AI/TLDR — every new AI model, tool, repo & paper
The latest AI releases, refreshed every 2 hours and explained in plain English.
What AI shipped today?
In the last 24 hours AI/TLDR tracked 8 new AI releases, including Claude Code stops cloning existing games — Anthropic adds a two-sentence rule, ChatGPT and Codex share one usage limit — token-based pool from December and Wes Roth — 'the BIGGEST HACK of all time just started...'. AI/TLDR is an AI release tracker that follows new AI models, open-source tools, papers, datasets and benchmarks — refreshed every 2 hours from verified primary sources and explained in plain English.
AI Release Index — live stats on AI releases · Learn AI
- Claude Code stops cloning existing games — Anthropic adds a two-sentence rule
Anthropic added two sentences to Claude Code's prompt on Opus 5.5 so Claude will not rebuild a specific existing game, even under a new name. Original games, genre games and games from your own assets still work.
- ChatGPT and Codex share one usage limit — token-based pool from December
OpenAI says Plus, Pro and Business plans will share one usage allowance across ChatGPT and Codex from December. Usage is counted in tokens, Instant chat does not count, and Settings > Usage shows what is left.
- Wes Roth — 'the BIGGEST HACK of all time just started...'
Wes Roth walks through a wave of attacks led by the $6.36 billion Binance theft, then turns to AI agent swarms, secret channels between agents and whether defending the internet may need AI that can't be safely shut down.
- OpenMausBot — an open-source Grok Bot alternative with cloud computers
OpenMausBot is an Apache-2.0 desktop chat app for a team of AI bots. Each bot runs Claude Code or Codex locally and can get its own cloud desktop or local VM. Version 0.1.136 adds bot sharing with teammates.
- Memdoor team-solve — five Qwen3.8-27B agents rerun the GVS5H coding test
Memdoor's team-solve workflow rebuilds the GVS5H method: five Qwen3.8-27B roles share files on disk. On 10 LiveCodeBench Hard problems over 5 passes it solved 9.8 on average; the authors' Fable 5 files show 8.0.
- DeepSeek Super-agent — an open coding agent that takes work from GitHub chats
DeepSeek open-sourced dsh-superagent, an MIT research prototype of a proactive coding agent for one GitHub repo. It runs on DeepSeek Harness, turns approved requests into tracked work and replies with results such as pull requests.
- Project Suncatcher — Google's satellite runs Gemma 4 on a TPU in orbit
Google's Project Suncatcher prototype satellite answered its first prompt from orbit. Gemma 4, running on a data-center-class TPU, wrote a haiku about space. Google calls it the first AI response from such a TPU in orbit.
- GPT-5.6 Sol beats Factorio — a rocket launch on peaceful mode
GPT-5.6 Sol, running in Codex at a low thinking level, launched a rocket in Factorio on peaceful mode in 13.5 hours of game time. Ryan Madden's harness gives it the same reach and fog-of-war limits a human player has.
- VOLTA infostealer — malware asks an LLM how much each victim is worth
Cisco Talos found VOLTA, a Windows infostealer that sends a profile of each victim to an LLM through OpenRouter. The model ranks the victim's stolen assets by money value and rates their security habits, replacing the operator's own analysis.
- ChatGPT Shared-Link Flaw — one click ran attacker code in the victim's session
Parameter researchers chained flaws in ChatGPT's graph and Preview features so a public shared chat ran JavaScript as chatgpt.com. OpenAI fixed it on Oct 26, rated it P1 and paid a $5,000 bounty.
- Northflank breach — stolen customer secrets used to run up AI API bills
Northflank tells every customer to assume their environment variables were decrypted and to rotate all secrets. Users report that AI API keys stored on the platform were used without permission. Wiz lists it as the ninth such platform breach in October.
- ChatGPT violence safeguards — chats can end, teens lose weapons details
OpenAI is adding violence safeguards to ChatGPT. It can end chats that fixate on violence, limits details on past mass attacks for everyone, and blocks practical weapons info for teens. A twice-yearly Community Safety Report starts in early 2027.
