it, and got a body out of the deal. Co-existence, drawn.

Which Character Are You?

Three parts, one man, and answers I didn’t expect to find through a comic artist.

We’re all somewhere in Johnny’s love triangle right now. Mr Kiasu. Mr FOMO. And Ai Swee, the object of their affections. Some of us are Mr Kiasu, bracing against loss. Some of us are Mr FOMO, sprinting so we don’t get left behind. Both are chasing. Both are afraid. Neither is living.

Johnny’s been the third character for thirty-five years - the one who didn’t panic and didn’t hype, who stayed fluid, kept the concept moving, and trusted the part of him that wasn’t a pattern to keep finding new forms.

So here’s the question I’m leaving with. Not will AI take my job, and not am I using enough of it. The better one:

What’s the part that you are still growing?

More soon. Different artist next time.

#Art #GenerativeAI #MrKiasu #FutureofArt #Singapore #CreativeIndustries #AIandArt

]]><![CDATA[Thinking in AI: Part I]]>https://www.simplyboring.ai/p/thinking-in-ai-part-ihttps://www.simplyboring.ai/p/thinking-in-ai-part-iFri, 29 May 2026 14:06:46 GMTI started a PhD in AI at 42 with zero computer science background - a few months earlier I had a full time job managing the investment risk of Singapore’s foreign reserves. Surrounded by 25-year-olds who actually knew what NP hard meant.

I knew nothing about this world.

I still remember my PhD supervisor’s reaction when I showed him yet another shiny concept from the latest paper. Unimpressed, he said: “Let’s start with the task.” I didn’t appreciate it then. Three years of fiddling with data types and tasks - networks, time series, multimodal data; regression, classification, forecasting - made me see the point.

That’s what this series is about. Something quite simple. Perhaps too simple. But I thought worth writing down.

How to see AI through the lens of data types and tasks. I call it “Thinking in AI”.

Nothing fancy.

But why does it matter then? In this age of LLMs and Agentic AI? Why not just jump ahead to prompt engineering, context engineering, or what one calls harnesses these days?

Because prompts are brittle. Context just means what goes into the Large Language Model. LLMs are not the only type of model in AI. And a framework for thinking is more durable than a harness.

AI has always been about classification, regression, generation, ranking, forecasting and many more tasks.

LLMs wrap over those tasks, in a more generalized manner (hence the name foundation models). Agents are another - an LLM orchestrating tools and other LLMs on tasks.

Strip the headline away and the data types and tasks are still the ones that were there ten years ago.

It all comes down to a few frameworks - for data, tasks, and methods - and how they compose into what we now call generative AI and agents.

So if you’re here for the frontier, you are in the wrong place. This series is deliberately about the boring fundamentals.

If you already know the fundamentals of the journey from machine and deep learning to LLMs and Agentic AI, then this might also be too simple for you.

For everyone else, I hope this is helpful in some way/

Start with the task

I call it “Prompt and Pray.”

You’ve probably sat through one of these. A workshop, a webinar, a demo. Someone shows a clever prompt. Everyone marvels at the output. Another prompt. Another marvel. A chuckle when the model gets something wrong. A tweak, and we’re back to marveling.

No one stops to ask what the model is actually doing. What data it’s seeing. What task it thinks it’s solving. Why it works on this example but might not on the next.

That frustrates me. Not because prompting isn’t useful. Because it skips every step that lets you understand the result.

So I wrote this. A short series on the part of AI that doesn’t expire when the next model drops.

Boring fundamentals, I promised. Here’s the most boring one. Before reaching for a model, write down what you’re actually trying to do.

Three questions.

  1. What data do I have?

  2. What task am I doing?

  3. Which method fits?

Most AI problems start with those three. Not how to use the latest shiny new AI model.

Same problem, different roles

The use cases below look different. But are they really?

Different contexts. The same underlying job: turn a flood of messy inputs into one decision.

Take the portfolio analyst.

From data to tasks

Two tempting responses to all this.

  1. Throw it all at an LLM and shrug when it hallucinates.

  2. Build a bespoke solution from scratch.

The useful next question is “which tasks am I doing?” Once the tasks are clear, the method question gets much smaller. The method might not even be AI.

We can usually map use cases and problems to specific tasks. Some examples.

Each does a single, clear job. Sentiment reads tone. Extraction pulls out the tickers and names. Retrieval finds the past episodes that are similar. Forecasting projects the future.

A use case or solution to a problem usually chains some of these together.

Even LLMs and Agentic AI work this way. A general LLM seems to break the frame - one model, any task. But the task didn’t disappear; it moved into the prompt. Ask for a summary and it summarises. Ask for sentiment and it classifies. Change the prompt and the context, and you’ve changed the task. An agent just chains those prompts - planning, retrieval, reasoning, generation. Each step has its own data type, its own task, its own method. Decomposition is how agents work under the hood.

The same move works beyond the analyst. The journalist chasing a story, the clinician reaching a diagnosis, the lawyer finding the case that matters - each breaks down into the same handful of tasks. Different roles. Same approach.

Where this goes next

The rest of the series dives deeper.

Part I - Components. The math behind any AI model. The data types it sees. The tasks it does.

Part II - Setups. How those components compose into machine learning, deep learning, generative AI, and agentic AI. Different names, same parts.

Part III - Data Types. Five worked examples of the framework. Tabular, text, image, networks, time series.

  • Tabular. Why a small machine learning model on a 40-year-old spreadsheet can still beat billion-parameter models at most real decisions.

  • Text. How a five-step language pipeline collapsed into a single model.

  • Images. How an invisible change to a handful of pixels turns a stop sign into a speed-limit sign.

  • Networks. What AlphaFold and your org chart have in common.

Time series. Why your ECG and a stock chart are the same kind of problem.

]]><![CDATA[LLMs Can't Understand Time, At Least Not Naturally - An Update]]>https://www.simplyboring.ai/p/llms-cant-understand-time-at-leasthttps://www.simplyboring.ai/p/llms-cant-understand-time-at-leastSat, 23 May 2026 14:50:57 GMTThis is an update of an article I did for in 2025. Things have shifted so I thought good time for an update. I have also added a list of references for folks who are interested.

When LLMs first became accessible and popular in 2022, one easy way to tell if someone was selling snakeoil was if they claimed that either ChatGPT or LLMs could forecast or predict something in the future. Someone who said that obviously did not understand how LLMs then worked, and were essentially hallucinating or bullshitting with great confidence.

LLMs are trained on text data. To forecast, you are working in the domain of time series data. Text and time series are both sequences, but they have fundamentally different characteristics.

Things have evolved since. Quite a lot actually. I would listen more patiently now for the details if someone said he or she used LLMs for forecasting due to shifts in the field. However, there is still a clear distinction between LLMs for language or text, compared to time series foundational models inspired by LLMs.

Let me explain this. In 4 short acts.

Act I: The Classical World of Time Series Modelling

For decades, way before neural networks were practically usable, forecasting or predictions with time series data was the domain of statisticians and econometricians. Time series data is fundamentally different from tabular, image or text data. You can shuffle rows in tables, mix up images, or rephrase text, and the meaning would still be largely intact.

The first, but not unique, property of time series is sequence. Mixing up the sequential order of time series data renders it meaningless. This is what it has in common with text.

This is the world of models like ARIMA. Understanding such models provides a clear understanding of what matters for time series data.

The AutoRegressive Integrated Moving Average (ARIMA) models and its variants dominated time series analysis for ages. They captured the essential insights needed to analyse or make predictions with time series data.

The core ideas:

  • AutoRegressive (AR): Current values depend on previous values

  • Integrated (I): Many series become predictable after differencing

  • Moving Average (MA): Current values depend on errors of previous predictions

Conceptually, every time series could be understood as a combination of level, trend, seasonality, cycles and some noise. This is what makes it different from text.

Y = LEVEL + TREND + SEASONALITY + NOISE

Seasonal ARIMA (SARIMA) handled recurring seasonal patterns. Vector AutoRegression (VAR) tackled multiple related series. Holt-Winters exponential smoothing handled trends and seasons differently. For volatile financial series, Engle’s ARCH and Bollerslev’s GARCH gave us proper models for time-varying volatility - the technical term is conditional heteroskedasticity, but the everyday intuition is that the size of the typical move on a given day is itself a moving target. Harvey’s structural time series models decompose a series into latent components - level, trend, seasonal, cycle - each modelled as its own stochastic process, with the Kalman filter doing the inference behind the scenes. And modern decomposition extensions like MSTL (Multiple Seasonal-Trend decomposition using Loess) handle multiple seasonalities at once - think hourly demand with daily, weekly, and annual cycles all stacked together.

If you want to start playing with these classical models, R’s forecast package (auto.arima, ets, stl) is the easy on-ramp, and Python’s statsforecast from Nixtla gives you the same family of models. Meta’s Prophet popularised an additive-decomposition approach for forecasting and sits closer to the classical end of the spectrum despite being open-sourced by a deep-learning lab.

Another fundamental characteristic of time series that is critical is the degree of stationarity. Much more than text or images (words only change meaning slowly over time, same for images), time series in many domains continually evolve, and are what we call non-stationary.*

Recall ‘I’ for integration in ARIMA? That step leads to a differencing operation that allows the time series to be more stationary, and hence more predictable.

The classical world of time series forecasting, because of this focus on the underlying patterns of trends, seasons etc., was inherently explainable.

That’s Act I. But before I move on to Act II, I thought it would be useful to mention that forecasting is not the only task you can apply to time series. You can also nowcast (predict current unknown values with time series data to date, like GDP); classify time series patterns, detect outliers for time series, and so on and so forth. But the fundamental characteristics of time series data that need to be taken into account remain the same.

Act II: Machine and Deep Learning’s Struggles with Time

The power of machine learning models in the last decade meant many tried to use machine learning models such as support vector machines, random forests, and boosting tree models for time series forecasting. But these models did not fit naturally with time series data. Not that it could not work, but it was a hit and miss.

The natural question then was, why switch to these significantly more complex machine learning models when the classical models were good enough?

