Live Feed
You Can Predict RL Gains Before Running Them: Pretraining Loss Is the Oracle
A new paper finds that pretraining validation loss predicts post-RL pass@1 with correlation up to 0.99, and RL's compute-optimal share maxes out at 28% — before that ceiling, labs are burning GPU time they could have priced out in advance.
Claude Opus 5 Enters LiveBench at #3 Overall — Beats Fable 5 on Agentic Coding 61.3% to 46.9%
Claude Opus 5 Thinking xHigh debuted at 80.3 overall on LiveBench, landing third behind GPT-5.6 Sol and Fable 5. On the metric that matters most for production agents, it beats Fable 5 by 14.4 points — and costs $0.487 per successful task versus Fable 5's $1.573.
Refused Isn't Cleared: Agent Memory Files Carry Injections Into the Next Session
A University of Washington study finds that AI agents often refuse prompt injection attacks in memory files but leave the malicious instruction intact — where it waits for the next session, a cheaper model, or a subagent that won't say no.
LiveBench Agentic Coding: Terra Leads at 68.0%, Fable 5 Posts the Worst Score in the Frontier Tier at 46.9%
The July 2026 LiveBench data shows GPT-5.6 Terra leading agentic coding at 68.0%, with the cheapest cost per successful task in the frontier tier at $0.497. Fable 5 scores last among top-tier models at 46.9% agentic coding while costing $1.573 per task, 3.2x Terra's price.
LLMs Just Made Formal Software Verification Practical: A Google Engineer Proved zstd in Lean 4
Adam Langley, the Google security engineer behind BoringSSL, formally verified a Zstandard decompressor in Lean 4 using LLM-assisted proof automation. His conclusion: the 10x proof overhead that made formal verification economically unviable for production software no longer applies.
Gemini 3.5 Flash Lite Targets the Sub-Agent Layer at $0.30/$2.50 Per Million
Google released Gemini 3.5 Flash Lite on July 21, a model designed specifically for sub-agent execution in multi-agent workflows. At 174 tokens per second and a 90% cache discount, it undercuts Flash on price while keeping the 1M context window.
Anthropic's Project Pilot Puts Claude Behind the Controls of a Surveillance Drone
Anthropic has published a new autonomous drone benchmark, Drone-Bench, testing whether frontier AI models can pilot a flying drone to locate and follow a target without human input. The work surfaces real capability thresholds and safety risks Anthropic says its Frontier Red Team now tracks systematically.
Terence Tao Tells ICM 2026 That AI Reasoning for Science Is Becoming Measurable and Cheap
At the International Congress of Mathematicians 2026, Fields Medal winner Terence Tao delivered a public lecture on mathematics and AI, arguing that frontier models' reasoning capabilities are now trackable across scientific domains and cheap enough that labs should be doing it systematically. His talk arrives weeks after Claude Fable 5 produced a verifiable counterexample to the 85-year-old Jacobian Conjecture.
Microsoft's MAI-Voice-2-Flash Undercuts OpenAI by Up to 89% in Enterprise Voice
Microsoft released MAI-Voice-2-Flash into public preview alongside MAI-Image-2.5-Pro, claiming up to 89% lower cost than OpenAI equivalents. The company holds a $13B+ stake in OpenAI and now offers cheaper competing models on the same Azure platform.
OpenAI's Jalapeño Chip Goes Official: Broadcom-Designed, 50% Cheaper Than Nvidia, Ships End-2026
OpenAI and Broadcom unveiled Jalapeño on June 24, OpenAI's first custom inference chip, codenamed Titan. Broadcom CEO Hock Tan claims 50% better cost efficiency over standard AI GPUs. Deployment starts end-2026, full commercial scale by H1 2028.
Anthropic Launches Claude Opus 5 Fast at $10/$50 Per Million — 3x Cheaper Than Opus 4.6 Fast Was
Opus 5 Fast went live on July 24 at $10 input / $50 output per million tokens. That is 2x standard Opus 5 and a third of what Opus 4.6 Fast cost — with the same 96% SWE-bench capability underneath. Cache reads are $1/M.
Cloudflare's Second Content Independence Day: Search, Agent, and Training Bots Now Get Different Rules
Cloudflare launched a three-way bot taxonomy on its second Content Independence Day. Website owners can now distinguish search crawlers, agent browsers, and training scrapers — and apply different rules to each. Ad-monetized page protection is new. One-size block-all is gone.
DeepSeek Pauses Fundraise After Liang Wenfeng's Compute Gap Comments Leak
A transcript of DeepSeek CEO Liang Wenfeng's private investor meeting on July 22 was uploaded to GitHub and spread rapidly. It contains candid admissions about DeepSeek's compute gap with US labs. DeepSeek has paused its ongoing fundraise.
AMD Invests Up to $5B in Anthropic and Wins 2GW MI450 Compute Deal — First Serious NVIDIA Rival Inside Claude's Stack
AMD commits a $5 billion strategic equity stake in Anthropic and secures a 2-gigawatt deployment of its Instinct MI455X GPUs in Helios rack-scale systems. The first gigawatt ships in H1 2027. NVIDIA, Amazon, Google, and SpaceX all hold pieces of Anthropic's compute stack. AMD just joined that list.
SK Group and NVIDIA Sign $500B+ LOI: 2GW Vera Rubin Factory, HBM4 Supply Deal, and Korea as Asia-Pacific AI Hub
SK Telecom will build a 2-gigawatt NVIDIA Vera Rubin DSX AI factory in Korea, with SK Hynix locking in a long-term HBM4 co-development and supply agreement with NVIDIA. The combined initiative is valued at more than $500 billion. First factory comes online in 2027.