This is not that impressive, there are numerous examples of browsers for training data to reference.

I don't buy this.

It implies that the agents could only do this because they could regurgitate previous browsers from their training data.

Anyone who's watched a coding agent work will see why that's unlikely to be what's happening. If that's all they were doing, why did it take three days and thousands of changes and tool calls to get to a working result?

I also know that AI labs treat regurgitation of training data as a bug and invest a lot of effort into making it unlikely to happen.

I recommend avoiding the temptation to look at things like this and say "yeah, that's not impressive, it saw that in the training data already". It's not a useful mental model to hold.

▲

deadbabe 4 hours ago | parent [-]

It took three days because... agents suck.

But yes, with enough prodding they will eventually build you something that's been built before. Don't see why that's particularly impressive. It's in the training data.

▲

simonw 2 hours ago | parent [-]

Not a useful mental model.

▲

deadbabe 2 hours ago | parent [-]

It is useful. If you can whip up something complex fairly quickly with an AI agent, it’s likely because it’s already been done before.

But if even the AI agent seems to struggle, you may be doing something unprecedented.

▲

simonw 2 hours ago | parent [-]

Except if you spend quality time with coding agents you realize that's not actually true.

They're equally useful for novel tasks because they don't work by copying large scale patterns from their training data - the recent models can break down virtually any programming task to a bunch of functions and components and cobble together working code.

If you can clearly define the task, they can work towards a solution with you.

The main benefit of concepts already in the training data is that it lets you slack off on clearly defining the task. At that point it's not the model "cheating", it's you.

	▲	aix1 an hour ago \| parent [-]
		Simon, do you happen to have some concrete examples of a model doing a great job at a clearly novel, clearly non-trivial coding task? I'd find it very interesting to see some compelling examples along those line.

▲

embedding-shape 7 hours ago | parent | prev | next [-]

Damn, ok, what should I attempt instead, that could impress even you?

	▲	anonymous908213 2 hours ago \| parent [-]
		Actually good software that is suitable for mass adoption would go a long way to convincing a lot of people. This is just, yet another, proof-of-concept. Something which LLMs obviously can do, and which never seems to translate to real-world software people use. Parsing and rendering text is really not the hard part of building a browser, and there's no telling how closely the code mirrors existing open-source implementations if you aren't versed on the subject. That said, I think some credit is due. This is still a nice weekend project as far as LLMs go, and I respect that you had a specific goal in mind (showing a better approach than Cursor's nonsense, that gets better results in less time with less cost) and achieved it quickly and decisively. It has not really changed my priors on LLMs in any way, though. If anything it just confirms them, particularly that the "agent swarm" stuff is a complete non-starter and demonstrates how ridiculous that avenue of hype is.

▲

usef- 5 hours ago | parent | prev [-]

What would be impressive to you?

	▲	deadbabe 4 hours ago \| parent [-]
		A browser so unique and strange it is literally unlike anything we've ever seen to date, using entirely new UI patterns and paradigms.