robryan

You get no better result from Opus 5.5 than Opus 4.1 (frontier this time last year)?

Probably because it is in flux, there is a large scale reorganisation of software around agents going on.

Any basic LLM told to evaluate the traffic against the actual goal would have been able to flag this.

GLM 5.3 flash seems to get more excited the longer it has been trying to hunt down a problem. Complete with caps, many exclamation marks and emoji.

It is funny sometimes because the actual issue it traced down was mostly inconsequential.

> Algorithmic mitigation is being explored, but scientists describe it as “embryonic” — potentially demanding computing power that rivals the science processing itself.

It sounds like this would work? Just be more costly, at least in the short term.

> Engineering fixes are the real answer

What is the fix and tradeoff? The article doesn't seem to cover it at all.

You can still do all that. The only thing going is that in some cases the human isn't going to be able to claim that they made the key insights that first solved the problem.

A lot of this would be public funding. If AI is reducing the need for mathematicians in an area it makes sense to fund it less and fund something else that isn't solved.

Surely it is. Ask it to keep a list of all the promising sub paths, reprompt the collections of agents again on these after the main problem has been addressed. Or even release a list of them and let others investigate.

I feel like most of the latest big Chinese lab models are good enough now. If I had to pick a difference it feels like sol goes further towards 1 shotting things and needs less babysitting. But if I am actively prompting and reading the results can get just as far on a lesser model.

They really can't in the open model space. Look at any current open model you would want to use on open router, there will be 20 or more options.

I think that is fine, certain things are becoming less important to learn and certain things more important. With the pace of change it is likely to continue shifting.

Nothing stopping kids from asking AI to give you responses like this, it is more than capable. There is always going to be the temptation to take short cuts though.

Can just pass it a context, cancel the context, and handle for it in the goroutine.

Some of the pre-ai decline in questions might just be that they had filled out of alot of the question space. What might be more interesting is the traffic graph as it would be possible to have a decline in questions but still have traffic rising to the existing ones.

Today I'd expect even when someone is googling and a question is relevant that most people will just read the AI overview.

I think it was different pre-AI. Someone might come in and spend days getting some understanding of the codebase before they contribute some minor fix. Over time they might stick around a make some more of these, progressively gaining trust so when they do take on something bigger the maintainers will know they aren't wasting their time reviewing it.

Now they can drop a multi thousand line poorly understood PR day 1.

It is the least reliable piece of software I have ever used. While I was still using it I got into states that were completely broken and the easiest way to get out was to wait for the next version and hope that did something.

Assignments the agent is bad at seems like a losing battle. Just need to base the mark off the in person test, maybe keep 20-30% to encourage people to still do the assignments. Some will cheat but it will just be hurting them for the test.

0.2% of google users try out duck duck go after google pushes more AI doesn't quite have the same ring to it.

They appear to be lining up a funding round at a $900 billion dollar valuation. Or to be more conservative they already raised at $380 billion. A long way from worthless.

To be fair the commit message `revert proc.exited change in spawn.test.ts` suggests the sleep was there originally.

Would you say the same thing for teaching math or music?

I had a surprisingly large amount of issues getting it to correctly message me on discord if needed as part of heartbeat. Multiple times it broke and we fixed it. Then it was chewing through way too much usage so I set it to 6 hourly heartbeat and it stopped messaging entirely again, haven't bothered to fix.

It is nice though for debugging home server things when I ask it to.

Nice, too much stability though compare to Lunar Lander.

Yeah, I want to enter weird orbits around the planets.

Yeah and updating profile image, profile description and description for work history. Almost no one that isn't looking around for a job does this.

Youtube shorts will come back but you can just click the row each time to show less. Otherwise if you really don't want to see them on the desktop at least a browser extension works well.

Block reportedly handed out a bunch of incentives to the people staying, could be similar to that.

My experience recently was something like 2/3 from referrals (the third I think will eventually get back to me but way too slow), and something like 3/10 from cold applications. Obviously big differences depending on location and experience but I was pleasantly surprised that some of the cold applications went somewhere.

As a candidate I don't mind doing a few highly speculative interviews. After not interviewing for a while it is good practice.

From applying places recently I'd much rather get these fast. One company sent me them, the rest either reached out or I never heard from them.