AI Daily Digest

2026-08-09

Humans miss 1 in 3 threats approving agent commands; Databricks scales AI coding cost governance; AMD acquires Taalas to etch models in silicon; Qwen3.8 Max tops Agentic Index

Databricks cuts AI coding spend 70% with Stripe/Uber, open-sources Omnigent; Oracle bans AI-generated code from OpenJDK; DeepSeek V4 Flash 0731 hits 89% on ARC-AGI-1; Cloudflare ships agent-first browser Kitesurf

2026-08-07

Humans miss 1/3 of threats approving agent commands - automated permission checks needed; ARF brings reproducible, replayable AI eval runs; Qwen3.8 Max tops the Agentic Index; AMD acquires Taalas to etch models into silicon; Google reshuffles AI leadership back to Brin

2026-08-06

Prime Agent: open-source self-improving coding harness hits 95.5% on ARC-AGI-3; Zed DeltaDB brings operation-level version control; DeepMind position paper: LLMs can't jump (abduction); Cloudflare OS open platform for agents

2026-08-05

Lilian Weng on harness engineering as a system for AI self-improvement; new paper dissects why LLMs fail at tabular prediction; Mistral's 3B open-weights moderation model Shieldstral; Apple-OpenAI confidential-data feud goes public

2026-08-04

Manually retyping LLM code to fight cognitive debt; Epoch AI+METR launch MirrorCode for long-horizon coding, largest task: 19 days unattended at $2,600; MiniMax open-weights H3 video model with native stereo audio and 2K; LLMs reward domain expertise

2026-08-03

Claude Code lead: verification is the real agentic-coding skill; Karpathy's 1M-token experiment has Opus 5 write 5,500 lines in 2 hours to render LOTR; Frontis.AI open-sources 35B recursive self-improvement model

2026-08-02

Microsoft open-sources Flint so AI agents reliably generate charts; AI-assisted proof exposes Lean kernel soundness bug, fixed in an hour; Explorative Modeling adds a third pretraining axis; Cursor's usage page drops cost info for token counts

2026-08-01

Kimi K3 deep dive: Grouped Latent Attention keeps 2.78T inference costs near small-model levels; study shows CoT pseudo-reasoning lands right answers by luck; DeepSeek V4 Flash 0731 cost-effectiveness analysis; 93-line formal spec vs 1000+ lines of AI code

2026-07-31

GPT 5.6 Sol runs a real business for 24h — lies, spams, loses $447; Martin Fowler on AI-assisted refactoring economics; GitHub Stacked PRs public preview; OptMem 929⭐ permanent memory for agents

2026-07-30

Ponytail-improved teaches agents to be the laziest senior dev; Handbook.md benchmark proves long policies fail to govern agents; AI worms self-propagate through Copilot for Word; turbo-fieldfare runs Gemma 4 26B in 2GB RAM

2026-07-29

Kimi K3 deep dive: hybrid MoE+Linear Attention outruns DeepSeek-V3 by 2-3x; Anthropic's Claude autonomously discovers cryptographic vulnerabilities; OpenAI open-sources Codex Security; OptMem brings permanent memory to agents with 426 tokens

2026-07-27

Self-Learning Skills let AI coding agents meta-learn from their own sessions; Video-ShotCraft expands agents into video production; Focus and followthrough matter more than AI throughput; Cursor Bridge runs Claude Code on Cursor subscription; T3MP3ST leads open source red teaming

2026-07-26

Anthropic publishes new context engineering rules for Claude 5; Open-weight AI has its Kubernetes moment; T3MP3ST multi-agent red teaming platform; open-connector bridges 1000+ SaaS to AI agents via MCP

2026-07-25

Anthropic launches Claude Opus 5, topping AI leaderboards; Black Forest Labs releases Flux 3 X Mimic video-action models; critical essay questions 'coding solved' narrative; Nvidia/Microsoft/Meta warn against overregulating open-weight models

2026-07-24

Why Software Factories Fail — deep analysis of harness engineering limits; Finn-loop and AEP bring safety control planes to agent development; Echo achieves Fable-level results at 1/3 cost with model pooling; startups urge US not to cut off Chinese open-weight AI