How do we guarantee that rays of light bend exactly the way you want them to for fisheye synthetic driving videos?
We model your camera as a precise mapping from the 3D world to pixels: intrinsics, extrinsics, lens distortion, field of view, and mounting geometry.
We then pass that info to our model to generate a video.
To verify it worked, we re-project known 3D points through the calibrated camera and measure whether they land where they should across the control and generated video.
[Sample Below]
🧵 1/4
Today, we’re announcing synthetic driving videos capabilities for ADAS.
High-fidelity, temporally consistent synthetic driving sequences designed around camera geometry, operating conditions, and safety scenarios that perception systems actually need.
A first look:
🧵 4/4
We’re also building towards programmable safety-critical scenarios: rare interactions, dangerous edge cases, situations that are difficult to capture reliably in the real world.
Synthetic video gives us a way to systematically scale these scenarios.
GPT on browser, Desktop and Codex all feel like more "complete" products than they did ~6 months ago.
UI is simple but professional, sign-in is insane (makes me excited to use the product), and they finally fixed the chat window icon alignment bug.
no no no no. your product is useless. nobody knows about it. why would you fix bugs. why would you ship features. just open your phone camera and start talking. that’s your entire growth strategy now.
you spent 6 months building something great? cool. some guy pointed at floating text with trending audio and got more signups in a day than you got all quarter.
and if someone asks ‘what’s your go to market?’ you say ‘i make reels and tiktok.’ that’s it. you got your seed round.
because who wins? not the guy with the best code. it’s the guy who did a chair spin saying ‘built this with AI in a weekend’ to 500k views.
every introverted engineer who got into this to avoid talking to people is now a content creator. nobody warned us. and the worst part? it actually works
Feels like Anthropic really shot themselves on the foot this time.
With the release of GPT-5.6 models (esp. Sol), they're going to have to release Fable full-time or they risk losing market share.
Sol is also a lot cheaper with comparable performance.
what happens when you prompt Fable to use up an entire week's worth of Claude Max 20x credits at once?
as it turns out, you end up with >50 playable games! 🤯
I realized that I had a Claude account which I hadn't used all week and its reset was the next day, so I wrote a one-shot prompt to parallelize dozens of Fable-5 agents and encouraged them to spend my entire weekly usage as fast as possible 🙃
the results were really impressive! everything was nice to look at, fun to play, and worked perfectly. many of the games bring a heavy dose of nostalgia for the Flash Games era
hard to imagine this being possible 6-12 months ago, especially from a single prompt!
PROMPT:
"ok i have a challenge for you! i need to use up my anthropic credits for the week in one day, and ONLY on fable 5!! lets see if we can't build something that leverages our claude code cli to build a bunch of epic demos for fable 5 of all sorts of different complex cool creative projects! leverage fable 5's full creativity and intelligence until our credits are gone!! tons of stuff in parallel!"
Finally switched back to @cursor_ai today and it’s phenomenal
It was the first agent IDE I used before switching to Claude Code for 9 months, then Codex for the past 3 months
I don’t want my coding workflow locked into one model company’s ecosystem anymore
I’m excited to see how much faster I can be with this new freedom
The Agents Window on @cursor_ai is not nearly good enough for when you want to keep a track of exactly what code you're generating + have complex environment requirements. Also the file management system is terrible.
I'm going back to the Editor Window (permanently).
We’ve designed and built our first AI chip: Jalapeño.
Designed from the ground up by OpenAI and brought to production with @Broadcom, Jalapeño is purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products.
Chips are foundational to the AI economy. Building our own expands our full-stack platform from products to models to infrastructure, and will help us scale intelligence, serve more people, and expand access to AI.