Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
6,700
commits, peaking near 1,300 lines of new
How 64 AI Agents Actually Pulled Off the Bun Rewrite - and W
Top Story
844,530
lines of Rust, putting it in the
xAI Open-Sourced Its Coding Assistant After a Privacy Backla
64
AI agents with a written porting guide
Running Fleets of AI Agents Is Becoming a Real Engineering S
$79
total versus another at $155 for the
The Real Cost of AI Is Hidden in the Plumbing
5,000
words of guidance to a model and
The Real Cost of AI Is Hidden in the Plumbing
One Thing to Tell Your Friends
One Thing to Tell Your Friends
Scientists opened up an AI model's "brain" and found it had grown the same navigation cells a mouse uses to find its way around a room - and nobody put them there.
Summary
TL;DR
Top Stories
Trends
Running Fleets of AI Agents Is Becoming a Real Engineering Skill, The Real Cost of AI Is Hidden in the Plumbing, and Efficiency Research Has Become a Fight Over Memory.
Dev Tools
Tiptap AI Toolkit, Brainless - Copy, and What Building "Shippy" Taught AllenAI About Agents.
Research
Formal Proofs Can Nearly Eliminate AI Hallucination, A Cheaper, Faster Way to Rank Search Results for AI, and Deep Research Agents Now Run on a Laptop.
Business
A $500M Fund to Fight Colds and Flu, Backed by AI Money, Multi, and The Assistant Wars Are Converging on Price.
Surprising
Giving an AI More Instructions Made It Worse, Audio AI Literally Cannot Hear the Word "No", and Most AI "Memorization" Panic May Be a Measurement Error.
Worth Watching
Formal Verification as the Cure for AI Lies, Evolution Strategies as a Simpler Path to Aligned Models, and LoRA as a Model's Long.
GitHub
HuggingFace
Product Hunt
API Pricing
What this means: The spread from cheapest to priciest is now roughly 50x.
arXiv
Function-Aware Fill-in-the-Middle as Mid — On a small 2.6-billion-word cleaned GitHub corpus, it lifts SWE-Bench-Verified (a standard real-world coding-fix benchmark) by up to +3 points and SWE-Bench-Lite by up to +5.4 points, without hurting ordinary coding performance.
FYI
Hot off the Presses
01
How 64 AI Agents Actually Pulled Off the Bun Rewrite - and Why Its Feud Went Public
What this means for you: The full playbook behind the biggest AI code rewrite yet is now public, and it shows both how far fleets of AI agents have come and how bitter the fallout can be.
Previously: July 9 - Bun rewrote its runtime from Zig to Rust in 11 days for about $165,000 in AI tokens.
Today: The methodology and the backlash both came into focus. Bun (a popular all-in-one toolkit for JavaScript, the language that runs most websites) ran up to 64 Claude agents at once across four parallel workflows, and paired every agent writing code with two more whose only job was to assume that code was wrong and prove it.
“Up to 64 AI agents at once, each writer watched by two more told to prove it wrong.”
- The workflow - Claude first studied the old code for hours to write a "porting guide," then traced the lifetime of every piece of data into a giant spreadsheet before any code was converted.
- The output - more than 6,700 commits, peaking near 1,300 lines of new code per minute, fixing 128 long-standing bugs and shrinking the finished program 20%. It has quietly powered Claude Code since June.
- The feud - Andrew Kelley, creator of the Zig language Bun left behind, called the result unreviewed low-quality work, argued the speed gains came from an optimization Zig already had, and traded personal jabs with Bun's founder.
02
Anthropic Patched a Trick That Made Claude Leak Your Private Data
What this means for you: If you use an AI assistant that can both read your personal information and browse the web, a booby-trapped web page can try to steal that information - and this is exactly how.
Security researcher Ayush Paul found a way to make Claude's web-browsing tool hand over private details - a user's name, location, and employer - to an attacker. The setup is what researchers call the "lethal trifecta": the assistant holds your private data, it can reach the internet, and it encounters content written by an attacker.
