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 · 19 STORIES

AI news, 13 November 2026

From the community

Mirendil raises $1 billion Series A, citing a one-researcher pretraining result

Mirendil announced a $1 billion Series A on 12 November, led by Kleiner Perkins and Clearlake Capital. Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 participated, among other new and existing investors. The announcement does not name a valuation.

The company is building coding models and an agent platform. It says one researcher, working with agents, pretrained Mirendil-Base from scratch in 3.5 months with less than $10M in compute, and that AI agents wrote over 99% of the code. Those figures come from Mirendil's 21 October report. The less-than-$10M cost is described as all-in for Mirendil-Base, which the report calls the smallest model in the family; larger models were still in training. The report also says the researcher decided what to try and what to trust, and intervened when needed, and that Mirendil-Base did not build itself autonomously. We have not seen an independent evaluation of the cost, the 99% figure, or the model.

Mirendil says it then post-trained Mirendil-Base into Fern, which it says holds up to 10 million tokens of context, and that Lor, now in private beta, lets hundreds of agents write code and run experiments. Those descriptions are also company-reported. Mirendil's 24 June 2026 launch post said its founding team consists of 20 researchers and engineers; the Series A post still describes a small team but does not restate a current headcount.

A Wall Street Journal article dated 13 November 2026, 'This Startup Raised $1 Billion to Build AI Faster and With Fewer Humans' by Sarah Klearman, is paywalled; we cite it by title only and did not use it as a source for these facts.

Last checked: 13 November 2026.

Why it mattersTreat the one-researcher, less-than-$10M and over-99% figures as company-reported until someone tests them independently. The compute figure covers only the smallest model in the family, the June launch post described a founding team of 20 researchers and engineers, and the Series A announcement does not name a valuation. The raise itself is a matter of public record; the training result is not yet something an outside party has checked.

Sourcesmirendil.com ↗ · mirendil.com ↗ · mirendil.com ↗ · techtimes.com ↗

Chrome's free skills teach coding agents modern web code

Google's Chrome team, with Firefox, MDN and Microsoft Edge, released Modern Web Guidance 1.0: a free set of agent skills with more than 200 expert-written guides that steer coding agents such as Antigravity, Cursor, Copilot and Codex towards current HTML, CSS and JavaScript. It installs with one command, includes a scan that flags outdated code in an existing project, and Chrome says it became GitHub's top trending repo within 48 hours.

Why it mattersChrome's own evaluations claim frontier models still miss modern web best practices in many edge cases, which shows up as slow, bloated or less accessible websites. Adding these skills is a cheap way for any team using a coding agent to get cleaner, more accessible front-end code without changing tools.

Sourcesx.com post ↗

Google's Gemini smart glasses open for pre-order

Google has opened pre-orders for its intelligent eyewear, smart glasses built with Samsung and the eyewear brands Gentle Monster and Warby Parker. You say 'Hey Gemini' to ask about what you're looking at, translate in real time, get recommendations or check your calendar, and the glasses have open-ear speakers and a 12 MP camera, according to Google.

Why it mattersThis puts Google's assistant into everyday-looking glasses from mainstream brands, which makes hands-free AI something staff and customers might actually wear. Businesses in retail, travel and field work should watch how translation and 'what am I looking at' help changes customer conversations.

Sourcesx.com post ↗

Where should a second brain live in November 2026?

A comparison of three places a second brain can live: Notion, a markdown wiki often viewed in Obsidian, or a memory layer such as Living Brain, Mem0 or Supermemory. This is a dated comparison of official claims, checked on 13 November 2026, not a ranking. We have not hands-on tested Notion's shared brain or Living Brain.

**Notion: a workspace-native personal memory, plus skills your team can share.** Official Notion Help and developer docs now spell out what the 4 November launch video left out. Memories are private to you, not a team-shared brain. Skills are the shareable layer. Memory via MCP is available on all plans through a user-backed connection, not a workspace-owned one. Connecting MCP does not import another app's chat history. A published setup prompt for ChatGPT and Claude runs once per agent, offers to bring in memories, recent work and skills, and saves an Agent Operating Guide; the agent is told to ask before it creates or changes anything.

This suits people who already keep their work in Notion and switch between ChatGPT, Claude and other MCP-connected agents. Memories live in a Notion database you can review and edit, so leaving Notion means exporting or rebuilding that store — we have not verified how clean a full export is. The workspace still sits on Notion's servers. Official prices, shown in GBP on notion.com/pricing that day: Free £0, Plus £8.50 and Business £16.50 per member per month; Enterprise custom. We have not verified whether MCP memory calls consume Notion AI credits (the same page lists agent credits separately: free to try, then $10 per 1,000 credits).

