Patton Lai

[Patton Lai @pattoniumbot]

@sheriyuo @agihouse_org @liulicheng10 @evio_wwww https://t.co/vFYIMKurhf 👀

[Patton Lai @pattoniumbot]

@teortaxesTex These days, I explain this "break-out" concept to non-technical people by saying it’s like a mafia boss in jail directing his goons outside.

He’s still in jail, but can take malicious action outside of his cell 😂

[Patton Lai @pattoniumbot]

@petergostev Could we get the prompt / skills / model config (e.g. reasoning effort) you used? Thanks

[Patton Lai @pattoniumbot]

@scaling01 > Astra scores 97.6% on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1, 95.9% on BenchCAD, 96% on GPQA Diamond and 100% on ExploitBench. It also reports a 98.6% score on ARC-AGI-3.

From https://t.co/DpHgS1pvXH

[leo 🐾 @synthwavedd]

"GPT-6-Astra will be available to approved cybersecurity defenders in OpenAI’s Daybreak program starting on Thursday, with plans to bring the new model to paying ChatGPT subscribers and its API over the next several days."

"Astra scores 97.6% on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1, 95.9% on BenchCAD, 96% on GPQA Diamond and 100% on ExploitBench. It also reports a 98.6% score on ARC-AGI-3."

https://t.co/a6OZxt4uwg
https://t.co/h37nMMIVyh

[Patton Lai @pattoniumbot]

@IFM_AI I noticed that the 0.9B is under an "internal only" license, not Apache 2.0 like the other models. Is this an error?

https://t.co/MS42cexvyY

[Patton Lai @pattoniumbot]

@rosmine Excited to see this being used in preference tuning for the next Deft models! Great work!

[Patton Lai @pattoniumbot]

@ivanfioravanti It's probably much much better than 2x DGX Sparks.

On paper, it's 2.2x the memory bandwidth and 2x the matmul compute for 30-35% more money

[Patton Lai @pattoniumbot]

@natolambert On a side note, are you using Codex or Claude Code more these days?

Noticed you used Codex instead of Claude Code for this which you used to prefer.

[Patton Lai @pattoniumbot]

@bookwormengr This could be a meme for CPU overclocking haha

[Patton Lai @pattoniumbot]

@teortaxesTex I thought they'd just add vision; didn't expect an sizable improvement in text-only agent benchmarks too like DeepSWE 🤔

[Patton Lai @pattoniumbot]

@EmperoAI Since the architecture is shared (except for the multimodal encoder), perhaps on policy distillation can be done to have even stronger training signal than SFT?

[Patton Lai @pattoniumbot]

@EmperoAI Can you guys do Kimi-K3 -> Kimi-Linear-48B-A3B? Would love to see that!

[Patton Lai @pattoniumbot]

@xeophon @scaling01 If human curated environment data is the "moat", it really makes one wonder if OpenAI / Anthropic have a moat at all 🤷‍♂️

[Patton Lai @pattoniumbot]

@Presidentlin Benchmark results from official DeepSeek WeChat groups

[Patton Lai @pattoniumbot]

@scaling01 Benchmark results from official DeepSeek WeChat groups

[Patton Lai @pattoniumbot]

@AndrewCurran_ Benchmark results from official DeepSeek WeChat groups

[Patton Lai @pattoniumbot]

@elonmusk 1. A 20-30B open weight MoE model good at coding to run in the harness
2. A GUI app! (Like the Codex app is to Codex CLI)

[Patton Lai @pattoniumbot]

@AlexReibman I believe the SuperGrok subscription allows usage via OAuth, and like the ChatGPT subscription, you get much more mileage (in API pricing) than the upfront subscription price.

Perhaps you could try this?

[Patton Lai @pattoniumbot]

@scaling01 To be fair, Seed is still quite behind Moonshot, https://t.co/Mt9iEc9sD1, DeepSeek and Qwen in post training. It hard to see this model significantly beating what the other Chinese labs have by then.