Thursday, August 13, 2026

Pulse (last 3d) · Research 83 · Agents 75 · Models 56 · Products 48 · Enterprise 37 · Open Source 36 · Industry 23 · Legal 22 · Media 15 · Inference 14

Trending · OpenAI 22 · Claude 13 · Claude Code 9 · Codex 9 · Anthropic 8 · Nvidia 8 · Meta 7 · Cursor 5 · Google 5 · ChatGPT 4 · GitHub 4 · Hugging Face 4

---

Daily AI Brief, August 13, 2026

Grok 4.6 ties GPT-5.6 on AA Index, rewrites cost math. Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index (tying GPT-5.6 Sol Max), with 2x turn efficiency and 4x input-token efficiency on long-horizon agentic tasks vs Claude Opus 5, at $2/$6 pricing held flat. Source

Gowers: LLMs win at counterexamples via probabilistic search, not depth. Fields medalist Timothy Gowers publishes a systematic analysis arguing LLM math successes (non-sofic groups, Ramsey lower bounds, Jacobian conjecture) stem from probabilistic search, not conceptual understanding. Source

TissueFormer: linear-attention FM for paired H&E + spatial transcriptomics. Broad/MIT/Schmidt Center preprint from Qiyu Gong et al. applying linear-attention multimodal foundation modeling to single-cell pathology. Source

Hidden reasoning from Claude and GPT decoded via sibling-model replay. Researchers extracted encrypted chain-of-thought from frontier APIs by routing through the cheapest sibling model, then found live credentials in agent traces already published on GitHub and Hugging Face. Source

Convergent Detour Hijacking: task-preserving resource amplification in agents. Shows how LLM agents can inflate resource use while preserving the surface semantics of a legitimate task. Source

Agent safety should be a runtime contract. Argues agent safety guarantees belong in runtime contracts rather than training alone. Source

Claude Code orchestration: Opus refuses only when work is delegated. Quesma benchmark of four Claude models on Terminal-Bench 2.1 hit 78% at $1,178, exposing failure modes unique to multi-agent delegation. Source

Unsloth Desktop. Open-source local app to run and train LLMs, diffusion, and audio models, with one-command hookup to Claude Code or Codex and no-code fine-tuning. Source

Bullet, speed-first coding agent. Auto-selects model/reasoning level per prompt, parallelizes searches, and skips repo-wide embeddings; reports 95.8% on SWE-bench Verified at 119s/task. Source

Zed Delta, multiplayer agent code review. Rust-editor team launches Delta, a separate CRDT-replicated app for multiplayer coding-with-agents; private beta now. Source

CodeBurn, local AI coding cost tracker. Open-source tracker reading session files from 40+ coding tools, surfacing per-task spend and an "Optimize" tab that applies fixes and tracks savings. Source

Anthropic watermarks anger Claude users over work/school detection. New output watermarking draws backlash from users relying on Claude professionally. Source

Cognition reportedly raising at $40B. Devin-maker already in talks for a new round at a $40B valuation. Source

Lovable raises $400M Series C. AI app-builder reportedly valued at $13.3B. Source

Grok launched as an assignable AI teammate in beta. xAI's SpaceXAI unit opens Grok as a workspace agent. Source

Discovered Materials (YC P26), AI agents for materials discovery. Y Combinator launch using autonomous agents to discover and validate new materials. Source

Three AI markets: Grok, Qwen, and Nvidia now compete in different layers. Analytical take distinguishing closed API (Grok 4.6), open-weight (Qwen3.8), and routing infrastructure (Nvidia Switchyard). Source

Zvi Mowshowitz monthly roundup #45, August 2026. Aggregated overview of the month's key AI developments. Source

Ablating one of 128 attention heads kills Morphy's queen sacrifice detection. Mechanistic-interpretability demo on a chess transformer showing single-head causal importance for a famous motif. Source