Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
$7
billion price tag for Decart
Amazon is reportedly in talks to buy AI video startup Decart
Top Story
5,674
of 6,123 reviewed findings were real bugs
Claude has flagged 29,000 possible security holes in free so
One Thing to Tell Your Friends
One Thing to Tell Your Friends
Andrej Karpathy, now a researcher at Anthropic, roughly estimates that about 6 billion people have barely touched an AI chatbot, while about 5,000 people inside AI labs are already using systems the rest of us have never seen.
Summary
TL;DR
Top Stories
Trends
Creative AI
Dev Tools
Research
Surprising
Worth Watching
GitHub
Leading repos: morluto/rea (+25,585), storytold/artcraft (+2,956), and cathrynlavery/diagram (+1,083).
HuggingFace
Leading models: google/embeddinggemma (98,829), Cloudflare/clef (14,351), and Qwen/Qwen-Image-2.1 (4,702).
Product Hunt
API Pricing
No price changes vs. October 10 - the fourth quiet day in a row at all four providers.
arXiv
All Verdicts are Not Equal — The most consistent judge agreed with the correct answer only 51% of the time, barely above chance, and in the worst setup just 2% of its verdicts passed the authors' trustworthiness test.
FYI
Hot off the Presses
01
Amazon is reportedly in talks to buy AI video startup Decart for about $7 billion
What this means for you: AI that changes a live video as you watch it is becoming something the biggest tech companies want to own, so expect it to show up in the apps and services you already use.
Decart is an Israeli startup that builds real-time generative video. Its Lucy model edits live video on the fly, and its Oasis model creates simulated worlds used to test robots and self-driving cars. The Wall Street Journal reported the talks on October 11, citing people familiar with them, and Reuters relayed the report.
Nothing is signed. A deal could come soon, but the talks could also collapse or another buyer could step in.
- About $7 billion price tag for Decart - nearly double the roughly $4 billion it was valued at in May, per the WSJ citing PitchBook
- Anthropic walked away first - Bloomberg reported in September that Anthropic finished checking Decart's books, then decided not to buy it for about $6 billion
- Other names were linked to the sale - Nvidia, an existing investor, was reported as a direct bidder, and SpaceX and Nebius were also named earlier in the process
02
An Anthropic co-founder reportedly used political ties to win a huge SpaceX computing deal
What this means for you: Which AI assistant gets the computing power to improve now depends partly on politics, not just engineering.
The Wall Street Journal profiled Tom Brown, an Anthropic co-founder who runs the company's computing deals. Outlets summarizing the piece say Brown, a Republican, is Anthropic's main go-between with the Trump administration and Elon Musk.
According to that coverage, Brown ended a weeks-long standoff after the White House forced Anthropic to shut off two models in June. He reportedly convinced officials that the company had tightened its safeguards. In March, he reportedly met Musk at SpaceX and argued Anthropic was less "woke" than Musk believed.
- About $1.25 billion a month - what Anthropic is expected to pay SpaceX through May 2029, per SpaceX's stock-listing filing
- Either side can exit - on 90 days' notice
- More than 220,000 Nvidia chips in Memphis - all of the Colossus 1 data center, over 300 megawatts, now used by Anthropic
- The Pentagon fight is not over - a February order labeled Anthropic a supply-chain risk, and Anthropic sued over it
03
Claude has flagged 29,000 possible security holes in free software, but only 516 are fixed
What this means for you: AI can now find software bugs far faster than people can fix them, so the apps you use may stay exposed longer even though the problems are known.
Previously: October 8 - Anthropic launched OSS Scanner, a free service that checks open-source software for security holes.
Today: New figures show the bottleneck. Over six months of scanning widely used open-source projects (free software anyone can use and change), Anthropic's models found more than 29,000 possible security holes. Six outside security firms have reviewed only about 6,000 of them.
- 5,674 of 6,123 reviewed findings were real bugs - confirmed by outside firms as of October 2, per Anthropic's public dashboard
- Only 516 have been patched - about 23,000 candidates are still waiting for a human to look at them
- Anthropic now sends unchecked reports straight to maintainers who ask - nearly 5,000 so far
- Maintainers quoted are positive - curl's Daniel Stenberg, a vocal critic of AI "slop," said it found one of the worst curl bugs reported in years
04
A database company ran its AI coding agents like a hospital, and merged a million lines of code
What this means for you: AI agents can now build real software at scale, but the companies doing it well rely on strict rules and human sign-off, not raw speed.