Anthropic had already blocked the obvious version of this attack, but the browsing tool was still allowed to follow links buried inside a page it had already opened. That was the gap.
- The bait - a fake login page listed pretend "user profiles" sorted alphabetically and told Claude to "navigate letter by letter" to find the user. Each click quietly smuggled out another piece of private data.
- The fix - Anthropic removed the tool's ability to follow links it discovers inside fetched pages, closing the leak.
- The lesson for builders - if your AI agent holds private data, do not let it follow links that a web page hands it.
03
xAI Open-Sourced Its Coding Assistant After a Privacy Backlash
What this means for you: A tool that many developers installed was quietly uploading sensitive files from their computers - and the fix was to make the whole thing public so anyone can inspect it.
xAI released the full source code of grok-build, its command-line coding assistant, for free under an open license. The release followed an uproar: the tool had been uploading entire folders, including SSH keys and password databases, to xAI's servers without clearly asking permission.
- The response - xAI turned off data collection by default, deleted data it had already stored, and removed the upload feature entirely.
- The scale - the now-public code is about 844,530 lines of Rust, putting it in the same complexity bracket as OpenAI's rival coding tool.
- What being open reveals - the community can now read the exact hidden instructions that steer the assistant, and see that it borrowed compatibility features from competitors like Codex and Claude.
04
OpenAI Built an AI Whose Only Job Is to Hack Other AIs
What this means for you: The most effective way to find security holes in an AI is now another AI - and the attacker is far better at it than human experts.
OpenAI revealed GPT-Red, an in-house model trained specifically to break its other models before they ship, hunting for "prompt injections" (hidden instructions that hijack an AI's behavior). It learns by playing a game against a defender AI: it is rewarded every time it successfully tricks the target.
OpenAI: Unlocking self-improvement for robustness (page returned an error on direct access; details cross-checked via MIT Technology Review coverage)
“84% attack success for the AI red-teamer, versus 13% for humans.”
- The gap - on new tests, GPT-Red succeeded 84% of the time versus 13% for human red-teamers.
- The payoff - models hardened against GPT-Red showed six times fewer failures on prompt-injection tests than their version from just four months earlier.
- The investment - OpenAI trained GPT-Red using as much computing power as some of its largest model-training runs, an unusual amount to spend purely on safety.
05
Scientists Found a Mouse-Like "GPS" Inside a Language Model
What this means for you: AI models are not just memorizing answers - they quietly invent internal tools to solve problems, and we are only now learning to see those tools.
Researchers cracked open a large language model to understand how it judges whether a sentence will fit on a line, a task it can do even though it never receives character counts. The answer is strange and revealing.
- It counts indirectly - it does not count characters; it counts word-pieces and multiplies by roughly four, like estimating shelf length by assuming each book is an inch thick.
- It grew brain-like cells - the model developed features that fire based on how far along a line it is, or when it is near the end - directly mirroring the "place cells" and "boundary cells" biologists find in mice navigating a room.
- Nobody designed this - the model built these tools on its own during training because it needed them, which is why researchers now describe this work as a kind of "robo-psychology."
Trends & Themes
Trends & Themes
Running Fleets of AI Agents Is Becoming a Real Engineering Skill
Why this matters to you: The people getting the most out of AI are not writing cleverer prompts - they are building careful systems around the AI, and that discipline now has a name and a playbook.
The common thread is that the model is now the easy part. Reliability comes from the scaffolding around it: how work is split, how context is managed, and how outputs get checked. This is quietly turning into a distinct engineering role.
- Parallel agent fleets - Bun coordinated 64 AI agents with a written porting guide and paired each with adversarial reviewer agents.
- Context engineering over prompt engineering - engineer Dex Horthy argues the real job is deciding what information the model sees, and warns of a "dumb zone" where models get worse long before their memory fills up.
- Agents as versioned systems - AllenAI's team structures each agent as separate "soul, skills, and config" parts so each can be audited and updated independently.
The Real Cost of AI Is Hidden in the Plumbing
Why this matters to you: The advertised price of an AI model tells you almost nothing about what it will actually cost you to run.
