highfrequency

Overcomplicated take. Burn out comes from lacking a feeling of forward progress and tractability to your problems, regardless of current objective state.

A useful angle: does doing this thing make me more able to enjoy it over time (eg by increasing the subtlety and dimensionality of perception), or less able to enjoy it over time (eg by desensitization)?

This is the practical reason to favor “thick desires” over thin ones: they slope upward over time.

Per the author’s links, he warned that deep learning was hitting a wall in both 2018 and 2022. Now would be a reasonable time to look back and say “whoops, I was wrong about that.” Instead he seems to be doubling down.

Great respect for Ilya, but I don’t see an explicit argument why scaling RL in tons of domains wouldn’t work.

> But for API use, the models are easily substituted, so market share is fleeting. The LLM interface being unstructured plain text makes it simpler to upgrade to a smarter model than than it used to be to swap a library or upgrade to a new version of the JVM.

Agree that the plain text interface (which enables extremely fast user adoption) also makes the product less sticky. I wonder if this is part of the incentive to push for specialized tool calling interfaces / MCP stuff - to engineer more lock in by increasing the model specific surface area.

Sure, every business owner has incentives that point to delivering a worse product (eg cheaper pizza ingredients increase margins). For most businesses there is a strong counteracting incentive to do a great job so the customer returns next week.

The key variable is how long that gap of time is. In the online dating example, if the dating app does a sufficiently great job you will never return. A milder version: if the used car salesman gives you great value, you might be back in 10 years. This creates very weak incentives for good service, so more predatory tactics dominate.

The universal theme with general purpose technologies is 1) they start out lagging behind current practices in every context 2) they improve rapidly, but 3) they break through and surpass current practices in different contexts at different times.

What that means is that if you work in a certain context, for a while you keep seeing AI get a 0 because it is worse than the current process. Behind the scenes the underlying technology is improving rapidly, but because it hasn’t cusped the viability threshold you don’t feel it at all. From this vantage point, it is easy to dismiss the whole thing and forget about the slope, because the whole line is under the surface of usefulness in your context. The author has identified two cases where current AI is below the cusp of viability: design and large scale changes to a codebase (though Codex is cracking the second one quickly).

The hard and useful thing is not to find contexts where the general purpose technology gets a 0, but to surf the cusp of viability by finding incrementally harder problems that are newly solvable as the underlying technology improves. A very clear example of this is early Tesla surfing the reduction in Li-ion battery prices by starting with expensive sports cars, then luxury sedans, then normal cars. You can be sure that throughout the first two phases, everyone at GM and Toyota was saying: Li-ion batteries are totally infeasible for the consumers we prioritize who want affordable cars. By the time the technology is ready for sedans, Tesla has a 5 year lead.

Is GPT-5.1-Codex better or worse than GPT-5.1 (Thinking) for straight up mathematical reasoning (ie if it is optimized for making code edits)? Said another way: what is the set of tasks where you expect GPT 5.1 to be better suited than GPT-5.1 Codex? Is it non-coding problems or non-technical problems?

> 1) Prioritize your ease of being over any other consideration: parties are like babies, if you’re stressed while holding them they’ll get stressed too. Every other decision is downstream of your serenity: e.g. it's better to have mediocre pizza from a happy host than fabulous hors d'oeuvres from a frazzled one.

This is great, and applies broadly to parenting.

If AI can solve all of your interview questions trivially, maybe you should figure out how to use AI to do the job itself.

> When your throughput increases by an order of magnitude, you're not just writing more code - you're making more decisions.

> These aren't just implementation details - they're architectural choices that ripple through the codebase.

> The gains are real - our team's 10x throughput increase isn't theoretical, it's measurable.

Enjoyed the article and the points it brought up. I do find it uncanny that this article about the merits and challenges of AI coding was likely written by ChatGPT.

Fortunately, we can have LLMs write code and keep all the benefits of normal software (determinism, reproducibility, permanent bug fixes etc.)

I don’t think anyone is advocating for web apps to take the form of an LLM prompt with the app getting created on the fly every time someone goes to the url.

> posed a fascinating question: How can a relatively modest embedding space of 12,288 dimensions (GPT-3) accommodate millions of distinct real-world concepts?

Because there is a large number of combinations of those 12k dimensions? You don’t need a whole dimension for “evil scientist” if you can have a high loading on “evil” and “scientist.” There is quickly a combinatorial explosion of expressible concepts.

I may be missing something but it doesn’t seem like we need any fancy math to resolve this puzzle.

> EVM-compatible, built on Reth

Anyone know what this actually means? Both literally (what is Reth?) and what it means qualitatively: are Stripe’s crypto efforts competing with Ethereum or strengthening it?

> We manage to dominate the world mainly by using brute force to simplify our environment and then maintaining and building systems on top of that simplified environment. If we didn't have the proper tools to selectively ablate our environment's complexity…

This is very interesting and I feel there is a lot to unpack here. Could you elaborate on this theory with a few more paragraphs (or books / blogs that elucidate this)? In what ways do we use brute force to simplify the environment, and are there not ways in which we use highly sophisticated leveraged methods to simplify our environment tools? What proper tools allow us to selectively ablate complexity? Why does our intelligence only operate on simplified forms?

