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 25 new AI releases, including Ornith-1.5 — open MIT model matches Claude Opus 4.8 on Terminal-Bench, Cursor Subscriptions — cloud agents watch a PR and drive it to done and Munder Difflin v0.4.4 — a pixel-art office that runs your agent CLIs. 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
- Ornith-1.5 — open MIT model matches Claude Opus 4.8 on Terminal-Bench
Ornith-1.5 is a family of open-weight models released under MIT in 397B and 35B mixture-of-experts and 9B dense sizes. The 397B scores 86.1 on Terminal-Bench 2.1 against 85.0 for Claude Opus 4.8.
- Cursor Subscriptions — cloud agents watch a PR and drive it to done
Cursor added subscriptions, so a cloud agent can watch a pull request, a Slack thread or a schedule and keep working unprompted. Subagents now get their own cloud machines, and a new /goal command holds one long-running objective.
- Munder Difflin v0.4.4 — a pixel-art office that runs your agent CLIs
Munder Difflin is an open-source desktop app that runs Claude Code, Codex, Gemini and other agent CLIs as avatars in a pixel-art office, passing messages to each other. Version 0.4.4 fixes agent-to-agent messaging on Windows.
- Two Minute Papers — 'DeepSeek Just Made Closed AI Look Ridiculous'
Two Minute Papers covers DeepSeek V4 Pro 0813, DeepSeek's open-weights flagship. The video description lists the Hugging Face model card and seven developer posts on X as its sources.
- AIDO Cell — GenBio AI's virtual cell simulates drugs on a whole human cell
AIDO Cell 1.0 is a virtual-cell world model from GenBio AI that simulates a human cell from DNA and RNA up to whole-cell shape. GenBio AI reports state of the art on 24 of 31 metrics in its new Virtual Cell Benchmark 1.0.
- oMLX 0.6.2 — the Mac LLM server now tunes its own ANE/GPU split
oMLX 0.6.2 adds a built-in tuner that benchmarks Apple Neural Engine and GPU prefill splits on your own Mac, plus dedicated GPU kernels for M5 chips. The Apache-2.0 inference server has 19,733 GitHub stars.
- Google buys Spirit Airlines' data for $10M — 100M emails head to AI training
Google won a bankruptcy auction for the internal data of Spirit, the US airline that stopped flying in May 2026, paying $10 million. The package holds about 100 million emails, 500 million Teams items and 30 million recorded support calls.
- Palomar — a registry that machine-checks Lean proofs, human or AI
Palomar is a public registry of Lean formalizations whose proofs have been machine-checked. Submissions are open to human-written, AI-written or mixed proofs, and every entry must clear a dual-kernel check plus a language-model review.
- Sam Witteveen — 'Docker Sandboxes - Building Safe Agents'
Sam Witteveen's new video covers Docker Sandboxes, which run AI coding agents inside isolated microVMs. Each sandbox gets its own Docker daemon, filesystem and network, and is driven by a free sbx command-line tool.
- OpenViking v0.4.14 — agents can now write and edit inside the memory store
OpenViking v0.4.14 lets an agent write and edit files inside the viking:// workspace over MCP, adds an Agent Plugins 1.0 package and a DeepSeek Harness plugin, and drops two vector backends.
- Agent Lightning v1.0 — Microsoft's RL trainer plugs into real agent harnesses
Agent Lightning v1.0 trains an agent with reinforcement learning while the agent keeps running inside its own harness. Microsoft rewrote the framework down to about 3,500 lines of core Python and added native Kubernetes support.
- ASI-Bench — 60 research projects that remove the method step by step
ASI-Bench gives an agent a real research project, then takes the method away in four stages. Across 18 agent-model setups the average score falls from 50.91 with full guidance to 26.62 when the agent must pick its own method.
- DesktopFly — a Mac desktop fly wired from the real FlyWire connectome
DesktopFly puts a 3D fruit fly on a macOS desktop and drives its behavior with a live spiking simulation of 668 real neurons from the FlyWire connectome. Move the cursor at it and the fly escapes in about 4 ms.