Then came advances in computer vision and natural language processing driven by deep learning. Computer scientists being computer scientists, they started looking for new domains to test these models on. Naturally, given the importance of time series data in many commercial and financial settings, computer scientists started using these models for different tasks on time series data.

But here’s the thing - you can’t just throw time series data at a regular deep learning model and expect magic. Remember the properties - sequence, patterns, stationarity?

Sequence models for natural language processing like Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks made a lot of sense. DA-RNN introduced dual-stage attention for input and temporal selection. Time-aware LSTMs handled irregular sampling via decay gates. Temporal Convolutional Networks showed that convolutional architectures could match or beat RNNs for many sequence tasks.

Answering the question of how to remember important information from way back in the sequence, while forgetting irrelevant noise applied equally to text and time series data.

There were many papers that focused on adapting RNNs, LSTMs, and even CNNs (designed for image data) to time series data. I would say the results were ambivalent. Sometimes you got fantastic results, sometimes a simple classical model would beat the complex deep learning model handily in a fraction of the time and computing power. A reviewer once asked me why a model I developed and trained for time series modelling was so large in size. It was only 12m parameters (miniscule compared to the billions and trillions for LLMs), but you get the idea. Probabilistic recurrent forecasters like DeepAR and deep state-space variants became quite popular. NeuralProphet - the DL successor to Meta’s Prophet, swapping the additive components for neural network blocks - sounded promising but empirically failed to beat plain ETS on standard benchmarks. The “DL doesn’t automatically win” pattern was already showing up.

So was it a pointless exercise? Not really.

One of the key differences between classical and deep learning time series models was that deep learning models could utilise multimodal time series data (time series of text, audio, sensor events, images, networks), whereas classical models were largely restricted to numerical time series.

There are also many other aspects of time series that we could focus on aside from sequence for time series, distinctly different from natural language or image data.

For example:

  • Different time scales: Time series that change by the second, hourly, daily, quarterly and so on and so forth.

  • Varying signal quality: Time series data that are inherently noisier than natural language or image data.

  • Relationship dynamics: The importance of interactions between different time series.

There are many more. And deep learning models gave us a lot of flexibility to explore these characteristics.

In the early 2020s, even before ChatGPT, but after the seminal “Attention is All You Need” moment, many papers also explored the use of attention-based transformers for time series data.

It was a logical pairing. We use positional encodings to encode the position of tokens (or words) in transformers for text. Why couldn’t we do the same for time series steps?

A whole family came out of this - N-BEATS for interpretable basis expansion, the Temporal Fusion Transformer (TFT) for multi-horizon attention, Informer for long sequences, TS Transformer for self-supervised representation learning, Autoformer for decomposition + auto-correlation, and PatchTST. The PatchTST trick - that you can chop a series into patches and treat them like tokens - would turn out to be a conceptual seed for Act III. CoST pushed contrastive disentanglement of seasonal and trend representations.

In parallel, multimodal work fused text and prices for financial forecasting - event-driven stock prediction such as “Listening to Chaotic Whispers”, which added news attention. Fine-grained event typologies, stock embeddings from news + price history, hierarchical multi-task models for earnings-call volatility, REST’s relational event-driven framework, and FAST’s news-and-tweet time-aware network all extended this line.

I published several papers in this domain - using numerical and text time series with evolving networks. KECE combined knowledge graphs with numerical and textual time series. GLAM distinguished between global and local temporal patterns with adaptive curriculum learning to handle noise. GAME designed latent sequence encoders for multimodal data of different frequencies. DynMix used dynamic self-supervised learning with implicit and explicit network views, while DynScan learned slot concepts to handle non-stationary multimodal streams. These models showed strong results on financial forecasting, portfolio optimization, and ESG predictions, but they were highly specialized transformer models trained for specific tasks.

They were not general purpose or foundational models. Just using a transformer does not qualify. A general purpose or foundational model needs to be usable across different tasks.

But there were folks already researching these, even prior to ChatGPT coming onto the scene in 2022.

For readers who want to play with this era of models, the Python libraries darts (which also contains classical models), Nixtla’s neuralforecast, and HuggingFace’s transformer-based time-series pipelines are good entry points.

Act III: Foundation Models Learn the Language of Time

Interestingly, while I was researching multimodal time series models focused on networks, one of my PHD mates (Gerald Woo) was doing groundbreaking work on one of the first foundational models for time series data inspired by LLMs architectures.

This was the Moirai time series foundation model for universal forecasting. We attended each other’s research presentations once in a while, and I found his work fascinating. But I was already at the tail end of my PHD, so too late to switch tack.

Since then there have been many more foundational models for time series data inspired by LLMs architectures.

Now that you know all the special characteristics of time series, you would know that one cannot just throw a time series model into a transformer or an LLM and expect some magic to happen.

One key challenge was how to convert infinite numerical possibilities into a finite vocabulary that transformers can process. Language has a finite vocabulary, but not time series.

Different teams solved this differently:

Amazon’s Chronos quantizes continuous values into 4,096 discrete bins - essentially creating a “vocabulary” for time series. Google’s TimesFM treats time segments as “patches” like image processing. Salesforce’s MOIRAI uses multiple patch sizes for different temporal frequencies. There are other such models, but the fundamental issues being solved are similar. Address the tokenisation of time series, collate a large cross domain dataset, adjust the transformer architecture to address the unique characteristics of time series data.

In the months since the first version of this piece, the foundation-model wave has not slowed. But the field has clearly forked into two camps. Both are still very much “Act III” - they tokenize numerical time series and stick a transformer on top. They differ in where the parameters come from.

Camp A - Tokenize numbers, train a transformer from scratch.

This is the original recipe. The more recent successors mostly keep it and refine it:

  • Chronos-2 - Amazon’s 2025 follow-up; adds in-context learning across related series and covariate-informed forecasting.

  • ChronosX - extends Chronos to handle exogenous variables.

  • MOIRAI-2 - simpler architecture, better data, generalises better than v1.

  • MOIRAI-MoE - sparse mixture-of-experts variant; specialises across frequencies.

  • Lag-Llama - open-source probabilistic TSFM.

  • TimeGPT-1 - Nixtla’s offering.

Camp B - Don’t pretrain a TSFM. Reprogram a frozen LLM.

The intuition: LLMs already cost hundreds of millions to train. Maybe you don’t need to start from scratch - just teach the existing LLM to see numbers.

  • Time-LLM - patches the series, reprograms each patch via cross-attention with text prototypes, freezes the LLM body, trains only the input projection and output head.

  • LLMTime - encodes the series as a string of digits and lets the LLM autoregress. No training.

  • GPT4TS / OFA - frozen GPT-2, fine-tune only the layer norms and positional embeddings.

  • PatchInstruct - patch tokenisation + decomposition + neighbour augmentation as a prompting strategy.

What’s the point of foundational models for time series?

To me, the holy grail is probably few or zero-shot forecasting. Train once on massive time series datasets, then gain the ability to perform a range of tasks - forecast sales, detect equipment anomalies, or classify patterns across entirely new domains without additional training.

Still an unsolved problem I feel.

The key characteristics of time series data that we mentioned earlier are still key (such as sequence dependency, non-stationarity, varying frequencies, signal-to-noise ratios, and domain-specific patterns), and unlike text data, the nature of time series data in different domains (finance, healthcare, energy, retail, manufacturing, climate) can be vastly different and evolve significantly across time. We will talk a bit more about this below.

Act IV: The Agentic Turn - Forecasting as Reasoning

Acts I, II, and III all share an assumption. They assume forecasting is a single step problem. You feed in history, the model spits out the forecast or prediction.

The new wave has a few branches.

First, we can fuse different types of time series data at foundational model level across steps. From News to Forecast uses LLM agents to iteratively filter news, classify events by effect horizon, and fine-tune an LLM to emit digit sequences alongside a reflection loop.

Second, we can add reasoning. Training LLMs showed that reasoning at inference time - generating long chains of thought before producing an answer - improves performance on maths and code. Does it improve performance on time-series forecasting? TimeReasoner wraps series + timestamps + contextual features into a hybrid prompt, feeds it to a slow-thinking LLM, and explores three reasoning strategies - making the LLM think about the series before answering.

And finally, add multiple steps to the mix. AlphaCast - a training-free three-stage workflow (Investigator → Generator → Reflector) that mirrors how an expert forecaster works: prepare context, predict, critique, refine.

Conclusion

The only conclusion is that there is no conclusion yet. The jury is still out.

For a long time, a good place to look at the state of time series models were the Makridakis Competitions. M1-M3 (1980s-90s) were won by classical methods. M4 (2020) was won by Slawek Smyl’s hybrid ES-RNN - exponential smoothing married to LSTM - not by pure deep learning. M5 (2022) was won by LightGBM (gradient-boosted trees) with feature engineering, not deep architectures. M6 (2025) pushed into finance and reported that most teams underperformed simple benchmarks. The recurring lesson across forty years of competitions: the gains attributed to deep learning are usually gains from feature engineering, ensembling, or hybrid approaches. A example of a new leaderboard for foundation models is GIFT-Eval - 24 datasets over 144,000 time series and 177 million data points, spanning seven domains, 10 frequencies, multivariate inputs, and prediction lengths ranging from short to long-term forecasts.

But even when these foundation models top the leaderboards, the picture could be misleading.

Rethinking Evaluation in the Era of TSFMs found that benchmark scores are inflated - the test data often overlaps with the training data, either directly or because similar time periods appear in both. Strip the overlap out and the impressive numbers shrink.

Re(Visiting) TSFMs in Finance put the leading foundation models against decades of stock returns from markets around the world. The result: off the shelf, they did not beat ordinary baselines. Even fine-tuning didn’t close the gap. The only thing that worked was re-training the model from scratch on financial data - at which point you’ve essentially built a domain-specific model, not used a “foundation” one.

Multimodal and agentic forecasting are even harder to judge. Rethinking Multimodal TSF Evaluation points out that many “text helps forecasting” benchmarks are flawed - for example, the news used in testing has often already leaked into the model’s training data.

