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 Sebastian Raschka — understanding positional embeddings in modern LLMs, Haru-4B — an open speech model that listens and talks at the same time and Jakub T. Jankiewicz — all AI-generated code should be open source. 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
- Sebastian Raschka — understanding positional embeddings in modern LLMs
Sebastian Raschka explains how LLMs encode token position, from GPT-2 learned embeddings and sinusoidal encodings to RoPE with PyTorch code, plus partial RoPE, NoPE and mixed RoPE/NoPE layers in 2026 open-weight models.
- Haru-4B — an open speech model that listens and talks at the same time
Haru-4B is higurashi's open full-duplex spoken dialogue model: Apache 2.0, built on Qwen3.5-4B, nine primary languages. It hears and speaks at once, with about 65 ms end-to-end latency on a MacBook.
- Jakub T. Jankiewicz — all AI-generated code should be open source
Jakub T. Jankiewicz argues that all AI-generated code should be open source. Models learned from open code and Stack Overflow, he says, so closing their output is "license laundering", and purely AI-written code may not be copyrightable.
- DeepSeek App API — web search, file input and safety checks built into the API
DeepSeek now offers its app's features through the API. deepseek-flash (V4.1-Flash) can search the web, read links, parse files and run safety checks on the server, and new OCR, ASR and TTS APIs are available too.
- Nathan Lambert — Jensen Huang's exit is a setback for US open models
Nathan Lambert argues that Jensen Huang's resignation as Nvidia CEO is a major setback for US open-weight models. He expects Huang's open-model funding to dry up and says the US needs its own truly frontier open model lab.
- Chrome DevTools MCP 1.13.0 — logpoints and network mocks for coding agents
Chrome DevTools MCP 1.13.0 lets AI coding agents set logpoints in a live page and mock its network requests. The mocks can block a request, serve a local file or fetch from another URL. The release also ships a long list of fixes.
- deepseek-v3-jax — DeepSeek V3 in 190 lines of pure JAX
deepseek-v3-jax is an MIT-licensed, 190-line JAX version of the DeepSeek V3 architecture. The 36.2M-parameter model trained in about 2.6 hours on a MacBook Air M5. The author calls it an architecture demo, not a strong language model.
- Plannotator — review your coding agent's plans and diffs in a browser
Plannotator is an open-source local review tool for coding agents. You annotate an agent's plan, diff or HTML output in the browser and send the feedback back. Version 0.28.64 shipped on 13 November as the repo trends on GitHub.
- WSJ: Nvidia backed buyers' loans — so they could keep buying its GPUs
The Wall Street Journal reports that Nvidia backed AI cloud firms and data-center builders with credit support so they could keep buying its chips. Interim CEO Colette Kress pushed to limit these deals before Jensen Huang left. Nvidia has not confirmed.
- EFF — AI wearables turn privacy into a problem for everyone nearby
EFF's Thorin Klosowski argues that AI smart glasses, Apple Watch Audio Intelligence and Amazon's Bee devices record and summarize people who never agreed to it, and that no technical safeguard protects those bystanders.
- No AI Slop Manifesto — AI engineers pledge to own what their agents ship
The No AI Slop Manifesto, written by speakers and attendees of AIE Code Summit 2026, asks engineers to own what they ship, disclose agent-written work and review output. Its signing gist has more than 330 signatures.
- Apple App Review — AI-made apps slow down music developers
Synth Anatomy reports that vibe-coded apps now bring an estimated 10–20 new iOS music apps a day, while known audio developers wait weeks in Apple's App Review or get rejected as spam, some suspecting AI-driven review.
- Antigravity /teamwork — Google's agent teams prove an open math conjecture
Google Research says Antigravity's /teamwork agents, working with human experts, proved the 2013 Courtade–Kumar conjecture in full and checked it in Lean 4. Google offers 10,000 academic researchers six months of free top-tier access.