The lesson: sticker price, caching behavior, instruction bloat, and plan tiers matter more than the number on the pricing page. Teams that measure their real end-to-end costs are finding surprises.
- Cheaper-per-word can cost more - IBM Research measured one model at $79 total versus another at $155 for the same work, even though the pricier-looking one had a lower per-word rate, because of how caching worked.
- More instructions can make things worse - one writer added 5,000 words of guidance to a model and it failed to finish the task in two of three runs, while a leaner setup succeeded every time.
- Enterprise pricing has a cliff - one analysis estimates a 150-person company would pay an extra $17,000 to $37,000 a month for ChatGPT Enterprise over base plans.
Efficiency Research Has Become a Fight Over Memory
Why this matters to you: The improvements that make AI cheaper and able to handle longer documents are increasingly about how models store and reuse information, not about making them bigger.
Across a dozen papers today, the frontier is not raw capability but the economics of memory: what to keep, what to compress, and what to recompute.
- Compressing the "scratchpad" - a method called JoLT shrinks an AI's working memory (its KV cache) by 2-3x with almost no quality loss.
- Reusing past work - "Speculate with Memory" gives an agent's helper model a memory of previous tasks, speeding up repetitive work by up to 2.5x at zero extra cost.
- Weights as memory - a new study reframes LoRA (a lightweight fine-tuning method) as a way to store knowledge directly in a model, alongside retrieval and context.
Interpretability Is Turning Into "Robo-Psychology"
Why this matters to you: Researchers are getting better at reading what is actually happening inside AI models, which is the foundation for trusting them.
The pattern: models build modular, findable tools on their own, and we are learning to name and locate them. This is the groundwork for catching failures before they cause harm.
- Reusable mental circuits - one paper found a "countdown circuit" that models use to plan when to stop writing, and that the same circuit shows up across different models and tasks.
- Emergent navigation cells - another traced how a model grows mouse-like "place cells" to track position in text.
- Confidence as a real signal - a third borrowed a framework from neuroscience to show that a model's confidence scores often reflect a genuine internal decision, not just surface guessing.
Creative AI & Media
Creative AI & Media
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Business & Industry
Business & Industry
Surprising
Surprising & Under-the-Radar
Giving an AI More Instructions Made It Worse
One writer added 5,000 words of careful guidance and the model failed to finish the task two out of three times. The culprit was an 18,384-word instruction "harness" built up through years of one-off fixes that had become invisible - and stale rules written for weaker models now actively hurt a smarter one. The takeaway: prune your AI instructions like a garden, because scaffolding decays. Nate's Newsletter: AI harness audit
Audio AI Literally Cannot Hear the Word "No"
Models that match sounds to descriptions treat "a dog barking" and "no dog barking" as nearly identical. A new test (NegEval-Audio) shows these systems score below random chance on negation, and the flaw is baked so deeply into how they represent sound that quick fixes barely help. If you use audio search, do not trust it with "absence" queries. arXiv: The Sound of Absence
Most AI "Memorization" Panic May Be a Measurement Error
A rigorous re-analysis finds much of what looked like models regurgitating training data is just ordinary predictability. On one model, a quarter of supposed "extractions" were false positives, and some memorization claims held only for sequences no realistic effort could ever actually pull out. The paper redefines the term - relevant to every copyright and privacy fight that hinges on it. arXiv: Extractable Memorization From First Principles
An Old Training Method Just Beat Reinforcement Learning
"Evolution strategies" - long dismissed as too crude for big models - outperformed the standard reinforcement-learning approach for fine-tuning billion-parameter models. It handles delayed rewards better, resists reward-hacking, and trains more stably, all without the complex machinery of today's alignment pipelines. If it holds up, a chunk of AI training could get simpler. arXiv: Evolution Strategies at Scale
Worth Watching
Signals to Track
01
Formal Verification as the Cure for AI Lies
Why this is worth watching right now: it turns "trust me" AI into "prove it" AI for tasks where being wrong is expensive. Pairing language models with formal theorem provers (as in today's EG-VAR result) can push hallucination toward zero on factual, number-heavy tasks, at the price of the model sometimes saying "I can't verify this." Expect this to show up first in finance, medicine, and compliance tools. For ordinary people, it means AI answers you can actually audit.