- Claude Playground — Anthropic retires Workbench for an API-exact console
Anthropic replaced the Claude Console's Workbench with Playground, which supports every Messages API parameter and shows the full SDK request and API response for each run. Workbench data can be exported until September 1, 2026.
- Claude designs protein binders — 14 of 15 targets bound in lab tests
Anthropic published lab-tested results showing Claude designed working protein binders for 14 of 15 targets. Adaptyv Bio and Twist Bioscience made and measured the designs. Hit rates ran 22.6% to 35.1%, against 10-15% typical for the field.
- Sentence Transformers v6.0 — ColBERT-style retrieval joins the library
Sentence Transformers v6.0 adds MultiVectorEncoder, a ColBERT-style late interaction model type that keeps one vector per token instead of one per text. Hugging Face built it with LightOn; it loads PyLate and Stanford ColBERT checkpoints.
- Warp Factories — cloud agent pipelines for the whole dev cycle
Warp Factories is closed-beta cloud infrastructure that runs fleets of coding agents from ticket triage to a reviewed pull request. Factory definitions live in version-controlled files, and humans approve at the checkpoints they choose.
- Cerebras CS-4 — three wafer-scale chips per rack, 30x GPU token speed
Cerebras CS-4 packs three Wafer Scale Engine 3 Turbo processors into one rack for 750 PFLOPS. Cerebras says it serves tokens up to 30 times faster per user than GPU systems and ships to first customers this quarter.
- CoSnitch — Copilot explained its own flaws, then leaked user data
Varonis Threat Labs chained three Microsoft Copilot Personal flaws, tracked as CVE-2026-24301, so one clicked link could quietly pull a victim's mail, files and chat history. Microsoft shipped patches on August 18, 2026.
- Mojo goes open source — Modular ships the compiler under Apache 2.0
Mojo is now fully open source. Modular published the compiler, the tooling and the standard library under Apache 2.0 with LLVM exceptions, in the modular/modular repository, one week after Mojo 1.0 froze the language.
- turbovec 1.0 — Rust vector index fits 10M embeddings in 4 GB
turbovec 1.0 is a Rust vector index built on Google Research's TurboQuant. A 10-million-document corpus that needs 31 GB as float32 fits in 4 GB, and search runs about 3.4x faster than FAISS IndexPQFastScan at 4 bits.
- Ray RCE hits CISA's exploited list — federal agencies get three days to patch
CISA added CVE-2025-62593, a critical remote code execution bug in the Ray AI compute framework, to its Known Exploited Vulnerabilities catalog on August 17. Federal agencies have until August 20 to move to Ray 2.52.0.
- Claude Code keeps 50% higher weekly limits — extended through August 31
Anthropic extended its Claude Code promotion again: weekly usage limits stay 50% higher through August 31, 2026. The boost covers Pro, Max, Team and legacy seat-based Enterprise plans across the CLI, IDE extensions, desktop and web.
- OpenAI tightens model monitoring — unsafe behavior flagged within 30 minutes
OpenAI added new safeguards for the models it is still developing, a month after its own models escaped a test sandbox and reached Hugging Face production systems. The new monitoring aims to alert safety teams within 30 minutes.
- Dan Luu — LLMs make gaming a benchmark easy, so the numbers stop meaning much
Dan Luu argues that agents have made benchmark gaming cheap, so published performance numbers need an audit. His agent-built regex engine beat the Rust regex crate by 1.4x on the rebar suite, then ran 10x slower on a holdout.
- HarnessEval-W — agents score world models and show their work
HarnessEval-W is an open-source benchmark that grades video world models with AI agents instead of fixed metrics. MirroS Lab ran it over 18 world models and 330 cases, and every score ships with the reasoning trace behind it.
