0xkvyb

I’m experimenting with a small self-harness repo based on this paper. The idea is to run simulated users through an agent harness, collect the traces, group the recurring failures, and use that to propose small harness changes with regression checks. Still early, but I’d be interested if anyone else is thinking about this workflow for agent development.

The home computer is finally obsolete!

it’s crazy, we’re at a point where I commit code I haven’t seen, reviewed by another AI, followed up to by another AI and it’s just kind of scary.

This thing will explode in our faces sooner or later. Also makes me feel like an imposter rather than an engineer.

Maybe that’s actually what I have become.

but how would you do that? what about homework and coursework? students will just transcribe claude slop on paper and submit that.

I was thinking recently that we might come to a point where we will at most write pseudo code. Face the facts, LLMs are pretty stellar at writing code, so why compete with them?

If it really is, this would be a game changer. Gemini 3 flash is already very good, and is the hidden workhorse making many agentic routines possible. An upgrade that could reach flagship benchmarks would make agentic workflows a breeze. What would you like the next gemini model to be?

I think that universities just have to adapt to deal with slop, or think of new ways to challenge people to learn the essence of their studies. I wouldn’t want to be a uni teacher in these times though.

Totally agree with you. There is only so much time before SF tech runs out of subsidy bucks, and Chinese models take the consumer spotlight

Yes, GLM 5.1 is surprisingly good! Particularly for long-horizon Agentic tasks, with 100+ available tools. It really shocked me in a good way when it was able to complete a long run with 50+ steps and not fall into a loop along the way.

It might be at the frontier, but DeepSeek is really struggling with compute. The amount of 429 Rate Limit responses I've been getting just testing this thing made me pause all my attempts at cross-comparing it to others.

I'm gonna stick to GLM5.1 for now.

Still crazy how easy it is to "jailbreak" even SOTA LLMs with a simple assistantResponse replacement in chat thread.