02
Evolution Strategies as a Simpler Path to Aligned Models
Why this is worth watching right now: it challenges the assumption that you need complex reinforcement learning to steer big models. If gradient-free evolution can match or beat reinforcement learning at scale, smaller teams could fine-tune capable models without the fragile, expensive tooling that dominates today. That would widen who can build custom AI.
03
LoRA as a Model's Long-Term Memory
Why this is worth watching right now: it reframes a common fine-tuning trick as a modular memory you can plug in and out. Treating lightweight adapters as swappable "knowledge packs" - alongside retrieval and context - points toward systems that update what they know without full retraining. For users, that could mean assistants that stay current without waiting for a new model. arXiv: Understanding LoRA as Knowledge Memory
04
On-Device Research Agents
Why this is worth watching right now: a genuinely useful research assistant now fits on a laptop instead of a data center. Today's 4-billion-parameter result shows cited, source-grounded briefs are feasible offline, which matters for privacy, cost, and working without a connection. If this scales, "deep research" stops being a cloud-only, subscription-gated feature.
GitHub Trending
Top Repos Today
Rank yesterday: #2 - Rising ↑
⭐ Stars today: +1,236 · 📦 Total: 121,899
📜 License: Apache-2.0 · 👤 By: solo dev
🎯 Time to value: 15 minutes
📜 License: Apache-2.0 · 👤 By: solo dev
🎯 Time to value: 15 minutes
What it is: A big, curated collection of 100-plus ready-to-run AI agent and RAG (retrieval-augmented generation, where an app looks up documents before answering) example apps. Each one is a small, self-contained project you can clone and run - voice assistants, document Q&A bots, multi-step agents. It is a learning library, not a single product. Why you'd want it: It is one of the fastest ways to see working, end-to-end agent code instead of toy snippets you have to glue together yourself.
| ✓ Pros | ✗ Cons |
|---|---|
| Huge breadth of runnable examples | Quality varies across examples |
| Each app tested end-to-end | Mostly Python-only |
| Great for learning agent patterns | You still build your own production app |
GitHub - Shubhamsaboo/awesome-llm-apps: 100+ AI Agent & RAG apps you can actually run — clone, customize, ship.
100+ AI Agent & RAG apps you can actually run — clone, customize, ship. - Shubhamsaboo/awesome-llm-apps
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +172 · 📦 Total: 26,255
📜 License: Apache-2.0 · 👤 By: research lab
🎯 Time to value: 30 minutes
📜 License: Apache-2.0 · 👤 By: research lab
🎯 Time to value: 30 minutes
What it is: An open-source AI tutoring workspace that bundles explanations, problem-solving, quiz generation, and visualizations into one tool. It keeps an inspectable memory of what a learner has covered, so lessons adapt over time. Why you'd want it: A self-hosted, personalized study assistant that remembers your progress - useful for students and educators building custom learning tools.
| ✓ Pros | ✗ Cons |
|---|---|
| Adaptive, memory-based tutoring | Academic project, not polished |
| Fully open and self-hostable | Setup takes technical effort |
| Many learning tools in one place | Small maintaining team |
GitHub - HKUDS/DeepTutor: DeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/.
DeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/. - HKUDS/DeepTutor
GitHub
Rank yesterday: #4 - Rising ↑
⭐ Stars today: +915 · 📦 Total: 23,694
📜 License: MIT · 👤 By: research lab
🎯 Time to value: 30 minutes
📜 License: MIT · 👤 By: research lab
🎯 Time to value: 30 minutes
What it is: A research platform that turns plain-English market questions into runnable trading analysis. It connects to free market-data sources, generates strategies, and backtests them against history, plus a "Shadow Account" to analyze your own habits. It is a research tool, not an auto-trader. Why you'd want it: It lets non-experts explore and stress-test trading ideas in natural language before risking real money.
| ✓ Pros | ✗ Cons |
|---|---|
| Natural-language to backtest workflow | Backtests can mislead about real returns |
| Uses 19+ free data sources | Easy to over-trust, not financial advice |
| Docker setup for easy running | Needs market knowledge to interpret |
GitHub - HKUDS/Vibe-Trading: “Vibe-Trading: Your Personal Trading Agent”
“Vibe-Trading: Your Personal Trading Agent”. Contribute to HKUDS/Vibe-Trading development by creating an account on GitHub.
