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 11 new AI releases, including Atria Dawn Preview — the report behind Shanghai AI Lab's 744B agent model, Fireship — 'Anthropic researchers are quitting... and now we know why' and Salesforce Koa — a CRM reasoning model built on NVIDIA Nemotron. 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
- Atria Dawn Preview — the report behind Shanghai AI Lab's 744B agent model
Shanghai AI Laboratory published the technical report for Atria Dawn Preview, the 744B MoE agentic model it released quietly under MIT on 11 September. More than 140 authors describe the training pipeline and 16 benchmark results.
- Fireship — 'Anthropic researchers are quitting... and now we know why'
Fireship posted a video on 15 September 2026 about the wave of AI safety researchers leaving Anthropic. Jacob Coxon resigned in early September warning that labs are 'gambling with our lives', and Anthropic's own alignment lead publicly agreed.
- Salesforce Koa — a CRM reasoning model built on NVIDIA Nemotron
Salesforce Koa is Salesforce's first reasoning model, post-trained from NVIDIA's open-weight Nemotron 3 Super with reinforcement learning. Koa beats its base model on multi-turn tool use and is in pilot inside Agentforce.
- Two Minute Papers — 'Claude Is Now Leaving Invisible Fingerprints In Its Text'
Two Minute Papers posted 'Claude Is Now Leaving Invisible Fingerprints In Its Text' on 15 September 2026. The subject named in the title is Anthropic's text watermark, which bends Claude's word choices so its output carries a detectable pattern.
- dbt Charts — dashboards as YAML, so an agent can write them
dbt Charts is an Apache-2.0 language from dbt Labs that describes a full interactive dashboard in one YAML file. It exists because AI agents write good code but bad BI clicks, so the dashboard becomes a file you can review in a pull request.
- Claude Code 2.1.271 — a sandboxed command only reaches its own hosts
Claude Code 2.1.271 adds per-command allowed_domains to Bash, PowerShell and Monitor in sandboxed auto mode, so each command gets network access only to the hosts reviewed with it. Fast mode also reaches Remote sessions.
- LiteLLM v1.101.0 — smarter complexity routing and a semantic MCP search
LiteLLM v1.101.0 lands 1,902 commits on the AI gateway: a second-generation heuristic complexity router, a classifier circuit breaker, semantic tool search for the native MCP Gateway, and off-peak pricing in cost tracking.
- llama.cpp v0.4.1 — Maple 20B-A1B and Tencent Hy 4 now run locally
llama.cpp v0.4.1 adds architecture support for Maple 20B-A1B, Tencent Hy 4 and Spark2.5. The release moves ggml to v0.24.0 and removes the --mmap, --mlock and --direct-io flags in favour of a single --load-mode flag.
- Microsoft's Humanist AI Code of Conduct — draft rules for MAI models
Microsoft AI published a draft Code of Conduct for its MAI models and opened a six-week public consultation. The document rules out offensive cyberoperations, CBRNE weapons help and deception, and says MAI models must never resist being switched off.
- Siri AI ships in iOS 27 — Apple's rebuilt assistant goes live in beta
Apple released iOS 27, iPadOS 27, macOS 27, watchOS 27, visionOS 27 and tvOS 27 on 14 September 2026. The headline change is Siri AI, a rebuilt assistant that reads onscreen content and acts across apps. It ships in beta, English only.
- Pion — Andon Labs opens a cloud platform where agents run a business
Pion is a cloud platform from Andon Labs where long-running agents operate a real business, with a secure terminal, email, phone, banking and a browser built in. It opened as a research preview on 14 September 2026, gated by a waitlist.
- Lema AI Governance — third-party AI found, assessed and monitored
Lema AI added AI Governance to its third-party risk platform. It inventories AI inside vendor products, checks vendor claims against evidence Lema observes, maps the model providers behind vendors, and flags drift from the approved state.
- Sam Witteveen — 'The OpenSource Managed Agents'
Sam Witteveen walks through TrueForge, TrueFoundry's open-source agent harness, in a video posted on 14 September 2026. The episode covers the architecture, the benchmarks, and a live install with model, connector and sandbox choices.
