Pinned Tweet
logical reasoning is one of the key capabilities AI should have before it could do useful works for us reliably. in feb I benchmarked the top line AI models with Zebra Logic benchmarks and o3-mini-high was the only model that could solve the 6*6 puzzles. 3 months past, a lot new models have been released, how are they fare against the Zebra Logic benchmark?
I have tested o3, o4-mini, Gemini-2.5-Pro and QWen 3 235B (released yesterday), and their results are ...
Pyre retweeted
Prime Agent is a general-purpose coding harness
On ARC-AGI-3, it scores 95.5%, surpassing the human-expert baseline, but the gain is not benchmark-specific.
We see major improvements across models when compared to their proprietary harnesses:
Pyre retweeted
Robert Scoble interviews me outside YC after the summer 2010 Demo Day. By that point we'd funded what seemed the enormous number of 208 startups, including 36 in that batch.
youtube.com/watch?v=UacbJ72d…
Pyre retweeted
1-bit GLM-5.2 GGUF vs. Claude 4.8 Opus vs. GPT-5.5
We gave 3 models the same prompt and compared one-shot outputs.
The 1-bit GLM-5.2 GGUF ran locally on a Mac Studio M3 Ultra with 256GB RAM at ~21.6 tok/s.
Which output do you like best?
GGUF: huggingface.co/unsloth/GLM-5…
GLM-5.2 can now be run locally!🔥
The 2-bit model retains ~82% accuracy after we shrunk it from 1.51TB to 238GB (-84% size).
Run on a 256GB Mac or RAM/VRAM setups.
GLM-5.2 is the strongest open model to date.
Guide: unsloth.ai/docs/models/glm-5…
GGUF: huggingface.co/unsloth/GLM-5…
Pyre retweeted
Actually insane that you can print this.
Flexible 3D printing resins are becoming stronger with Formlabs new Flexible 80A Resin V2💪
You can see that this geometry, printed in Flexible 80A Resin V2, snaps back into place and withstands repeated bending, tension, and compression. Compared to its previous version, the new material delivers 2x higher tear strength, 2x higher elongation at break, 4x higher rebound, and improved aesthetics and aging performance.
Pyre retweeted
Introducing autoresearch for arXiv papers
Change 'arxiv' to 'autoarxiv' in any paper URL
An agent deploys to resolve setup issues on the codebase, run a minimal reproduction, and estimate full replication cost. Read more below
Pyre retweeted
It trained on 10,000 real combat videos before it ever flew.
Ukraine's SkyFall has unveiled the P1-SUN Long - an AI-enabled interceptor drone trained on more than 10,000 real combat interception recordings. It detects aerial targets up to 800 meters out, then tracks and pursues them automatically. A human operator still has to confirm before any engagement.
This is what drone warfare looks like four years in: systems that learn from their own combat history to get more precise.
Pyre retweeted
I'm going to show this to every founder who refuses to launch early.
xitter.cf/broadglow/status/20670…
Introducing LifeSciBench, a benchmark for measuring and improving how well AI supports real-world life science research.
Developed with 173 scientists from biotechnology and pharmaceutical research, LifeSciBench includes 750 expert-authored tasks across seven biological research workflows.
openai.com/index/introducing…
Pyre retweeted
X should put this label on fully generated written posts, not just images
Same idea. Transparency about AI-generated content protects trust, rewards authenticity and originality, & prevents timeline pollution
Anything over Pangram 80% automatically gets the label
Pyre retweeted
Yet another study shows a 24% reduced risk of dementia after the Shingles vaccine. This one in over 500,000 participants with a recent skilled nursing facility stay, adding to 4 huge natural experiments in 4 countries (US, Canada, Wales, and Australia)
acpjournals.org/doi/10.7326/… @AnnalsofIM
Pyre retweeted
The world’s first seaglider factory is complete!!
A quarter million square feet dedicated to bringing our product to the world.
Today: A blank canvas
By this time next year: Viceroy and Squire seagliders on the assembly line
1 Seaglider Way
Game on
Pyre retweeted
A lot of engineering goes into making autonomy look simple
Here’s a 2 hour timelapse of F.03 repeatedly walking up and down stairs. Figure HQ is full of tests like this, each one helping push robots closer to fully autonomous systems
Pyre retweeted
We've always told startups to launch early, but I realized there's a powerful new reason to.
Before you launch, the speed you can build is now mainly limited by your imagination in what you tell AI.
After you launch, the AI can watch your users and make improvements on its own.
Pyre retweeted
Western intelligence gave Kyiv 72 hours on Feb. 24, 2022.
Ukraine is now at day 1,569 of the war — longer than World War I.
Today, it fields the world's first Unmanned Systems Forces and produces about 4 million drones a year, The Telegraph. 1/
Pyre retweeted
One interesting pattern with Fable 5 is that it will often say things that are gibberish when I use it for coding. Things like "The morning's slim-scan fix cured the scan hang", "this is a latent-drift API-shape wrinkle", etc.
When I ask why it does this, Fable explains that it invents codenames while reasoning about the problem, then fails to realize they're meaningless to me. Its neuralese is blending into its output because of a theory-of-mind failure about what's in its head vs. mine.






