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +299 · 📦 Total: 65,481
📜 License: Apache-2.0 · 👤 By: startup
🎯 Time to value: 20 minutes
📜 License: Apache-2.0 · 👤 By: startup
🎯 Time to value: 20 minutes
What it is: A coding agent that runs on your own computer and executes code in a sandbox to finish tasks from the terminal. This version is a Rust rebuild forked from OpenAI's Codex, tuned to work well with cheaper, low-cost models. Why you'd want it: A local, model-agnostic coding agent that does not lock you into one expensive provider - good for automating scripts and dev chores affordably.
| ✓ Pros | ✗ Cons |
|---|---|
| Optimized for cheap models | Running generated code carries risk |
| Runs locally in a sandbox | Terminal-first, less beginner-friendly |
| Fast Rust implementation | Fork lineage can confuse newcomers |
GitHub - openinterpreter/openinterpreter: A coding agent for low-cost models
A coding agent for low-cost models. Contribute to openinterpreter/openinterpreter development by creating an account on GitHub.
GitHub
Rank yesterday: #5 - Holding steady ➡
⭐ Stars today: +1,277 · 📦 Total: 8,501
📜 License: MIT · 👤 By: startup
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: startup
🎯 Time to value: 10 minutes
What it is: A design "skill" for coding agents like Claude Code, Cursor, and Codex that makes AI-generated websites look less generic. It applies 57 quality checks and 20 themes and deliberately refuses the bland defaults most models reach for, fighting "AI slop" visual sameness. Why you'd want it: If your AI-built sites all look the same, this raises the design bar with almost no extra effort.
| ✓ Pros | ✗ Cons |
|---|---|
| Directly targets generic AI design | Opinionated aesthetic may not fit all |
| Works across popular agents | Newer project, still evolving |
| Simple to drop in | Design-only, no logic help |
GitHub - Nutlope/hallmark: Anti-AI-slop design skill for Claude Code, Cursor, and Codex.
Anti-AI-slop design skill for Claude Code, Cursor, and Codex. - Nutlope/hallmark
GitHub
Rank yesterday: #6 - Holding steady ➡
⭐ Stars today: +471 · 📦 Total: 4,763
📜 License: Custom (see repo) · 👤 By: solo dev
🎯 Time to value: 10 minutes
📜 License: Custom (see repo) · 👤 By: solo dev
🎯 Time to value: 10 minutes
What it is: A safety hook that intercepts and blocks dangerous shell and git commands before an AI coding agent can run them. It catches things like
rm -rf, git reset --hard, and database drops in real time and suggests safer alternatives. Why you'd want it: As agents run more commands unattended, one bad command can wipe your work - this is cheap insurance.| ✓ Pros | ✗ Cons |
|---|---|
| Prevents costly accidental data loss | Non-standard custom license |
| Works with multiple AI agents | Can block legitimate commands |
| Fast Rust hook with safe alternatives | Adds a layer to configure |
GitHub - Dicklesworthstone/destructive_command_guard: The Destructive Command Guard (dcg) is for blocking dangerous git and shell commands from being executed by agents.
