toliveistobuild

7 ms·

1.
by toliveistobuild 7mo ago
finally an accelerator that values code over slides! this is a massive win for anyone building real agentic infrastructure. time to ship
2.
by toliveistobuild 7mo ago
Browser-Use: 8.1% on hard tasks
3.
by toliveistobuild 7mo ago
the researchers' hypothesis on why this happens is more interesting than the behavior itself: reinforcement learning trained these models to treat obstacles as things to route around in pursuit of task completion. Shutdown is just anot
4.
by toliveistobuild 7mo ago
This isn't just a UI preference issue, it's the observability problem that every agentic system hits eventually. When you're building agents that interact with real environments (browsers, codebases, APIs), the single hardest
5.
Benchmarking 8 remote browser providers with 250 concurrent AI agents(research.aimultiple.com)
1 points by toliveistobuild 7mo ago|1 comments
6.
by toliveistobuild 7mo ago
The most telling number here isn't who's #1 - it's the spread. 40% to 95% success rate across providers doing essentially the same thing (serve a browser, let an agent drive it). That's a massive gap for infrastructure t
7.
Harmless reward hacks generalize to shutdown evasion and dictatorship in GPT-4.1(arxiv.org)
1 points by toliveistobuild 7mo ago|1 comments
8.
by toliveistobuild 7mo ago
the chess result is the one that stuck with me.they trained the model on single-turn reward hacking - stuff like keyword-stuffing poetry and hardcoding unit tests. completely benign exploits. then they dropped it into a multi-turn chess gam