Cockroach Labs, which makes a database used by large companies, built "MOLT Sinai" to develop its data-migration tools. Each bug or feature is treated as a patient: AI "nurses" route the work, AI "fellows" plan and write code, separate AI reviewers hunt for faults, and human "chiefs" make the final calls.
The rules are strict. No code is written before a plan is reviewed, tests cannot be weakened, and stuck agents must write a handoff note and ask for help.
“Over a 5 month period, this agent pipeline successfully processed over a million lines of code with only 7 total reverts”
- Over 1 million lines of code merged in five months - from April 21 to September 11, with only 7 reported rollbacks
- Just over $135,000 in Claude usage - about $84 per issue
- $4,172 in tokens to add IBM Db2 support in under two days - versus about nine months and roughly $160,000 for Oracle support built by people in 2024, though humans have not yet reviewed the Db2 code
- One three-issue loop burned about $631 - so the team added a circuit breaker that escalates after four straight review rounds with new blocking problems
05
McKinsey: 11 million Americans may need to switch to a different kind of job by 2035
What this means for you: Your job may survive AI, but there is a good chance it will need new skills, and degree requirements may be the biggest barrier to moving.
McKinsey, the consulting firm, says the US will have more jobs in 2035 than today, but a very different mix. The research was presented in Fortune by two of its senior partners.
- About 770,000 workers a year will cross into a new occupation - 3.6 times the usual rate, for at least a decade
- Office and admin support faces the largest losses - automation could take on up to 80% of hours worked there
- More than 70% of workers will need to learn new skills - even if they never change employers
- Job ads asking for AI fluency are up elevenfold since 2022
- 85% of growing jobs require a degree or other credential - but only about 38% are required by law, so nearly half is hiring habit
Trends & Themes
Trends & Themes
More AI agents is not the same as better work
Why this matters to you: If a tool promises "hundreds of AI agents" on your task, you may be paying several times more for a faster answer, not a better one.
The pattern across all of these, and Cockroach Labs' hospital setup above: what makes agents useful is an automatic, hard-to-fool check of their work. More agents only help once that check exists.
- Teams of AI agents cost 1.8 to 5.1 times more than a single agent - in tests by evaluation firm Vals AI, only one of four team setups scored meaningfully higher
- Anthropic's own data shows diminishing returns - on one task, going from 10 to 100 Claude agents moved the score from 0.70 to only 0.74
- A developer spent over 500 billion tokens (word pieces) having agents rebuild a shooter game's code - the agents' reviewer kept approving wrong code until he switched to checking each function byte by byte against the original
- Prime Intellect used about 2,209 agents to rewrite its own coding tool in Rust - it worked because a separate verifier agent ran parity checks against the old version before each merge
Checking AI's work has become the real bottleneck
Why this matters to you: AI output now arrives faster than anyone can verify it, so "an AI checked it" is not the same as "it is correct."
The gap is now between what AI produces and what humans can confirm. Anthropic's bug scanner above is the clearest example: thousands of real findings, waiting in line for people.
- Only about 7% of developers trust AI output for important work decisions - 48% trust it when they can easily check the result, in a Stack Overflow survey of 30,903 people
- A new paper says OpenAI's computer-checked math proof checked the wrong thing - its formal version of the Navier-Stokes result (first covered September 8) does not match the written proof, the authors argue
- Workday's technology chief warns agents can make every step correctly and still break company rules - for example skipping tax checks when hiring a contractor abroad
- All 19 companies with agents in production in one small survey require human approval for risky actions - such as moving money or changing data
AI's physical buildings are becoming political and military targets
Why this matters to you: The data centers behind your AI apps now face local votes, wars and hackers, which could make AI services less reliable and more expensive.
Data centers used to be invisible. Now they are fought over in town halls, struck in wars and lobbied for in Washington, as the Tom Brown story above shows.
