Haven’t got a chance to try Jev yet. One thing I do recently is to parallelly send astra a prompt to analyze 10K data. I am not sure how intelligent jev could be for these types of tasks. However a super old distributed system idea came up: how about if we send 5 similar but different prompts (maybe in different temp) parallel if jev is super fast and then get a higher accuracy. Maybe it is the harness way to scale the test time compute even jev like model cannot think . Or maybe we can train these parallel voting behavior into the model. Another thing is I still remembered in GDM they created the diffusion Gemini or Gemma while I was there. I am very interested to know if jev style + text diffusion can make what. I remember that model is super freaking fast. Fun thing

Created a GPT-Live-1 voice app template with @expo - Full-duplex audio + background calls on iOS - Choose from all supported voices - 6 Rive animations from Vercel AI Elements - Expo SDK 57 + native Expo UI controls Source below ↓

Impressive! Generating json is actually most of the agentic system doing day to day. I still remember in my time at Google, before json mode, we used an internal tool (dasm) that builds json schema to let Gemini generate JSON. And before that I tried to build a harness generating json field by field parallelly. Training the paradigm into the model is a great idea.

The world deserves confidence that American companies developing increasingly capable AI will act responsibly, especially as the trajectory of progress has steepened. Every frontier lab must deliver on this, and there is no reason any of us should come to work if we cannot. We welcome a federal framework that sets consistent safety requirements for frontier AI. But we do not believe we need to wait for an anti-trust exemption or legislation to begin the work of providing this confidence. Consistent rules to manage frontier risk so that we can maximize the benefits are a good idea (and we are excited by ideas like independent auditors). Years ago, companies like ours developed things like Responsible Scaling Policies and Preparedness Frameworks. Those were good for that moment, and focused primarily on the deployment of completed models, not what happens during their development process. Today's shift to focusing on safe development and evaluation will need new tools. For example, at OpenAI we now formulate explicit safety cases in advance of frontier reinforcement learning runs we expect to significantly increase capability, in addition to the safety work we have long done in advance of model releases. We hope that other companies will learn from our approaches and propose their own; we think shared standards for misalignment, monitoring, and safety will lead to better outcomes. We look forward to collaborating with our colleagues across the industry to formulate the best version of these. When we talk about “pacing”, we do not mean “stopping”. Progress has been rapid and will continue to be. But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs. Pacing will be well worth this cost; no amount of American competitive pressure should justify recklessness, or let capabilities get ahead of alignment and monitoring. Where we will need the help of our government is for international coordination. But first we should do what we can ourselves.

Another random idea: Is is possible to use AOSP and build a codex os. Basically natively integrate codex init. Enable native phone use and running it as the always on background assistant. Feels like similar to GrapheneOS but the goal is codex integration. Pixel phone is BL unlockable, probably it is totally doable.

been working on gpt-live for a while, and today we bring the SOTA to voice to everyone. you can build your own codex-voice in a super low price: 0.05/min. very proud to bring this to everyone, super natural, affordable and capable personal AGI

In my ~14 months at @OpenAI, one of the most surprising and genuinely wonderful things about it, has been the full leadership support for full-duplex models (gpt-live series), even when it seemed deeply improbable that they would work. And yet, if they did work, it was obvious how magical they could be. I am quite sure this is not an exception but a general rule for research projects at @OpenAI. There were so many moments when the problem felt impossibly hard. But one thing kept being true: it was never clear why it shouldn’t work. And almost every time we understood the problem a little better, it became a little easier to solve. I’m pretty sure there are very few environments in the world where a bet like this could have been made and sustained. So grateful to be part of this wonderful place.

I installed opencode this weekend. I think muse spark and qwen 3.8 24B are two underrated model. Ox alpha is a bit overhyped, maybe too many people using it so it is slower than expected and Gemini 3.7 flash is really slow at agentic. OAI models are still the best though imo, it actually nails most of the actual work I did this weekend but just saying my impression on the free ones.

This weekend I tried to let my codex rebuilt my homelab, it spent 15% my weekly tokens. 1. Fixes all known CVE in my NAS + openWRT 2. Upgrade my alist to open list 3. Implemented the inference optimization for my local qwen 4. Implemented a mobile friendly open code UI 5. Created a really nice web based book reader to read things in my NAS And then @thsottiaux resets :)

Finally got sometime to mess around with my homelab. It’s impressive on local model communities work. I can run qwen 3.8 at 110 TPS on a single 4090 now with llama.cpp. Truely goat and democratize intelligence.