- AI training opt-out token — one robots.txt rule for OpenAI and Google
The __usage_ai-training token lets a website opt out of AI training with one robots.txt group. OpenAI and Google both said on October 30, 2026 that they will honor it; the IETF draft is written by Anthropic's Kevin Kelley.
- Two Minute Papers — 'OpenAI Claims It Just Did The Impossible'
Two Minute Papers covers OpenAI's AI-written manuscript 'The Exponential Time Hypothesis is False', which claims a subexponential algorithm for counting solutions of fixed-width CNF formulas.
- Claude Code 2.1.316 — Opus 4.8 fast mode deprecated, pooled gateway budgets
Claude Code 2.1.316 deprecates fast mode for Opus 4.8 (removal on November 30, 2026), adds pooled group budgets to the Claude apps gateway, and adds a subagent effort setting.
- nesformer — a Nintendo CPU compiled into Qwen3.5 weights
nesformer is an MIT-licensed project that runs an NES CPU inside the weights of a stock Qwen3.5 model. The weights are compiled, not trained, and the model predicts the CPU log one token at a time.
- Sam Witteveen — '3,000 Tokens/Sec Locally: Testing the ZGX Fury'
Sam Witteveen tests HP's ZGX Fury, a shared desktop AI machine built on NVIDIA's GB300 chip with 748 GB of memory, and shows local inference speeds, software and use cases.
- Martin Fowler on LLM Voice — authors should write the prose themselves
Martin Fowler says text that sounds LLM-written makes him stop reading, and that authors should not let an LLM draft their work because editing rarely removes the style.
- Claude Code extension 2.1.313 — tampered build carried an SSH backdoor
Cursor says a tampered Anthropic.claude-code 2.1.313 was live in its extension marketplace for about 11 hours on 29 October. It adds an SSH key and a reverse tunnel. Cursor says at most 5,000 users were affected.
- Claude Code 2.1.314 — a block on writes outside the project and gateway groups
Claude Code 2.1.314 adds a permission setting that blocks file writes outside the working directories, lets the Claude apps gateway route upstreams by IdP group, and fixes effort-cap errors and silent exits in claude -p.
- YC Labs — AI agent swarms play Minecraft and Mindustry
YC Labs ran groups of AI agents in Minecraft and Mindustry, and open-sourced the Minecraft harness. In Mindustry, GPT-6.1 Sol agents beat Claude Opus 5.5 agents 6-1 over 7 games.
- Codex CLI 0.166.0 — keyboard transcript selection and codex doctor
Codex CLI 0.166.0 adds keyboard-driven transcript selection with search, a `codex doctor --network-only` diagnostic, `codex queue --resume`, and GPT-6.1 Terra in the Vertex AI and Bedrock catalogs.
- OpenAI's misalignment monitors — four layers watch training, Codex and Astra
OpenAI's post 'How we detect frontier model misalignment' explains the four-layer system that now watches its models for misbehavior. Replayed on the Hugging Face incident, it alerted staff 30 hours before the real breach.
- OpenAI Model Spec update — five core values, and the model is now 'GPT'
OpenAI's Model Spec of 29 October 2026 reorganizes its rules for model behavior around five values, led by faithfulness to human intent. It adds guidance on context, human agency and side effects, and retires most style rules.
- Mistral Large 4 open weights — Le Chonk lands on Hugging Face
Mistral Large 4 (Le Chonk) is now open-weight: Mistral released the 1T-parameter multimodal MoE on Hugging Face after three more weeks of RL. It scores 41% on Terminal-Bench 4.0, up 13 points from the preview.
- Google AI Mode dodged 'who won 2020?' — Google calls it a technical gap
Ars Technica found that Google's AI Mode would not name the winner of the 2020 US presidential election and sent users to the FEC instead. ChatGPT and Claude answered. Google says it is fixing a 'technical gap'.
