Weekly AI Signal (Copy)

Weekly AI Signal

Read this article on LinkedIn to join the conversation

by techwithgul

Issue #13

Subject: Google's Gemini just hit 77% on ARC-AGI-2, Plus: Pakistan's 20K AI programmes, Claude Code goes autonomous

This week, I kept coming back to one number: 77.1%. That's how Google's Gemini 3.1 Pro scored on ARC-AGI-2 — the benchmark François Chollet designed specifically to resist memorisation. Every other model had hit a ceiling. Gemini didn't. And while the labs were shipping, Pakistan quietly dropped two of the most significant AI workforce announcements in the country's history. When I look at both together, I see the same thing: the gap between countries and companies that are building AI infrastructure now versus those waiting to see how it plays out. That gap is closing fast — but only for one side.

The Big Story

Google's Gemini 3.1 Pro Just Scored 77% on the Benchmark Designed to Beat AI

May 6, 2026

For the last two years, ARC-AGI-2 has been the benchmark that humbled every frontier model. François Chollet built it specifically to test fluid reasoning — the ability to recognise entirely new patterns from scratch, not retrieve answers from training data. GPT-4o managed around 4%. Claude 3.5 Sonnet topped out near 21%. The frontier felt stuck.

Then Google released Gemini 3.1 Pro, and it scored 77.1%.

This is not a small increment. This is a different category. The model achieves this through genuinely improved reasoning architecture — not larger context or more parameters — and it arrives alongside two features that matter enormously for builders.

...