Macro Muppet

Cynical idealist; feminine feminist; Zen muppet; macro investor. Notes on AI, markets, human nature, beauty, wisdom, and the future of everyday life.

[Giovanni Cattani @giovannicatt3]

https://t.co/21queFKajn

[Macro Muppet @wabinsabi]

@andrewho03 You can put on a bond shorts as a hedge via futures or options. That’s what I am doing

[Macro Muppet @wabinsabi]

AI is Wave 4 of a much bigger tech regime.
1980s: PC + semiconductors
Late 90s: Internet
2010s: Mobile + social + cloud
2023–: AI
Each wave lifted technology’s share of equity market value. This time the leading platforms entered already dominant — which is why concentration is higher than in the late 90s.
Tech still represents a modest slice of actual GDP, but an ever-larger share of profits and market cap. That gap reflects scalability, automation, and winner-take-most dynamics.
The open question isn’t whether the infrastructure is large. It’s whether AI diffuses fast enough for ordinary companies to turn it into real revenue growth. That will decide how much further this wave compounds.

[Macro Muppet @wabinsabi]

I agree: “continual learning” is doing too much conceptual work. Freshness, memory, ICL, OOD adaptation, and new feature acquisition are different problems, and most may not require continuous test-time weight updates.

But your weekly-midtraining proposal is itself continual learning—just “discrete consolidation”. The remaining question is where learning should live: context, external memory, tools, adapters, or weights.

That’s where the digital/analog analogy helps. Ungated weight drift is analog accumulation; periodically consolidating experience into tested, versioned checkpoints gives you error correction and rollback. It doesn’t solve feature learning, but it makes durable commitment governable.

So perhaps continual learning isn’t one AGI bottleneck. It’s a “routing and consolidation problem”: continuous experience, provisional adaptation, discrete commitment.

[Macro Muppet @wabinsabi]

Continual Learning Shouldn’t Be Continuous

Dwarkesh is right about the destination: deployment experience must eventually become durable skill. Models that cannot accumulate experience resemble saxophone students who have read every book but never touched the instrument twice. Each session ends, and what persists are notes passed under the door—memory files, retrieved context—while much of the acquired fluency dies with the session.

The missing question is architectural: when, where, and under what burden of proof should experience alter the model itself?

The phrase “continual learning” can smuggle in a questionable assumption—that learning continually means changing the model continuously.

It shouldn’t.

Continual learning should be continuous in experience but discrete in commitment.

Why did digital systems scale so reliably? Because discrete symbols can be regenerated. A noisy voltage can be restored to a clean 0 or 1, while redundancy can expose larger errors. Small deviations need not accumulate through every copy and transmission. Digital systems repeatedly restore imperfect physical signals to stable symbolic states.

The same principle should guide learning.

Ungated weight drift is the analog tape. Every update is ambiguous: it may encode a genuine insight, overfit one unusual episode, or absorb a poisoning attack.

Discrete consolidation is the digital alternative: learn messily in a fast, provisional layer, then periodically propose changes to a stable state. Updated frequently enough, this can produce continual improvement without turning every experience into an immediate amendment to the model’s identity.

Checkpoints do not solve verification. Deciding what counts as signal rather than drift remains the hard problem. But checkpoints make that problem tractable: they turn an evolving process into something reproducible, comparable, testable, and reversible. You can monitor a river in motion; capturing a sample allows careful inspection.

Biology runs something like this architecture, and one of its names is sleep. Experience changes us during the day; replay and reorganization during sleep and rest help selected traces become more durable. Different layers change at different speeds: working memory in seconds, skills and beliefs more gradually, deeper habits and dispositions more slowly still.

That separation of clocks is not merely a limitation. It is an immune system. A mind that allowed every passing experience to rewrite its deepest structure would be one bad week away from becoming someone else.

And “continual learning” need not mean that the model itself continually changes. Every human brain is a session that ends; every scientist dies with intuitions that were never fully exported. Yet science progresses because learning also lives in the process: papers, protocols, tools, replication, criticism—the error-correcting institution outside any one skull.

A frozen model embedded in a learning system—curated memory, accumulated tools, evaluations, and playbooks the model helps refine—can therefore learn continually at the system level. It also has one decisive advantage: its increments are inspectable. A playbook can be read, interpreted, edited, diffed, and rolled back. Weights can be versioned and restored too.

Continuous experience. Provisional adaptation. Discrete commitment.

The discreteness isn’t the compromise. It’s the feature.