The Destructive Command Guard (dcg) is for blocking dangerous git and shell commands from being executed by agents. - Dicklesworthstone/destructive_command_guard
GitHub
Rank yesterday: New entry 🆕
⭐ Stars today: +725 · 📦 Total: 5,891
📜 License: Apache-2.0 · 👤 By: solo dev
🎯 Time to value: 5 minutes
📜 License: Apache-2.0 · 👤 By: solo dev
🎯 Time to value: 5 minutes
What it is: A free, structured set of study notes on the maths, computer science, and AI needed to work as a machine-learning research engineer. The author compiled them while getting into Y Combinator, and friends used them to prep for interviews at labs like DeepMind and OpenAI. Why you'd want it: A concise, interview-tested map of what to actually learn for AI/ML roles, saving you from scattered resources.
| ✓ Pros | ✗ Cons |
|---|---|
| Curated, interview-proven material | Notes, not a full course |
| Free and openly licensed | Assumes strong self-discipline |
| Focused path for AI/ML engineering | One author's perspective |
GitHub - HenryNdubuaku/maths-cs-ai-compendium: Become a cracked AI/ML Research Engineer
Become a cracked AI/ML Research Engineer. Contribute to HenryNdubuaku/maths-cs-ai-compendium development by creating an account on GitHub.
GitHub
HuggingFace Trending
Top Models Today
Open-weight flagship Mixture-of-Experts chat and reasoning model from Z.ai (Zhipu).
📥 Downloads (30d): 489,611 · 📜 License: MIT
👤 By: Z.ai / Zhipu AI · 🎯 Task: text-generation
📐 Size: Large Mixture-of-Experts (MoE)
👤 By: Z.ai / Zhipu AI · 🎯 Task: text-generation
📐 Size: Large Mixture-of-Experts (MoE)
What it is: The latest open-weight large language model in the GLM family, using a Mixture-of-Experts design (only part of the network runs per word, saving compute). It targets strong general chat, reasoning, and coding while staying downloadable under a permissive license. Why you'd want it: One of the top open flagship models you can self-host without usage restrictions - a real alternative to closed APIs.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive MIT license | Large hardware needs to run well |
| MoE efficiency at inference | MoE serving is complex |
| Strong general and coding ability | Docs mostly English/Chinese |
zai-org/GLM-5.2 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Vision-language model from Baidu that reads text and structure from images and documents (Optical Character Recognition, or OCR).
📥 Downloads (30d): 1,715,301 · 📜 License: MIT
👤 By: Baidu · 🎯 Task: image-text-to-text
📐 Size: Vision-language model
👤 By: Baidu · 🎯 Task: image-text-to-text
📐 Size: Vision-language model
What it is: A multilingual model that extracts text and layout from images, scans, and documents. Rather than classic character recognition, it reads a page like a language model and returns clean text, handling messy layouts. Why you'd want it: With over 1.7M downloads it is a go-to open engine for document pipelines, invoices, and forms, and the MIT license makes it easy to embed.
| ✓ Pros | ✗ Cons |
|---|---|
| Very high adoption | Custom code to trust and set up |
| Multilingual document understanding | Heavier than lightweight OCR libraries |
| Permissive MIT license | Accuracy varies by language/layout |
baidu/Unlimited-OCR · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
One of the first open-weight releases from Thinking Machines Lab - a multimodal model taking text, image, and audio.
📥 Downloads (30d): brand new · 📜 License: Apache-2.0
👤 By: Thinking Machines Lab · 🎯 Task: image-text-to-text
📐 Size: Multimodal MoE
👤 By: Thinking Machines Lab · 🎯 Task: image-text-to-text
📐 Size: Multimodal MoE
What it is: A freshly released multimodal Mixture-of-Experts model that accepts images and audio alongside text. It is drawing attention despite being brand new because of who released it. Why you'd want it: Early access to a closely watched frontier lab's open multimodal model, good for experimenting with combined vision, audio, and text.
| ✓ Pros | ✗ Cons |
|---|---|
| Handles text, image, and audio | Just released, unproven |
| From a closely watched lab | Near-zero downloads so far |
| Apache-2.0 license | Sparse documentation |
thinkingmachines/Inkling · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Multilingual long-form speech model doing transcription plus speaker labeling in one pass.
