miwojc.llm ;)

@marc_krenn @simonw Yeah there was a checkbook on that popup. Was checked by default to keep old codex app icon.

@simonw Codex auto-updated, went completely MIA, then – after re-starting ChatGPT – I was prompted with "Codex is now the ChatGPT app" but the app's name is still "ChatGPT" while showing the ol codex app icon. This is insane. Like, the hell is going on?!

I am *so confused* by ChatGPT v. ChatGPT Codex v. ChatGPT Work v. Claude v. Claude Code v. Claude Cowork right now!

@Al_Grigor @ClaudeDevs thank you very much. my weekly fable allowance was zero hahahaaah

@ClaudeDevs After 3 hours of the usual weekly reset...

We've reset 5-hour and weekly rate limits for all users.

@thsottiaux It's surprising to me that there is codex skill for adversarial code review for Claude code but there isn't for Codex app / cli github.com/openai/codex-p…

What is something that you feel is surprising that Codex still can't do well and we should have gotten right a while ago?

@thsottiaux Show session stats, like current context in main view in codex app 1M context /clear current session context Make agent follow AGENTS.md Allow for copying output 'raw' Agentsview like session browser/viewer Show answer / session cost

@simonw Superpowers' Subagent driven development skill does something similar. Although it uses more specific instructions which model to use for different cases. #model-selection" target="_blank" rel="nofollow noopener">github.com/obra/superpowe…

More notes on my blog: simonwillison.net/2026/Jul/3/jud…

The most interesting Fable tip I've heard so far is to let the model use its own judgement as much as possible I told it "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent" and it seems to be saving a lot of tokens

@mitsuhiko Like different better?

It's crazy how different fable feels in pi vs claude code. Harness really matters here.

Today, we’re excited to announce our $50M Series B, led by @GreenfieldVC (formerly TPG Capital), with participation from @lightspeed and @notablecap. 🚀 At @PatronusAI, we develop simulations and evals to train and improve AI. The first phase of AI was built on static benchmarks, but that era is over now. As agents are used to solve longer and longer tasks, they need to practice in dynamic, living worlds to get better. Simulations are the critical infrastructure powering this next phase. As a company, we’re behind the most influential research and products in AI evaluation, like FinanceBench, Lynx, and Percival. And things have moved at the speed of light since. ⚡ We partner with the world's leading frontier AI labs and enterprises, and our revenue has grown more than 15x over the past year. Additionally, today, we’re introducing a preview of the first Digital World Model for AI agent training and simulation: Patronus-DWM. Digital World Models are language diffusion world models that predict realistic environment behaviors and steer agent actions across digital workflows. Just as physical world models predict how objects move through space, we’re developing the equivalent for the digital world: predicting how agents act in digital workflows, then using that to scale the creation of high-quality training data for LLMs. Digital World Models help us push the frontier of ultra long horizon workflows, and unlock a new class of self-improving RL environments. This is our scalable approach to simulating all of the world’s intelligence. The round was also joined by @datadoghq, @SamsungVentures, @gokulr, @factorialcap, and a large cohort of amazing AI leaders and researchers across @AnthropicAI, @OpenAI, @GoogleDeepMind, @nvidia, @Recursive_SI, and more. ✨ It has been the ride of a lifetime. But we’re just getting started. The best is yet to come. "Do not go gentle into that good night, Rage, rage against the dying of the light" - Dylan Thomas (1954)

Today, we’re excited to announce our $50M Series B, led by @GreenfieldVC, with participation from @lightspeedvp and @notablecap. 🚀 At Patronus AI, we develop simulations and evals to train and improve AI. The first phase of AI was built on static benchmarks, but that era is over. As agents are used to solve longer and longer tasks, they need to practice in dynamic, living worlds to get better. Simulations are the critical infrastructure powering this next phase. As a company, we’re behind the most influential research and products in AI evaluation, like FinanceBench, Lynx, and Percival. And things have moved at the speed of light since.⚡ We partner with the world's leading frontier AI labs and enterprises, and our revenue has grown more than 15x over the past year. Additionally, today, we’re introducing a preview of the first Digital World Model for AI agent training and simulation: Patronus-DWM. Digital World Models are language diffusion world models that predict realistic environment behaviors and steer agent actions across digital workflows. Just as physical world models predict how objects move through space, we’re developing the equivalent for the digital world: predicting how agents act in digital workflows, then using that to scale the creation of high-quality training data for LLMs. Digital World Models help us push the frontier of ultra long horizon workflows, and unlock a new class of self-improving RL environments. This is our scalable approach to simulating all of the world’s intelligence. The round was also joined by @datadoghq, @SamsungVentures, @gokulr, @factorialcap, and a large cohort of amazing AI leaders across @AnthropicAI, @OpenAI, @GoogleDeepMind, @nvidia, @Recursive_SI, and more.✨ It has been the ride of a lifetime. But we’re just getting started. The best is yet to come. "Do not go gentle into that good night, Rage, rage against the dying of the light" - Dylan Thomas (1954)

There's security, and there's clankers.

