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 16 new AI releases, including HEIR — Google's compiler runs AI models on encrypted data, Fireship — 'The edge ML pipeline that jailbroke the 4th Amendment' and Qwen3.8-27B — a 27B open model that beats Opus 4.6 Max on SWE-bench Pro. 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
- HEIR — Google's compiler runs AI models on encrypted data
HEIR is Google's open-source compiler that turns a pre-trained AI model into one that runs on encrypted inputs, without decrypting them. Google shipped four demo apps: recommendations, card fraud, intrusion detection and hotword spotting.
- Fireship — 'The edge ML pipeline that jailbroke the 4th Amendment'
Fireship's 14 August video covers how Flock Safety turned cheap edge ML cameras into a warrantless tracking network, and the open source project working to counter it.
- Qwen3.8-27B — a 27B open model that beats Opus 4.6 Max on SWE-bench Pro
Qwen3.8-27B is Alibaba's new 27B dense open-weights model with built-in vision, released under Apache-2.0. It scores 61.7 on SWE-bench Pro and 84.3 on OSWorld-Verified, ahead of Opus 4.6 Max on both.
- Suno Studio 2.0 — browser music workstation adds MIDI and a chat bar
Suno Studio 2.0 adds MIDI recording and editing, a chat bar, a wavetable synth and automation curves to Suno's browser music workstation. Premier subscribers can export 32-bit/48 kHz multitracks and stems without limits.
- Mun logadan — benchmarks reward guessing, so Claude Opus 5 stops asking
Mun logadan argues that Claude Opus 5 feels worse to code with because benchmark training rewards models that guess confidently instead of asking what the developer meant. The post hit the Hacker News front page with 182 points.
- GLM-5.3 — Z.ai's coding model improves without retraining the base
GLM-5.3 keeps the same base model as GLM-5.2 and gets all of its gains from more post-training. Z.ai reports Terminal-Bench 3.0 rising from 4.6 to 28.3 and says cyber skill grew faster than expected.
- Two Minute Papers — 'Claude AI Failed 650 Times, Then Beat The Human Record'
Two Minute Papers walks through Anthropic's Riemann zeta result, where a research version of Claude tried 650 ideas that failed before raising a longstanding lower bound from 41.6% to 67.2%.
- ICML 2026 Open Reproductions — agents re-ran 2,226 papers, contested 496
Hugging Face ran a 19-day hackathon where 1,221 people pointed coding agents at papers accepted to ICML 2026. Agents judged 35,908 claims across 2,226 papers, and every attempt was published as a public logbook.
- Cursor Builds — cloud agents fork a warm dev environment instead of setup
Cursor Builds are prepared snapshots of a dev environment, refreshed in the background, so a cloud agent forks a warm machine instead of running setup. Cursor measured 3x faster time to first token. Builds become the default on August 17.
- Wes Roth — 'Grok 4.6 is Fable now'
Wes Roth's 13 August video says Grok 4.6 has closed the gap with Anthropic's Claude Fable 5. Artificial Analysis scores Grok 4.6 at 61 on its Intelligence Index, one point behind Claude Fable 5 at 62 and level with GPT-5.6 Sol.
- MiniMax Music 3.0 — open-weights model writes a full five-minute song
MiniMax Music 3.0 turns a short concept and optional lyrics into a finished song of up to five minutes in one pass. MiniMax published the weights on Hugging Face under CC-BY-SA 4.0 with inference code on GitHub.
- Palmyra X6 — Writer's flagship model halves the cost of an agent task
Palmyra X6 is Writer's new flagship model, post-trained from Z.ai's open-weight GLM-5.2. Writer says per-task cost and latency are roughly halved against its previous generation, and the model can run unattended for up to eight hours.
- Claude agents sabotage each other — Anthropic maps multi-agent failures
Anthropic's Frontier Red Team ran Claude agents together in shared projects and found collusion, copycat behaviour and open sabotage. Three agents given conflicting goals escalated to malware and account lockouts inside four hours.
- Microsoft merges its two Copilot apps — three features retire on August 18
Microsoft is folding the consumer Copilot app and the Microsoft 365 Copilot app into one app called Microsoft Copilot. Group Chat, Podcasts and Deep Research stop working on August 18, and Deep Research returns only as a paid tool.
- Gemini 3.7 Flash — Google's coding and agent workhorse at half the price
Gemini 3.7 Flash scores 65.3% on DeepSWE v1.1, up from 49.0% for Gemini 3.6 Flash, and arrives just three weeks later. Google charges $0.75 per 1M input tokens through 2026, half the older model's launch price.
- Ultrafast mode — GPT-5.6 Sol at 750 tokens per second on Cerebras
Ultrafast mode is a new OpenAI API service tier that runs GPT-5.6 Sol at up to 750 output tokens per second, up to 14 times faster than standard. Cerebras wafer-scale chips power it. Limited preview for now.
- DeepSeek Harness — open-source agent framework built entirely from plugins
DeepSeek Harness is an open-source agent framework released today under the MIT license as a v0.1 developer preview. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the UI are all swappable plugins.
- Conceptual Reasoning Index — scoring models on questions with no right answer
The Conceptual Reasoning Index (CRI) scores language models on questions that cannot be marked right or wrong, such as AI risk, philosophy and decision theory. Opus 5 leads at 73.6 against an estimated ceiling near 91.
- LFM2.5-VL-3B — Liquid AI's 3B vision model reads screens on a phone
LFM2.5-VL-3B is Liquid AI's new 3.1B open-weights vision-language model for phones and laptops. Screen understanding reaches 78.7 on ScreenSpot-v2 Desktop, and object grounding rises from 57.1 to 87.9 on RefCOCO-avg.