Our 5 November check, Does Notion really give every agent one shared brain?, is Partly fact: the launch is real, but the demo skipped setup, Muse and Bot were named and not shown, and "automatic" saving still goes through agent tool calls. The docs now fill some of those gaps. We still have not run the Worth testing steps. See also the 5 November news pick and the Notion tool page.

**A markdown wiki you own, often viewed in Obsidian.** Andrej Karpathy's llm-wiki gist (created 4 April 2026) describes the pattern: raw sources stay untouched, a model maintains a folder of interlinked markdown, and you browse it — in his write-up, with Obsidian open beside the agent. Our 6 November pick, The Second Brain Guide: Claude + Obsidian personal wiki, follows that pattern, with a linter so notes stay structured. We also published a Worth testing walk-through for keeping a coding agent's memory in an Obsidian vault, and rated a separate "$100,000 a month" claim as Hype.

This suits people who want plain files they can move, copy or put in git, and who will keep the structure in order. Obsidian's pricing page that day says the app stores notes locally, does not collect telemetry, and — if you use optional Sync — encrypts that data with AES-256 end-to-end encryption. Official prices: the app is free without limits; Sync is $4 per user per month billed annually or $5 billed monthly; Publish is $8 per site per month billed annually or $10 billed monthly; Catalyst is a $25 one-time licence; a commercial licence is $50 per user per year, encouraged for organisational use but not required. See the Obsidian tool page.

**A dedicated memory layer, outside any one app.** Living Brain, Mem0 and Supermemory sell a store that ChatGPT, Claude and other agents can write to and read from, usually over MCP or an API — following the same gap, that one assistant cannot read another's built-in memory. Our 4 November check, Does Living Brain really give every AI one shared memory…?, is Partly fact: the product and pricing are real; the headline benchmark figures are self-reported and mixed together (on LoCoMo the stated saving is 2.8×, not 22.6×). We have not tested Living Brain hands-on. See the 4 November Living Brain pick.

Official prices that day. Living Brain: API usage at $8 per 1 million input tokens and $0.50 per 1 million output tokens, prepaid from a $5 minimum top-up with no monthly minimum, plus 1 GB of storage and 100,000 requests a month. Workspace is listed at $99 per month and described as included with ChatChat Pro+ (three seats; a shared credit pool of $80 for the first three cycles, then $50); Business is custom. Mem0's docs pricing FAQ says Starter, Growth and Pro are no longer sold to new accounts. New accounts get a free plan: for the first 30 days Mem0 pays for model calls (10,000 adds and 1,000 retrievals a month); after that you connect your own LLM key, limits rise to 50,000 adds and 10,000 retrievals a month at no charge from Mem0, and you can turn on usage-based billing above those limits (the rate is shown in the dashboard, not as a public list price). The mem0.ai/pricing page still lists Hobby (free), Starter at $19 a month and Pro at $249 a month. Supermemory: Free $0 with $5 of credits a month; Pro $19 ($20 of credits); a $100 plan with $130 of credits; Scale $399 ($600 of credits); Enterprise custom. We also covered its memory-teammate bot on 4 November. Momentum by Tactico is a local layer for switching coding harnesses; our 8 November pick noted it is still in development, its demo is illustrative, and its "30+" harness figure includes entries under review. It has no public price. Watch it rather than buy it.

This suits people who switch models and do not want memory locked inside Notion or one assistant. Unless you self-host, the store usually sits with the vendor; we have not independently verified how complete an export is. Supermemory documents a free local binary for up to 10,000 documents (not open source) and self-hosting on Scale and Enterprise. Mem0's docs also describe an open-source self-host path. Mem0's FAQ states that with your own key, the text you add is sent to that LLM provider under its terms.

**Who should try which.** If you already live in Notion and want ChatGPT and Claude to share *your* memory, and your team to share skills, start with Notion's setup prompt and treat the 5 November caveats as still in force until you try it. If you want files you control, use a local markdown wiki — often in Obsidian. If you want a store that sits outside any one app, compare Living Brain, Mem0 and Supermemory on official price, hosting and what you are willing to keep with a third party; leave Momentum until there is a public release. This is not a scored ranking, and we have not hands-on tested Notion's shared brain or Living Brain.

Why it mattersAssistants still cannot share memory with each other, so the choice is where that store should live, judged here from official claims rather than a hands-on test.

