Lead OpenAI researcher: "There a lot of things that we have internally that we are doing, that allow the model to work for much longer. We still didn't figure out the right product to deploy them, but the models can think for 30 minutes, hour, two hours on certain types of

Claude swooping in

Lead OpenAI researcher: "GPT-5, in some way, can be considered o3.1 - iteration of the same concept" "What I'm after right now is something next, what would be a significant jump to how we interact with models, that are more capable, think for even longer and interact with even

Claude Haiku progress: 3 > 3.5 > 4.5

Don't overthink it? No, GPT-5, please do

One is Microsoft AI another is PI ai - can you tell which one is which?

Fruit Machine simulator - testing Haiku 4.5 and all other Claude models over the last 18 months. Remember how Sonnet 3.5 was everyone's favourite? (you embarrassed now? 😉) Haiku 4.5 (last in the video) is a huge jump from Haiku 3.5 but quite a bit behind Sonnet 4.5 and Opus

Nvidia's DGX Spark goes on sale today and @lmsysorg have done a brilliant bit of benchmarking vs other systems. In short, it is a very usable system for smaller models and closer in performance to Apple's devices (e.g. Mac Mini M4 Pro), but it is priced at $4,000 vs $1,400 for

From getting ChatGPT Pulse notifications every morning, I learned that OpenAI engineers think that everyone wakes up at 8am

Nvidia's GB200 is a beast on all dimensions - raw performance, price vs performance or tokens per megawatt. The last one is particularly important in a world of constrained energy - if a data centre has only 1 GW of power, you would be 50-100% more from GB200 GPUs than any

OpenAI Infra team: would you like to go with Nvidia, AMD or a custom chip? @sama: yes

Can't tell if this is good or bad for my productivity

In case you want to listen to it 😀 voiceover with @SunoMusic, video with Sora, subtitled with @veedstudio

My short essay on LLMs & AGI: Have we reached the 'AGI of flight'? Not even close. Our planes can carry hundreds of passengers across the globe in just a few hours, yet nothing we build matches a godwit flying non-stop for eleven days without rest or the agility of a bat

Imagine one day you are working for a cool fast growing tech company, then @billpeeb pops up with Sora 2 and calls your social media behemoth a 'legacy platform' 🫠

Something I haven't thought about too much, but inference prices for open source model shift quite a bit over time - e.g. I've taken 6 models for which I had snapshots of data in August and October 2025 and variance has been quite large: ~30% drops for OpenAI models and ~60%

Lots of people never created any longer form AI videos, you can absolutely maintain consistency and with much crappier models that this one. The key feature that is missing is start & end frame, it is a pretty regular feature in many other models and I'm sure OpenAI can add it if

It normally costs ~$60m to make a Hollywood movie, with Sora 2 Pro HD it would cost you $2,700 for a 90-minute feature (or $540 for Sora 2 standard) Even if you assume 1% clip success rate, the cost would still only be $270k (or $54k). Sora-2 feels like is the closest to

It was cool to see one of my slides featured on the 'Open Models' talk by @dkundel at the OpenAI DevDay Btw, really worth checking out this and other talks they have uploaded

ChatGPT Agent: "Go to the Agent Builder and create a meme generator agent". It was magnificent, 20 minutes of work, checking documentation, created 9 notes, custom prompts, MCPs... None of it worked or was even close to working but I appreciated the effort