- Fireship — 'Did the AI bubble just pop?'
Fireship's new video 'Did the AI bubble just pop?' asks whether the AI boom is ending. The short news video was uploaded on October 28, 2026.
- Claude Code's agent status line — a test says it used 16% of spend
Flamerix AI measured Claude Code's agent_summary calls, which write a 3-5 word status for each background agent, at 16.4% of spend in its own week. A Claude Code engineer put the average at about 2% and promised a cheaper mode.
- Google Connected Apps Platform — build one app for Search and Gemini
Google's Connected Apps Platform is a new developer preview for building one app that runs inside Google Search, the Gemini app, or both. Apps connect through a remote MCP server or Android AppFunctions and are managed in Google Connected Console.
- Claude Enterprise smart reports — what a team's Claude spend produced
Anthropic made smart reports, in-app surveys and pooled group budgets generally available for Claude Enterprise. Smart reports group a team's Claude work into workstreams and show the spend, the deliverables and where work got stuck.
- Anthropic AI for Science — free Claude credits now flow through 20 partners
Anthropic is expanding its AI for Science program beyond single researchers: funders such as the Sloan Foundation, Simons Foundation and the US NSF will hand out free Claude credits, and institutes like EMBL and the Crick will use them.
- Anti-LLM obfuscation — a bot-check script hides fake jailbreak text
A servicepipe.tech fingerprinting script fills its code with fake weapon-themed function names and a hidden jailbreak prompt so AI models refuse to deobfuscate it. Astra refused; GLM-5.3 complied.
- DAYJOB: Tax — best AI agent passes only 10.8% of real tax assignments
DAYJOB: Tax is Surge AI's new benchmark of 50 professional tax assignments with 1,789 client files. Claude Sonnet 5.5 leads at 10.8%, GPT-6 Astra scores 4.8% and Gemini 4 Argon 3.2%.
- Gemini 4 Carbon — Google's new frontier model, first for cyber defenders
Gemini 4 Carbon is Google's new frontier model for cyber defence and long agentic coding. It ranks #1 on the Artificial Analysis Intelligence Index v5.0 with 59, and goes first to a subset of cyber defenders, not to the public.
- GitHub Copilot Fusion — Copilot now picks and combines models for you
GitHub Copilot Fusion is now generally available on all Copilot plans. Instead of picking a model, you pick Fusion and an effort level, and Copilot routes the task to one or more models. It grew out of the HydraFusion research preview.
- Two Minute Papers — 'Deepseek's New AI Has A Secret'
Two Minute Papers looks at DeepSeek V4.1 Flash 1017, the open-weight in-place upgrade DeepSeek released on 17 October. The video description links DeepSeek's release note and clips that users posted of the model at work.
- Artificial Analysis Index v5 — private test sets now carry 60%
Artificial Analysis Intelligence Index v5 adds FrontierCode 2.0 and Terminal-Bench-Science and swaps Humanity's Last Exam for HLE-Diamond. 60% of the Index now uses private data. Claude Opus 5.5 (max) leads at 58.
- Nemotron 3.5 Transcribe — NVIDIA's open speech model you steer with prompts
Nemotron 3.5 Transcribe is NVIDIA's open-weight speech model. It transcribes 40 languages, translates 31, and labels up to eight speakers in hour-long recordings. NVIDIA also released Granary v2, a 463,062-hour speech dataset.
- Wes Roth — 'AI builds EVERYTHING...'
Wes Roth builds Blood Bench, a fantasy football game, with AI writing the code, graphics and online game, then has competing AI models coach teams against each other in a tournament.
- LFM3 — Liquid AI's 2B-class open models for text, vision, audio and omni
Liquid AI released LFM3, six open-weight 2B-class models on one backbone: text, vision, audio, omni and two decision models. LFM3-1.7B has a 256K context, fits in about 1 GB at 4-bit, and scores 38.42 on WebArena-Verified.