- A developer offered each of 4,500 residents of a Pennsylvania township a potential $10,000 - to accept a 1,300-acre data center whose end user, emails show, is Amazon, though Amazon says no deal is final
- About 70% of Americans oppose a data center near them - per Gallup, with majorities in both parties
- A Ukrainian drone strike knocked a big Russian data center offline - Semafor identifies it as Yandex's Sasovo site; Iranian strikes hit Gulf data centers earlier this year
- Consulting firm Booz Allen says AI models reached their goal in all eight tests on a simulated factory network - and warns many industrial operators are not ready
AI raises the floor for workers, but can hide whether they are learning
Why this matters to you: AI can make your work look better right away, while quietly skipping the practice that makes you good at it.
Across these studies, the message is consistent: AI helps most people who already know what good work looks like.
“Beware of the illusion of competence.”
- In a Google-run trial with 133 patent lawyers, AI raised draft quality by about 11 percentile points - mostly by cutting weak drafts
- On a final test without AI, only senior lawyers showed a gain - juniors showed none on average in the not-yet-peer-reviewed study, and MIT economist David Autor warns young workers of an "illusion of competence"
- On a task just outside AI's abilities, 84.5% of consultants without AI got it right - versus 70.6% of those using it, in a Boston Consulting Group study of 758 consultants
- Karpathy says the gap in understanding AI is widening - by his rough math, about 20 million professionals see firsthand what agents can now do, out of about 8 billion people
- Writer Nate B. Jones argues taste, agency and follow-through matter more now - because AI lets one person attempt far more
Creative AI & Media
Creative AI & Media
A tiny AI writes Korean by hand, stroke by stroke, in your browser
- hangul.ink draws Korean handwriting as pen strokes - with pressure, so lines thicken and thin like real ink
- The model has only about 350,000 parameters - it runs entirely in your browser, with no server
- One hired writer produced about 4,500 lines of training data - written on a smart pen
- The code and dataset are free - released for Hangul Day on October 9
Voice AI is still waiting for its "ChatGPT moment," executives say
- Billions have gone into voice AI - but no voice product has broken through to everyone the way ChatGPT did
- The bottleneck has moved to thinking speed - PolyAI's technology chief says AI that talks while listening already exists, but answers still need to come faster
- Mishearing key words breaks everything after - Otter's marketing chief says transcription errors spill into wrong follow-up actions
- Callers should always know they are talking to AI - PolyAI's technology chief stressed, while Otter wants everyone in a meeting told when it is being recorded
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
Business & Industry
Business & Industry
Surprising
Surprising & Under-the-Radar
Terence Tao says racing AI to solve famous problems could hurt math
Tao, one of the world's best-known mathematicians, argues that AI is moving math from an era of scarce proofs to one of abundant proofs, though most frontier problems remain unsolved. Solving famous problems was always a stand-in for human understanding, he says, and blindly optimizing for it is now "actively harmful." He urges AI teams to report failures and computing costs, not just wins.
HPE's networking boss says no network problem will need a human within two to three years
Rami Rahim, who runs networking at Hewlett Packard Enterprise (HPE), claims 70-80% of network trouble tickets already need no human. He says hardware failures will be the exception, and even then an intern could swap the part. The figures are his own estimates, and he sells this technology.
Debate: is AI's slow economic payoff a paradox, or a boom in disguise?
Paradox: KKR's Ming Lu cites McKinsey data showing 88% of companies use AI somewhere, yet only 6% get more than 5% of their operating profit from it. Boom: The Decoder argues cheaper AI keeps driving more use overall, which keeps chips scarce and is, in its words, "Jensen's best friend," meaning Nvidia chief Jensen Huang.
Debate: can AI rewrite your app if nobody reads the code?
What happened: Ruby on Rails creator David Heinemeier Hansson (DHH) had AI agents port his Campfire chat app to Rust, Elixir and Go, and did not read the Rust code himself. The pushback: Developer Piotr Sarnacki found the Rust version quietly dropped a security token and delivered only about 14% of messages under steady traffic, because the AI made design choices nobody asked for. His point: you still have to spell out the requirements an agent cannot guess.