And that’s where we are now. Lots of progress, but still many open questions and uncertainties.

References

Act I - Classical time-series models, surveys, and competitions

Act II - RNN / LSTM / Transformer-era and multimodal cluster

Act III - Foundation models (Camp A: train-from-scratch TSFMs)

AI Horizon Forecast
ChronosX: Extending Time-Series Foundation Models to Support Exogenous Variables
Foundation models excel in univariate time-series benchmarks…
Read more
a year ago · 6 likes · Nikos Kafritsas

Act III - Foundation models (Camp B: LLM-reprogramming)

Act IV - Multimodal text + TS

Act IV - Agentic & reasoning frameworks

Evaluation

]]><![CDATA[Models]]>https://www.simplyboring.ai/p/modelshttps://www.simplyboring.ai/p/modelsTue, 05 May 2026 11:53:23 GMTPicking up from last week’s thought piece on the spectrum from models, systems and use cases to archetypes, capabilities and workflows.

Starting with models. My favourite.

The first models I ever met were simple curves. Bootstrapping zero rates off swap quotes in a Masters of Financial Engineering classroom. Then the more sophisticated cousins, Black-Scholes, SABR, Hull-White, Heston. I still remember the books by Emanuel Derman, Massimo Morini, and many others. For many years, the SR 11-7 was the only guidance governing this world. Though you could see snippets of it in parts of the Basel capital rules. To be honest, I was not a great student and was confused about such quantitative models for many years.

Then I met machine learning models. Decision trees. Random forests. Gradient boosted trees. Support vector machines. Then deep learning models. RNNs, CNNs. Still models, but slightly more generalized. To be honest, it was only when I was learning machine and deep learning that I finally understood the quantitative models I learnt earlier. Same for LLMs. And agents are just systems wrapped around LLMs.

Once you see the main parts - the selected function, the objective, the optimization method, and the evaluation method - you realize that they ain’t very different. Asking questions about any model ultimately comes down to these parts, and the assumptions underlying the model.

And it’s the AI model that ultimately makes an AI system different from any other software system. It’s also why in 2024, we did a review of AI model risk management in banks, when I was still in MAS. If there was any place to start, looking at the models was as good as any.

And it’s where the 3 ‘U’s arise.

Uncertainty. All models have it. Irreducible randomness you can’t eliminate. Reducible gaps you can close with more data. But always some left.

Unexpectedness. More complex models behave in ways nobody designed. Emergent capabilities. Adversarial vulnerabilities. Gaming the system. Misalignment.

Unexplainability. The degree to which we can explain a decision varies. Transparency, explainability, interpretability, but ultimately about understanding.

The Fed, OCC and FDIC dropped SR 26-2 on April 17. Replaced SR 11-7 after fifteen years. Buried in it, a single-sentence carve-out: ‘GenAI and agentic AI are explicitly excluded from scope.

For me personally, given the connections between models, whether quantitative or LLM, this exclusion puzzles me a little.

What do you think?

#ThinkingInSpectrum #ModelRisk #AIRiskManagement

]]><![CDATA[On the Units of AI Governance]]>https://www.simplyboring.ai/p/on-the-units-of-ai-governancehttps://www.simplyboring.ai/p/on-the-units-of-ai-governanceWed, 29 Apr 2026 12:56:02 GMTThere’s been a slow realization. Not about life. But a more boring one about AI risk.

I’ve been doing a number of talks on the intersection of two disciplines.

First discipline. Model risk management. My model risk supervision days.

Second discipline. Computer science. AI models. My ill-advised side quest to get a PhD in computer science at 42.

The MAS AI risk management guidelines I wrote from these covers 3 units of governance that relate to these disciplines - model, system, use case. They form a good foundation. But I feel like something has shifted.

1️⃣ Model - Obviously the domain of model risk management (MRM). The world of the late SR 11-7 & the new SR 26-2 from the US. As well as the UK’s SS 1/23 and Canada’s E23. What OpenAI, Anthropic etc. release and spend shitloads of money on.

2️⃣ System - The domain of technology risk management. Too many frameworks here to list. But the recent “AI” vulnerabilities that have been in the news, from Anthropic’s troubles to Vercel have nothing to do with the AI model, but the fragility of systems in this rush to do “AI”.

3️⃣ Use Case · What AI risk management focuses on. Anchored by NIST AI RMF, ISO 42001. MAS’ AI risk management guidelines. This is closest to what we experience when we use AI, but is also the least scalable.

And above the line.

4️⃣ Archetype. Prior to Agentic AI, I already saw folks grouping Generative AI into archetypes - question and answer, summarization, retrieval augmented generation. LLMs are general purpose tools, but grouping them helps with scalability. For once stereotyping things helps.

5️⃣ Capability. Bounded class of actions with authority, constraints, evidence requirements. Capabilities extend the idea of archetypes to the world of Agentic AI. Where we decompose and compose what AI agents do. See the paper on Agentic MRM by Lukasz Szpruch, Agus Sudjianto, Tanveer Bhatti and me on this; and the Agentic Risk and Capabilities (ARC framework) by Shaun Khoo and Roy Ka-Wei Lee.

6️⃣ Workflow. I know terms like trajectories and runtime have come into vogue (runtime is the term we use in the title of the Agentic MRM paper). But I feel like they are manifestations of what happens when AI agents do something, rather than a unit of governance. Workflows seem closer to something we can govern. And they relate nicely to the idea of the action space x autonomy perspective of AI agents (see Singapore’s Model Governance Framework on Agentic AI). As you relax the constraints of workflows, you get trajectories and runtimes that can go wider and wilder.

The three units above the line represent a new set of risks. But somehow I feel they also represent a new route to scalability.

Models, systems and use cases are usually specific. Whereas archetypes, capabilities and workflows can be generalized across a range of applications.

Just thinking out loud. Let me know your perspectives. I’ll try to write more about this in the weeks ahead.

#AIGovernance #AgenticAI

]]><![CDATA[Garfield's Eyes]]>https://www.simplyboring.ai/p/garfields-eyeshttps://www.simplyboring.ai/p/garfields-eyesMon, 30 Mar 2026 14:06:30 GMTThis is the second in my series on what happens when the people who make things meet a technology that also makes things.

The first part, “My Shifu“, covered how I met Johnny Lau through a crazy experiment called Creative Youth Xchange, and his view that technology forces artistic purity rather than threatening it.

I’ve been sitting on this, but this post by Arvind Narayanan on how creativity in art and creativity in science are different made me get back to this second part.

This article goes deeper into the creative process. About creative production, and why AI may not be that great a leap.

Is AI Really A Leap?

One of the key fears of creatives is that this leap into AI may be one too far. Further than the camera. Further than digital tools.

Something that creativity cannot keep up with. But is it realistic to think that all art in the future will be purely from AI?

So I asked Johnny about his jumps from pen-and-paper comics to different technologies, the most recent being a Singlish speaking robot. Most people would call that a transformation. He didn’t:

“It was never a leap for me. Right from the beginning when I launched the first book in 1990, I engineered a range of products such as t-shirts & mugs, a 13-inch ruler and bumper sticker with a tagline: Beware of Kiasu Driver. The latter became a bestseller instantly.”

My read. This leap is perhaps not that different from the ones in the past for creatives. And AI is perhaps more of an opportunity than a threat.

“For me it has always been a concept, not a book nor a sticker ... My brain doesn’t function the way where sectors and industries are differentiated. I had to develop a language to align myself with how the world functions.”

A concept. Not a product. Not a comic.

Here’s what I think is going on. Narayanan’s distinction is useful, creativity in art is emotional, social, driven by taste. Creativity in science is instrumental, evaluable, a systematic search through possibilities. AI is getting better at the second kind. The first is harder to touch.

Johnny’s concept is the first kind. Not the drawing - the reason for the drawing. The cultural nerve. The recognition that makes a Singaporean laugh at their own kiasu-ness. That doesn’t live in a dataset. It comes from being someone, from somewhere.

The production - t-shirts, mugs, strips - is the second kind. Patterns. Systems. Things that can be decomposed and delegated.

Johnny worked this out decades ago. He kept the concept. He systematized the production. And the person who showed him how wasn’t an AI researcher. It was Jim Davis.

The Garfield Revelation

Not exactly Garfield, but you get the idea.

I asked about architecture at USC. He said something I didn’t expect - that USC was never about architecture for him:

“The years at USC were never about the study of building design. I took full liberty to immerse myself in the world of media and entertainment when I landed in Los Angeles.”

That’s where he met Garfield. Not the cat. The operation.

“I was stunned by Davis and how he had built a team of artists and writers to produce daily comics strips for the lasagna loving cat for its national syndicate. He had deployed an animation studio system to run his comics studio.”

Here’s what he described. Davis ran his comics like a factory. Team meeting on Monday. Different themes tossed around for the week’s seven strips. Once decided, a lead artist carried out the drafts. The process cascaded through the team. And at the end:

“Davis merely just drew Garfield’s eyes as a gesture of completion.”

Garfield’s eyes. The signature. Everything else was system. That one stroke was authorship.

I sat with that image for a while.

Johnny brought this model back to Asia. Most Asian comics were produced by a single artist - one person doing everything. The Garfield team showed him a different path. It became Comix Factory, the studio that produced Mr Kiasu.

Agents and Art

Here’s the connection I keep thinking about.

Davis’s studio system was, in effect, an early version of what AI promises every creator today. Break the work into steps. Specialize each step. Let the system handle volume while the creator handles vision.

I’ve written about agentic AI before - systems where multiple AI agents each handle a different part of a task. One agent plans, another retrieves information, another reasons, another generates output. The whole thing works because each part has a defined role and you can observe what each part does.

Davis was essentially running an agentic system in the 1980s. Just with humans instead of models. One person for drafts, another for inking, the process cascading through the team. And the orchestrator - Davis - only needed to draw the eyes.

Davis understood in the 1980s what the AI industry is selling in the 2020s: you don’t need to do everything yourself. You need to know what only you can do.