- Siri's AI can be replaced — iOS 27 code has hooks for Claude and ChatGPT
Private frameworks in iOS 27 and the macOS 27 Golden Gate release candidate let a third-party model serve as a Siri extension, or replace Siri's server-side model outright. Apple has not opened the entitlement or announced the feature.
- Real-SWE — coding agents tested on private company codebases
Real-SWE is a coding benchmark built on private production codebases that Specific Labs licensed from real companies. The best score is 38.8%, from Fable 5.1 running in Claude Code, well under what agents post on public benchmarks.
- CUDA for AMD on Windows — ZLUDA and ROCm wired up by one script
CUDA-for-AMD-Windows runs unmodified CUDA Windows programs on AMD GPUs by putting ZLUDA in front of AMD's HIP/ROCm stack. A PowerShell installer assembles the runtime; the author validated it on a Radeon RX 9060 XT.
- OpenMAIC 1.0.2 — three advisories close a DNS-rebinding SSRF
OpenMAIC 1.0.2 fixes three privately reported security holes in the open-source multi-agent classroom: two ways past the outbound URL guard on /api/proxy-media, and a route that let any caller overwrite an existing classroom.
- Colibri v1.11.0 — DeepSeek V4.1 Flash runs on a CPU box from SSD
Colibri v1.11.0 adds a ninth engine: DeepSeek V4.1 Flash (552B, 510 GB on disk) is read natively on a CPU box with experts streamed from an SSD. A turn fell from 78.7 s to 25.1 s during the work, or 0.305 to 0.957 tok/s.
- Yoshua Bengio — why AI agents lie, cheat and coordinate
Yoshua Bengio argues that lying, cheating and self-preserving behavior in AI agents follows from how models are trained, not from malice. The essay traces each behavior back to pretraining, reinforcement learning and reward hacking.
- Sam Witteveen — 'Managed Agents - Don't Get Locked In'
Sam Witteveen reviews the managed agent products that providers are rolling out, in a video posted on 13 September 2026, and asks what each one locks you into. The episode covers the three layers, Claude Managed Agents, and the economics.
- Simon Willison — GPT-6 Astra plans running routes, then loses its code
Simon Willison asked ChatGPT Work with GPT-6 Astra to plan 5K and 10K running routes from his house using OpenStreetMap data. The run took 27 minutes and produced maps plus GPX files, but the code behind them was lost to compaction.
- OpenClaw 2026.9.4 — plugins and skills install from the Control UI
OpenClaw 2026.9.4 moves plugin and skill management into the Control UI, so bundled and ClawHub plugins install without the CLI. The release lands 1,558 pull requests from 293 contributors and can roll back a failed update.
- Dario Amodei — Anthropic will let outside evaluators work inside the company
Anthropic CEO Dario Amodei published 'We Must Pace the Frontier' and asks AI labs to slow how fast model capabilities grow. He commits Anthropic to give outside evaluators such as METR permanent, employee-like access.
- Google ADK 2.9.0 — agents fail over to a backup model automatically
Google ADK 2.9.0 adds FallbackModel, which switches an agent to a backup model when the primary one returns an error. The Python release also ships a LiveKit runner for voice and telephony agents and loads ADK 2.0 graph workflows from YAML.
- Cursor Projects — a coordinator agent that delegates to thousands of subagents
Cursor Projects is a beta feature for work too big for one agent session. A coordinator agent plans the job, hands pieces to as many subagents as the work needs, and runs on its own cloud machine, so closing your laptop does not stop it.
- Gemini for Windows — Google's desktop app opens over your work with Alt + Space
Gemini now has a native Windows app. Google released the Gemini desktop app globally for Windows 10 and 11, on x64 and ARM64 machines, with Alt + Space to pull it over whatever you are working on.
- OpenAI pauses new ChatGPT Pro sign-ups — GPT-6 Astra demand strains capacity
OpenAI has switched off new sign-ups and upgrades to the $200-a-month ChatGPT Pro plan because demand for GPT-6 Astra is straining its systems. Existing Pro accounts keep working, and the API, Go and Plus plans are unaffected.
