Shao retweeted
We accelerated one of the most optimized NVFP4 decode kernels while also making it more accurate
Fewer bits typically mean more speed but less accuracy. We did the opposite: more bits, but up to 10% faster generation
Already used in Ampersand’s RL and Boba serving
Year 2000, I ordered a book from Amazon and it showed up.
Year 2026, I ordered a boba from DoorDash using personal agent and it showed up.
Both are mind blowing
Shao retweeted
No one changes the world alone
Introducing Ampersand - the first model built for us. Ampersand is a collaborator trained to help people achieve their individual and collective goals. In multi-month simulations, users collaborating with Ampersand accomplished more than with Fable and other frontier models
Available to everyone through our API
Comparing main chat + side chat UI in Codex and Claude Code:
* one of them has the input box aligned
* the other doesn't
I prefer the aligned version, you?
Time to first approval for PRs
- 100 lines, 1 hour
- 1,000 lines, 10 mins
- 10,000 lines, 1 min
- 100,000 lines, 1 second
“OpenAI enterprise customers can now use existing OpenAI commitments for open models served by Baseten within Codex or via the Responses API.”
This went even further than I thought - baseten.co/blog/baseten-open…
If things play out well, this might meaningfully accelerate tokens consumed per day (TPD) through OpenAI and hopefully improves margin as well
Honestly congrats to Cognition team to show everyone what is possible - you don’t need an in house model to get to $1B ARR
Just realized that my agent is using more browser tabs than myself. They will probably come for my doom scroll time next.
Grok 4.7 is the weakest among Grok 4.7/Opus 5.5/Sol 6, yet I keep looking at my Grok bot usage and considering when do I need to upgrade.
Similarly, people will pay to use Muse and indirectly, a weaker Muse Spark 1.3 model.
Maybe this is when people start paying for the product and not necessarily the raw intelligence ?
Unpopular opinion: advancements in frontier coding models are having a smaller and smaller impact. Product innovations are having a larger impact.
Take me for example. Switching between Fable/Opus/Sol/Astra has maybe a 10% delta. A lot of that delta is likely due to prompts, skills, and harness when I am using those models, not the model itself, ie,
- Codex desktop has great computer use
- ClaudeCode is aggressive in sub-agents
On the contrary, my day-to-day life is impacted a lot more by the product innovations:
- Grok Bot that takes care of the outer loop work, incident triaging, CI/CD babysitting
- Cursor Projects promises a single coordinator, which I am extremely excited about
Coupled with big labs' call to slow frontier model progress down, the tide might be turning from model to product.
It is good news for startups without a fleet of GPUs. It's back to the classic act of talking to customers, building, and iterating.
I am hopeful for the next wave.
This is how I connect with my family in 2026 -
I know they’re around because my music stops the moment they open Spotify
Camping over the long weekend made me realize again that most people outside the tech bubble don’t worry about AGI in 2026 and instead, they focus on spending quality time with their families and friends in the nature.
From that perspective, the golden era of tech might be ahead of us for many years to come, since it would take a while to slowly enable everyone to have a better daily life with AI.
Word on the street is Gemini has been very Flash-y lately:
OpenAI: GPT-5.6
Google: Gemini 3.6 Flash!
Anthropic: Opus 5
Google: Gemini 3.7 Flash!!
Someone: Ox Alpha. Is that you, Google?
Google: Gemini 3.8 Flash!!!!!