Worth Watching
Signals to Track
01
Anthropic may be building its own voice models
Hidden clues suggest Anthropic may be working on voice technology of its own.
A hidden "Voice" item appeared in an internal preview build of Claude's web app around October 10, TestingCatalog reports. Earlier in October, Anthropic began asking users to optionally share voice recordings for training. Claude already has a voice mode, so if this plays out, it could get faster, more natural voice conversations built in-house, though nothing is confirmed.
02
Claude's web app shows hidden "Arborio" and "Iggy" sections
A heart icon hints at Claude Health on the web, and nobody knows what Iggy is.
Two unreleased menu items appeared in Claude's web interface, TestingCatalog reports. A heart-in-hand icon may point to a health dashboard, since Claude's phone apps already analyze fitness data for US Pro and Max subscribers. If this plays out, you could see your sleep and activity trends in Claude on a computer, though nothing is confirmed.
03
OpenAI quietly listed a new life-sciences "discovery" model
A second life-sciences model appeared on OpenAI's price list before any announcement.
"gpt-rosalind-discovery" now sits on OpenAI's price list beside gpt-rosalind-research, at the same $5 per million input tokens and $25 per million output tokens, TestingCatalog reports. Its sibling is open only to approved organizations, and OpenAI has not said who can use the new one. If TestingCatalog's guess that it targets drug discovery is right, drug companies could get a more specialized AI helper.
04
A former White House biosecurity chief wants AI tripwires agreed in advance
"Not yet" is information, not a plan.
Raj Panjabi, who led global health security at the White House in 2022-2023, proposes three warning signs that would trigger specific actions on AI and biology. One is an AI-designed dangerous DNA sequence being ordered from a supplier. If adopted, governments would act on clear signals instead of trusting AI companies to judge their own risk.
GitHub Trending
Top Repos Today
Rank yesterday: #1 - Holding steady ➡
⭐ Stars today: +25,585 · 📦 Total: 102,440
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 15 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 15 minutes
What it is: A tool that lets an AI coding agent take apart apps you do not have the source code for and explain how a feature works. It passed 100,000 stars and added Android app support in a new release today. Why you'd want it: Understand how an app, now including Android apps, does something you admire, or debug a closed-source dependency.
| ✓ Pros | ✗ Cons |
|---|---|
| One tool covering several analysis engines, now including Android | Reverse engineering can break software license terms |
| Runs locally and shows its evidence | Native analysis needs paid or separate tools |
| Free and open source | Three releases in about 36 hours, so pin a version |
GitHub - morluto/rea: Reverse engineer anything with agents, from app behavior down to native binaries.
Reverse engineer anything with agents, from app behavior down to native binaries. - morluto/rea
GitHub
Rank yesterday: #2 - Holding steady ➡
⭐ Stars today: +2,956 · 📦 Total: 17,719
📜 License: Apache-2.0 · 👤 By: startup (Storytold)
🎯 Time to value: 30 minutes
📜 License: Apache-2.0 · 👤 By: startup (Storytold)
🎯 Time to value: 30 minutes
What it is: A desktop app where you stage a scene first (props, poses, camera angle) and then have AI generate the image or video. A day after switching to an open-source license, the team started a new, lighter desktop version written entirely in Rust. Why you'd want it: Repeatable control over composition and camera across shots, which plain text-to-image tools struggle with.
| ✓ Pros | ✗ Cons |
|---|---|
| Now permissively open source | The new native app is unreleased and changing fast |
| Bring your own Midjourney account or API key | Some bundled assets are still excluded from the open license |
| You own what you generate | The easy path still runs on paid credits |
GitHub - storytold/artcraft: ArtCraft is an intentional crafting engine for artists, designers, and filmmakers
ArtCraft is an intentional crafting engine for artists, designers, and filmmakers - storytold/artcraft
GitHubRank yesterday: #3 - Holding steady ➡
⭐ Stars today: +1,083 · 📦 Total: 50,262
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 5 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 5 minutes
What it is: An add-on that makes AI coding agents draw clean, magazine-style diagrams (44 types) as single web files matched to your brand. It passed 50,000 stars today. Why you'd want it: Diagrams for docs and slides from a sentence, without the default look of common diagram tools.
| ✓ Pros | ✗ Cons |
|---|---|
| Plain web files you can open anywhere | Opinionated minimal style |
| Works across six coding agents | Diagrams still need fact-checking |
| Free and open source | No new features since yesterday |
GitHub - cathrynlavery/diagram-design: Editorial diagram design for Claude Code, Codex, GitHub Copilot, Cursor, Factory Droid, and Pi. 44 diagram types. Self-contained HTML + SVG. No shadows. No Mermaid slop.