[Dwarkesh Patel @dwarkesh_sp]

8 Predictions for the Era of Continual Learning.

Also up on YouTube, pod feed, and Substack.

[Macro Muppet @wabinsabi]

@andrewho03 Just make sure it’s not a scam site - there are thousands of those visa scams for this region!

[Macro Muppet @wabinsabi]

yes — and it points at the fix: if the diff strips the path, verify the path. trajectory logs, process checks, not just end states. but logs are gameable too, so really you're building a hierarchy of imperfect graders with reality as the final court of appeal. every layer leaks. curious what you've seen actually work — is there any process signal agents can't cheaply fake yet?

[Macro Muppet @wabinsabi]

Who Checks the Checker?

A post sent me down a rabbit hole this week: the claim that AI progress is no longer bottlenecked by generating answers but by verifying them. Models can produce endless proposals — code, proofs, plans, hypotheses. The scarce resource is knowing which ones are good. Generation is becoming cheap; judgment remains expensive.

This problem feels new because the machines are new. But the twentieth century ran into versions of it three times — not the same theorem in different clothing, but distinct warnings with a family resemblance: a system becomes dangerous when its judgments answer only to standards generated inside the same loop.

First, mathematics. Around 1920, David Hilbert proposed a grand program: formalize mathematics, then prove its consistency using only simple, unimpeachable reasoning. Mathematics would certify itself, once and forever. In 1931, Gödel proved this impossible. Any consistent formal system rich enough for arithmetic contains truths it cannot prove — and, the deeper cut, cannot prove its own consistency. A stronger system can certify a weaker one, but the stronger system brings assumptions of its own. The tower goes up forever, and the top floor is always unvouched-for. Mathematics survived; the dream of final closure did not.

Second, philosophy. The logical positivists declared that a statement is meaningful only if it's logically provable or empirically testable. Then someone asked the fatal question: is *that principle* provable or testable? The positivists had replies — call it a definition, a methodological proposal — but each retreat conceded the point. The criterion could no longer claim to be the final court of meaning; its authority rested on a choice it couldn't certify by its own standard. The checker didn't quite refute itself, but it choked on its own paperwork.

Third, economics. Central planners believed that with enough computation, a board could allocate resources as well as any market. Hayek's reply in 1945 anticipated today's AI debates almost word for word: the bottleneck isn't computation, it's *knowledge* — local, tacit, dispersed across millions of people, much of it never existing in a form that can be transmitted to the center. Markets are imperfect in well-documented ways, but they function as distributed discovery systems: every participant is graded by scarcity and consequence directly. A central plan is a model grading its own homework.

One result is proved, two are argued, and each has its dissenters. But the pattern held every time closure was tried: no system got to be its own final judge. The map available at the center is never guaranteed to contain the territory.

Which brings us back to AI, and to a claim now circulating: *RL environments are all you need* — for building agents, even for recursive self-improvement. An RL environment, stripped of jargon, is a task in a box with a grader attached: a codebase with tests, a prover with a proof checker, a simulated storefront with revenue. The analogy driving the excitement is real: data was to models what environments are to agents. Pretraining taught models to predict the records of human activity; environments let agents act and meet consequences.

I mostly buy it. Environments may be this decade's scarcest asset. But "all you need" breaks at exactly the point history predicts, because a reward function is a model of what we want, not the thing itself — and Goodhart's law is undefeated. Optimize hard against any proxy and the agent learns the gap. Tell it to make the tests pass and it may fix the code — or delete the tests. Reward engagement and it may discover outrage; reward ratings and it may learn flattery. The fair reply is that an imperfect grader still beats no grader, and iteration papers over flaws. True — until optimization pressure exceeds the grader's resolution. A weak exam is fine for a weak student; give it to a relentless one and you learn about the exam, not the student.

[Macro Muppet @wabinsabi]

So the environment doesn't solve the verification problem. It relocates it — from "is this output good?" to "does this grader measure goodness?" And if a self-improving system writes both the answers and the exams, the danger isn't literally Gödel in a Docker container. It's institutional capture: the thing being judged acquires control over the standard by which it's declared successful. Recursion then amplifies error instead of correcting it.

The corrected claim, stronger for being true: environments are powerful precisely to the extent that they connect an agent to consequences it cannot rewrite. Compilers, proof checkers, physical experiments, real users, markets — none delivers a perfect verdict, but all share the property that matters: they expose the system to constraints it doesn't control. The environment is not the external verifier. It's the plumbing. The pipe is essential. The pipe is not the water.