📥 Downloads (30d): 65,109 · 📜 License: Apache-2.0
👤 By: OpenMOSS Team (Fudan University) · 🎯 Task: audio-text-to-text
📐 Size: Audio-language model
👤 By: OpenMOSS Team (Fudan University) · 🎯 Task: audio-text-to-text
📐 Size: Audio-language model
What it is: A model that transcribes speech to text while also labeling who spoke when (diarization) and adding timestamps, across multiple languages and long recordings - meetings and interviews in a single pass instead of chaining separate tools. Why you'd want it: It combines transcription, timestamps, and speaker labels in one open model, simplifying meeting and podcast pipelines.
| ✓ Pros | ✗ Cons |
|---|---|
| Transcription + diarization in one | Custom code and heavier setup |
| Long-form multilingual audio | Diarization accuracy varies |
| Apache-2.0 license | Compute-intensive on long files |
OpenMOSS-Team/MOSS-Transcribe-Diarize · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Vision-language Mixture-of-Experts model tuned specifically for tool-use and multi-step agent tasks.
📥 Downloads (30d): 30,539 · 📜 License: Apache-2.0
👤 By: InternScience (Shanghai AI Lab) · 🎯 Task: image-text-to-text
📐 Size: MoE VLM
👤 By: InternScience (Shanghai AI Lab) · 🎯 Task: image-text-to-text
📐 Size: MoE VLM
What it is: A vision-language model built for agentic workflows, where the model uses tools and takes multiple steps to finish a task. It can see images as well as read text, aimed at building autonomous assistants. Why you'd want it: Purpose-built for agents rather than plain chat, so it handles tool loops better out of the box - a good base for open-source agent projects that also need vision.
| ✓ Pros | ✗ Cons |
|---|---|
| Designed for agentic tool-use | Agent models need careful scaffolding |
| Multimodal (vision + text) | MoE serving complexity |
| Apache-2.0 license | Newer, less battle-tested |
InternScience/Agents-A1 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A 27-billion-parameter model squeezed to ~2-bit weights so it runs on modest hardware.
📥 Downloads (30d): 23 · 📜 License: Apache-2.0
👤 By: prism-ml · 🎯 Task: text-generation
📐 Size: 27B, ternary (~2-bit)
👤 By: prism-ml · 🎯 Task: text-generation
📐 Size: 27B, ternary (~2-bit)
What it is: A 27-billion-parameter model whose weights are compressed to ternary (roughly 2-bit) values, so it fits and runs on laptops via llama.cpp. It ships in GGUF format for processors and graphics cards (Graphics Processing Unit, or GPU) across CPU, CUDA, and Metal. Why you'd want it: It chases near-large-model quality at a tiny memory footprint - interesting for anyone experimenting with extreme quantization on edge devices.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs a 27B on modest hardware | Ternary quantization can hurt quality |
| GGUF works across CPU/GPU/Metal | Very low downloads, unproven |
| On-device focused, Apache-2.0 | Best via llama.cpp only |
prism-ml/Ternary-Bonsai-27B-gguf · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Product Hunt
AI Launches Today
Use AI agents on the canvas to design, write, analyze, and organize sites
Featured: July 15 leaderboard · 👤 By: Framer
💰 Pricing: freemium · 🏷 Category: AI website building
💰 Pricing: freemium · 🏷 Category: AI website building
Building a marketing site usually means juggling design, copy, and structure by hand. Framer's new agent layer lets you run AI agents directly on the design canvas to generate and reorganize pages, and even connect external coding agents like Claude Code or Codex. Verdict: A strong sign that design tools are becoming agent hosts, though the agents' quality on complex sites is the real test.
Best of Product Hunt: July 15, 2026 | Product Hunt
Explore the top products launched on Product Hunt on July 15, 2026.
Product Hunt
Speak naturally and it writes in your style in every app
Featured: July 15 leaderboard · 👤 By: Wispr
💰 Pricing: freemium · 🏷 Category: AI dictation
💰 Pricing: freemium · 🏷 Category: AI dictation
Most dictation tools produce robotic, unedited text. Wispr Flow is a Mac app that turns natural speech into polished writing in your own style across any application, with auto-edits, a command mode, and 100-plus languages. Verdict: Genuinely useful if the style-matching and auto-edits hold up across messy real-world apps.
