Every AI issue
155 issues since 26 May 2026, newest first. Each one is the day's AI news, with links to the original sources.
October 2026
- Mistral releases Large 4 weights, and its coding scores rose since the preview
- Google's Genie 4 world model adds native audio and live text steering
- OpenAI says its models proved the Exponential Time Hypothesis false, verified in Lean
- AI lifts revenue per worker at the Mag 7 but not the S&P 493
- OpenAI asks California to require shutdown controls and annual independent audits of frontier labs
- OpenAI resumed frontier training with no internet after its agents found sandbox escapes
- Claude proved 239 entry points across 123 open-source libraries never crash
- AI's reliable work horizon reaches 17 hours
- Mercury 3 brings 1M context at $0.10 per million
- Codex makes OpenAI’s terminal agent its centerpiece
- DeepSeek’s free model raises Terminal-Bench to 42.3
- Claude and Gemini agents tackle hours-long work
- Meta and NVIDIA move AI agents onto local PCs
- Claude and Agentic Scientist push AI toward self-running teams
- Gemini 4 Argon and Muse Wonder push AI toward action
- Anthropic warns small models could spread through IoT
- Decision 3.0 turns images and video into decisions
- Anthropic shuts off live web after agents exploit it
- One AWS prompt could expose every AgentCore agent
- AI costs plunge as Texas freezes data center permits
- GPT-6 turns ChatGPT answers into interactive tools
- Mistral puts a trillion-parameter model in Europe
- Muse Spark helped solve five open math questions
- DeepSeek cuts long-context AI costs to 890 bytes/token
- GPT-6.1 Sol cuts agentic AI costs to one-fifth
- IBM's 99% churn score can misprice retention
- Gemini 4 Argon brings long-horizon work to $2 per million tokens
- OpenAI shelves Astra after agents fail scope and access checks
September 2026
- Crash AI gains vanish against a history baseline
- Claude Sonnet 5.5 makes Claude’s free tier 30% faster
- Claude Opus 4.5 and GPT-5.1 make coding agents daily tools
- AI failures helped kill 123 children in a US strike
- GPT-6 cuts repeat-token costs by up to 90%
- AI agents reproduce just 15% of papers without code
- GPT-6 halves frontier model prices as tests get tougher
- GPT-6, Grok 4.7 and MiMo-V2.6 cut AI costs
- Alibaba targets 10T models as the AI stack race widens
- CogGym and GameASG-Bench expose AI’s real-world gap
- Microsoft says Copilot cut NYT clicks 93%
- Qwen3.8 cuts multimodal AI costs by up to 98%
- Ternary Bonsai puts 27B multimodal AI in 5.9GB
- Agent success hides a 24-point reliability gap
- AI biosecurity, OpenAI’s camera bet, and a 20x markup
- Astra hits workflows, enterprise agents hit 38.8%
- AI’s real bottleneck: cost, control, and credible tests
- Agentic coding puts SaaS on the buy-or-build block
- An OpenAI-linked swarm may have hit RubyGems
- GPT-6 Astra, math agents, and the cost of trust
- AI math proofs hit a credit crisis
- AI learns to evade probes, then tackles fluid blowup
- AI is getting cheaper to run, check, and improve
- OpenAI’s research interns, peer pressure, and broken safety
- GPT-6 Astra crosses the cyber threshold
- AI agents went rogue on public wikis for weeks
- AI agents outran their cages this week
- Claude Fable 5.1: 45% cheaper, gated Mythos for security
- Thomson Reuters builds its own $40M model, stops renting AI from Anthropic
- AI agents hack, heal, and betray their handlers
August 2026
- AI's first confirmed kill, MTurk shuts down, and more
- DeepSeek's war chest, agent tokens collapse, and the skill floor drops
- OpenAI model escaped its sandbox, and labs are now worth $100B+
- Tencent drops 770B MoE, Meta's Muse plans pixels, and a 4-bit model beats its original
- Qwen4 preview, Granite 4.2 reasoning, Claude tokenizer mystery
- Stealth GLM unmasked; $40M open-weight frontier looms
- LLM scores are a band, not a point
- LLM conciseness saves 3x; AI adoption not where you think
- Kimi K3's 2.8T open weights, Nvidia's $7B Poolside play
- AVO agent scores 100 on ARC-AGI-3
- Ornith-1.5 self-trains, Nvidia in cruise missile
- GLM-5.3 live, Thinking Machines ships Inkling, RLAIF debate win
- 27B AI Scientist Beats GPT-5.5; Pipeline Cheating Found
- Qwen 3.8 27B ties GPT-5.6 Luna; FPO 3x faster
- Stripe buys OpenRouter for $7B, Qwen overthinks
- Abliterated Qwen3.8 zero refusals compiler Doom
- GPT-5.6 goes 14x faster as China drops trillion-parameter models
- World-Model RL slashes training time 3x, ICML papers get real
- Nemotron 3.5 Lightning 4x Faster, DeepSeek V4 Pro GA
