Live Feed
Google Ships Gemma 4 QAT Checkpoints: E2B Drops to 1GB on Mobile With Custom Quantization Schema
Google DeepMind released quantization-aware training checkpoints for the full Gemma 4 family today, cutting the E2B model's memory footprint to 1GB with a purpose-built mobile schema. QAT variants outperform standard post-training quantization at the same bit width across all sizes.
EU Formally Launches Tech Sovereignty Package: Chips Act 2.0, €200B Data Centre Build, and €2B for Open Source
The European Commission unveiled four linked legislative proposals on June 3 targeting a tripling of EU data centre capacity to 65GW, a doubling of EU semiconductor market share, and the first €2B open-source funding strategy framed explicitly as sovereignty infrastructure — not a cost-saving measure.
rsync's Creator Used Claude to Rewrite Its Test Suite. Then Came 329 Comments and a Death Threat.
Andrew Tridgell — the engineer who wrote rsync in 1996 — used Claude, Codex, and Gemini to rebuild rsync's test suite and fix CVEs. The regressions in 3.4.3 triggered one of open source's ugliest AI debates. An independent analysis tried to separate the data from the rage.
Pentagon's 'La Tilde' Ran AI-Generated Propaganda Across Latin America — Disclosed June 2
The US Department of Defense operated an AI-powered content generation program called La Tilde targeting Latin America, The Intercept revealed. The disclosure arrives as the Trump administration pushes aggressive military AI expansion while senior uniformed commanders urge guardrails first.
WIRED: Meta Silently Embedded Face-Recognition Code in Millions of Phones for Its Smart Glasses
WIRED found unreleased facial recognition code built into Meta's smart glasses platform — internally called 'Name Tag' — already deployed in apps on millions of phones. Meta previously paid $7B+ in biometric privacy settlements and explicitly planned to launch during political turmoil.
LEAP Solves All 12 Putnam 2025 Problems in Lean — General LLMs Beat Specialist Provers With Agentic Framework
Google's LEAP paper uses only general-purpose Gemini 3.1 Pro with an agentic DAG framework to achieve 100% on Putnam 2025 and 70% on a new IMO-style formal benchmark — surpassing every specialised theorem-proving model without any fine-tuning on formal corpora.
Anthropic Data: AI Task Horizons Doubling Every 4 Months, 8x Code Output — Global Pause Urged
Anthropic's Institute publishes previously unreported internal data showing Claude-powered engineers shipping 8x more code per quarter than the 2021-2025 baseline, with autonomous task horizons doubling every four months. The WSJ reports Anthropic is calling for a global pause.
Arena Launches Agent Leaderboard With Causal Tracing — Grok 4.3 Ranks Last, GPT-5.5 High Leads
Chatbot Arena has replaced pairwise voting with causal tracing from real in-the-wild agent sessions, producing rankings that diverge sharply from chat leaderboards. Grok 4.3, which places high on chat, finishes last on agent tasks. GPT-5.5 High and Claude Opus 4.7 Thinking lead.
OpenAI Ships Dreaming 2.0: Asynchronous Memory Synthesis Fixes ChatGPT's Staleness Problem at Scale
OpenAI is rolling out a revamped memory architecture for ChatGPT that synthesizes memories automatically in the background, tackling the staleness, correctness, and scale failures of saved memories. Plus and Pro users in the US get it today; Free and international rollout follows.
Alphabet's AI Capex Is Collapsing Its Free Cash Flow by 89%. Amazon's Goes Negative. The $2 Trillion Revenue Gap.
The five largest hyperscalers will spend $700-900B on AI infrastructure in 2026 while AI revenue runs at roughly one-sixth of what Sequoia calculates is needed to justify the buildout. Alphabet FCF drops from $73B to an estimated $8B. Amazon projects negative free cash flow for the first time in its cloud era.
Meta Launches Business Agent to All 1B+ Daily WhatsApp Business Threads — the Enterprise AI Play Nobody Was Watching
Meta's Business Agent goes globally available on June 3, deploying AI across WhatsApp, Messenger, and Instagram where more than 1 billion customer threads already run daily. The platform connects to hundreds of enterprise systems including Shopify and Zendesk. Free to start.
Reve 2.0 and Ideogram 4.0 Both Ship Spatial Layout Control on the Same Day — Image Generation Finally Knows Where Things Go
Two leading image labs shipped bounding-box layout control on June 3. Reve 2.0 claims the world's best 4K model with precise spatial editing. Ideogram 4.0 went open-weights and debuted at Arena #8 overall, #1 among open models, with outsized gains in text rendering and commercial design.
Alibaba's Fun-Realtime-TTS Takes AA Speech Arena #1 at Elo 1,219 — Edges Google by 5 Points
Alibaba's Fun-Realtime-TTS moved to the top of Artificial Analysis's Speech Arena leaderboard with an Elo of 1,219, edging Gemini 3.1 Flash TTS (1,214) and Inworld Realtime TTS-2 Research Preview (1,209). It is Alibaba's first #1 finish on the Speech Arena and the second major leaderboard this week where a Chinese lab has taken the top spot.
35% of Berkeley CS 10 Students Failed in Spring 2026 — 1,100 Faculty Sign Petition as AI Cheating and Math Gaps Hit Together
UC Berkeley's CS 10 failure rate tripled to 35.3% in a single semester, with Professor Dan Garcia naming AI-assisted cheating as the primary driver. Simultaneously, over 1,100 faculty across the UC system signed an open letter citing students arriving unable to do middle school math — two compounding failures that won't fully surface in graduation statistics for years.
Qwen3.7 Plus Hits ScreenSpot Pro 79.0: Alibaba's GUI+CLI Hybrid Agent Enters Frontier-Tier Computer Use
Alibaba launched Qwen3.7 Plus on June 1 with ScreenSpot Pro 79.0 — placing it inside the frontier band for GUI grounding — and Terminal-Bench 70.3 for agentic coding. At $0.40/M input via OpenRouter, it is the cheapest path to frontier-class computer use in the current market.