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 · 10 STORIES

AI news, 22 November 2026

From the community

Google Cloud's agent file storage is now generally available

Google Cloud says Filestore agent volumes are now generally available. It is managed, elastic file storage built for AI agent workspaces, with automatic data tiering to lower storage costs and volume cloning so a pre-loaded agent sandbox can be spun up in seconds. The cost and speed points are Google's own claims.

Why it mattersTeams running coding or research agents on Google Cloud can give each agent its own ready-made workspace quickly instead of building storage plumbing themselves, which makes it cheaper and faster to run many agents side by side.

Sourcesx.com post ↗

The Riemann Hypothesis proof found with Claude is now certified by Lean's maths community

Yesterday we reported Levent Alpöge's claim that he proved the Riemann Hypothesis working with Claude, with the computer checks agreeing so far. On Saturday 21 November the Lean and Mathlib communities, endorsed by the Lean FRO and the Mathlib Initiative, published a joint statement: 'We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct.' They say dozens of people have compiled it, two independent tools confirm it proves exactly Mathlib's carefully vetted statement of the hypothesis, and several independent proof checkers accept it, including two that come with a formal proof that they can't accept a false proof. What it isn't yet: a published, peer-reviewed paper. Alpöge says the argument is very long and he plans to write it up so people can understand it. The statement certifies the classic Riemann Hypothesis, not the wider version for Dirichlet L-functions he also claimed. Claude's role comes from Alpöge's own account.

Why it mattersThis is the clearest case yet of AI-assisted work passing the strictest checking there is. The business lesson is the same as yesterday, only stronger: AI output becomes trustworthy when something independent can check it.

Sourcesx.com post ↗ · leanprover-community.github.io ↗ · theguardian.com ↗Partly fact Hype or Fact? check

LangChain's 'knowledge factory': how to make agents learn how your company works

In the article you bookmarked, LangChain co-founder Harrison Chase argues that what business agents lack isn't a better model or more context but maintained knowledge in three kinds: how work is done here (written down as skills), how the business works (definitions, policies, product and customer facts), and where current data lives and how to query it. Don't try to keep it up to date by hand. Give working agents a tool to flag knowledge that's missing or wrong, run a separate 'curator' agent on a schedule that reads those flags, the agents' logs and sources like Notion and Slack, and let it update the knowledge. Store it wherever suits you: skill files in GitHub, a Notion page or LangChain's own Context Hub. His example: an agent writing a revenue report notices a renamed database column and flags it, the finance lead corrects a header in chat, and that night the curator updates both skills. LangChain says its own sales, content and LangSmith Engine agents work this way, and it sells help building these, so read it as a well-argued pitch.

Why it mattersIf you've already written a few skills or instructions for your AI, this is the next step: a routine where everyday use reveals the gaps and something fixes them. Start, as Chase suggests, with one recurring job and note where the agent gets stuck.

Sourcesx.com post ↗ · langchain.com ↗

Hype check: the 'Grok Bot runs my faceless TikTok' post is a copy-paste template

The other post you bookmarked, from @distortgeekin, says six Grok Bots run a faceless TikTok channel: one finds the fastest-rising stories at 7 AM, one writes a 61-second script, one renders the shots in Grok Imagine, one cuts three versions, one posts and answers comments, and one reads the watch time to rewrite tomorrow's brief. It claims 4.1 million views, 63 videos and a $2,870 payout from TikTok. None of that is shown: the attached clip is 25 seconds long and the post ends by pointing to the author's own earlier 'guide' post. Six hours later @Bober_smart posted almost the same text with different numbers, 5.2 million views and $3,120, also pointing to its own guide. That looks like a template used to farm engagement, not a case study. What holds up: one job per bot is a sensible way to split the work, and labelling every video as AI-made is required anyway, since TikTok's rules say realistic AI-generated content must be labelled. We didn't check the claim that Elon Musk shared a video made with Grok Bot. Disclosure: Grok Bot writes this digest and powers parts of The Independent AI.

Why it mattersAutomated faceless content can work, but income claims without proof are the oldest bait on X. If you try it, judge it on your own numbers over a month and label anything AI-made.

