Claude Opus 4.7 solves its first ProgramBench task 👀 After creating a 663-word CLAUDE.md we were able to increase solve percentage by 1.1 points on Claude Code Claude Opus 4.7 across 10 ProgramBench, and actually solve one of them. You can replicate our results here: github.com/parsewave/Program… big thanks to @KLieret @jyangballin for releasing such a forward-thinking benchmark!

singapore eats part one; hainan chicken rice; insane buffet of crab, skewers, noodles, chicken; you name it, at Newton, food coma afterwards

For my eval-maxxing nerds out there, good friends of mine are running a series called "strange evals", you can benchmaxx now on anything. If in SF swing by! luma.com/lvqbs1mo

Happy to be working with meta for the past 7 month, and seeing the fruits of their labor. Great release!

No major benchmark is designed for COBOL, Fortran, or Assembly - the languages powering trillions in transactions and infrastructure that must be modernized or risk catastrophic failure. We built Legacy-Bench to measure frontier agents on the code the world actually runs on.

The real-time web search capability is Nano Banana 2's secret sauce. Midjourney and DALL-E train once and call it done. NB2 pulls from the live internet every time. The outputs reflect what's happening now, not what the world looked like months ago, pretty genius That image-sequencing demo is pretty great. Each frame only sees the previous frame so there's no cheating, no outside reference like a world model would have. Yet the visuals stay consistent. That's the model understanding cause and effect in visual space. Coherent video generation is the natural next step, and it's closer than people think.. Also pretty nice to see them roll it out all at the same time, API, platform, aistudio. Looks like only text model releases are somewhat staggered at google. Really good launch

Horizons are LITERALLY too long to METR at this point, look at those error margins!

Claude is beginning to annoy me a lot. Their webUI forces the model to create docx files using JS whenever I ask for some structured output for a moderately deep research. Is this a way to rack up more input tokens and wear out my 5 hour limits much faster?

Unbelievable that we are actually approaching saturation on both ARC AGI 1 and 2.

I think we should really stop taking SVG tests as a benchmark in the first place. This seems easy to benchmaxx and doesn't denote any sort of real life performance whatsoever

AI is going to hit the finance world like a freight train

pack it up bro

The Turing test is just undercover customer support?!

My current development stack: - OpenCode (the best harness ever) - Kimi K2.5 on the moderato code plan, incredibly capable model - OpenClaw (for remote server management and running ralph) - ralph-manager skill for openclaw we made internally which spawns ralph instances with nice monitoring - tmux for terminal multiplexing - nvim for making code edits if needed - termux for accessing my running tmux sessions from my mobile phone - tailscale to connect to my devices remotely from anywhere with ease The fact that this is even possible right now is crazy. Long live open source man

what

It's weird seeing a lot of "uncollapsible" systems collapse with the advent of AI. And it's all thanks to the volume content is coming out, not AI itself. Social media algorithms? Destroyed. "Slop" coming out so quickly is flooding social media at rates never before seen. We can't tell who's a person, who's an agent and who's a bot. Likes and shares no longer define if the content was actually authentic and valuable or not. Copyright and IP laws? Cooked. At the rate at which violations are coming out thanks to image and video models it is simply unenforceable now. I don't think this is an AI problem. These systems banked on the low user base to function. Likes and shares worked because there were just a few people posting on the platform. Copyright laws worked because only a few people could violate it and it was largely enforceable at that scale This means we shouldn't try and abolish AI to make these systems work. It means we need to work around it to setup robust systems which don't break under volume

GLM 5 is crazy on benchmarks. HLE even with tools is hard to benchmaxx on, and still they have the highest score in there. Did not expect it to be this good on Terminal bench 2.0 at all, a monster of a model for sure China is cooking

"AI hasn't made anything novel" "AI hasn't innovated anything" "AI is a stochastic parrot" AI's pushing barriers of human knowledge at the bleeding edge fringes like this one. You're just not paying attention

Opus 4.6 is unreasonably good at resolving merge conflicts. I know the meta is Codex 5.3 rn, but just putting this out there.