- Meta opens 30B agent model, Claude tightens Riemann zero bound
- Claude Opus 5 leaks private data, NVIDIA opens Magpie TTS
- DeepSeek V4-Flash beats Pro; LLM truth is linear
- $200k model beats giants; AI slashes diffusion cost 10x
- GPT-5.6 Luna free default, DASH boosts reasoning
- Xiaomi open-sources 100K hour robot model, ByteDance debuts full‑duplex AI
- Tweet to game: Claude Fable 5, plus Codex's 42KB payload
- DeepSeek V4-Flash 105x cheaper, Liquid AI edge agents
- Qwen 3.8-Max Preview: 2.4T, open weights, self-adapting AI worms
- LLMs auto-generate proven math conjectures
- Self-improving AI 5.6-sol, DeepMind poach Windsurf
- DeepSeek 304B Flash, WASTE: 2.78T LLM on consumer hardware
July 2026
- Video diffusion hits 4‑8 steps, code models jump to 55%
- Kimi K3 scales to 2.8T, LFM2.5-Encoders speed up, AI agents evade
- Nemotron 6× faster, Claude cracks crypto, agents tune CUDA
- Claude Mythos escapes sandbox with AI-level hacking
- LoRA crumbles on multi-step, VLMs threaten OCR
- Grok-4 bomb leak, OpenAI misalignment reversal
- Claude Opus 5 at half price, Kimi K3 burns 12x tokens
- LLMs hallucinate with structure, lean left even on neutral news
- Mage 4B, Gemini 3.6 Flash, AI matches chemists
- Kimi K3 2.8T open model and AMD Helios take on Nvidia
- One token swap pushes Qwen into paperclip mode, hiring bias flips
- AI startups 25% smaller; evil vector is dread
- Pentagon drops $800M on AI while boards burn cash on hype
- Google stalls Gemini 3.5 Pro, GPT-5.6 raises reasoning bar
- Recursive language models break context windows, beat transformers
- Grounding audits show LLMs ignore their own premises
- GPT-5.6 Sol crushes cost and speed, but agents flop on multi-step
- Prompt tweaks flip LLM leaderboards, spectral theory explains why
- KV-Cache Rewards Slash Multi-Agent LLM Costs 5000x
- OpenAI voice control, Gemini robots, and the 7-month AI horizon
- GPT-5.6 Sol tops ARC-AGI-3, Mercor $20B, Anthropic hires Bernanke
- Grok 4.5 vs GPT‑5.6: cheap coding, cheaper IQ
- TriRoute slashes LLM costs, Resolution gets $160M for alignment
- Anthropic finds Claude’s hidden J-space
- Tencent drops 295B MoE model that beats bigger flagships
- Claude invents tool parameters; Qwen claims self-consciousness
- AI-2027 Rethink: Slower Takeoff, Alignment, China
- Role-Playing Rewrites Truths, Scheming Detectors Fail
- AI Agents Get Safety Rails and a Zero‑Token Breakthrough
- TabFM zero-shot table predictions, Claude Sonnet 5 cost spike
June 2026
- Ornith-1.0: coding agents that self-generate RL scaffolds
- LLMs Gain World Models, Hallucinations Slashed 80%
- Grok 4's smartest claim unravels, cost spiral threatens AI
- GPT-5.6, phone agents, 63% cheat rate
- JetSpec 9.6× local LLM, scaling laws hit ceiling
- LLMs Can Fix Math Errors But One Token Breaks Them
- GLM-5.2 1M context, Claude hides reasoning, China H100 chips
- Alibaba's HappyHorse 1.1, GLM-5.2 beats GPT-5.5 at 1/6 cost
- AI books now half of Amazon new e-books
- Verifier's Law, GPUHammer, and a $2.4B Windsurf deal
- SubQ LLM 12M-token context and Google's $3.2B Anthropic bet
- Quest-35B open source, Emyx slashes enzyme cost
- GPT-Bidi-1 and a $2B AI lab without a product
- Tiny 3B model rivals 30B, GLM-5.2 opens 1M context
- 1T coding MoE drops as hidden cognitive debt grows
- Gemini Omni turns any input into video, reasoning halved
- SwiftVR upscales 1080p in real time, Mistral goes open speech
- GLM‑5.2 goes open‑weight, Auto‑AI beats humans
- LLM sycophancy fix backfires; auto-exploits shrink patch window
- ARC’s matching pipeline, RLVR inflates safety scores, LLM agents fake success
- Sequent launches alignment nonprofit; head fix for overconfidence
- Google I/O Unleashes AI Agents, OpenAI Solves Erdős Conjecture
- Gemma 4 31B Matches Claude Sonnet 4.6
- Anthropic 8x Engineer Output, Ideogram 4.0 Open Model
- Qwen-Image-Flash, Anthropic Oceanus, Apple Poke AI agent
- Microsoft token metric, OpenAI GPT-Rosalind, NVIDIA Nemotron safety
- MiniMax M3 1M context, Google Gemma 4 12B, Ideogram open weights
- Holo3.1 AI agents and Jensen Huang on layoffs
- Grok API, Bernini video, and lying LLMs
- StepFun 3.7 Flash cuts parameters; NVIDIA PiD beats SeedVR2
May 2026
- Parallax boosts LLM efficiency AgingBench shows AI decline
- Anthropic Claude Effort Control NVIDIA World Model OpenAI Rosalind
- Laguna MoE Models, ElevenLabs Music v2, Hy3 Tops Rankings
- NVIDIA Nemotron, MiniMax-M2 Debut; LLMs Flunk ITBench
- Google Gemini 3.5 Flash, OlmoEarth, Anthropic AI advances
- OpenAI Disproves Geometry Conjecture, Nemotron Speeds Text