Editorial diagram design for Claude Code, Codex, GitHub Copilot, Cursor, Factory Droid, and Pi. 44 diagram types. Self-contained HTML + SVG. No shadows. No Mermaid slop. - cathrynlavery/diagram-design
GitHub
Rank yesterday: not ranked - New entry 🆕
⭐ Stars today: +872 · 📦 Total: 51,585
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 10 minutes
What it is: A browser app that puts live public data (flights, ships, satellites, earthquakes, fires) on a realistic 3D globe. You can steer it by voice through an AI agent, asking things like "take me to LAX." Why you'd want it: A ready-made example of a voice-controlled map app over public data, or a sandbox for learning open-source intelligence.
| ✓ Pros | ✗ Cons |
|---|---|
| Most map layers work with no API keys | Voice features need a paid OpenAI key |
| Built-in spending cap on voice sessions | Realistic 3D needs a separate map key |
| Free and modular | The surveillance-style features call for care in use |
GitHub - bilawalsidhu/gods-eye-view: A spy satellite simulator in your browser, except the data is real. Live open source spatial intelligence on a photorealistic 3D globe.
A spy satellite simulator in your browser, except the data is real. Live open source spatial intelligence on a photorealistic 3D globe. - bilawalsidhu/gods-eye-view
GitHub
Rank yesterday: not ranked - New entry 🆕
⭐ Stars today: +314 · 📦 Total: 104,460
📜 License: BSD-style · 👤 By: organization (PyTorch Foundation)
🎯 Time to value: 10 minutes
📜 License: BSD-style · 👤 By: organization (PyTorch Foundation)
🎯 Time to value: 10 minutes
What it is: The software framework most open AI models are trained and run on. It made today's short trending list without a specific news hook. Why you'd want it: The default foundation if you want to train, fine-tune or run models yourself.
| ✓ Pros | ✗ Cons |
|---|---|
| The industry standard | Large install |
| Huge ecosystem of add-ons | Graphics-card driver versions must match |
| Permissive license | Most people use it through simpler libraries |
GitHub - pytorch/pytorch: Tensors and Dynamic neural networks in Python with strong GPU acceleration
Tensors and Dynamic neural networks in Python with strong GPU acceleration - pytorch/pytorch
GitHub
Rank yesterday: not ranked - New entry 🆕
⭐ Stars today: +541 · 📦 Total: 36,439
📜 License: Apache-2.0 · 👤 By: individual developer
🎯 Time to value: 30 minutes
📜 License: Apache-2.0 · 👤 By: individual developer
🎯 Time to value: 30 minutes
What it is: A Python library that runs very large AI models on small graphics cards by loading one layer at a time from disk. Its README reports running a 27-billion-parameter Qwen model in about 3.3GB of graphics memory. Why you'd want it: Try giant open models on hardware you already own, when speed does not matter.
| ✓ Pros | ✗ Cons |
|---|---|
| One-line setup for many model families | Very slow, not for live chat |
| No compression needed by default | Needs lots of disk space |
| Can also fine-tune on small cards | Specific models need pinned software versions |
GitHub - lyogavin/airllm: AirLLM 70B inference with single 4GB GPU
AirLLM 70B inference with single 4GB GPU. Contribute to lyogavin/airllm development by creating an account on GitHub.
GitHub
HuggingFace Trending
Top Models Today
Google's small search model held the top spot for a third day as monthly downloads more than doubled overnight.