- Claude outage — login failures, then API errors hit claude.ai and Claude Code
Claude had two outages on 13 November. Logins to claude.ai and Claude Code failed from 09:58 to 11:58 UTC. From 13:09 to 15:38 UTC, claude.ai, the Claude API, Claude Code and connectors returned elevated errors.
- Unsloth Dynamic v3.1 — Qwen3.8-27B GGUF quants gain 5-10% accuracy
Unsloth updated its Qwen3.8-27B GGUFs to the Dynamic v3.1 preview recipe. Unsloth says it improves accuracy by 5-10% at the same file size on most benchmarks. Most 2-bit to 6-bit quants were replaced.
- OpenAI disbands its robotics team — general models take over robot work
OpenAI has disbanded its standalone robotics team and moved the researchers to other groups, Fortune reports. OpenAI says robotics stays a core part of its long-term vision.
- 600 JEPA world-model experiments — subgoals, not physics, limit planning
Haddy Alchaer ran about 600 experiments with the LeWM JEPA world model on Push-T. Its learned predictor plans as well as true physics, while adding subgoals lifts long-horizon success from 22% to 80%.
- Gödel's notebooks — Claude finds 663 theorems in 1,009 shorthand pages
Helsinki logician Jan von Plato used Claude to read his transcriptions of Kurt Gödel's 1940s workbooks. Claude listed 663 theorems; in the Coq proof checker about 400 held as stated, 200 needed repair and 50 were false.
- Claude Code 2.1.325 — sandboxed Bash commands now ask for permission
Claude Code 2.1.325 makes sandboxed Bash commands follow the normal permission flow by default and fixes a security issue in zstd response decoding. It also closes several gaps in Bash permission checks.
- Microsoft 365 outage — Copilot and AI connectors down for 20+ hours
A Microsoft 365 outage that began about 18:00 UTC on November 12 took down Exchange Online, Teams and Copilot for more than 20 hours. The Microsoft 365 connectors in Claude and ChatGPT failed with it.
- Wes Roth — 'Gemini 4 Swarm is out of control...'
Wes Roth reviews Google's week of releases: Gemini 4 Carbon, Gemini 4 Hydrogen, Gemini Decision-1, AlphaProof Code and /swarm in Antigravity. Gemini 4 Carbon edited the whole 30-minute video on its own as a test.
- ChatGPT Wallet — OpenAI's agents can now buy things for you
ChatGPT Wallet lets agents in ChatGPT Work, Codex and dots complete purchases. You save a card once and approve each order with Face ID or a fingerprint. It starts with US Pro users over 18, built with Amex, Mastercard and Visa.
- llama.cpp v1.0.0 — first stable release, with semantic versioning
llama.cpp and ggml reach v1.0.0 after 3.5 years. Breaking API changes now bump the major version. The release removes 36 deprecated functions and adds an FP8 KV cache, Kimi-K3.1 support and a text-to-speech server endpoint.
- Transformers v5.23.0 — LFM3, Inkling 2 and a rebuilt model export
Transformers v5.23.0 adds Liquid AI's LFM3 and LFM3-VL, Thinking Machines' Inkling 2 Small, Dots 3 Note Preview and InternLM3. Exports now return a callable ExportedModel, and expert parallelism gets a selectable implementation.
- ZONOS3-Mini — Zyphra's open TTS model clones voices in real time
ZONOS3-Mini is Zyphra's open-weight text-to-speech model: an 800M-parameter MoE with about 100M active, Apache 2.0, 43 languages and zero-shot voice cloning. It runs in real time on laptops, phones and in the browser.
- Brave Search API — AI agents can now pay per request with x402
Brave Search API now accepts pay-per-request payments through the open x402 protocol. Agents need no account or API key: they pay in USDC on Base, at $0.005 per request on the web, news, image, video, local and LLM Context endpoints.
- Codex CLI 0.169.0 — side-chat tabs, task search and collapsible turns
Codex CLI 0.169.0 lets the terminal workspace hold several side-chat tabs that survive a restart, searches older tasks in the agent command center, and can fold finished activity under each prompt in the transcript.