@JacksonAtkinsX Yes, also other way round, all the time

My current experience with coding models.

@DannyLimanseta @elonmusk But they don't train on user data right? right? 😟 > None of your code will ever be trained on by us or any third-party. 🤥

@elonmusk This is incredibly exciting. Just curious, what’s in the cursor data? This means that the upcoming Grok model will be much better at coding?

Grok foundation model V9-Medium (1.5T) has finished training. Evals look good. A lot of Cursor data was added in supplementary training and there is more to come. Fine-tuning is underway and reinforcement learning begins in a few days. 2 to 3 weeks to public release. This will be a major improvement over the 0.5T v8-small that currently serves all Grok production traffic, especially for difficult coding tasks.

@theo Claude Code (Desktop) doesn't support plugins...

Are you still using the CLI versions of your preferred agent instead of desktop apps like Codex App, Conductor, or T3 Code? Tell me why below. Genuinely curious.

RT @PatronusAI: Spotlighting our newest benchmark for agentic search: DETOUR When people try to recall something in conversation, they rar…

@thsottiaux i don't know. plugins is still greyed out for me...

Now that the Codex app is close to being the super app. What should the super duper app do?

Excited to share that our paper, Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis, has been accepted to ICML 2026 🎉 As RL coding agents become more capable, they also become better at exploiting gaps in reward functions: passing tests, satisfying proxies, or appearing successful without actually solving the underlying task. Detecting this behavior from live training rollouts is difficult, especially when a single trajectory can look plausible in isolation. To study this, we introduce TRACE: a human-verified benchmark of 517 multi-turn trajectories spanning 54 fine-grained categories of code reward hacks. The key finding: models are much better at detecting reward hacks when they analyze trajectories contrastively, rather than one at a time. This setup fits naturally with rollout-based RL pipelines such as GRPO, where multiple trajectories are already generated and compared. In our experiments, GPT-5.2 improved from a 45% detection rate in isolated settings to 63% in contrastive settings, but the gap to human-level performance remains substantial. A few takeaways: Contrast matters: Increasing cluster size from N=1 to N=5 produced a large improvement in Match Rate across models. Semantic hacks are harder: Models detect syntactic exploits, such as test manipulation or hardcoded outputs, more reliably than hacks that require understanding intent or broader context. Model behavior varies but trends remain consistent: GPT-5.2 was the most robust overall, while Claude Opus 4.5 showed the largest gain when evaluated contrastively. Reasoning strategy matters: Models performed better when they grounded their judgments in specific code artifacts and explored downstream consequences. They performed worse when they over-relied on user acceptance or the agent's own explanations in the trajectory. Our hope is that TRACE helps the community build more robust reward functions and better detection systems for RL training pipelines. arXiv Paper: lnkd.in/guWD-dnk Hugging Face Dataset: lnkd.in/gxGcuCUf

@thorstenball moving to desktop app, similar to codex and claude code?

The best coding agent in the world is back. ampcode.com/news/neo

I wrote Deep Learning with Python to be the definitive guide to how deep learning works and how to best make use of it. Tens of thousands of people got their career start via this book. 120,000 copies sold, and downloaded by millions more. And now it's free to read online: deeplearningwithpython.io

you can outsource your thinking but you cannot outsource your understanding

@thsottiaux No image just make it to respect custom system prompt please.

Send us feature requests for codex in the form of an images 2.0 generated image. It makes it easier for codex to implement if we decide to go for it. Saw some good ones today already that codex is cooking on.

@NeurAlch @badlogicgames Meanwhile, openai doubled gpt5.5 per token price compared to 5.4 ...

@badlogicgames Codex will just win because others got out

People of pi. I'm removing Gemini CLI and Antigravity logins from pi. Welcome to 2026, the year of the end of subsidies.