AI Daily Digest — August 1, 2026: ARC-AGI-3 Harness Discovery, EU AI Gigafactories, Devin SWE-1.7

Short summary

OpenAI found GPT-5.6 Sol's poor ARC-AGI-3 score was a harness problem, not a model problem — enabling retained reasoning and context compaction tripled the score to 38.3%. The EU opened calls for up to seven AI gigafactories backed by €10B+ in public funding, barring non-EU entities from consortia. Cognition launched SWE-1.7, a frontier coding model scoring 42.3% on FrontierCode 1.1, alongside Devin Security Swarm which found 72% of real vulnerabilities in a 50-CVE test set.

• •ARC-AGI-3 scores depend heavily on harness settings; retained reasoning tripled GPT-5.6 Sol's score
• •EU committing €10B+ public funding for up to 7 AI gigafactories, restricted to European consortia
• •Cognition's SWE-1.7 hits 42.3% on FrontierCode 1.1; Devin Security Swarm finds 72% of real CVEs

Generated with AI, which can make mistakes.