- OpenCode 2.3.0 — subagents run in the background by default
OpenCode 2.3.0 runs subagents in the background by default, so several can work at once while you stay in the parent session. The TUI groups tool calls into a collapsible row, and the desktop app can jump to earlier prompts.
- Daybreak Red Team and Defense Factory — OpenAI's agents test and patch apps
OpenAI launched two Daybreak security products. Daybreak Red Team, in early access, uses agents to pen-test web apps. Daybreak Defense Factory finds, checks and fixes flaws in code and hands engineers a pull request to approve.
- Self-hosted Claude Managed Agents — run the whole agent service in-house
Anthropic now lets companies run the full Claude Managed Agents service on their own cloud or on-premise servers, in private beta. Agents, credentials and session history stay inside; model calls can go to Bedrock, Google, Foundry or a gateway.
- Claude Meetings — Claude joins video calls as a participant
Claude Meetings is a beta feature where Claude joins Google Meet calls as a visible participant. It answers questions, builds docs and whiteboards live, and writes notes and action items. For Team and Enterprise plans.
- Edge0-145B — an open 145B MoE model that runs on one consumer GPU
Edge0-145B is an open-weight 145B Mixture-of-Experts model with about 10B active parameters, built on Qwen3.8-Flash-Next. Edge0 reports 29 GB peak memory and 80+ tokens/s on a single consumer NVIDIA GPU.
- Mirendil raises $1B Series A — ex-Anthropic lab building AI R&D agents
Mirendil, the AI lab founded by former Anthropic researchers, raised a $1B Series A led by Kleiner Perkins and Clearlake Capital. The money backs tools that let small teams do frontier AI research, including its Lor agent environment.
- Codex CLI 0.168.0 — side-by-side chats, audio clips and animated GIFs
Codex CLI 0.168.0 can show a side conversation next to the main chat, accept audio recordings for audio-capable models, and make animated GIFs. A new Sessions menu picks the default model provider for new chats.
- Checkpoint forensics — Chinese open models carry a Blackwell GPU fingerprint
Will DePue's checkpoint forensics reads the GPU layout left in open model weights. He finds a Blackwell pattern in six models from DeepSeek, Moonshot, MiniMax and Ant Group, including DeepSeek-V4-Flash and Kimi K3.
- smolmaxxing — a 16B MoE model runs on a 1 GB TV box and a 2013 iPad
smolmaxxing is an MIT llama.cpp patch for 1 GB devices. It runs Qwen3.5-2B at 7.7 tokens/s and the 16B Ling-mini-2.0 MoE at 5.1 tokens/s on a Xiaomi TV box, and 1-bit Bonsai-1.7B at 24–26 tokens/s on a 2013 iPad Air.
- Claude Code 2.1.323 — Code Mode lets Claude call its tools from JavaScript
Claude Code 2.1.323 adds Code Mode, an experimental opt-in where Claude writes JavaScript to call its tools. Unknown model IDs now get a 1M-token context window by default instead of 200K.
- Fireship — 'Nvidia just got caught with its pants down...'
Fireship breaks down Jensen Huang's resignation as Nvidia CEO and the scandal behind it, including a Singapore company that bought $4.8 billion of Nvidia GPUs.
- Google Intelligent Eyewear — Gemini glasses open for preorder from $395
Google Intelligent Eyewear, built with Samsung, Warby Parker and Gentle Monster, is open for preorder. The Gemini glasses start at $395, go on sale December 10 and work with Pixel, Galaxy and iPhone phones.
- VS Code 1.146 — the Local agent harness is deprecated, Copilot harness stays
VS Code 1.146 deprecates the Local agent harness, which will be removed in February 2027, so agent sessions move to the Copilot harness. It also adds an Inbox for agent sessions, skills from MCP servers and a /code handoff from Copilot CLI.