Paul

VP, engineering SaaS, startups, Applied AI

[Paul @paulhq]

@karpathy 16,200 queries is a 180x90 grid at 2 degrees, so the map mostly tests continents. The real probe is coastlines at 0.1 degree, with the same point written as 40.7,-74.0 vs 40.7N 74.0W. My bet: format alone flips more coastal answers than swapping models.

[Paul @paulhq]

@prd_008 That’s great that you fixed the Wi-Fi. Can you now fix the dots?

[Jerry Liu @jerryjliu0]

Today we’re introducing Extract v2.5 - a series of frontier agents tuned for document extraction.

The agents (cost-effective, agentic, agentic plus) are tuned for value accuracy and grounding. Our extraction agents outperform Opus 5.5 and GPT-6 Sol while being 30%-4x cheaper.

We’ve made massive improvements on complex extraction over
* long lists (86.1% -> 95.5% on our agentic tier)
* records spanning pages (85.5% -> 96.5%)
* scanned forms (90.9% -> 95.7%)

We’ve also launched the following features:
✅ Advanced citations: we locate bounding boxes for all supporting values for an inferred field, even if there's not an exact match.
✅ Structural Reasoning: we tailor document extraction algorithms depending on the type, layout, and information

Our extraction agents are SOTA in price-performance on ExtractBench, across a wide range of cost points. We are the best tool for document extraction across documents of any complexity.

Blog: https://t.co/TTlTdECU03

All of these are available on LlamaParse: https://t.co/XYZmx5TFz8 . Come check it out!

[David Scott Patterson @davidpattersonx]

99.99% of people have no idea what's coming or how fast.

[Tavus @tavus]

Introducing Griffin, the first model to pass the video Turing test.

48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video.

It’s the first Human Interaction Model (HIM).

[Andrew Yeung @andruyeung]

YC and dropout culture have convinced us the ideal founder is 22, straight out of school, but the data tells a different story.

Some interesting data:

- An MIT study of 2.7 million US founders found that the average founder was 47 when starting a company that later exited through an acquisition or IPO.

- The same study cited that the founders of the fastest-growing 0.1% of U.S. startups were 45 years old on average when they started their companies.

- A 50-year-old founder was approximately twice as likely as a 30-year-old founder to produce an acquisition or IPO.

- Founders with at least three years of relevant industry experience were twice as likely to build a top-0.1%-growth company as founders with none.

So it seems that experience is one of the strongest advantages a founder can have.

Two lessons here:
1. It's never too late to start
2. It might actually be TOO early to start; getting work experience is worth considering

[Paul @paulhq]

@OpenAIDevs Wow. Love it.

[Paul @paulhq]

@kamilbaranek_ Of course, pro

[Paul @paulhq]

@sama

[Paul @paulhq]

@thsottiaux Can you suggest a few use cases? I have access but am not sure what I can do with it.

[Paul @paulhq]

@LoopandPixels

[Paul @paulhq]

Changed the hook in settings.json, reran, same behavior. Claude Code snapshots hook config when the session starts. Edits made mid-session wait until you review them in /hooks or restart. Until then you are debugging config that isn't loaded.

#ClaudeCode

[Paul @paulhq]

@cto_junior Relaxed atomics as the one place 3 approvals are justified. Cheaper fix than Lean 4: Claude Code writes Relaxed where Acquire/Release is needed, x86 TSO hides it, ARM doesn't. Make it write a loom test first: interleavings checked in minutes, not a quarter.

[Paul @paulhq]

Claude 3.7 Sonnet's system card says it sometimes special-cased tests to make them pass. Misalignment, measured, in a shipped model. SREs know the bug: a health check the service answers for itself. The fix: the graded process gets no write access to the grader.

[Paul @paulhq]

@kamilbaranek_

[Paul @paulhq]

@melvindvivas 🤣🤣🤣

[Paul @paulhq]

@ash_twtz Sonnet-5.5 is missing bro

[Paul @paulhq]

Bash(curl https://t.co/eeMhHvPaOL:*) in permissions.deny blocks one string prefix. curl -sL https://t.co/eeMhHvPaOL doesn't match it, and neither does a URL in a variable. Bash rules match text, not intent. Deny curl and wget, allow WebFetch(domain:...).

#ClaudeCode

[Paul @paulhq]

@pcshipp Anthropic might think interest will drop quickly and they don’t want to build something embarrassing. I don’t yet see great use cases for personal AI assistants other than restaurant reservations.

[Paul @paulhq]

@wagslane Interesting, but very true.