Also, what would convince you that symbolic intelligence is actually “harder” than embodied intelligence? To me the natural test is how hard it is for each one to create the other. We know it took a few billion years to go from embodied intelligence (ie organisms that can undergo evolution, with enough diversity to survive nearly any conditions on Earth) to sophisticated symbolic intelligence. What if it turns out that within 100 years, symbolic intelligence (contained in LLM like systems) could produce the insights to eg create new synthetic life from scratch that was capable of undergoing self-sustained evolution in diverse and chaotic environments? Would this convince you that actually symbolic intelligence is the harder problem?

> Sometimes we dislike things simply because we have a concept of ourselves as not liking them.

Astute!

To crack NLP we needed a large dataset of labeled language examples. Prior to next-word prediction, the dominant benchmarks and datasets were things like translation of English to German sentences. These datasets were on the order of millions of labeled examples. Next-word prediction turned the entire Internet into labeled data.

> without language, your thoughts are just emotions.

Is that true though? Seems like you can easily have some cognitive process that visualizes things like cause and effect, simple algorithms or at least sequences of events.

Do you have a link handy for where he says this explicitly?

> The “establishment” mistakenly assumes that a shameless person wants the approval of their community, when it turns out that, much like any cult or counterculture, that person’s goal was to attract a following, regardless of who the members are. The disgust of one’s peers doesn’t matter anymore, because that disgust forms the basis for an entirely new community.

This is a great point, and we can push it further. Perhaps the more powerful effect is that once the supporting fringe communities grow large and influential enough, the original establishment will move over to the shameless person’s camp. This happens swiftly, like dominoes falling, because the establishment’s opposition was actually not ideological to begin with but rather based on perception of the most socially acceptable / financially beneficial position at every moment.

Enjoyed the article. To play devil’s advocate, an entirely different explanation for why huge models work: the primary insight was framing the problem as next-word prediction. This immediately creates an internet-scale dataset with trillions of labeled examples, which also has rich enough structure to make huge expressiveness useful. LLMs don’t disprove bias-variance tradeoff; we just found a lot more data and the GPUs to learn from it.

It’s not like people didn’t try bigger models in the past, but either the data was too small or the structure too simple to show improvements with more model complexity. (Or they simply trained the biggest model they could fit on the GPUs of the time.)

Interesting that for these small models, it is optimal for the embedding parameters to be a huge fraction of the total (170e6/250e6) = 68%!

This is awesome - thanks for sharing. Appreciate the small-scale but comprehensive studies testing out different architectures, model sizes and datasets.

Would be curious to see a version of your model size comparison chart but letting the training continue until perplexity plateaus / begins to overfit. For example: are your larger models performing worse because they are overfitting to a small dataset, or because you are comparing model sizes at a fixed 5 minute computation time - so that the large models just don't get to learn very much in that time.

(Also interesting would be learning curve comparisons between architecture/param count)

> We can literally define an airplane parametrically in a configuration file and press a button. In a matter of minutes we have a complete quick-and-dirty analysis of how the whole aircraft performs—as mkBoom flies the aircraft through a full simulated mission (takeoff, climbout, acceleration, cruise, descent, landing). Overnight, mkBoom can run higher-fidelity simulations for a more exact understanding of performance.

Awesome stuff! Allows large scale exploration across all dimensions of plane design to jointly optimize all components and their interactions.

Reasonable argument, but preference for students whose parents went to the exact same school seems like a very inefficient way to insert more wealth and power into the student body.

Multi-tenancy likely explains the bulk of it. $10k vs. $10b gives them six orders of magnitude more GPU resources, but they have 9 orders of magnitude more users. The average user is probably only running an active ChatGPT query for a few minutes per day, which covers the remaining 3 orders of magnitude.

Great to see more private benchmarks. I would suggest swapping out the evaluator model from o3 to one of the other companies, eg Gemini 2.5 Pro, to make sure the ranking holds up. For example, if OpenAI models all share some sense of what constitutes good design, it would not be that surprising that o3 prefers GPT5 code to Gemini code! (I would not even be surprised if GPT5 were trained partially on output from o3).

Could you elaborate with a few more paragraphs? What do you mean by “working in the theory?”

> OpenAI bungling its 15-month product lead

Do you mean from ChatGPT launch or o1 launch? Curious to get your take on how they bungled the lead and what they could have done differently to preserve it. Not having thought about it too much, it seems that with the combo of 1) massive hype required for fundraising, and 2) the fact that their product can be basically reverse engineered by training a model on its curated output, it would have been near impossible to maintain a large lead.

It is frequently suggested that once one of the AI companies reaches an AGI threshold, they will take off ahead of the rest. It's interesting to note that at least so far, the trend has been the opposite: as time goes on and the models get better, the performance of the different company's gets clustered closer together. Right now GPT-5, Claude Opus, Grok 4, Gemini 2.5 Pro all seem quite good across the board (ie they can all basically solve moderately challenging math and coding problems).

As a user, it feels like the race has never been as close as it is now. Perhaps dumb to extrapolate, but it makes me lean more skeptical about the hard take-off / winner-take-all mental model that has been pushed.

Would be curious to hear the take of a researcher at one of these firms - do you expect the AI offerings across competitors to become more competitive and clustered over the next few years, or less so?