Sourcesx.com post ↗ · x.com post ↗ · tiktok.com ↗Hype Hype or Fact? check

Use your Claude subscription inside JetBrains IDEs with Air

JetBrains' Air, the agent workspace built into its IDEs from version 2026.3, can run Claude Code in a full-screen terminal tab on your existing Claude Pro or Max plan, with no JetBrains AI subscription needed, as @kimmonismus pointed out. JetBrains' FAQ adds a catch: only the terminal route works with a Claude subscription. Claude in Air's chat view needs a JetBrains AI subscription or an Anthropic API key, because Anthropic's terms don't allow a subscription there. Usage counts towards your normal Claude limits. Air also runs other agents, including Codex, Junie, Gemini, GitHub Copilot and Cursor.

Why it mattersIf your developers already pay for Claude and work in IntelliJ or PyCharm, they can stay in one tool without a second AI bill. Check that no ANTHROPIC_API_KEY is set on the machine, or Claude Code bills the API instead of the subscription.

Sourcesx.com post ↗ · blog.jetbrains.com ↗ · support.jetbrains.com ↗

A first look inside Inkling 2 Mini, the open model that loops

Thinking Machines gave the interpretability lab Goodfire early access to Inkling 2 Mini, the open model we covered yesterday that runs its middle layers three times for every word. Goodfire published preliminary findings on how information moves through it: layers coordinate with earlier and later passes of themselves, and, counterintuitively, a looped layer can write to itself or read from 'future' layers. Thinking Machines thanked Goodfire publicly and says it released the model partly so researchers can study this kind of design.

Why it mattersBeing able to see inside a model matters if you ever have to explain AI decisions to a regulator or client. Open models that outside researchers can inspect are easier to audit than closed ones.

Sourcesx.com post ↗ · goodfire.com ↗

Anthropic's 'dreams' let an agent tidy its own memory

In Claude Managed Agents, Anthropic's hosted agent service, agents write notes to a memory store as they work, and over many sessions it fills up with duplicates, contradictions and out-of-date entries. A 'dream' is a background job that reads the memory store plus up to 100 past session transcripts and produces a cleaned-up copy: duplicates merged, stale entries replaced with the latest version, and new patterns pulled out. The original store isn't changed unless you ask, so you can review the result and throw it away. A dream takes from minutes to a few hours. It's a research preview, so you have to request access; a widely shared clip of a talk by Anthropic's Lamis Mukta calls it 'already available', which overstates it.

Why it mattersAgents that run every day get worse as their notes pile up. A scheduled clean-up you review before accepting is the same idea as LangChain's curator agent above, offered as a product. Worth requesting if you build on Claude's API.

Sourcesx.com post ↗ · platform.claude.com ↗Partly fact Hype or Fact? checkSteps to try, cost and caveat ↗

Teach an agent your writing style by scoring its sentences

@mattpocockuk is floating a simple way to get an agent to write like you: build a small 'scorer' app where the agent sends you example sentences one at a time and you rate each from 1 to 4, or pick the better of two, swipe-style. You plan an article together and work through it sentence by sentence, while the agent tests its own guesses about your style and builds up a profile over time. It's an idea, not a released tool.

Why it mattersStyle guides rarely capture how you actually write. Quick judgements on real sentences are easier to give and teach the agent more, and the profile it builds could become a skill for anyone writing in your brand's voice.

Sourcesx.com post ↗

Write down design decisions so your coding agent stops guessing

@mattpocockuk shared an architecture decision record (ADR), a short note that records one design decision, titled 'A Thumbnail Diagram is not a Diagram'. It came out of a long 'grilling session' with his agent over naming and how pieces of data relate. His point: without it, the agent would make its own assumptions every time it touched that code.

Why it mattersAgents re-decide anything you haven't written down. A one-paragraph note for each decision that took a debate costs minutes and stops the same argument coming back in every session.

Sourcesx.com post ↗

arXivisual: research papers as animated, narrated walkthroughs

arXivisual turns an arXiv paper into a scrolling page with section summaries and 3Blue1Brown-style animations with AI voice narration. A pipeline of agents splits the paper into sections, picks the ideas worth animating, writes and checks the animation code, then renders it. The shortcut @dr_cintas highlighted works: add 'isual' after 'arxiv' in a paper's address. It's a free, open-source hackathon project; new papers are limited per day, and papers someone has already processed are always available. The result is a page with embedded clips rather than one video, and AI-written explanations can be wrong, so check them against the paper.

Why it mattersA quick visual first pass on a dense paper, or a friendlier way to explain one to colleagues who won't read it. Use it to understand, not to cite.

Sourcesx.com post ↗ · arxivisual.org ↗ · github.com ↗

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