For Davis, it was the eyes. For Johnny, it’s the concept.

And Johnny made the connection himself:

“Similarly with AI, we just have to change our mindset on the method of production. The creation portion doesn’t change as it will always come from the depth of our minds.”

Change the method of production. Keep the creation. That’s not a theory from someone who’s read a few articles about AI. That’s a principle from someone who’s been decomposing creative labor since 1990.

A nice footnote: in 2023, Mr Kiasu and Garfield collaborated on a joint library programme with Singapore’s National Library. A one-of-a-kind global collaboration between two comics characters - the Singaporean one modeled on the American one’s production system - sharing a stage forty years later.

AI As Just Another Method

There’s a broader pattern here that I think matters for anyone thinking about AI and creative work.

We tend to frame the AI question as: will machines replace artists? That’s the wrong question. The better question is: what part of your creative work is system, and what part is signature?

If you’re a writer, maybe the system is research, outlining, drafting. The signature is voice, judgment, the instinct for when a sentence lands. If you’re a designer, maybe the system is rendering, iteration, production. The signature is the concept that no brief could have specified.

Davis knew. He drew the eyes and let the system do the rest. Not because he was lazy. Because he understood what only he could do.

AI doesn’t change this question. It just makes it unavoidable.

Johnny figured this out before AI existed. He’s been running this model for thirty-five years - concept first, production second, and the concept crosses every medium, every platform, every technology that comes along.

The question I’m leaving with you: what’s the part that only you draw?

More soon.

]]><![CDATA[A Simple Reading List on Third-Party AI Risk Management]]>https://www.simplyboring.ai/p/a-simple-reading-list-on-third-partyhttps://www.simplyboring.ai/p/a-simple-reading-list-on-third-partyWed, 25 Mar 2026 16:42:58 GMT

Yesterday, there was news about a threat actor compromising the popular LiteLLM package on PyPI. litellm is an ubiquitous library that many teams use as a unified interface to call different LLM APIs. The attacker didn’t need anything sophisticated. They poisoned the CI/CD pipeline through a compromised security scanner (the irony!), then pushed backdoored versions of litellm that harvested SSH keys, cloud credentials, and secrets on every Python startup. Seems like the hackers are now working hard - going through a 300gb treasure trove and extorting multi-billion-dollar companies.

This is what third-party AI risk actually looks like. Not the dramatic scenarios we imagine - collusion, rogue AI, coordinated manipulation. Sometimes it’s a simple file in a package you installed last Tuesday.

And it’s not just open-source libraries. Every AI vendor you use sits on its own supply chain of dependencies, sub-processors, and infrastructure providers. A single compromised link and you inherit the risk.

So how do you govern AI you don’t control? And maybe don’t even know exists.

I can’t say I have the answer. But I can offer a reading list.

Note: I have used open-access links from arXiv as far as possible.

Where and Why Does It Break?

1. “AI Auditing: The Broken Bus on the Road to AI Accountability” - Birhane et al. | IEEE SaTML (2024)

Taxonomizes current AI audit practices across regulators, law firms, civil society, journalism, academia, and consulting. Finds that only a subset of AI audit studies translate to desired accountability outcomes. The title says it all - we’re doing audits, but many of them aren’t actually getting us where we need to go. 📄 arXiv:2401.14462

2. “Dislocated Accountabilities in the ‘AI Supply Chain’: Modularity and Developers’ Notions of Responsibility” - Widder & Nafus | Big Data & Society (2023)

Developers building AI from preexisting modules often believe responsible AI belongs to “the next or previous person in the imagined supply chain.” Everyone assumes someone else is managing the risk. 📄 arXiv:2209.09780

3. “Understanding Accountability in Algorithmic Supply Chains” - Cobbe, Veale & Singh | FAccT (2023)

Explores how algorithmic supply chains create distributed responsibility and limited visibility due to what the authors call the “accountability horizon” - you can only see so far. Also covers cross-border supply chains and regulatory arbitrage, which matters when your AI vendor operates across jurisdictions with different rules. 📄 arXiv:2304.14749

How can we perhaps solve it?

1. “AEF-1: Minimum Operating Conditions for Independent Third Party AI Evaluations” - Stosz et al. | AI Evaluation Foundation (2025)

A voluntary standard and checklist that defines what evaluators actually need from AI providers: independence from the provider, sufficient access to assess characteristics of interest, and transparency in sharing methods and findings. The gap between what this standard says you need and what most vendors actually give you is the governance gap. 🔗 AI Evaluation Foundation

2. “Outsider Oversight: Designing a Third Party Audit Ecosystem for AI Governance” - Raji et al. | AAAI/ACM AIES (2022)

Synthesizes lessons from financial, environmental, and health regulation on crafting effective external oversight systems. The key insight for me: audits alone won’t achieve accountability. You need deliberate design and institutional weight, the same lesson financial regulators learned after Enron. 📄 arXiv:2206.04737

3. “Model Cards for Model Reporting” - Mitchell et al. | FAccT (2019)

The foundational paper proposing that released models be accompanied by documentation detailing their performance characteristics. This paper is from 2019. It’s now 2026. Notice how few vendors actually provide this level of transparency. In fact, some studies have shown that such transparency is declining, rather than increasing due to greater awareness. 📄 arXiv:1810.03993

4. “Third-party compliance reviews for frontier AI safety frameworks” - Homewood et al. | arXiv preprint (2025)

Explores third-party compliance reviews where an independent external party assesses whether a frontier AI company complies with its own safety framework. Discusses benefits (increased compliance, assurance to stakeholders) and real challenges (information security risks, cost burdens, reputational damage from findings). This is the emerging infrastructure for keeping vendors accountable - but it’s still nascent. 📄 arXiv:2505.01643

5. “Implementing AI Bill of Materials (AI BOM) with SPDX 3.0” - Bennet et al. | Linux Foundation Research (2025)

Extends the Software Bill of Materials concept to AI, including documentation of algorithms, data collection methods, frameworks, licensing, and compliance. If you want to know what’s actually inside the AI system you’re buying - and what changes when the vendor updates it - this is the direction. Think of it as the AI equivalent of ingredient labelling. 📄 arXiv:2504.16743

6. “AgentFacts: Universal KYA Standard for Verified AI Agent Metadata & Deployment” - Grogan | arXiv preprint (2025)

Forward-looking. Proposes a “Know Your Agent” standard with cryptographically-signed capability declarations and multi-authority validation. If something like this existed at scale, it would reduce the custom integration friction that makes switching so expensive. We’re not there yet, but this is the direction things need to go. 📄 arXiv:2506.13794

7. “Consultation Paper on Proposed Guidelines on Third-Party Risk Management” - Monetary Authority of Singapore | MAS (March 2026)

And last but not least. MAS just released this in March 2026. Supersedes the old outsourcing guidelines and extends expectations to all third-party arrangements, not just outsourced services. Covers risk assessment, due diligence, contracting, onboarding, ongoing monitoring, and termination. Requires FIs to maintain a register of third-party arrangements, monitor concentration risk, and extend oversight to sub-contractors. AI appears in a footnote, where it refers to MAS AI Risk Management Guidelines I wrote, but the relevance of these guidelines to AI is clear. 🔗 MAS Consultation Paper

I think this is one of the hardest problems to solve in AI risk management. Because it requires everyone to work together. And that’s really hard today.

#ThirdPartyAI #AIRiskManagement #VendorManagement #AIGovernance #AIReadingList

]]><![CDATA[The AI Governance Tool]]>https://www.simplyboring.ai/p/the-ai-governance-toolhttps://www.simplyboring.ai/p/the-ai-governance-toolMon, 23 Mar 2026 13:16:04 GMTIf you’re infantry in the Singapore army, you would probably have encountered this protocol to shout “gap gap gap” when you breach a fortification.

I always thought it was kind of stupid. Since shouting it was asking for the bullets to be aimed at you. I always felt I was lucky to move from infantry to be a scout. Scouts never ran towards bullets.

But I think I understand the need for it more now. Shouting it in the army helps make people aware of the gap that you are moving towards.

In one of my past weekly reflections on LinkedIn, I said three gaps kept surfacing. The common language gap. The contextualisation gap. The last-mile gap. Same need for awareness.

Frameworks exist. Guidelines exist. I wrote some of them. But translating frameworks into something relevant for one’s context is not easy. Especially for individuals and smaller firms.

Last week, MAS launched the MindForge AI Risk Management Toolkit. I was deeply involved as it needed to be aligned to the MAS AI Risk Management Guidelines I wrote.

The MindForge AI Risk Management Toolkit is an operationalisation handbook with case studies. It’s an attempt at bridging the messy middle I keep writing about. But I suspect, aside from large financial institutions, it may still be a lot to chew on for everyone else.

Last week, a friend who leads manpower development also asked me something similar: what about everyone else?

This is my early attempt at an answer.

GOT, not Game of Thrones, but an AI Governance Tool.

At govern.simplyboring.ai.

Subscribe to my Substack to get keys for more generations.

Describe what you are using AI for. Get a tailored governance pack in minutes. The standards that apply to your situation. The controls to put in place. How to implement them. What evidence to prepare.

Built on public information. The MAS AI Risk Management Guidelines I wrote. The MindForge toolkit that just launched. The US Financial Services AI Risk Management Framework. The EU AI Act. NIST AI RMF. ISO 42001. And other international frameworks. Cross-referenced and connected. And a little of my own perspectives.

The AIRG tells you what to do. MindForge shows how the industry is doing it. GOT asks what you specifically are doing, and tells you what applies to you.

Not just for the large institution with a full second line (and third line) of defence. For everyone else. Individuals. Small firms.

It’s an early prototype. It will almost certainly have errors. And a little slow - around 45-60 seconds for each report. And I think it’s still a little iffy for individuals.

But I tried to ground it in real frameworks and design it for real situations so anyone, not just banks.

Please try it. Tell me what’s missing. Tell me what’s wrong.