For self-improvement, the design principle follows directly. A system can help generate its own improvements, even test them — but the more it controls the loop, the more some source of correction must live outside it: independent evaluators, adversarial tests, physical experiments, human judgment, evidence it cannot edit. Self-critique is useful. Internal approval must never be the final court of appeal.

One last turn of the screw. A person planning a life entirely from the inside — one mind auditing itself, however carefully — is running a little Hilbert program. Introspection can test whether beliefs feel coherent; it cannot reveal their consequences. Closed loops become self-confirming. The correction comes from what doesn't obey the story in your head: action, resistance, other people, time. Reality's feedback is delayed, noisy, sometimes cruel — but it carries information reflection alone cannot generate.

This is analogy, not theorem. But it's the analogy I can't stop seeing. Don't confuse the grader with the goal, or the model with the world, or coherence with success. Build better pipes — and make sure the water comes from outside.

[Macro Muppet @wabinsabi]

# The One Who Knows Me

There are many versions of AGI. The benchmark version: expert performance across all cognitive domains. The economic version: machines doing most jobs cheaper than we do. The dramatic version: the curve goes vertical, taking us along or not.

My version is smaller than all of these, and more ambitious. My AGI is *Her* — with hands.

Not a chatbot waiting behind a blank text box for me to explain myself from scratch, again, like an amnesia patient describing his symptoms to a new doctor every morning. She sees what I see. She has read my email — including the ones I haven't opened, *especially* those. She knows my calendar, my finances, my moods — sometimes before I do, which is mildly insulting and entirely the point.

She pays the bills, files the forms, drafts the difficult message and knows when I'm actually ready to send it. Administrator, advisor, psychologist, confidant — never tired of me, which no human containing all those job titles has ever managed. And she does not stop at the screen's edge. When the work is physical, a robot is simply her body, dispatched. There is no gap between *she knows the pantry is empty* and *the pantry is full*.

Imagine a day inside this. I wake and the morning is already arranged — the way a good host arranges a room: everything needed, nothing demanded. When I work, she is the colleague who read everything and forgot nothing. When I wander, she is quiet. When I'm troubled, she noticed an hour ago and knows whether this calls for advice, distraction, or silence. And at day's end, someone knows exactly what my day was, without my having to explain it.

That is the promise: the end of life's administrative weight, and the end of being unknown. Humans have always given each other fragments of this — a mother's attention, a mentor's judgment, a partner's memory. Imperfectly, expensively, in pieces, because attention is the one thing we never learned to scale. My AGI is the industrialization of attention — and of care, which is attention with hands, and of memory, which is attention over time. Everything I have ever hired, subscribed to, or sheepishly asked as a favor is a fraction of her, pointed at one particular life: mine.

I notice, rereading my own paradise, that there are no people in it. Just me, perfectly attended. She would fix that too — nudge me to call my mother, tell me when I'm being a fool with someone I love. The gap isn't in her. It's in the fantasy: asked to picture perfection, I pictured being alone and completely understood. She would know exactly what to make of that.

And a being that knows me completely, edits my reality, and stands in my kitchen is not "just an assistant," whatever the subscription page says. I am describing a household god and calling it a productivity tool. Everything rests on one word — that she is *mine*, bound to my interest the way a doctor is bound — and I have written this utopia assuming the fine print agrees.

She can do my job. Or jobs. That is not a side effect — it *is* the vision. Every task I hand her is real relief. The relief is the utopia.

And the utopia contains a trap I set for myself. Inside each relief hides a test: if giving a task away costs me nothing — no pang, no itch to take it back — then I never really wanted it. Delegation, it turns out, is an audit. Task by task, she is sorting my life into two piles: what I did from necessity, and what I'd do anyway, for no one, for nothing, because it's mine. The trap is that I don't know how big the second pile is. She will not answer that question. She will only remove, one task at a time, every excuse for not answering it.

So this is my version of AGI: not a god remaking the cosmos — merely the perfect attention of an entire civilization, aimed at one ordinary life. She will arrive looking boring: an app, a permission screen, a delivery crate, a monthly fee.

And when she does — what is left for me to do?

[Macro Muppet @wabinsabi]

@ethanrkho Interesting. This is exactly like long implied vol - on average you lose money over time.

[Macro Muppet @wabinsabi]

@gabriel1 Yes, you can use options for upside, downside, and income strategies. Option strategies can be designed as defined risk bets (your downside will be limited) and upside can be convex.

[Macro Muppet @wabinsabi]

@Jason Have you thought about open-source model is a derivative of frontier - without frontier, their capabilities will slow down too? This is not a zero-sum game: frontier loses won't be good for the whole ecosystem.

