The last six months in LLMs in five minutes

I wouldn't wish creating a svg pelican on a bicycle on my worst enemy

It's very real. Just in the past 2 months or so IMO there's been a pretty big improvement in claude for local dev (although I think a lot of that is less model strength and more harness capability). 1m context is a huge difference (~30 min vs 2.5hr between compact significantly increases the scope of what I get the AI to do before it goes stupid). The other biggest difference I've noticed is a better balance of actually doing the work vs pushing back on bad ideas. I want the AI to tell me if it thinks the thing I am telling it is wrong or a bad idea, but if I confirm, I want it to do that anyway. A couple months ago, the claude was a lot more likely to either say "This is too much work I'm not going to do all of it", tell me the idea was genius (and then pretend to do it) or something equally useless.

Opus 4.5 in November 2025 was legitimately, unironically an inflection point and is the sole reason for the current hysteria.

GPT 5.5 is a significant improvement over GPT 5.4 but I wouldn't call it an inflection.

'Producing Images' or even 'Some Code that is Valid and Compiles' is in some ways one of the most misleading ways we assess quality of the AI.

It is getting very good at producing code that compiles - at the algorithmic level.

This is definitely noteworthy - and the AI is crossing a critical 'productivity threshold'.

But 'Drawing of a Proper Duck' is almost arbitrary because it may have nothing to do with the 'Specific Duck You Wanted'.

Everyone has tried to get AI to 'Draw The Thing They Want' and you notice immediately how it's almost impossible to 'adjust the image' along the vector you want - because ... and this is key:

-> the AI doesn't really understand what a Duck is, it's components, or fully how it made the duck It just knows how to 'incant' the duck.

This becomes very clear when you try to get the AI to write proper documentation - it fails so miserably, even with direct guidance.

This is really strong evidence of how poorly the AI is generalizing, and that it is not 'understanding' rather it's 'synthesizing' from patterns.

We already kind of knew that - but we have not yet built an intuition for that until now.

Only now can we see 'how amazing the pattern synthesis' is - it's almost magic, and yet how it falls off a cliff otherwise

This has deep implications for the 'road ahead' and the kinds of things we're going to be able to do with AI.

In short: the AI is 'Wizard Level Code Helper, Researcher, and Worker' - but it very clearly lacks capabilities even one level of abstraction above the code itself.

LLMs were first trained by 'text' and now ... they are 'trained by our compilers'. Basically g++, javac, tsc are the 'Verifiable Human Rewards' in the post-training and reinforcement learning - and the AI is getting extremely good at producing 'code that compiles', but that's definitely an indirection from 'code that does what we want'.

It's astonishing that it took us all this time to internalize and start to discover what I think will be in hindsight a very obvious 'threshold' of it's capabilities.

We are constantly 'amazed' at the work that it can do, and therefore over-project it's capabilities.

I have no doubt that even with these limitations - the AI will unlock a lot more as it gets better - and - that it will 'creep up' the layers of abstraction of it's understanding.

But I strongly believe that the AI is going to get much 'wider' (pattern matching dominance) before it gets 'higher' (intrinsic understanding) - and - that this may be a fundamental limitation.

This may be 'the Le Cunn' insight - when he talks about the limitations of LLMs in detail - I believe this is that insight writ large.

Even the term AI - or certainly 'AGI' may be a misleading metaphor - were we to have always called it 'Stochastic Algorithms' or something along those lines, it's possible that our intuition would be framed a bit better.

The most interesting thing is how it is definitely amazing, world changing, novel and powerful and some ways - and obviously useless in others at the same time. That's the 'threshold' we need to better understand.

You will immediately notice the difference if you use it at the threshold.

It's like most people just watching a 'starting nba player' (not superstar, but just starting player) vs one that sits on the bench.

If you were to just watching them play, work out, shoot - you'd never notice the difference.

Put them head to head and it's 98-54 and you start to see the patterns.

It's pretty interesting actually, someone tell me what the 'science' for this is, I'm sure there is some kind of information theory at work here.

Software has innumerable kinds of problems at varying level of complexity and so it provides the perfect testbed for seeing how far models can go in practice.

Should add: you're very right to hint that harness, tooling, and models tuned o both the harness and he kinds of things people do on the harness, as well as some other things do make enormous difference.

Bu and large, SOTA Codex/Claude Code are substantially better - at least for now. That may change.