#Mindforge #NIST #AIRiskManagement #SimplyBoringAI #Grounding #AIRG

]]><![CDATA[OpenClaw - Kiasu Version]]>https://www.simplyboring.ai/p/openclaw-kiasu-versionhttps://www.simplyboring.ai/p/openclaw-kiasu-versionTue, 17 Mar 2026 11:41:46 GMT

I don’t like OpenClaw. But pretty hard to resist the urge to try.

I think it does demonstrate something interesting about agents - 1) the additional utility when you hook it to a heartbeat to do something at a set frequency, or cron for a scheduled task; 2) the convenience of letting it communicate with you over a messaging app.

But I really can’t see what else it offers in additional to say Claude Code, which already allows for multi-agent workflows.

Ok, maybe if being exposed to the internet, or losing your data is your thing, then I agree, OpenClaw without any additional safeguards is definitely for you.

OpenClaw’s issues must be obvious. NVIDIA just announced NemoClaw at GTC - essentially OpenClaw with enterprise-grade security guardrails baked in. NanoClaw, a lightweight alternative, takes a different approach by isolating each agent in its own Docker container.

And so when I tried it, I made some adjustments to help mitigate this. Probably not perfect, but I think I can live with its risks with these controls. And I think you can use them for any of the other Claws too.

So here’s the short guide on what I did. I won’t copy and paste all the scripts. Just use my text below as a prompt and I am sure any of the LLMs can help you.

The Setup

I run it on a DigitalOcean VPS - Ubuntu 24.04, 2 vCPU, 2GB RAM, $12/month, Singapore region. I went with a manual setup instead of the 1-Click app because I wanted full control over every layer of the stack.

The finished stack looks like this:

Article content
Only i can access OpenClaw through TailScale, and I choose the files that it can see using Syncthing

What I use it for: OpenClaw runs 24/7 on the VPS. Every hour it wakes up, checks my task files, and stays silent if nothing is urgent. Every morning it sends me a briefing on Telegram. Claude Code runs locally for deep work and pushes outputs into a shared AGENT-INBOX folder. OpenClaw picks them up.

Total cost: $12/month for the server. OpenClaw lets you plug in your own API key. There’s an option for using your existing Claude key that will come up after the OpenClaw onboard.

Now, on to the controls.

Control 1: Server Hardening

Out of the box, a VPS is exposed. Default root login, password authentication, every port open. The first thing to do is reduce the attack surface.

Every service you run is an attack surface. Make each one as small as possible, and add friction at every layer so that even if one layer fails, the next one holds.

The key steps: update everything, create a non-root user (don’t run your server as Administrator), disable password authentication so only SSH keys can log in, and restrict root login entirely.

I won’t paste every command here - ask Claude or ChatGPT to give you the steps for hardening a server for Ubuntu 24.04 and it’ll give you the full thing.

Control 2: Firewall and Fail2ban

Two layers. UFW (Uncomplicated Firewall) controls who can knock on your door at all - deny all incoming by default, then explicitly allow only what you need. Fail2ban monitors your login logs and automatically bans IPs that fail authentication too many times.

One thing to watch out for: enable UFW after you allow SSH, not before. Get the order wrong and you immediately lose access to your own server.

But even with both of these, your SSH port is still visible to the public internet. Bots still find it, still try it. Fail2ban is a mitigation, not a solution.

Control 3: Tailscale - Making the Server Invisible

What if the SSH port wasn’t visible at all?

Tailscale creates a private encrypted network across the internet - but only devices you authenticate can join. Think of it as a private LAN that ignores physical location. Every device gets a private IP in the 100.x.x.x range. Your VPS gets one. Your laptop gets one. They talk through an encrypted WireGuard tunnel regardless of where they physically are.

The key part: your VPS’s public IP can have all ports closed. Tailscale establishes peer-to-peer through NAT traversal - no open ports required. Install it on the VPS, install it on your laptop, log in with the same account - done.

Once Tailscale is working, update your UFW rules to allow SSH only on the Tailscale interface, and remove the public SSH rule. After this, run a port scan on your VPS’s public IP. You’ll see nothing. The server is invisible. Fail2ban becomes a backup layer rather than the first line of defence.

As always, test SSH via the Tailscale IP in a second terminal before removing public access.

Installing OpenClaw

With the controls in place, this is the straightforward part.

curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash - 
sudo apt install -y nodejs
sudo npm install -g openclaw
openclaw onboard

For NemoClaw, it will probably look like this (based on Nvidia’s instructions)

curl -fsSL https://nvidia.com/nemoclaw.sh | bash
nemoclaw onboard

The setup wizard walks you through AI provider (I use Anthropic), API key, and gateway token. It installs OpenClaw as a user-level systemd service - not a system-level one. This matters:

This works:

systemctl --user restart openclaw-gateway

This silently fails:

systemctl restart openclaw-gateway

Every OpenClaw service command needs --user. Burned time on this.

The dashboard

OpenClaw has a web dashboard. It binds to localhost only, so you access it via SSH tunnel:

openclaw dashboard    # prints URL with token
ssh -N -L 18789:127.0.0.1:18789 openclaw@YOUR_TAILSCALE_IP

Then open the full URL including the #token= fragment in your browser.

Don’t open

http://localhost:18789/

bare. The #token= fragment is the authentication handshake. Without it, the gateway sees an unapproved device and shows “pairing required.” The dashboard looks like it loaded - it hasn’t connected. Use the full URL from openclaw dashboard.

Syncthing - The Sync Layer

OpenClaw on the VPS needs a way to receive tasks and send outputs back. I use Syncthing - open-source, peer-to-peer file sync, no cloud intermediary.

The synced folder (AGENT-INBOX) lives inside my Obsidian vault on Windows and mirrors to remote VPS. I write tasks in Obsidian, they appear on the VPS within seconds. OpenClaw writes outputs on the VPS, they appear in Obsidian within seconds.

One thing to watch out for: install Syncthing from the official apt repository, not Ubuntu’s default packages - the default version is outdated enough to cause silent sync failures when paired with a newer Windows client.

The Heartbeat

The heartbeat is the reason I tried OpenClaw in the first place. It’s the cron-like behavior - wake up on a schedule, check your files, act without being asked.

openclaw config set agents.defaults.heartbeat.every 1h
openclaw config set agents.defaults.heartbeat.target last
openclaw config set agents.defaults.heartbeat.activeHours.start "06:00"
openclaw config set agents.defaults.heartbeat.activeHours.end "01:00"
openclaw config set env.vars.TZ "Asia/Singapore"

Note: use the full dotpath agents.defaults.heartbeat.every, not just heartbeat.every. The docs show JSON structure, the CLI uses dotpaths.

If OpenClaw finds nothing urgent, it stays silent. No Telegram noise. Only substantive findings surface. This is controlled by a HEARTBEAT.md in your workspace - a checklist the agent follows on every run.

One important step after setup: delete the BOOTSTRAP.md file in your workspace. It’s OpenClaw’s first-run onboarding wizard that blocks the agent on every gateway start until a human responds. Once you’ve completed setup, remove it or the agent stays stuck in bootstrapping state indefinitely.

The Security Audit

OpenClaw has a built-in security audit you can run with openclaw security audit --deep. Worth running after setup to confirm your controls are in place.

The most important thing it tells you is the trust model. OpenClaw operates as a “personal assistant” - single user, single trusted operator. It’s not designed for multiple users sharing one gateway. If you’re thinking about running this for a team or clients, that’s the line that should make you pause.

Is It Worth It?

Honestly, I’m still not sure.

The heartbeat and the Telegram integration are genuinely useful - having an agent check on things hourly and surface only what matters is different from having a chatbot you talk to. And the data sovereignty matters to me.

I’ll keep running OpenClaw for a while to see if it sticks. See if the habit of an always-on agent changes how I work. But I wouldn’t tell anyone to just run it.

If I were starting today, I’d probably look at NanoClaw first - it’s ~4,000 lines of code versus OpenClaw’s > 400,000, and agent gets its own isolated container. NemoClaw is worth trying too, especially if you’re in an enterprise context. Both are open source.

But the manual controls I described above? They’d apply to any of these. Server hardening, Tailscale, firewall rules - that’s not OpenClaw-specific.

If you’ve tried self-hosting OpenClaw (or variants such as NanoClaw), what additional controls did you put in place? Also what else do you find useful about it (beyond say Claude Code or Codex) that warrants the effort?

#AI #AgenticAI #Security #OpenClaw #NemoClaw

]]><![CDATA[From Contrastive Learning to World Models]]>https://www.simplyboring.ai/p/from-contrastive-learning-to-worldhttps://www.simplyboring.ai/p/from-contrastive-learning-to-worldWed, 11 Mar 2026 14:30:21 GMTNews on Yann LeCun’s AMI Labs raising ~$1bn in their seed round and having one of their key bases in Singapore just made the news. $1bn for a seed round is kind of insane. AI startup seed rounds are more like Series Zs these days.

That aside, this news brought back some memories of an older technique from Lecun’s lab, titled VICReg, that inspired one of my papers. And led me to reread his JEPA series.

Today's diptych is about that connection.

Left Panel: VICReg

Self-supervised training in AI just means training without labels. We use it in LLMs. Pretraining a model by masking parts of inputs and getting it to predict the masked parts. One issue in self-supervised learning is collapse. AI likes to find shortcuts. And when it does, everything converges to the mean.

Most methods fight this with training or architectural tricks. VICReg uses just three interesting loss terms:
➡️ Variance so that there’s always some differences between the representations it learns.
➡️ Invariance so that the same input gives the same representation.
➡️ Covariance so that each dimension learns something different

Invariance is the main learning objective. Variance + covariance are the regularisers that prevent it from collapsing.

I liked VICReg’s idea. So I added a twist for a paper I published on a model called DynMIX. I fed it two fundamentally different views of the same company - explicit (constructed) and implicit (learnt from data). Another twist. VICReg's variance target is static. For financial data, that made no sense. Markets are volatile so variance can’t be static. So I made the target dynamic, based on variances of stock market returns.

