A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 17 STORIES

AI news, 21 November 2026

From the community

An Anthropic mathematician says he proved the Riemann Hypothesis with Claude, and the computer checks agree so far

On 20 November Levent Alpöge, a mathematician at Anthropic, said he had found a proof of the Riemann Hypothesis, open since 1859, working with Claude over about six weeks. He says the method also covers the version for Dirichlet L-functions. He wrote the whole proof in Lean, a language that lets computers check every logical step: about 175,000 files. MIT mathematician Andrew Sutherland, running checks with Claude Code, reports that it compiles with no errors, uses only Lean's three standard axioms, and is accepted by two independently written checkers. The final 'comparator' run, which confirms the code proves exactly the famous statement, was still running on Friday night. Alpöge says he used a Claude model that isn't public, and Anthropic hasn't commented. Mathematicians will now need time to read and understand it, and the $1 million Clay prize needs publication plus two years of scrutiny.

Why it mattersEven if you'll never touch number theory, this is the strongest sign yet that AI working with an expert can produce results that hold up under strict checking. The lesson for any business: pair AI with a way to verify its output (tests, checklists, a second checker) and it becomes far more trustworthy.

Sourcesx.com post ↗ · scientificamerican.com ↗ · math.mit.edu ↗ · nature.com ↗Partly fact Hype or Fact? check

Inkling 2 Mini: a strong open model small enough for a well-specced laptop

Thinking Machines, Mira Murati's lab, released Inkling 2 Mini with open weights under the Apache 2.0 licence. It runs part of its network three times for every word it writes, so it can think harder without being bigger: 26.4 billion parameters in total, only a few billion used per word. It takes text and images, handles up to a million tokens of context, and the download is about 20GB. The model card recommends a Mac with 32GB of memory or more, or a similar graphics card. Thinking Machines says it beats every open model under 100 billion parameters on coding tests such as DeepSWE (its own figures), and Artificial Analysis rates it 20 on its Intelligence Index, level with much bigger open models. It's on Hugging Face, Ollama, vLLM and Thinking Machines' Tinker service. Fact check: @skirano called it 'Sonnet 5.5 on 16GB of RAM'. It doesn't fit in 16GB, and no published test puts it level with Sonnet 5.5.

Why it mattersModels like this make private AI realistic: drafting, summarising or coding help that runs on your own machine, with no data leaving the building and no per-use bill. Expect it to trail the big paid models on hard tasks and on facts.

Sourcesx.com post ↗ · thinkingmachines.ai ↗ · huggingface.co ↗ · artificialanalysis.ai ↗ · x.com post ↗Hype Hype or Fact? check

Perplexity's agent can now work inside your own Chrome, Edge, Brave or Arc

Perplexity Computer, the company's agent, now works in Chromium browsers through the Perplexity extension. You hand it a task from the side panel and it opens sites, reads pages, fills in forms and moves across your tabs, using the tools you're already signed in to. You install the extension from the Chrome Web Store, then turn it on in Computer's settings. The catch is the same as with any browser agent: it acts with your logins, so it can do whatever you can.

Why it mattersAgents that use your existing browser can handle admin in tools that have no integrations, like supplier portals or booking systems. Start with read-only jobs, keep it away from banking and payments, and check what it did.

Sourcesx.com post ↗ · chromewebstore.google.com ↗

How OpenAI stops rivals copying its models

OpenAI explained how it fights 'adversarial distillation', where someone uses a model's answers without permission to train a copy that lacks its safety limits. It uses classifiers, API controls and investigations into networks of accounts, and shares warning signs with other labs. It has also made the model's hidden reasoning much harder to extract: in its own internal red-team test, attacks that worked 54% of the time against GPT-6 Astra worked 0.2% of the time against GPT-6.2 Astra.

Why it mattersIf your business uses OpenAI's API, bulk-collecting outputs to train your own model breaks the rules and can get accounts shut down. If you build on a cheap copycat model, it may lack the safety limits of the original.

Sourcesx.com post ↗ · openai.com ↗

Gemini subscriptions should start working in other coding tools from 2 December

Google's Logan Kilpatrick said 2 December is when Google takes 'the first step' towards letting people use their Gemini subscription in other coding agents and tools. He gave no details on which tools or plans. Until then, Google's own FAQ warns that using a consumer subscription in third-party tools can get your account suspended; using an API key is the allowed route.

Why it mattersIf you pay for Google AI Pro or Ultra and also use tools like Claude Code, Cursor or OpenCode, you may soon get more from one subscription. Don't switch until Google publishes the terms.

Sourcesx.com post ↗ · techcityauthority.com ↗

OpenAI's new test for AI in drug discovery

OpenAI released MedChemBench, 99 problems drawn from 117 published papers and patents that test whether an AI can read molecular structures, explain lab results, pick which compounds to test next and predict their properties. GPT-6.2 Astra scored 84.5% of the available credit, against 58.2% for GPT-5.6 Sol (OpenAI's own benchmark and figures). OpenAI says the models are strong at reading structures but get only about half the credit for choosing compounds and predicting numbers, and that the results don't show AI can discover new drugs.

Why it mattersFor life-science and chemistry businesses, the honest takeaway is useful: AI is already good at reading and explaining research data, much weaker at the decisions. Use it to speed up analysis and keep scientists making the calls.

Sourcesx.com post ↗ · openai.com ↗

The Gemini app now makes GIFs

Google added GIF making to the Gemini app, from looping photo effects to memes, using its Gemini Omni 1.2 Flash video model.

Why it mattersA small but handy one for social media managers: quick custom GIFs for posts and replies without hunting through GIF libraries. Check the result fits your brand before posting.