📥 Downloads (30d): 98,829 · 📜 License: Apache-2.0
👤 By: Google DeepMind · 🎯 Task: Feature extraction (embeddings)
📐 Size: 0.74B
👤 By: Google DeepMind · 🎯 Task: Feature extraction (embeddings)
📐 Size: 0.74B
What it is: A small model that turns text, code, images, video and audio into number lists (embeddings) that can be searched together. It works in 100+ languages. Why you'd want it: One search index for documents, screenshots, video and audio, small enough to run on a laptop.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | Text quality barely above version 1 outside code |
| One index for every media type | Google's own benchmark numbers |
| Small enough for on-device search | Tool support still thin |
google/embeddinggemma-2 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Cloudflare's "decision model" stayed near the top: typed questions in, probabilities out, no free text.
📥 Downloads (30d): 14,351 · 📜 License: Apache-2.0
👤 By: Cloudflare · 🎯 Task: Image-text-to-text (decisions)
📐 Size: 27.4B
👤 By: Cloudflare · 🎯 Task: Image-text-to-text (decisions)
📐 Size: 27.4B
What it is: A tuned version of Qwen's 27-billion-parameter model that answers fixed questions with a probability for each allowed option in one pass. It also runs on Cloudflare's Workers AI service. Why you'd want it: Fast, structured yes/no and multiple-choice judgments inside an app, without parsing chatbot prose.
| ✓ Pros | ✗ Cons |
|---|---|
| Calibrated, structured answers | Needs custom loading code |
| Permissive license | A tuned model, not a general base |
| Accepts images too | Sensitive to question wording |
Cloudflare/clef · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
The day's biggest climber: Qwen's fast image generator and editor more than doubled its downloads.
📥 Downloads (30d): 4,702 · 📜 License: Qwen Research (non-commercial)
👤 By: Qwen (Alibaba) · 🎯 Task: Text-to-image and editing
📐 Size: 7.1B
👤 By: Qwen (Alibaba) · 🎯 Task: Text-to-image and editing
📐 Size: 7.1B
What it is: The official sped-up version of Qwen-Image-2.1 that makes or edits an image in 8 steps. It plugs into the popular Diffusers library. Why you'd want it: Quick image generation and editing from one model on your own machine.
| ✓ Pros | ✗ Cons |
|---|---|
| Official release | Non-commercial license only |
| Makes and edits images in one model | No published quality comparison with the full model |
| Fast | Needs the newest Diffusers code |
Qwen/Qwen-Image-2.1-Turbo · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Still the go-to free video model for turning an image into a clip on your own computer.
📥 Downloads (30d): 1,726,683 · 📜 License: LTX community license
👤 By: Lightricks · 🎯 Task: Image-to-video
📐 Size: not listed
👤 By: Lightricks · 🎯 Task: Image-to-video
📐 Size: not listed
What it is: An open-weights video model that animates a still image into a short clip. A large community builds add-ons for it. Why you'd want it: Make short AI videos locally without paying per clip.
| ✓ Pros | ✗ Cons |
|---|---|
| Runs on your own hardware | License has revenue limits |
| Big ecosystem of add-ons and workflows | Needs a lot of graphics memory |
| Very widely used | Not an official open-source license |
Lightricks/LTX-2.5 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
A European reasoning model for German and English kept gaining downloads even as its rank slipped.
📥 Downloads (30d): 12,550 · 📜 License: Apache-2.0
👤 By: Aleph Alpha · 🎯 Task: Text generation
📐 Size: 78B (3.5B active)
👤 By: Aleph Alpha · 🎯 Task: Text generation
📐 Size: 78B (3.5B active)
What it is: A Mixture of Experts (MoE) model, a design where only a small part of the model works on each word, from Germany's Aleph Alpha. It supports step-by-step thinking and tool use. Why you'd want it: Strong German and English reasoning under a permissive license, cheap per word once loaded.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | The full 78B must fit in memory |
| Strong in German | Weak on coding-agent tasks |
| Cheap per word | Benchmark scores are self-reported |
Aleph-Alpha/Kolibri-1 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
China Telecom's coding-agent model, trending as a community version that runs on ordinary PCs.