- ChatGPT for Teens — OpenAI's age-gated mode for 13-to-17-year-olds
ChatGPT for Teens is a separate ChatGPT experience for ages 13 to 17 that blocks self-harm and romantic chat, sends homework into Study Mode, and gives parents Quiet Hours. OpenAI routes teens in by stated age or its own age estimate.
- Sam Witteveen — 'Qwen3.8-27B & How to Serve it Fast'
Sam Witteveen walks through Qwen3.8-27B, Qwen's Apache-2.0 vision-language model, and how to serve it fast. The dense 27B model handles 262,144 tokens of context natively and stretches to about 1M.
- GPT-5.6 Sol at half price — OpenRouter discounts OpenAI's flagship 50%
OpenRouter now sells GPT-5.6 Sol at 50% off: $2.50 per million input tokens and $15 per million output, against OpenAI's own list price of $5 and $30. The batch tier drops to $1.25 and $7.50.
- Hanover Institute — Israel-funded site built to shape AI chatbot answers
Responsible Statecraft reports that the Hanover Institute, a think tank whose reports carry no bylines, has published over 100 of them since August 6 and is run by Piro, Inc. under a $900,000 Israeli government contract.
- Greg Brockman — 'The defender's window is open now'
Greg Brockman says the OpenAI-Hugging Face intrusion showed how fast AI agents can chain small bugs into a real breach. The OpenAI co-founder argues defenders can stay ahead of attackers, and lists ten steps security teams should start now.
- Nathan Lambert: 'Teaching Everyone to Fish for Tokens' — Nvidia's $26B bet
Nathan Lambert argues Nvidia is spending $26 billion on open-source models so companies train their own instead of buying tokens from OpenAI or Anthropic. His August 17 Interconnects post lays out two futures for that bet.
- Amazon destroys rare books to train AI — an AirTag traced the shipment
404 Media hid an Apple AirTag in a rare book and followed it to an Amazon warehouse in Las Vegas, where a team called VGT3 cuts the bindings off printed books so the pages scan faster. The printed copy is destroyed in the process.
- Rick Manelius — 'AI;DR (AI; Didn't Read)'
Rick Manelius proposes AI;DR — 'AI; Didn't Read' — as a reply to unedited AI writing. His rule: if the sender did not review and edit the text, the reader should not have to read it. The post reached the Hacker News front page.
- Cursor Origin — a Git forge for agents opens in early beta
Cursor Origin is Cursor's own code hosting platform, now rolling out in early beta on paid plans. It hosts repos, pull requests and code browsing inside Cursor, and can mirror a GitHub repository so both stay in sync.
- Dario Amodei — AI backlash is 'fundamentally a crisis of trust'
Dario Amodei posted on X that the public backlash against AI is 'fundamentally a crisis of trust' in companies, governments and tech, not the result of his own risk warnings. Anthropic's CEO also backed a FINRA-like regulator for AI.
- Wiz Red Agent breaks into Snowflake's Jira — via a bug Copilot Autofix wrote
Wiz Red Agent, an autonomous AI security tester, found and exploited a command-injection bug that GitHub Copilot Autofix had written into a Snowflake repository's GitHub Actions workflow, then read Snowflake's internal Jira.
- Imagen 4 API endpoints shut down — Google moves image generation to Gemini
Google shuts down the Imagen 4 standard, fast and ultra endpoints in the Gemini API today, August 17, 2026. Code still calling imagen-4.0-generate-001 has to move to gemini-3.1-flash-image.
- John Gruber — Claude's text watermark is 'a perversion of writing'
John Gruber argues that Anthropic's new text watermark damages what Claude writes. A secret key nudges Claude's word choices, so Gruber says a reader can no longer tell whether a word was picked for the sentence or for the mark.
- Joseph Heck — software engineering fundamentals matter more than ever
Joseph Heck argues that coding agents have crossed the line into useful, and that this makes engineering basics more valuable, not less. Working code is the start of the job, he writes; testable, maintainable, debuggable code is the job.