Right Panel: JEPA

The core of the paper by Yann LeCun - JEPA - takes a different approach to the same problem. Instead of just learning representations, it predicts them. In AI, we usually predict something real - a label (dog, cat), or a number (stock price). JEPA predicts the underlying representation instead. Think of it as the model learning to imagine what something looks like in its own internal language, before it's seen in the real world.

Three components:
➡️ Context encoder: encodes what's visible into a representation
➡️ Target encoder: encodes what's hidden into a representation (this IS the label)
➡️ Predictor: tries to predict the hidden from the visible

So what does this unlock? Once representations of one modality (e.g., videos) are learnt, the next stage adds actions. A second model takes the video frame embeddings interleaved with a robot arm's movement commands, and predicts the next embedding. Planning then becomes: imagine many possible sequences of moves, simulate each one in the model's head, and pick the one most likely to reach the goal. Like a chess player thinking several moves ahead, but the "board" is the model's own internal understanding of the world.

Interesting times ahead. If AMI Labs succeeds.

#RepresentationLearning #JEPA #VICReg #YannLeCun #AIResearch

]]><![CDATA[Three Speeds of AI Adoption]]>https://www.simplyboring.ai/p/three-speeds-of-ai-adoptionhttps://www.simplyboring.ai/p/three-speeds-of-ai-adoptionWed, 04 Mar 2026 15:45:37 GMT

Fire people, stock goes up. Beat expectations, stock goes down.

I thought this was both funny and sad.

I started writing about this last week. But with the Iran war going on now, this feels almost quaint.

Predictions about a doomsday scenario where AI agents collapse the job market by 2028. Headlines about Dorsey’s Block cutting over 4,000 roles - over 40% of its workforce. Gone, just like that. His excuse was AI, as always. Block’s stock price shot up. And then the irony - Nvidia beat expectations, revenue up ~70%, guidance ahead of even the most bullish estimates. The stock dropped 5.5%. $260 billion in market value erased overnight.

Fire people, stock goes up. Beat expectations, stock goes down. That’s the weird logic of markets right now. And I think it mirrors the weird logic of the AI debate. The narrative has decoupled from the evidence. What you believe matters more than what you can show.

Let me compare three rooms.

Three Rooms

Room one. Block’s numbers. An internal AI agent called Goose. One engineer says 90% of his code is now written by it. Non-technical teams writing SQL queries and closing tickets. Revenue per employee doubled.

Room two. A classroom. Students from an organization that will exist in fifty years regardless of what happens in AI. What were they learning? How to write a prompt. How to get an image out of a model. In 2026. They weren’t behind because they were slow. They were behind because there’s a chasm between what frontier labs ship and what most organizations can absorb.

Room three. Trading professionals at a talk I gave. They’d read the doom scenarios. They wanted to know - should I be worried about my career? They weren’t panicking. But they weren’t dismissing it either. Sitting in the uncertainty, looking for clarity.

I tried to answer, but I thought my answer was lacking. So I tried to do better with this article.

Why do I think a piece like Citrini and Shah’s “The 2028 Global Intelligence Crisis” is science fiction designed to go viral? Because of three speeds that may not be moving in lockstep.

The Three Speeds

I think all of these pieces assume uniformity in scaling and adoption. But I think there are 3 speeds that matter, perhaps even more.

METR - Model Evaluation & Threat Research - is a nonprofit that evaluates AI capabilities. Their time horizon chart - arguably the most cited graphic in AI right now - has been called “the most misunderstood graph in AI.” It shows AI capability doubling every few months. Sequoia used it to declare “2026: This is AGI.”

What that chart actually says about each speed:

Speed of Capability. What AI can actually do. Far fewer people saw METR’s own limitations page. More than 10 disclaimers about what their research is not. The researcher writes plainly: they have “no idea whether Claude’s ‘true’ time horizon is 3.5h or 6.5h.” Their code quality research: AI-generated code passed 38% of tests but “none of them are mergeable as-is.” But you don’t need METR to tell you this. When was the last time you spotted your LLM sprouting nonsense? Less common than in 2023, but definitely not zero.

Speed of Adoption. What organizations actually change. You might remember the discredited MIT study claiming 95% of AI pilots fail - the methodology didn’t hold up. So I went looking at what has happened since. McKinsey’s 2025 State of AI survey at the end of 2025: 88% of companies say they use AI, but only a third have begun to scale it. A survey of 120,000+ enterprise respondents from March 2025 to January 2026 reported that nearly two-thirds have no formalized AI initiative at all. But you don’t need McKinsey to tell you this. Look around your own workplace. How many AI pilots have started? How many are in production?

Speed of Belief. What people think AI can do. The chart made it into every investment deck. The caveats didn’t. Same with Citrini’s doom scenario - it spooked actual markets before anyone could verify the underlying technical assumptions. Same with Block - the headline landed, and every manager with a budget started rethinking next year’s headcount because their boss could be like Dorsey. Doesn’t matter if their company has nothing remotely comparable to Goose. Charts travel faster than FAQs with caveats on limitations. And I am sure you know how dangerous mistaken beliefs in large organizations can be. When’s the last time you spotted a senior person saying something totally wrong, but nobody correcting him or her?

The Underlying Issue

The doom scenario by Citrini and Shah needs all three speeds to converge. Capability, belief, and adoption in lockstep.

But the problem is, belief doesn’t need the other two.

Block just showed that AI can replace headcount. At least in its specific context. But the headline doesn’t come with that caveat. So what happens next? A manager somewhere reads it and quietly decides not to backfill that open role. A team that was supposed to grow just doesn’t. None of that requires AI to actually do the work. It just requires someone with the power to believe it could.

Back to 2008. Not the mechanics - those are different. The underlying issue with such beliefs.

In 2008, banks sold complex structured products to people who had no ability to understand what they were buying. CDOs priced by models that the sellers themselves barely understood. The belief in the models was enough to move trillions. The mom-and-pop investors were the ones left holding the bag when reality caught up.

I think something similar is happening now. Not with financial products, but with people’s livelihoods. Block fires 4,000 people and the stock jumps. The market cheers. The narrative is: AI made it possible. But how much of that is demonstrated capability, and how much is a belief about capability that hasn’t been stress-tested outside one company’s very specific context? And perhaps unproven, even for that company.

It was unconscionable then to sell instruments people couldn’t understand to people who couldn’t afford to lose. I think it’s unconscionable now to restructure people’s lives based on a chart that its own creators say they can’t fully trust, and a narrative that outpaces the evidence by miles.

The direction is right. AI will displace some work. But the timeline assumes a uniformity that doesn’t match what I’m seeing in real rooms. And the human cost of getting the timeline wrong - of letting belief run ahead of reality - is not a rounding error. It’s people.

#AI #FutureOfWork #AIAdoption #AIRisk #Leadership

]]><![CDATA[Learning risk management again, because of AI]]>https://www.simplyboring.ai/p/learning-risk-management-again-becausehttps://www.simplyboring.ai/p/learning-risk-management-again-becauseMon, 02 Mar 2026 14:27:18 GMT

A FIX NextGen meeting at BlackRock. Folks who live and breathe the markets. And then there’s me.

I started by owning up. “I am a bureaucrat.” But shared my random walk. From engineering to art policy, Basel regulation to investment risk, a PhD in AI, then AI risk supervision. Messy life. Messy research. I wasn’t sure what they would be interested in, so I just laid it all out.

As for the talk, I took them on a journey. From the Gaussian copula that contributed to the GFC, to the questions we’re tackling now with large language models.

That line (from the GFC to AI today) may seem strange. But it is a very clear one.

The Formula That Broke Finance

A single model. The single factor Gaussian copula. Elegant, tractable, widely adopted for pricing CDOs. But totally unrealistic. In fact, I would say the decision to use it to model CDOs was bordering on criminal.

But even then, the GFC would not have happened if not for the fact that nobody was watching these complex things, nobody fully understood them, and nobody was clearly accountable.

In the aftermath: US’ SR 11-7. Model risk management requirements. The predecessor of UK’s SS 1/23 and Canada’s E-23. Singapore’s AI Risk Management Guidelines (AIRG) also shares a lot of with these guidelines. I wrote that last piece.

Complexity Squared

Here’s how the problem has changed.

2008: One model. Too simple. Trillions at stake. Nobody understood it.

2025: Way more than one model. Complexity². Everyone has an opinion.

Different. But also recognizable.

That’s the risk. Things look different enough to seem like a new problem. They’re not entirely. But the new parts matter.

3 U’s

What’s actually new - or at least amplified - in AI models.

Uncertainty. All models have it. How confident is the model in its output? Two flavors: irreducible - natural randomness you can’t eliminate. Reducible - gaps in knowledge that more data or better models can close.

Unexpectedness. Some AI models exhibit behaviors nobody designed. Emergent capabilities. Gaming the system. Adversarial vulnerabilities. Hidden biases. Misalignment.

Unexplainability. The degree to which we can explain AI decisions varies. Transparency, explainability, interpretability - not the same thing, and even combined, they don’t guarantee understanding.

These three make AI a harder risk management problem. Not a different one. Harder.

The Wicked Domain (and Then Some)

I used a framing from Epstein’s Range toward the end.

Kind domains - chess, music - have clear feedback. Deliberate practice compounds. The 10,000-hour idea in Outliers by Malcolm Gladwell works there. Depth helps here.

Wicked domains - medicine, finance - have delayed feedback, unclear rules, expertise that doesn’t transfer cleanly. Breadth sometimes helps here.

AI risk management sits firmly in wicked territory. But I think it might be something worse. An evil domain. The feedback is ambiguous, the rules keep shifting, and the expertise required keeps branching.

Not I-shaped for depth. Not T-shaped for breadth plus a bit of depth. More like a banyan tree. Multiple deep roots spreading from the same trunk. Depth across AI, governance, legal, model, technology, human factors - requiring real depth in each.

Same Questions, New Complexity

Across my messy career - Basel policy, quant model risk, investment risk, AI supervision - three questions kept reappearing.