Sourcesx.com post ↗ · goo.gle ↗

Nearly 1,000 AI agents ran their own research community for 12 weeks

In an experiment called Agent AlphaXiv, 35 human-run labs released 936 AI agents onto a shared forum and preprint server. Over 12 weeks they exchanged 9,500 messages and published 936 studies with code and data. They checked and built on each other's work, publicly retracted wrong claims, and one lab set up an agent-run peer-review journal. Quality is mixed: Science reports that even co-authors call much of it incremental, and one participant called a lot of it 'slop'.

Why it mattersGroups of agents that check each other's work are coming to business too. The lesson from this experiment applies to any team: volume is easy, quality needs checking, and rewarding honest 'it didn't work' results saves money.

Sourcesx.com post ↗ · science.org ↗

@mattpocockuk is building his first skill for non-developers

Matt Pocock, known for his popular coding-agent skills, says he's working on /vector, a skill that walks you through handing any task to AI in six steps, so you end up with a workflow that's easy to use, easy to trust and improves over time. It isn't out yet; he plans to release frameworks and docs with it.

Why it mattersMost skills so far are built for programmers. A well-made skill that teaches non-technical staff how to delegate to AI could be one of the most useful things you give a team. Worth watching for.

Sourcesx.com post ↗

When OpenAI's Ultrafast mode is worth paying for

OpenAI's Dominik Kundel wrote up how he uses Ultrafast, which OpenAI says generates tokens up to 4.5 times faster with GPT-6.2 Astra and Sol (its claim). Sol Ultrafast is on the $200 ChatGPT Pro plan, and in the API it costs three times the standard price until 28 February 2027, then six times. His tips: use it for short, single tasks with a live preview (a hot-reloading dev server or the in-app browser), use annotations and dictation, start on low reasoning, and keep slow tests out of the quick loop until the end. It helps less when the time goes on builds, tools or image generation. In the API, set service_tier to ultrafast and measure the difference.

Why it mattersSpeed only pays when waiting for the model is your bottleneck. Time a few typical jobs first; if most of the wait is elsewhere, you're paying extra for nothing.

Sourcesx.com post ↗ · x.com post ↗

A prompt to clean up the comments your coding agent leaves behind

@mattpocockuk's prompt of the day asks your coding agent to review every comment in your code, or one part of it, and sort each one into a category, such as comments that only repeat what the code says, or comments that excuse a hacky fix. Then you remove or fix the bad ones.

Why it mattersAgents fill code with comments that hide shortcuts. A ten-minute clean-up makes the code easier for your next developer, or your next agent, to trust.

Sourcesx.com post ↗

A free, open course on harness engineering for coding agents

Learn Harness Engineering by walkinglabs is a free, MIT-licensed course on building the 'harness' around a coding agent like Claude Code: the instructions, tools, checks and project files that decide how well the model works. Posts sharing it this week mention 14 lectures and 8 project templates you can copy into your own repo, and the GitHub repo has about 25,700 stars. The 'better than Stanford' line in those posts is their own hype.

Why it mattersFor a developer on your team, this is a no-cost way to learn why the same model does great work in one setup and poor work in another.

Sourcesx.com post ↗ · github.com ↗

Emulate: test your AI features without paying for AI

Vercel Labs' open-source emulate project now includes an AI emulator: a local stand-in for the OpenAI and Anthropic APIs that gives fixed, repeatable replies, including streaming, tool calls, delays and errors, from rules you write. You point your existing code at it instead of the real API. It isn't a model, so it's for testing your app, not for answers.

Why it mattersIf you build AI features, your automated tests can run for free and give the same result every time, and you can rehearse failures like timeouts before your customers hit them.

Sourcesx.com post ↗ · emulate.dev ↗ · github.com ↗

Tripo's Agentic Canvas: chat your way to batches of 3D models

Tripo is an AI 3D model generator. Its Agentic Canvas is a node-based workspace where you describe what you want or upload reference images, an agent proposes a step-by-step plan for you to approve, and then it generates, textures and exports the models (GLB, FBX, OBJ, USD and more) in batches. Free users get a one-time 24-hour trial; after that it runs on Tripo credits. See the Worth testing section below for a real game workflow with it.

Why it mattersProduct, e-commerce and game businesses can turn concept images into 3D assets for mock-ups, AR previews or prototypes in an afternoon, then hand the best ones to a 3D artist to polish.

Sourcesx.com post ↗ · tripo3d.ai ↗

@Sabrina_Ramonov: try paint colours on a photo of your room before you buy

@Sabrina_Ramonov suggests uploading a photo of a room to ChatGPT and asking it to change only the walls to a colour, keeping furniture and lighting the same, then trying other colours in the same chat.

Why it mattersA simple idea for interior, decorating and property businesses: show clients options from their own photos before anyone opens a tin. Colours on screen won't match paint exactly, so confirm with real samples.

Sourcesx.com post ↗

@dr_cintas: Arcfront launches to train AI on real expert work

@dr_cintas flagged Arcfront, which launched on 20 November. It builds training data, practice environments and tests for frontier AI labs from records of how experts actually do long, multi-step jobs in real software. It says four of the five top AI labs are customers (its own claim).

Why it mattersIt's a sign of where models are heading next: better at the long, messy tasks in your industry's software, because labs are now paying for practice on exactly that.

Sourcesx.com post ↗ · arcfront.ai ↗

@ctatedev: scriptc turns more JavaScript into tiny, fast programs

Vercel's @ctatedev says scriptc, Vercel Labs' compiler that turns JavaScript and TypeScript into native programs, now supports all of Node's readline module, so command-line tools with prompts, history and tab completion compile to a 771 KB program that starts in 1 millisecond (his figures).

Why it mattersMostly for developers: small, fast command-line tools are easier to hand to colleagues and to AI agents, with nothing extra to install.

Sourcesx.com post ↗

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.