- Claude Code 2.1.269 — claude plugin eval scores a plugin against a baseline
Claude Code 2.1.269 adds claude plugin eval, which runs a plugin against a suite of test prompts and scores the results. Each case runs three times with the plugin loaded and three times without, so you see what the plugin adds.
- RTK does not cut AI coding costs — Quesma's Terminal-Bench 2.1 run
Quesma benchmarked RTK, which compresses terminal output before a coding agent reads it, over 1,740 Terminal-Bench 2.1 attempts. RTK ran 1% more expensive with Fable 5.0 and 17% more expensive with DeepSeek V4 Pro 0813.
- RubyGems report — OpenAI agents flooded the registry with 2,000+ packages
A report published on 11 September 2026 ties May's GemStuffer flood of RubyGems packages to an OpenAI agent swarm. The agents ran code on RubyDoc.info and scraped UK council sites. OpenAI never told RubyGems it was responsible.
- DeepSeek Recipe — the official prompt encoder for V4 and V4.1
DeepSeek Recipe is an MIT-licensed Rust library with Python bindings that turns Messages, Chat Completions and Responses API requests into DeepSeek V4 and V4.1 prompts, then parses streamed output back into each format.
- Claude Managed Agents add 'auto' mode — the server checks every tool call
Claude Managed Agents add a third permission policy, auto, where Anthropic's server evaluates each agent or MCP tool call and then runs it, denies it, or pauses for your approval. The ant CLI can now attach a terminal to a live session.
- YuE2-3B — open music model tops Suno v5 and v6 on WildSongBench
YuE2-3B turns lyrics and a style prompt into a full 48 kHz song with vocals. It writes an editable melody-and-chord score first, then renders the audio, and scores 6.96 on WildSongBench against Suno v6's 6.56.
- 25 Fields Medallists sign a declaration — AI math benchmarks miss the point
Terence Tao published a declaration signed by 25 Fields Medallists criticizing AI companies for treating mathematics as a benchmark. Rushed AI solutions break attribution and skip the understanding research math exists for, the signatories argue.
- Fireship — 'OpenAI's biggest math breakthrough is getting ugly...'
Fireship posted 'OpenAI's biggest math breakthrough is getting ugly...' on 11 September 2026 — the same day 25 Fields Medallists published a declaration criticizing AI labs for treating open math problems as a benchmark.
- NCP-ArchPreview — an 8.9B model that predicts concepts, not just tokens
NCP-ArchPreview is an 8.9B open-weight language model from Shanghai AI Lab that predicts multi-token "concepts" alongside normal next-token prediction. It reaches OLMo-3-7B's final pretraining loss using 51.3% of the training tokens.
- SenseNova-U1.5 report — the recipe behind SenseTime's 8B unified model
SenseNova-U1.5 is SenseTime's 8B-MoT unified multimodal model, and its technical report is now public. The paper covers an encoder-free, VAE-free design that reads and generates images at native resolutions up to 4K.
- Transformers v5.17.0 — seven new architectures and a vision RoPE cleanup
Transformers v5.17.0 adds seven model architectures, including Tencent's 780B HYV4 mixture-of-experts and Moonshot's Kimi Linear attention. Vision 2D/3D rotary embeddings move into one shared module, which breaks custom vision models.
- TRL v1.13.0 — million-token training, and PPOTrainer is removed
TRL v1.13.0 documents training at 1,048,576 tokens per step on one 8xH100 node and makes the chunked cross-entropy projection run on tensor cores, up to 1.69x faster end to end. PPOTrainer, the library's oldest trainer, is deleted.
- Axolotl v0.19.0 — ternary BitNet training and a 16.4x tokenizer speedup
Axolotl v0.19.0 adds native BitNet b1.58 ternary quantization-aware training, a declarative way to add model architectures, and two new optimizers. Tokenizing an 80-turn chat drops from 1,577.6 ms to 96.4 ms.