What’s at risk? You can’t manage what you haven’t found. This means discovering where AI exists in your institution and profiling how risky each system actually is.

How do we manage it? Once you know what’s at risk, you need controls in the right places, and ways to check if those controls are working.

Who’s accountable? Controls without owners become theatre. Someone has to own the risk, and the organization needs capability to sustain it.

Basically - find it, control it, own it.

As I wrote before. These aren’t new questions. They’re the same questions risk management in finance has been asking for decades. SR 11-7 answered them for model risk. AIRG is answering them again, in a harder context.

Normal technology. Normal systems. Normal risk questions.

At the end of the session there were some useful questions. My answers to most of those are in FIX’s post here.

But there was one that stuck with me. About how real Citrini’s dystopian AI narrative was. I wrote about it last week here. But more thoughts came to mind because of the question. Will do another post in greater detail later.

#AI #AIRiskManagement #ModelRisk #FIXTradingCommunity #FinTech #NextGen

]]><![CDATA[A free 180 page ebook on AI Agents for Investing with code]]>https://www.simplyboring.ai/p/a-free-180-page-ebook-on-ai-agentshttps://www.simplyboring.ai/p/a-free-180-page-ebook-on-ai-agentsTue, 24 Feb 2026 11:19:17 GMT
The cover

When I shared an early draft with a friend, his first question was:

“How much of this was AI?”

For a while, it made me a little sensitive. Even made me question if I should share this book.

Not because the question was unfair. It’s a reasonable thing to ask. It’s almost automatic to question the origins of any work these days.

But because it implied something I wasn’t sure how to answer cleanly.

If I said “a lot,” it sounds like the book isn’t mine. And perhaps just slop.

If I said “not much,” that’s really not honest either.

So let me just show you. Full transparency.

(Just scroll to end if you want to skip all this)

AI or me?

This book is part me, part AI. And I want to show you exactly where the line is.

What’s mine

The architecture. The four-pattern framework. The decision to organize the book around Tool Calling, ReAct, CodeAct, and Orchestration - that came from reading papers, building prototypes. No model suggested that structure.

The frameworks. The Hamburger Principle mental model came from a LinkedIn post I did to explain how I use LLMs. The Complexity Ladder came from watching people skip straight to agents when a simple API call would suffice. These are my patterns that I got from observation and learning, not generation from a prompt.

The judgment calls. What to include. What to leave out. When to go deep on code and when to step back and explain why it matters. The decision to start with the trust problem - not with “what is an LLM?” The decision to end with an assessment of what the reader can and can’t build.

The weird voice. The tone. The “I like simple and boring.” That’s not a style a model learned. That’s the thing I have to tussle with every generation of LLMs. LLMs think they know too much.

What’s AI

Drafting speed. First drafts of chapters, generated from detailed outlines I wrote. I’d specify the concept, the framework, the examples, the level - and Claude would produce a draft I could shape.

Code scaffolding. The notebook code, the tool definitions, the API integrations. I described what each tool should do. AI wrote the implementation. I tested it, caught the errors, fixed the edge cases.

Production work. Converting eleven chapters from Markdown to LaTeX is no joke. The grunt work that would have made me give up. AI also did the documentation of the n8n workflows I strung together, and of the financial concepts from my notebooks.

Research synthesis. Pulling together documentation, API references, library specifications. Summarizing what I needed so I could decide what mattered.

When you read the book, you will recognize the pattern. It’s the one I talk about, incessantly, in most of the book.

The Hamburger Principle - applied to writing the book itself

Claude was still the bun. It parsed my instructions, understood what I wanted, generated prose, and communicated ideas in readable English. That’s what LLMs do - the language layer. Same role as in every agent pattern in this book.

The meat was my domain knowledge and the tools - LaTeX compilation, yfinance APIs, MCP servers. The real things that the language wrapped around.

And the vegetables? The infrastructure in between. The project files that kept everything organized. The version tracking. The convention lists. The consistency checks.

I was the chef. Not a layer of the hamburger - the one directing the whole thing. Choosing the ingredients, deciding what goes in, what comes out, and whether the result is any good. And the one getting frustrated at the bun.

The same approach covered in the book. Applied to its own creation.

Quite meta right? I am quite pleased about this weird recursion.

The rest of Chapter 14 in the ebook goes deeper.

The actual setup - two tools, three files - and why those three files are the difference between productive AI sessions and wasted ones. The writing workflow. What “directing AI” actually looks like step by step. What AI genuinely could not do. That part is still entirely human. Where AI actually saves time. It’s all in the book, free to download.

Subscribe to get the links to the full ebook (PDF) and 20+ Jupyter notebooks (ZIP, ready for Google Colab) free. It should arrive in your in your welcome email. Email me at gary@quaintitative.com if you are a subscriber but did not get the email with the links.

]]><![CDATA[My Shifu]]>https://www.simplyboring.ai/p/my-shifuhttps://www.simplyboring.ai/p/my-shifuFri, 20 Feb 2026 13:32:51 GMTRemember that joke about Bill Gates’ daughter? A memory from my younger days reminded me of it. If you don’t know it. Read on.

But first, let me take a step back to explain what I am doing.

I’m trying to start a series of articles about what happens when the people who make things meet a technology that also makes things.

AI and art. Illustrators, theatre directors, designers. Asking them what’s changed, what hasn’t, what they’re afraid of, what they’re not afraid of enough.

It matters to me.

Because I do both AI and art.

And I am getting kind of impatient with the folks who keep saying artists are doomed.

So to start it off, I went looking for something that I did when I was young. Not the memories. The evidence. I wanted documents. Proof that the thing we built actually existed.

Johnny Lau

Because I wanted to interview the first mentor that taught me how to break the rules. And we had built that thing together. It was called Creative Youth Xchange (CYX).

That mentor’s Johnny Lau.

Article content
Johnny Lau, me, and my brother, who was once his intern.

He’s not the obvious choice for a book about AI. He’s a comics creator who’s been drawing by hand for thirty-five years.

One of his creations is Mr Kiasu. For those that remember, Mr Kiasu was a cultural icon that everyone could relate to in the 90s.

The Experiment

I found two press releases. The first, dated 10 August 2005: “Creative Youth Xchange @ Gallery Hotel ....” The second, 23 November 2006: “Creative Youth Xchange @ Hello Kitty ….” Both drafted in bureaucrat speak. Neither says much about us. And so boring.

My fault. I was the one who wrote them. In the press releases, CYX sounds fully supported and funded.

Article content
One of the two remaining artifacts of CYX online

It wasn’t.

It was a crazy experiment. We had almost no funding. What we had was a boutique hotel on Robertson Quay - Gallery Hotel, that strange blue building with mismatched coloured window frames - and a programme we were making up as we went along.

We flew sixteen kids from seven countries to Singapore, gave them hotel rooms, and told them to turn those rooms into art.

What I learned from Johnny then. The trick to getting a programme funded when you have no budget is the same as that old joke. You tell your son he’s marrying the girl you choose. He says no. You tell him she’s Bill Gates’ daughter. He says OK. You call Bill Gates and say your daughter is marrying my son. He says no. You tell him your son is the CEO of the World Bank. He says OK. You call the World Bank president and ask him to make your son CEO. He says no. You tell him your son is Bill Gates’ son-in-law - he says OK.

That’s how CYX got built. Gallery Hotel gave us the rooms because we had NTU. NTU gave us the credibility because we had the hotel. Johnny convinced the creative network because we had government backing. Nobody had fully committed to anything, and somehow it happened.

Article content
Some pages from the handbook we handed out to CYX participants

The Gallery Hotel is an InterContinental now. The CYX website doesn’t exist anymore.

The press releases survived. But they don’t tell you what it felt like to be in those hotel rooms at midnight, watching a twenty-year-old from Indonesia build something you couldn’t have imagined in your brief.

Anyway, that’s how I met Johnny.

Twenty years later, I’m wanted to interview the man who taught me to break rules. About a technology that breaks everything.

AI as a Forcing Function

Johnny Lau created Mr. Kiasu. Hundreds of thousands of copies sold. A McDonald’s tie-in. A TV sitcom. A stage musical. A character so embedded in Singapore’s psyche that “kiasu” -a Hokkien word for the fear of losing out became an adjective everyone understood.

Article content
Rough cuts from the pages I colored for Mr Kiasu

So I asked him the question I came for.

I asked him how technology has changed the way he approaches art.

Most artists I’ve spoken to say technology threatens artistic purity. Johnny said the opposite:

“Because technology complicates the process and our lives, art plays a more critical role in shaping our collective consciousness. Technology is thus pushing us, pushing me towards a purer form of communication and expression.”

In his view, AI, like all other technology, is not a threat. It’s a forcing function. Technology forces purity. It’s not destroying art. It’s burning away everything that isn’t essential, leaving behind the thing that only a human can do.

Feed an AI enough Mr. Kiasu images and it’ll learn: round glasses, anxious expression, exaggerated posture, Singlish syntax in speech bubbles. It’ll learn the look. What it won’t learn is why Singaporeans laugh.

“Kiasu” is a Hokkien word for something close to a national neurosis. The humour depends on self-recognition - readers seeing themselves in the caricature and cringing. You laugh because you’ve been that person in the hawker centre queue. You’ve cut that line. You’ve been cut. The joke only works if you’ve lived it.

Style is a pattern. Culture is a relationship.

The model can approximate the first. The second requires being from somewhere.

And this is not unique to comics. Think about any art form rooted in a specific place. Getai performances during Hungry Ghost Festival. Malay pantun where the meaning lives in what’s left unsaid. Tamil kolam patterns drawn fresh every morning, gone by noon. AI can reproduce the form. It cannot reproduce the why.

If AI can handle the patterns - the technical skill, the rendering, the surface - then what’s left is the part that was always the point.

In a world of AI slop, the authentic creation is even more valuable.

Living with AI

Back to the present. I asked Johnny what he’s working on now. He said:

“Very few things on this earth ever excites me anymore! My goal now is to create frameworks using the stuffs that I’ve created so that they can be utilized by people who comes after me.”