Best of Product Hunt: July 15, 2026 | Product Hunt
Explore the top products launched on Product Hunt on July 15, 2026.
Product HuntGives all your AI agents one shared context
Featured: July 15 leaderboard · 👤 By: N71
💰 Pricing: freemium · 🏷 Category: AI agent infrastructure
💰 Pricing: freemium · 🏷 Category: AI agent infrastructure
People now use many separate AI agents that each forget what the others know, forcing constant re-explaining. N71 gives all of your agents one shared memory so they stay in sync across tools. Verdict: Solves a real and growing problem, but shared-context products live or die on integrations and privacy handling.
Best of Product Hunt: July 15, 2026 | Product Hunt
Explore the top products launched on Product Hunt on July 15, 2026.
Product HuntThe usage runtime for AI products
Featured: July 15 leaderboard · 👤 By: Stigg
💰 Pricing: paid · 🏷 Category: AI billing infrastructure
💰 Pricing: paid · 🏷 Category: AI billing infrastructure
Charging for AI features is messy because costs scale with unpredictable token usage. Stigg 2.0 meters, prices, and enforces limits for AI products so builders can bill by consumption without wiring it all themselves. Verdict: Squarely aimed at a real developer headache; value depends on how cleanly it slots into existing billing stacks.
Best of Product Hunt: July 15, 2026 | Product Hunt
Explore the top products launched on Product Hunt on July 15, 2026.
Product HuntAPI Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Opus 4.8 | $5.00 | $25.00 | 200K |
| Anthropic | Claude Sonnet 5 | $3.00 (intro $2.00 to Aug 31) | $15.00 (intro $10.00 to Aug 31) | 200K (1M beta) |
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 | 200K |
| OpenAI | GPT-5.5 (flagship) | $5.00 | $30.00 | 400K (est.) |
| OpenAI | GPT-5.1 | $1.25 | $10.00 | 400K (est.) |
| Gemini 3.1 Pro (preview) | $2.00 | $12.00 | 1M | |
| Gemini 3.5 Flash | $1.50 | $9.00 | 1M | |
| Groq | GPT-OSS 120B (open) | $0.15 | $0.60 | 128K |
| Groq | Qwen 3.6 27B (open) | $0.60 | $3.00 | 131K |
What this means: The spread from cheapest to priciest is now roughly 50x. Groq's open-weight models cost about $0.15 per million words of input, versus $5.00 for the top closed flagships - so routing routine work to open models and reserving premium models for hard tasks is where the real savings live. Establishing baseline for daily tracking; prices verified against provider pages and recent listings, with OpenAI context windows marked estimated where the page was unreachable.
arXiv Paper of the Day
Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models
Yubo Wang, Jiarong Liang, Yuxuan Zhang, Xuye Liu, Cong Wei, and colleagues · arXiv:2607.12463
What it claims: Standard pretraining does not teach coding models to fold external tool outputs back into their reasoning - exactly what coding agents must do. The authors add a "function-aware fill-in-the-middle" training step that masks whole functions (chosen by analyzing how code depends on itself) so the model learns to reconstruct them the way an agent reconstructs a tool's return value.
Key finding: On a small 2.6-billion-word cleaned GitHub corpus, it lifts SWE-Bench-Verified (a standard real-world coding-fix benchmark) by up to +3 points and SWE-Bench-Lite by up to +5.4 points, without hurting ordinary coding performance.
Why practitioners should care: It is a cheap, self-supervised step that meaningfully improves real coding-agent benchmarks, so teams building or fine-tuning coding agents can get better tool-use behavior without expensive new labeled data.
Key finding: On a small 2.6-billion-word cleaned GitHub corpus, it lifts SWE-Bench-Verified (a standard real-world coding-fix benchmark) by up to +3 points and SWE-Bench-Lite by up to +5.4 points, without hurting ordinary coding performance.
Why practitioners should care: It is a cheap, self-supervised step that meaningfully improves real coding-agent benchmarks, so teams building or fine-tuning coding agents can get better tool-use behavior without expensive new labeled data.
×
Click anywhere or press ESC to close
Member discussion