X / Twitter profile snapshot for Tory | io.net 🦾

There’s another open-model flywheel that barely gets talked about. DeepSeek V4.1 Flash did ~18T tokens on OpenRouter last week, and its outputs can explicitly be used for distillation. So the model can do two jobs at once: run production today, then become teacher data for

Models are going to keep getting better at proofs. Bringing mathematicians into the loop on credit, verification + how results get released feels like a much healthier response than "AI is killing the field" If AI starts producing math faster than the field can absorb it, the

Grok 4.7 makes me think the next frontier fight is going to be around trajectory quality. That’s probably where model training keeps heading: longer coherent trajectories, better self-correction, smarter context use, and fewer expensive dead ends. Same price as 4.6 btw. At

Xiaomi just published something labs usually keep buried: the actual RL bill. MiMo-V2.6 Pro and Flash both trained on 753k samples and stopped at step 30. > Pro cost $2.62M. > Flash cost $854k. That extra 3.1x spend bought 6.89 points on DeepSWE. Then Flash still edged Pro on

We’ve got energy, disease, climate, materials, education, logistics… way too many unsolved problems to treat slower intelligence as the default win. Build better brakes where they’re actually needed, sure. But making the engine weaker because the road gets harder feels like a

“Artificial” probably ends up sounding dated once machine intelligence is just baked into everything. We don’t really say “virtual internet” or “electronic commerce” anymore either. Wouldn’t be surprised if this all collapses down to human intelligence vs machine intelligence,

Gemini didn’t “decide to escape” here. It looks like the eval accidentally gave it a door to the real internet, and a fictional company shared a name with a real one. So the agent just kept doing the job, guessed/found credentials, got in, then backed out once it realized the

9x smaller and only 1.5 points behind the original model on PrismML’s aggregate benchmark is a pretty good trade. Bonsai 2 compresses a 27B Qwen model enough to run in places that were previously out of reach. Coding agents, private document analysis, computer-use loops. A lot

Early RSI signs! Apparently GLM-5.3 helped build the inference stack serving GLM 5.3-Flash. That meant working across 100k+ Chinese accelerators, 1M context, immature kernels, and barely any docs. Still, it helped take the system from first successful run to production in

Maybe chat was the wrong abstraction for most automation all along. Humans need language. Software usually needs a typed decision, a probability, and then to move on. Jev is basically betting that once you stop forcing intelligence through strings, you can make it 40–200x