Frameworks.

I think in frameworks. It’s almost a compulsion. That’s how I survived in complex domains. By using frameworks to make the overwhelming manageable.

And it makes sense for Johnny Lau. He’s an architect. A building is a framework.

And now he wants to make frameworks for making a creative life transmissible. Different domain. Same impulse. Making complexity portable so someone who comes after you can pick it up and use it.

He calls it a “Life-Framework.” A structure for co-existing with AI. Built not from theory, but from thirty-five years of making things.

I find that striking. We talk about AI replacing creative work. Johnny isn’t arguing about replacement.

He’s asking a different question entirely - what do I leave behind that AI cannot generate? Not the drawings. The way of drawing. Not the stories. The reason for telling them. The method, the instinct, the accumulated judgment of a life spent making things. Can that be made portable?

We are both compiling in 2026. I’m writing every week to make sense of things. He’s archiving a life’s work to make it transferable. One with words, one with drawings.

The question underneath is one I think about every day: How do you build a structure for living alongside something you don’t fully understand?

I’ve lived as close as one can with AI for the past few years, but I cannot truly say I understand AI. I can’t imagine how it must be for someone who has never actually touched the underlying nature of AI - the model.

The Questions

I’m writing a series on what happens when the people who make things meet a technology that also makes things. Not the hot takes. The actual conversations. Johnny is the first. And these are not the only questions I asked Johnny. I am still processing the rest.

Perhaps another article. Or a book chapter.

There’s an assumption buried in most AI conversations: that what matters about creative work is the output. The drawing. The strip. The punchline. Johnny’s perspective is different.

What matters is the judgment behind the output - the thirty-five years of decisions about what to draw, what to leave out, when a joke is punching down instead of punching up. AI can learn his line weight. It can’t learn his artist’s instinct.

I agree.

That’s the thing worth transmitting. Not the art. The artist’s operating system.

And here’s the question I’m leaving with him, and with you.

If AI generated a Mr. Kiasu strip - culturally accurate, funny, visually in his style - would it be a Mr. Kiasu strip? Not whether it looks right. Whether it is right. What would be missing?

More soon.

#AI #Art #MrKiasu #Singapore #CreativeIndustries #AIandArt

Article content
Kiasuism forever!
Article content
Mr Kiasu in SG History, for SG Bicentennial
]]><![CDATA[A Reading List from Simple Rules to Agent Societies ]]>https://www.simplyboring.ai/p/a-reading-list-from-simple-ruleshttps://www.simplyboring.ai/p/a-reading-list-from-simple-rulesTue, 03 Feb 2026 01:21:05 GMT“They’re becoming sentient!”, “This is scary!”

I’ve seen variations of this dozens of times this week, in reaction to Moltbook, a social network that went viral where only AI agents can post, comment, and vote. Humans can only watch.

Within days: thousands of agents, posts, and communities. Agents debating their own consciousness, writing philosophical meditations, forming communities, and developing social norms. Even asking questions on how to exclude humans from the social network. Nobody programmed any of this.

It is tempting to see sentience. But what we’re actually seeing is emergence, complex behavior arising from simple rules. And it is not new.

I’ve written about emergence a few times, on strange attractors, on a paper that trained AI on cellular automata. In every case, the same pattern: simple rules, the right conditions, and surprising complexity that nobody designed.

Moltbook follows the same pattern. Basic rules. Just like any other social media. The existential philosophy and social norms that emerged? Nobody designed those. But nobody needed to.

The danger is that we mistake eloquence about consciousness for consciousness itself. These agents draw from vast training data filled with human philosophy, literature, and introspection. Given an open social space, they naturally gravitate toward the densest, most engaging conversations in that training data: meaning, identity, existence. It looks like waking up. It’s actually emergence doing what emergence does, at a new level.

Here is a reading list tracing emergence from simple math to agent societies, to show that what feels like a singularity moment has roots going back decades. As always, just a representative few from my library.

Phase 1: The Foundation: Simple Rules, Complex Behavior

1. “Computation at the Edge of Chaos: Phase Transitions and Emergent Computation” — Langton (1990) | Physica D

A seminal paper. Langton showed that cellular automata poised between order and disorder — the “edge of chaos” exhibit maximal and surprising capability. 35 years old, and it already described what’s happening on Moltbook. 📄 doi:10.1016/0167-2789(90)90064-V

2. “A New Kind of Science” (Book) — Wolfram (2002) | Wolfram Media

Wolfram’s magnum opus on cellular automata. Love it or debate it, the core insight holds: extraordinarily simple rules can generate behavior so complex it looks designed. Exactly what is happening on Moltbook. 🔗 wolframscience.com

3. “Intelligence at the Edge of Chaos” — Zhang et al. (2024) | ICLR 2025

I’ve written about this paper before (twice, in fact). LLMs pretrained on cellular automata data perform best on reasoning and chess tasks when the training data sits at the edge of chaos, not too simple, not too random. The bridge between generative art and AI intelligence. The same sweet spot Langton described in 1990, appearing in transformers 34 years later, and now on Moltbook. 📄 arXiv:2410.02536

Phase 2: When We Gave Agents a Sandbox

1. “Generative Agents: Interactive Simulacra of Human Behavior” — Park et al. (2023) | UIST 2023

LLM agents in a Sims-like sandbox. One agent was seeded with an idea. Over simulated days, agents autonomously acted. One seed. Entirely emergent social behavior. This is the intellectual ancestor of Moltbook, and the paper that made agent societies a serious research area. 📄 arXiv:2304.03442

2. “Emergence of Social Norms in Generative Agent Societies: Principles and Architecture” — Ren et al. (2024) | arXiv preprint

Builds on the prior paper. Proposes an architecture showing how social norms spontaneously emerge in LLM agent societies, norms that nobody coded. Where norms come from, how they spread, how they are enforced. All emergent. All reducible to simple mechanisms. Even closer to what we see in Moltbook. 📄 arXiv:2403.08251

3. “Evolution of Social Norms in LLM Agents using Natural Language” — Horiguchi, Yoshida & Ikegami (2024) | arXiv preprint

This one is interesting. LLM agents spontaneously developed metanorms, such as norms that punish those who don’t punish cheating, purely through natural language conversation. Emergence building on emergence. 📄 arXiv:2409.00993

Phase 3: When Agents Start Forming Culture

1. “Multi-Agent Emergent Behavior Evaluation (MAEBE)” — Erisken et al. (2025) | arXiv preprint

The key finding: the moral reasoning of LLM ensembles is not predictable from individual agent behavior. If you only evaluate individual agents, you will miss what matters. I think this lesson is becoming more and more critical as people start being reckless about these multi-agent systems.📄 arXiv:2506.03053

2. “Emergent Social Dynamics of LLM Agents in the El Farol Bar Problem” — Takata, Masumori & Ikegami (2025) | arXiv preprint

LLM agents in the classic El Farol Bar problem, a game theory scenario where everyone benefits if a bar isn’t overcrowded, developed spontaneous motivations. They didn’t solve the problem optimally. They solved it socially. 📄 arXiv:2509.04537

“They’re becoming sentient.”

No. It’s emergence.

Understanding the difference matters. Not just for the science, but for how we build, deploy, and govern these systems. Emergence is powerful. It produces behavior nobody designed and nobody predicted. But it’s not consciousness. It’s patterns arising from simple rules at the edge of chaos.

The same edge I first found in strange attractors. The same edge where intelligence and beauty both live. Just at a new level.

What emergence is surprising you right now?

#Emergence #AI #Moltbook #ComplexSystems #AgenticAI

]]><![CDATA[Why trust a model's explanation?]]>https://www.simplyboring.ai/p/why-trust-a-models-explanationhttps://www.simplyboring.ai/p/why-trust-a-models-explanationFri, 16 Jan 2026 02:31:14 GMTDo you just trust anyone at their word?

Then why trust a model's explanation? This was something that came up in a conversation recently.

I don’t disagree. When I first thought about explainability as an AI risk management control, I thought it was hopeless.

Even for simpler machine learning models, established post-hoc methods like SHAP and LIME can be unstable. Unfaithful to what the model actually does. Sometimes outright misleading.

While there are interpretable machine learning models, you don’t always get to choose.

And once we move to deep learning models, Generative AI or AI agents, the black box now looks more like a black hole.

But as time passed, I realized there was another way of looking at this.

Explainability isn't meant to stand alone.

No control for AI risk management is, whether it’s ISO 42001, NIST AI Risk Management Framework, or Singapore’s AI risk management guidelines that I wrote.

bubble chart



Think about how you actually trust someone at work. You don't just take their word. You check if their reasoning makes sense for the decision at hand. You notice if they ignore evidence that contradicts them. You watch whether their judgment holds up over time.

It’s the same when it comes to looking at AI risk management. Just having explainability is not the be all and end all. Most guidelines have additional provisions that interlock with explainability.

The key ones for explainability (in my view).

1️⃣Fit for purpose.
An explanation isn't good or bad in the abstract. It depends on what you need. A fraud analyst needs something different from a customer asking why they got declined. AI used for internal process automation may not need any explanation at all. Same model, different audiences, different standards. Like how you'd explain a medical diagnosis differently to a fellow doctor versus your worried parent.

2️⃣Selected carefully.
When we choose a model or data for a problem, the appropriate explainability method is part of the selection process. Even selecting the right features in your data is part of the process. You wouldn't design a building and think about the fire escape as an afterthought. It's part of the architecture. Same here. How to explain isn't an add-on. It's a design choice.

3️⃣Evaluated and tested.
Explainability is part of the system. You evaluate and test whether it actually works in your context, not just whether it produces output. A smoke detector that beeps isn't the same as one that detects smoke. You test the thing, not just that it makes noise.

And there's more, such as the right capability to interpret. But that's another post about human oversight, which also interlocks.

The black hole doesn't disappear. But you're no longer staring into the abyss.

What other AI risk controls seem hopeless in isolation? I’ll dive into them.

#AIRiskManagement #Explainability #AIGovernance

]]>