- Meta AI's July 7 post details the first non-invasive MEG brain-to-text pipeline to match implant-tier accuracy on freely composed sentences.
- 61% average word accuracy across participants, 78% for the best subject — the previous non-invasive benchmark sat at 8%.
- Trained on ~22,000 sentences from 9 volunteers, each recorded for 10 hours typing under a MEG scanner.
- HN thread hit 98 points and 52 comments; the standing catch is that MEG hardware is still a room-sized scanner, not a headset.
- Commission announced July 7 in Brussels; tech chief Henna Virkkunen confirmed the plan will not be accompanied by any new legislation.
- Three pillars: safe/responsible AI use, stronger cyber defense via NIS2 and the Cyber Resilience Act, and scaled European AI capabilities.
- ENISA gets a mandate to build a 'European Blueprint' for secure access to advanced AI systems plus a test platform for energy, transport, health, finance, and public administration.
- Launches an EU Grand Challenge on AI for cybersecurity to seed native tooling — a hedge against the bloc's dependence on U.S. frontier models.
- TeraWulf press release: 401 MW critical IT campus at the Justified Data site in Hawesville, Kentucky, built on a former aluminum smelter.
- Initial capacity in service second half of 2027, full 401 MW ramp by early 2028; deal value exceeds TeraWulf's ~$12B market cap.
- TeraWulf simultaneously divests its 50.1% stake in the Abernathy (Texas) JV to a Fluidstack-led group to focus on the Anthropic build.
- Second megawatt-scale Anthropic compute deal in a week after last week's SpaceX partnership — Anthropic is buying capacity ahead of Fable 5's post-grace credit-only pricing that starts today.
- Kyiv's Ministry of Digital Transformation told Reuters on July 7 it will procure only self-hostable models for government, business, and military use.
- Chief AI Officer Roman Kyslyi cited the U.S. Commerce Department's forced pause on Anthropic access as the trigger; remote-only APIs are 'interim' at best.
- Diia government app currently runs on Gemini via EU-hosted servers with personal data stripped before every query.
- Kyivstar is building a domestic model on Google's open-weight Gemma for autumn rollout across government, private sector, and defence.
- First peak/off-peak schedule on a frontier open-weight API: 9am–12pm and 2pm–6pm China time bill at 2× the off-peak rate.
- V4-Pro is 1.6T total / 49B active params; V4-Flash is 284B/13B — both stay MIT-licensed, both graduate from the April preview.
- Preview cost curve set the tone for the 'margin collapse' thesis making the rounds on HN: V4-Pro API is roughly a fifth of Opus 4.8 pricing at similar SWE-bench Verified numbers.
- Legacy names deepseek-chat and deepseek-reasoner shut off July 24 — migrate before then.
- Martin Alderson essay argues Z.ai's open-weight GLM 5.2 matches Opus/GPT quality on agentic coding at ~15–20% of the price.
- Thesis: task-routing across GLM 5.2, DeepSeek V4, and Qwen 3.6 captures the margin the frontier labs currently earn on Opus-only workflows.
- Caveats acknowledged in-thread: no vision, higher token counts from heavier thinking, slower interactive latency.
- HN discussion is the developer counterpart to today's frontier-lab compute megadeals — the closed labs are locking in supply just as open weights close the quality gap.
01
Meta Ships Brain2Qwerty v2 — Typing From Thought at 61% Word Accuracy, No Implant
research ai.meta.com
02
EU Unveils Cybersecurity + AI Action Plan — Implementation Only, No New Law
industry commission.europa.eu
03
Anthropic Signs $19B, 20-Year Kentucky Data Center Lease With TeraWulf
industry terawulf.com
04
Ukraine Says It Will Only Buy AI Models It Can Run Itself
industry reuters.com
05
DeepSeek Locks In V4 GA for Mid-July With Peak-Hour API Pricing
models deepseek.com
06
'GLM 5.2 and the Coming AI Margin Collapse' Hits HN Front Page
community news.ycombinator.com