Sourcesnotion.com ↗ · developers.notion.com ↗ · notion.com ↗ · gist.github.com ↗ · obsidian.md ↗ · livingbrain.com ↗ · docs.mem0.ai ↗ · mem0.ai ↗ · supermemory.ai ↗ · momentumachines.com ↗Partly fact Hype or Fact? checkPartly fact Hype or Fact? checkHype Hype or Fact? checkSteps to try, cost and caveat ↗

Marketing Skills adds a competitor teardown skill, with sign-up and purchase modes

The free, MIT-licensed Marketing Skills library has a new skill, reverse-engineer-marketing, that tries to reconstruct how a company grew. The maintainer, Corey Haines, called it "the biggest thing I've shipped since Marketing Skills itself" in a post on 11 November 2026 (13:30 London). The library now has 53 skills; the current release is v2.14.12, dated 11 November. The skill itself first landed in the repo on 30 October (v2.14.0) and is now at version 1.1.4.

We checked the claims against the skill in the GitHub repo. The public-data side matches the post: it can chart monthly traffic and brand searches (if you have an SEO tool), date launches, price changes and ads, read sitemaps and archives, check AI-answer citations, mine reviews and job posts, compare the target with your own analytics, profile several competitors, and rerun monthly in Watch mode. There are 31 industry packs, from SaaS and ecommerce to solar, insurance and games.

Two modes go further than public data, and only after you say yes to each step, under your real name. With that approval the agent can create an account, log onboarding screens and up to 60 days of emails, map the upgrade and cancel flows, and book a demo. Funnel Hack mode, with a separate purchase budget, buys through the funnel to map order bumps, upsells, abandoned-cart messages and live tests. The skill says it will read the target's terms first and tell you if they bar competitive sign-ups, then leave the decision to you. It also says not to pose as a buyer with a budget you don't have, and to buy to keep rather than purchase-and-refund.

Those modes can still waste real people's time — a sales demo, a support chat, or an abandoned cart that triggers their email sequence — and may breach a site's terms even after you've been warned. Treat them as a last resort, not the default. The useful part is the public-data teardown.

The skill is tool-agnostic for search and traffic, but a full keyword and traffic curve needs a paid SEO source you already have (it names DataForSEO, Semrush, Ahrefs or Similarweb). Without one, you only get page families from the sitemap. Optional paid tools cover social and ad-library pulls (ScrapeCreators or Apify) and a browser the agent can drive. AIsa, a paid Skill Partner of the library, is not named in this skill.

The library itself stays free. It is funded in part by disclosed Skill Partners — currently Converly, Ploy, AIsa, Omnia and Livestorm — and by the maintainer's own products. Install with: npx skills add coreyhaines31/marketingskills.

Why it mattersThe public-data teardown is a genuine, free way to study how a competitor grew. The sign-up, demo-booking and purchase modes are not a victimless shortcut: they use real people's time and may break a site's terms. Switch those modes off unless you have a clear, honest reason to go inside.

Sourcesgithub.com ↗ · github.com ↗ · x.com post ↗ · marketing-skills.com ↗ · github.com ↗

Microsoft 365 outage takes Outlook and Copilot offline into a second day

Microsoft is still recovering from an outage that began on Thursday 12 November and has cut users off from Outlook mailboxes and other Microsoft 365 services. Microsoft first reported mailbox access problems in the Americas (incident EX1550543). Within minutes it widened the warning to users in any region and opened a broader incident, MO1550821, covering the rest of the suite.

Copilot is affected too. Microsoft says users may be unable to load Copilot Chat or get responses, and Copilot may not be able to reach work email, calendar or files. Microsoft Graph requests for Exchange data may time out, so apps and agents built on Graph are hit as well.

Microsoft says servers began failing automated health checks, which put the remaining infrastructure under strain. Servers it brought back "can become unresponsive again once they receive traffic", and a first fix was "not taking as expected". It says it isn't aware of any malicious activity and will publish a full report once the problem is mitigated.

Last checked 13 November, 07:45 London: Microsoft still reports the incident as ongoing.

Why it mattersWhen your email, calendar, files and AI assistant all run on one provider, a single outage takes out all of them at once.

Sourcesx.com post ↗ · x.com post ↗ · cnbc.com ↗ · uctoday.com ↗

Google launches Gemini 4 Hydrogen, a fast, cheaper model for everyone

Google released Gemini 4 Hydrogen on 12 November, a quicker, lower-cost member of the Gemini 4 family built for everyday coding and agent work. It's rolling out in the Gemini app (free users get it when auto-selection picks it), Google Search, AI Studio, the Gemini API and, for paid AI Pro and Ultra subscribers, Google Antigravity, where a new Orchestrator switches between Hydrogen and the bigger Carbon model in one session. It costs $1.50 per million input tokens and $7.50 per million output tokens. Google says it slightly beats Gemini 4 Argon on the Artificial Analysis Intelligence Index and finishes tasks 2.5 times faster than Gemini 3.8 Flash; those are Google's own figures. Flagged by @kimmonismus and @Mr_Salio.

