And, you know, maybe also some funny business. I think it's good to be a little suspicious of a model that happens to shoot upwards in performance on a specific benchmark while also kind of keeping up with the pack on a bunch of other benchmarks.

From experience I can say that it seems like a lot more common sense to simply replace the front wheel of your bike with a wheel containing an electric motor, rather than attaching whatever Asus is selling to your bike.

Anyway, the actual theory is that Google and Meta have fallen behind because they've been playing by this playbook of focusing on scale and training data, whereas OpenAI and Anthropic have done so well because they are likely doing much more interesting things to improve their models over time. It makes sense when you realize that one of Google's key strengths, besides talent, is that they have an incredible amount of data they can use for training due to being both the world's leading search engine as well as having all that video data from YouTube. Scaling the training data makes more sense to them than it does to Anthropic and OpenAI, who are both relatively data-disadvantaged.

You can kind of see this when you look at the Gemini 3 scorecard when it came out (https://blog.google/products-and-platforms/products/gemini/g...) and notice that while it wasn't as good as Claude And GPT at coding, it scored higher on a bunch of other non-coding benchmarks, and I think the reason why is simply because of Google's data advantage.

If true, I feel even more vindicated for believing that the "scale is all we need" narrative was bullshit.

This blog post seems to fundamentally misunderstand the nature of solo-developer SaaS, but then again arguably mostly software developers also fundamentally misunderstand it.

So I think that's maybe the other differentiator between web experiment and art, because demoscene has a very distinct but difficult to describe cultural element that makes me identify it as art.

People are experimenting with what you can do on the web, but the experiments aren't very "aesthetically inspiring". For that reason I'm kind of lukewarm on neal.fun.

EDIT: so I think a better way to describe it is that when artists experiment with technology, you get something like cachemonet. When developers experiment with technology, you get a web experiment that challenges conventional notions of what you can do with the web, but with varying degrees of creativity. I think terra.layoutit.com is best appreciated by other web devs who can appreciate the sheer amount of work required to figure out how to render a terrain map in CSS, but otherwise it's basically just a tool to generate terrain height maps, and not a particularly good one. Generating terrain maps in CSS is not a feature, but a handicap.

This is most evident in the way things collide.

It’s a little bit crazy, I know, but it’s odd to me that evolutionary forces would produce a mushroom that makes you have some specific hallucinations, rather than simply make things swirl together or simply produce intense feelings of euphoria or dread. I mean, marijuana just gets you high and that’s that.

Not all examples they gave were like this. The example they gave of the word "Typography" would have fooled me as human-made. The infographics stood out though. I would have immediately noticed that the String of Turtles infographic was AI generated because of the stylistic choices. Same for the guide on how to make chai. I would be "suspicious" of the example they gave of the weather forecast but wouldn't immediately flag at as AI generated.

Similar note, earlier I was able to tell if something was AI generated right off the bat by noticing that it had a "Deviant Art" quality to it. My immediate guess is that certain sources of training data are over-represented.

I'm going to take his statements at face value and assume that he really does have faith in his own predictions and isn't trying to fleece us.

My gripe with this statement is that this prediction is based on proxies for capability that aren't particularly reliable. To elaborate, the latest frontier models score something like 65% on SWE-bench, but I don't think they're as capable as a human that also scored 65%. That isn't to say that they're incapable, but just that they aren't as capable as an equivalent human. I think there's a very real chance that a model absolutely crushes the SWE-bench benchmark but still isn't quite ready to function as an independent software engineering agent.

So a lot of this bullishness basically hinges on the idea that if you extrapolate some line on a graph into the future, then by next year or the year after all white-collar work can be automated. Terrifying as that is, this all hinges on the idea that these graphs, these benchmarks, are good proxies.

And if they aren't, oh wow.

Haven't gotten the chance to ask, but I imagine managing a team of AI agents would feel a little too much like their day job, and consequently, suck the fun out of it.

That said, looking back, I think the reason why generative AI is so fun for so many coders is because programming has become unnecessarily complex. I have to admit, programming nowadays for me feels like a bit of a slog at times because of the sheer effort it can sometimes take to implement the simplest things. Doesn't have to be that way, but I think LLM copy-paste machines are probably the wrong direction.

If managing a bunch of AI agents is a very un-fun way to spend time, then I don’t think it’s the future. If the new way of doing this is more work and more tedium, then why the hell have we collectively decided this is the new way to work when historically the approach has been to automate and abstract tedium so we can focus on what matters?

The people selling you the future of work don’t necessarily know better than you.

Unless the main area of improvement was tools and scaffolding rather than the model itself.