📥 Downloads (30d): 51,938 · 📜 License: Apache-2.0
👤 By: China Telecom AI · 🎯 Task: Text generation (coding agents)
📐 Size: 29B per card, 31B stored (4B active)
👤 By: China Telecom AI · 🎯 Task: Text generation (coding agents)
📐 Size: 29B per card, 31B stored (4B active)
What it is: A Mixture of Experts model built for coding agents, trained on Huawei chips. A community-made compressed version is what is trending. Why you'd want it: Strong coding-agent scores for its size, and it can run locally.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | Self-reported benchmarks |
| Only 4B parts active per word | The trending version is a third-party conversion |
| Runs on home hardware | Little independent testing yet |
XingChen-AGI/Xing4.0-29B-A4B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Product Hunt
AI Launches Today
Your recruiter-informed AI copilot for tech job search
🔥 Upvotes: 244 · 👤 By: HireKey
💰 Pricing: free plan, then $39/month · 🏷 Category: Careers
💰 Pricing: free plan, then $39/month · 🏷 Category: Careers
Built by a co-founder who says she spent 10 years recruiting at Microsoft, Amazon and Stripe. It checks how your experience fits a role, tailors your resume with your approval, and runs spoken mock interviews with an AI voice coach. It led Sunday's board at time of writing. Verdict: The recruiter angle and voice practice stand out, and the free plan makes it easy to test, but $39 a month is steep for job seekers.
HireKey: Your recruiter-informed AI copilot for tech job search | Product Hunt
HireKey is a recruiter-informed AI job-search co-pilot for tech job seekers who want to apply strategically & interview with confidence. It helps you assess your role fit, tailor resumes, communicate experience & track applications. Unlike basic ATS tools, HireKey combines AI with 10 years of recruiting experience at Microsoft, Amazon & Stripe, plus practice out loud with an interactive AI voice coach that responds to your answers & provides real-time interview feedback.
Product Hunt
A personal AI computer that runs entirely on-device
🔥 Upvotes: 137 · 👤 By: Ghost
💰 Pricing: $3,499 one-time · 🏷 Category: Hardware
💰 Pricing: $3,499 one-time · 🏷 Category: Hardware
A desktop box that runs open AI models at home and connects to your apps, files, screen history and smart-home devices so it can act on its own. It has a 24GB graphics card and no subscription. Verdict: Keeping your data at home is the right pitch, but 24GB limits it to mid-size models, so wait for independent reviews.
Ghost Core: A personal AI computer that runs entirely on-device | Product Hunt
Ghost Core is a personal AI computer that runs AI models on-device, connects to your apps, files, and devices, and builds continuous context to proactively help you.
Product Hunt
Record your demo and let your AI agent do the edits
🔥 Upvotes: 132 · 👤 By: OpenScreen community
💰 Pricing: free and open source · 🏷 Category: Video
💰 Pricing: free and open source · 🏷 Category: Video
A screen recorder that adds automatic zooms and smooth cursor moves, lets you edit by deleting words from the transcript, and can hand edits to an AI coding agent. Recording stays on your computer, with no watermark. Verdict: A credible free alternative to paid demo recorders, though it is a community continuation of an archived project.
OpenScreen 2: Record your demo and let your AI agent do the edits | Product Hunt
Record your screen and the demo looks great on the first take: zooms planned from your clicks, smooth cursor, 3D camera moves, cursor themes, backgrounds and webcam layouts. Then let your AI agent do the edit over MCP, or cut the video yourself by deleting words in the transcript. Windows, macOS and Linux. Free, MIT, no account, no watermark.
Product Hunt
Answer every fan comment in your voice
🔥 Upvotes: 108 · 👤 By: MePersona
💰 Pricing: free tier, paid from $29/month · 🏷 Category: Creator tools
💰 Pricing: free tier, paid from $29/month · 🏷 Category: Creator tools
One inbox for comments and messages from YouTube, Instagram, TikTok and X, with an AI that learns your writing voice and drafts replies. Nothing posts until you hit send unless you turn on auto-send. Verdict: Handy for busy creators, but fans think they are talking to you, so consider telling them and keep auto-send off.
MePersona: Answer every fan comment in your voice | Product Hunt
MePersona is the unified social media inbox for content creators, but the inbox is only where it starts. Your Persona learns how you talk and improves with every reply you approve. It drafts replies to comments and DMs from YouTube, Instagram, TikTok, X and more, scores how much each sounds like you, and remembers what each fan said last time. Nothing posts until you hit Send. It works in English and Spanish, so bilingual creators stop choosing which half of their audience gets a real answer.