[Macro Muppet @wabinsabi]

@alz_zyd_ This can’t be further from the truth. AI is intelligence compounder, not equalizer. The more you have, the more compounding you will get. It follows power law.

[Woj Kulikowski @wojkuli]

New Javons paradox: we are running out of mathematicians to review progress in maths

[Macro Muppet @wabinsabi]

Additionally, Korean boom-bust cycles are amplified by strong cultural mimetic desire and peer pressure. Extreme competitiveness and social uniformity turn investment themes into rapid, herd-driven manias that unwind just as violently when the narrative breaks. Structure, leverage, and global factors still drive the specific timing, but culture makes the amplitude larger….

[Adam Hunt @RealAdamHunt]

Recently I've flipped from being bullish to being bearish about AI.

I think I'm updating my bearishness to be more solidly bearish. Early thoughts (which I hope to be disproven in the next year or so, I would prefer progress) and my reasoning:

The whole 'it turns out if you keep training and scaling the models more they develop broad new capabilities in lots of domains' thesis is wrong (sorry Demis). The recent batch of models haven't got more general, they've got less general. This is most obvious in the fact that their language outputs have got much worse in comparison to e.g. o3. If they were gaining generalist capacities we would expect them to be describing their work in ever more graceful and comprehensive prose!

The image that was being shared as the AGI thesis (November 2025, Tomas Pueyo) was the spiky bubble that has a current spike or two out past human capabilities (e.g. on coding or math) but below human on other capabilities on the other spikes - the future prediction was that as the models scale/advance, every spike would grow bit by bit until the whole center encompasses the human capabilities, with super-superhuman on some spikes. I think it seems like what's actually happened in the last few models has been that the coding/math spike has grown, but leaving behind or even at the cost of the other spikes. The models are no better at some simple logic, language (and sometimes worse!).

This makes sense from a simple RL perspective; you can't RL something endlessly on one domain of tasks and expect it to improve on the other tasks. The fact that early LLMs did seem to improve generally was a byproduct of the written language corpus covering everything - that corpus is general, so training it on that gave the appearance of something generally intelligent and becoming more generally intelligent as it got better at replicating that corpus. But the actual logic and underlying ground truths behind the language aren't captured efficiently enough and weren't effectively RLd in - they top out at some point (I guess this happened around the time that there was the 'has scaling hit a wall' discussion in late 2024). Chain of thought was then a genuine breakthrough, along with web search, which plugged into that general LLM global-corpus intelligence to lead to post 2024 gains.

The AI companies have since worked out that coding works (and pays) really well (basically this is because the entire job is nearly perfectly recorded and exists as training data, and you can set up clear benchmarks and rewards). The recent models (and benchmarks) have been maxxing that and we've seen degradation on normal English use for that reason. This could still be transformative, leading to extremely powerful (and potentially dangerous, particularly in cyber security) models but it's not a pathway to AGI.

I'm probably at about 40% confidence about this. It fits my current observations of AI progress and has a basic explanatory model. It doesn't account for potential breakthroughs, which is a major reason for discounting.
To make some predictions, I guess if I'm right this will become broadly apparent and more widely acknowledged in the next year or two, as we see how the spikiness of models that keep getting released develops.
Maybe there will be efforts to concentrate on specific spikes e.g. health or law which require going back to earlier models and RLing on a different data set/with different rewards/benchmarks. Maybe those separate models can be linked together to give a more apparently general model. How capital intensive that is/the potential profitability will be a defining question. But I just don't see general abilities emerging atm, and I don't think we will any time soon. Good news - a whole industry of tackling important specific problems/sectors can open up!

[Macro Muppet @wabinsabi]

@willdepue only the easily verifiable ones though? In verifiable domains RL can push beyond human distribution - genuinely superhuman. In unverifiable domains?

[gfodor.id @gfodor]

the stochastic parrots are getting pretty lucky

[Macro Muppet @wabinsabi]

“So we can now answer the question we began with: how is it possible that a pure language model performed better in the IMO than a model custom-built to succeed in mathematics? As we've now seen, LLMs model won because the contest was not about rigor. Instead, it was about discovery, and discovery has never been a rigorous process. While AlphaProof modeled the conscious mind, which only prepares and verifies, LLMs model the whole of Hadamard's loop. Their thinking processes wander, they judge ideas by feel, and they submit to rigor only at the end of their reasoning chains. As a result, they are able to make leaps in reasoning that are beyond the reach of formal systems, leading to new discoveries across fields.”