Max S retweeted
Tokenizer Myths and Where to Find Them!
Slides from talk last night - preprint forthcoming - feedback welcome 🙏
1/11
I recently left OpenAI.
I had an amazing time on the computer-use team. We took the models from having a hard time clicking the right date when booking a flight to shipping computer-use in the mainline series and, this week, Astra. I learned a lot from a lot of very smart people and made good friends along the way, so leaving is bittersweet.
I'm now co-founding a company that puts agents to work inside companies. If anyone is interested in accelerating agents deployed in the real world, get in touch!
Really exciting to see the Astra launch and super proud of the work we did together on computer-use agents. Huge congrats to the team!
All of the demos in this video were recorded in real-time with Astra ⭐️ It’s been so much fun making our models faster and safer at CUA with @BrookeaJoseph @fjord41 @magusnn @msieb1 @yihengxu_ @yilongqin @.alz @realtsmc @tejalpatwardhan @mckbrando @AriX and so many others!
Max S retweeted
If you tell GPT 5.6 Sol not to waste food,you'll find it eating its way through the city's dump three hours later
Max S retweeted
Like most things in life, Project Mariner only happened thanks to a legendary team of dreamers who were willing to work through the stack to make Gemini truly agentic and capable of collaborating with computers in the same way as humans! <3
Replying to @sundarpichai
We are investing in the frontiers of agentic capabilities with a few early prototypes. Project Mariner is built with Gemini 2.0 and is able to understand and reason across information - pixels, text, code, images + forms - on your browser screen, and then uses that info to complete tasks for you.
When evaluated against the WebVoyager benchmark, it achieved a state-of-the-art result of 83.5% working as a single agent setup.
One example of regulation going wrong -- removing the Maps shortcut after performing a Google search in Europe. Often I google a place to quickly find it in Maps, but now I have to manually go to Maps first.
thelocal.com/20240307/why-go…
To sanity-check for delusion, one should ask themselves what evidence or facts they would need to be presented in order for them to change their opinion on something. If the answer is the null set, this should make you question.
Max S retweeted
Gemini 1.5 Flash-8B is wild:
- 1,000,000,000 cached tokens for $10
- 1 million token context window
- Native multimodal support (image, audio, video)
- Code Execution, Structured outputs, etc
Truly in a class of its own
Max S retweeted
Today, we’re excited to release two new, production-ready versions of Gemini 1.5 Pro and Flash. 🚢
They build on our latest experimental releases and include significant improvements in long context understanding, vision and math. → dpmd.ai/4ebNJ5x
Never deleting this app xitter.cf/chatgpt21/status/18289…
Replying to @adcock_brett
You have no product? That is even remotely at minimal manufacturing, and I still can’t get a flight on any of archers products. How dare you compare figure to Anthropic and open ai and say you look “cheap” 😭 ship one robot to a consumer then say that
Max S retweeted
Everyone knows that @NoamShazeer is a genius, but one thing that most people don't know is how kind and honest he is. His treatment of *all* employees in the events today was exceptionally fair and generous. 10/10 would work for him again.
I don't get all the hate I see about @paulg . His essays are probably the most widely used source for long context LLM needle-in-haystack reasoning benchmarks.
@Tesla The parking garage is a static environment. If Autopark cant reliably navigate in that scenario, at least abort the maneuver if you cannot guarantee collision avoidance. Stop offering those features until they work. And don't trick customers into thinking they work.
