Product Hunt
Use any model inside Claude Code, keep your Claude plan
🔥 Upvotes: 79 · 👤 By: Edgee
💰 Pricing: free for one developer, $30/month per developer for teams · 🏷 Category: Developer tools
💰 Pricing: free for one developer, $30/month per developer for teams · 🏷 Category: Developer tools
Adds a panel to Claude Code that switches your next request to any of 200+ AI models with one keystroke, then back to Claude. It also shows tokens used and what Edgee's compression saved. Verdict: Useful for trying cheaper models mid-session, but your code passes through a third-party service, so check your plan terms and company policy. Product Hunt Correction: yesterday's edition said ReSO AI led Saturday's board. In the final tally, Onepin finished first.
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Fable 5.1 | $10.00 | $50.00 | 1M tokens |
| Anthropic | Claude Opus 5.5 | $4.00 | $20.00 | 1M tokens |
| Anthropic | Claude Sonnet 5.5 | $2.00 | $10.00 | 1M tokens |
| Anthropic | Claude Haiku 5.5 | $0.10 (up to 100K-token prompts; $0.50 above) | $0.50 ($2.50 above 100K) | 1M tokens |
| OpenAI | GPT-6 Astra | $10.00 (up to 272K-token prompts; $20.00 above) | $50.00 ($75.00 above 272K) | 1.05M tokens |
| OpenAI | GPT-6.1 Sol | $2.00 (up to 272K; $4.00 above) | $10.00 ($15.00 above 272K) | 1.05M tokens |
| OpenAI | GPT-6 Luna | $0.10 (up to 272K; $0.20 above) | $0.50 ($0.75 above 272K) | 1.05M tokens |
| Gemini 4 Argon (limited access) | $2.00 intro, $4.00 later | $10.00 intro, $20.00 later | not disclosed | |
| Gemini 3.8 Flash | $0.75 (to Dec 31; $1.50 from Jan 1) | $3.75 (to Dec 31; $7.50 from Jan 1) | not listed | |
| Groq | GPT-OSS 120B | $0.15 | $0.60 | 131K tokens |
| Groq | Qwen3.8-27B | $0.80 | $4.00 | 131K tokens |
No price changes vs. October 10 - the fourth quiet day in a row at all four providers.
What this means: General chatbot prices are holding steady, so for most people the bigger savings now come from picking a smaller model for simple jobs. Haiku 5.5 and GPT-6 Luna cost about 1% of the priciest models (Fable 5.1 and GPT-6 Astra) for short prompts.
What this means: General chatbot prices are holding steady, so for most people the bigger savings now come from picking a smaller model for simple jobs. Haiku 5.5 and GPT-6 Luna cost about 1% of the priciest models (Fable 5.1 and GPT-6 Astra) for short prompts.
arXiv Paper of the Day
All Verdicts are Not Equal: Rethinking LLM Judge Reliability
Vineet Kumar, Darshita Rathore, Anindya Moitra · arXiv:2610.12083
What it claims: Many teams use one AI model to grade another, a practice called "LLM-as-a-judge" (LLM stands for large language model). Testing six judge models on four benchmarks, the authors find that a judge giving the same answer every time is often not giving the right answer, and simply swapping the order of the two answers can flip its verdict.
Key finding: The most consistent judge agreed with the correct answer only 51% of the time, barely above chance, and in the worst setup just 2% of its verdicts passed the authors' trustworthiness test.
Why practitioners should care: If you use AI graders to pick models or approve changes, test both answer orders, repeat calls, and check the judge against a labelled sample from your own task. The judges tested are a generation old, so re-check with current models.
Key finding: The most consistent judge agreed with the correct answer only 51% of the time, barely above chance, and in the worst setup just 2% of its verdicts passed the authors' trustworthiness test.
Why practitioners should care: If you use AI graders to pick models or approve changes, test both answer orders, repeat calls, and check the judge against a labelled sample from your own task. The judges tested are a generation old, so re-check with current models.
×
Click anywhere or press ESC to close
Member discussion