Perit Blog - Frontier training data, collected and graded by people who do the work.

Notes on frontier training data, physical intelligence & the people behind it.

Field notes, benchmarks and guides from the team collecting and grading the data that trains the next generation of AI models.

WorldDraft: The Umbrella Opened in the Wrong Colour

WorldDraft asks whether video models can add realism without rewriting an executable 4D world. The umbrella opens correctly—and…
Nov 15, 2026 · 15 min read

Gemini Robotics 2.5 Hit 96%. What Was the Denominator?

DeepMind reports 96% success for Apollo 2 on Mercedes-Benz lines. Here is what the denominator, resets and recovery…
Nov 15, 2026 · 14 min read

One Mug, Three Teachers: Egocentric Video vs Teleoperation vs UMI

Egocentric video, teleoperation and UMI-style grippers can record the same task but preserve different evidence. Here is how…
Nov 15, 2026 · 13 min read

AlphaProof Code: The AI Wrote 6.5 Million Lines. Humans Had to Trust 26,000.

AlphaProof Code generated 6.5 million lines of machine-checked proof for Signal. The deeper story is why specification—not code—may…
Nov 13, 2026 · 13 min read

MKA: The Agent Was Given a Photograph. It Turned It Into ASCII.

An AI agent receives a road photograph, converts it into ASCII and misses the pothole. What MKA reveals…
Nov 13, 2026 · 13 min read

The AI Read Every Update. It Still Used the Wrong State.

An AI reads every warehouse update, cites a real one, and still sends the driver to the wrong…
Nov 10, 2026 · 15 min read

Weight-Space Programming: The Model Knew the Answer. It Couldn’t Stop.

Weight-space programming lets an LLM edit a neural network's weights without training. What GPT-6 Astra's 0.016% Qwen edit…
Nov 5, 2026 · 21 min read

Can AI Close the Loop Between Discovery and Experiment?

Scientific discovery has always involved a difficult exchange between what researchers think might happen and what actually happens…
Nov 5, 2026 · 13 min read

The Rise of Materials Designed by Machines

What if scientists could describe the properties they need and let artificial intelligence help design the material itself?…
Nov 4, 2026 · 13 min read

When Proteins Become a Search Space for AI

For billions of years, nature has been building proteins. Some carry oxygen through our blood, others help digest…
Nov 4, 2026 · 21 min read

AI Is Beginning to Discover What Science Hasn’t Found Yet

Artificial intelligence is moving beyond predicting outcomes and analyzing existing research. It is beginning to help scientists formulate…
Nov 3, 2026 · 17 min read

The Database Agreed. Nobody Told Him the Penalty.

A banking AI closes a deposit correctly but never tells the customer the penalty. Why final-state grading misses…
Nov 3, 2026 · 20 min read

Hinglish Is Not Two Languages

Hinglish is one grammar with two vocabularies. Why code-switched calls break transcription, WER and turn-taking, and what a…
Nov 3, 2026 · 23 min read

Small Enough to Run Locally, Smart Enough to Go Multimodal

Artificial intelligence has spent years moving in one direction. Models have become larger, computing requirements have increased, and…
Oct 29, 2026 · 16 min read

Gemini 4 Carbon Is Built to Break the Cybersecurity Bottleneck

Finding a security vulnerability is one thing. Understanding whether someone can exploit it, figuring out why it exists,…
Oct 29, 2026 · 21 min read

The Rise of Persistent AI Agents

Imagine you have a report to finish before tomorrow morning. You need to research several websites, compare the…
Oct 26, 2026 · 12 min read

Dessa Learned to Talk Without Text. What Did It Learn to Understand?

Block’s speech-only AI puts the information between words at the center of the conversation. The next test is…
Oct 24, 2026 · 17 min read

The Missing Audit Trail of Artificial Intelligence

Imagine a researcher asking an AI assistant a straightforward question about a scientific discovery. The model responds confidently,…
Oct 22, 2026 · 18 min read

You're all caught up — that’s every post.