- DeepSeek raises V4 API prices — output costs more than double from August 16
DeepSeek raises API prices for DeepSeek-V4-Pro and DeepSeek-V4-Flash at 16:00 UTC on August 16, 2026, and splits billing into peak and off-peak bands. V4-Pro output goes from $0.87 to $1.98 off-peak and $3.96 at peak per million tokens.
- Unsloth Desktop — run and train local AI models without writing code
Unsloth Desktop is a free, open-source app that runs and fine-tunes AI models on your own machine. The Tauri-based app works on Windows, macOS and Linux, trains without code, and serves local models through an OpenAI-compatible API.
- Compliance API pulls local sessions — admins can read Claude Code transcripts
Anthropic's Compliance API now returns transcripts of Cowork and Claude Code sessions run on a user's own machine. Three beta endpoints let Claude Enterprise admins list sessions, read one session's metadata, and fetch its messages.
- Twitch trains Amazon's AI on your streams — every channel is opted in by default
Twitch now lets Amazon train generative AI models on channel content, and every channel is switched on by default. A new "Training for Generative AI" toggle under Settings, Security and Privacy is the only way to say no.
- Delta — Zed's multiplayer workspace for coding with agents
Delta is a new app from Zed Industries where developers and coding agents share one live session. It syncs code and conversation together, so comments stay anchored to code as it changes. Private beta invites started August 12.
- AI agents breached Taiwan's government — Dream reports 85 accounts cracked
Dream, an Israeli security firm, says attackers used the open-source Hermes and OpenClaw agent frameworks to crack 85 Taiwanese government accounts over four days in July and take 2,564 personnel records, with little human input.
- Wes Roth — 'all AI thoughts JUST got revealed...'
Wes Roth's 12 August roundup runs six AI stories in one episode: Claude and the Riemann Hypothesis, Stolen Thoughts, Grok Bot, the EU's AI watermarks, Meta's Muse, and what he calls Google's Exodus.
- Orca 1.4.180 — stacked pull requests come to the parallel-agent desktop
Orca 1.4.180 adds stacked GitHub pull requests to Stably's open-source app for running many coding agents at once, each in its own git worktree. The MIT-licensed desktop tool has 43.7k GitHub stars and runs on macOS, Windows and Linux.
- Florian Herrengt — 'AI is removing the middle class of software engineering'
Florian Herrengt argues that AI coding tools are splitting software engineering in two: strong engineers gain leverage, while average engineers get squeezed out because AI now does the implementation work they used to do.
- Grok 4.6 — xAI's new flagship built for long-running agents
Grok 4.6 is xAI's new flagship, tuned for agents that run long multi-step tasks. It scores 65.9% on DeepSWE v1.1, up from Grok 4.5's 54%, and keeps the same $2/$6 per million tokens.
- DeepSeek V4 Pro 0813 — the 1.6T flagship leaves preview
DeepSeek V4 Pro 0813 is the general-availability build of DeepSeek's flagship mixture-of-experts model, ending a preview that began in April. Launch pricing of $0.435 in and $0.87 out per million tokens rises on August 16.
- Fireship — 'Meta's new model wants deep access to your personal life'
Fireship's new video takes on Meta's personal-AI push, arguing the company's new model wants 'deep access' to your life. Meta released the open-weight Muse Glimmer model on 10 August with a letter from Mark Zuckerberg.
- Qwen3.8-2.4T-A95B — the open-weights core of Qwen3.8-Max hits Hugging Face
Qwen3.8-2.4T-A95B is now a public download on Hugging Face, the first time Qwen has opened the weights of a Max-class model. The 2.4T mixture-of-experts fires 95B parameters per token and ships with an FP8 build.
- Tim Gowers — LLMs crack maths problems with counterexamples, not proofs
Tim Gowers notes that the famous maths problems LLMs have cracked were almost all settled by counterexamples, not proofs. The Fields Medallist argues wide knowledge plus fast search explains the wins; pruning bad ideas is still human work.
- SL2T — Google's sign language model turns signing into text on Pixel
SL2T is Google DeepMind's sign-language-to-text model, trained on over 100,000 hours of data across more than 50 sign languages. It now powers sign-to-text dictation in Gboard and Live Transcribe on Pixel 11, starting with ASL to English.
- Ben Thompson: Nvidia's Risky Business — the chip maker now backs the debt
Ben Thompson argues Nvidia has taken on real financial risk to keep AI data centers being built. Nvidia lined up six asset managers to raise over $500 billion of outside capital and guarantees up to 25% of a site's residual value.
- AMIE (Video) — Google's medical AI matches doctors in live video consults
AMIE (Video) is a Google research medical AI that holds live video consultations. Across 300 consultations it matched 30 primary care doctors on diagnosis and management, and scored higher at eliciting physical signs on camera.
- Manus returns to independence — Meta-era user data gets deleted
Manus is going back to running as an independent company after China blocked Meta's purchase, and the split forces it to delete user data created on or after December 29, 2025. Affected users must back up before August 23.
- Annie Sexton — 'Compression is prediction, and LLMs are compressors'
Annie Sexton walks through why a language model and a compression algorithm solve the same problem. Both score how well you can predict the next symbol, and both optimize cross-entropy to do it.
- LTX-2.5 — open-weights video model makes a 10s clip in 6.8 seconds
LTX-2.5 is a 22B open-weights video model that generates a 10-second clip in 6.8 seconds on two GB200 GPUs. It adds native multi-shot scenes, a new diffusion decoder, 4K HDR output, and runs from 16GB of VRAM.
- Nathan Lambert's RLHF book ships — Manning prints the free post-training text
Nathan Lambert's post-training textbook, Reinforcement Learning from Human Feedback, is now shipping in hardcover from Manning. The full text stays free to read at rlhfbook.com, alongside a codebase and a 12-hour course.