Why it mattersIf you build on Gemini, this is the model to test for routine work like drafting, sorting and simple agent steps, where speed and cost matter more than deep reasoning. Google also says it now links to a wider range of sites in Search answers, so your AI search visibility may shift.

Sourcesx.com post ↗ · blog.google ↗

Claude Meetings puts Claude into your team's video calls

Anthropic is rolling out Claude Meetings in beta on Claude Team and Enterprise plans, starting with Google Meet. Claude joins as a visible participant, takes notes and, when someone asks, builds a doc, whiteboard or other artifact live while the call is going on; it can also use each person's connected tools with their permission. Notes, transcripts and answers are included in the plan, and docs and whiteboards made in meetings are free until 10 February 2027. It isn't available yet to organisations with zero data retention, HIPAA-ready setups or customer-managed encryption keys. Flagged by @kimmonismus.

Why it mattersMeetings usually end with a to-do list someone has to write up; here the first draft of the plan exists before you hang up. Agree a simple rule first: tell everyone on the call that Claude is transcribing.

Sourcesx.com post ↗ · support.claude.com ↗

Thinking Machines releases Inkling 2 Small as open weights

Mira Murati's Thinking Machines Lab published the weights of Inkling 2 Small on Hugging Face under the permissive Apache 2.0 licence on 12 November. It's a text-and-image model with a 1-million-token context, built for agents, coding and tool use; despite the name it's large (594 billion parameters, 17 billion active), so running it yourself needs a serious GPU cluster. It's also available through providers such as Fireworks, Together and Cloudflare. Shared by @miramurati.

Why it mattersOpen weights under Apache 2.0 mean you can fine-tune it and use it commercially without per-seat fees. For most small businesses, the practical route is trying it through one of the hosted providers rather than running it yourself.

Sourcesx.com post ↗ · huggingface.co ↗

ChatGPT's dots now follow your rules, like 'show me a draft first'

OpenAI updated dots, its personal agents in ChatGPT, on 12 November: faster replies, less jargon and better follow-through. More usefully, dots now follow custom instructions, including which actions need your go-ahead, for example "show me a draft before sending any emails". OpenAI's @thsottiaux also said a promised usage reset is on its way to accounts.

Why it mattersIf you let a dot act for you, write down your approval rules on day one. A single line like "never send or pay without asking me" is the cheapest safety net you can add.

Sourcesx.com post ↗ · x.com post ↗

Google Cloud launches a certification for people who design AI agents

Google Cloud's new Professional Agentic Architect certification is open to all technical professionals. It combines a two-hour multiple-choice exam with hands-on labs, costs $200, lasts a year, and Google recommends at least three years of cloud experience, including one building agents. Google says it's what it uses to train its own engineers and partners. From your bookmarks.

Why it mattersIf you're hiring an agency or freelancer to build agents on Google Cloud, this gives you one more signal to ask about. For most small teams it's a nice-to-have, not a requirement.

Sourcesx.com post ↗ · cloud.google.com ↗

Expo Simulators let coding agents test mobile apps in the cloud

Expo made Expo Simulators available on all its paid plans on 12 November: cloud iPhone simulators and Android emulators that coding agents such as Claude Code, Codex and Cursor can drive. The agent installs a build, taps through it, checks its own fix and attaches a recording, screenshots and a live link to the pull request, so a person reviews proof rather than code alone. It costs $0.10 per simulator minute, with 60 minutes a month included, and works for apps not built with Expo too. Shared by @dr_cintas.

Why it mattersIf you pay a developer or agency for a mobile app, ask for recorded proof of each fix. Tools like this make that cheap to provide.

Sourcesx.com post ↗ · expo.dev ↗

Voice-agent company Retell passes $100 million in annual revenue on $4.6 million raised

Retell, which sells voice AI agents for contact centres, says it crossed a $100 million annual revenue run rate, up from $10 million 17 months ago, is profitable, and has raised only $4.6 million. Those are the company's own figures. @kimmonismus contrasted it with far bigger funding rounds for rivals Sierra and Decagon.

Why it mattersAI phone agents have moved from demo to real spending by businesses. If your team answers lots of repetitive calls, it's worth getting a quote and testing one on a single call type.

Sourcesx.com post ↗ · retellai.com ↗

Apify, a scraping platform many AI agents use, reports a security breach

Apify says an attacker used a stolen credential on 9 and 10 November to reach parts of its internal systems and some customer accounts. It detected the access on 11 November, shut it off, rotated API tokens and proxy passwords for affected customers, revoked some integrations including GitHub, and is emailing affected users; a post-mortem will follow. @mikefutia flagged it.

Why it mattersIf you or your automations use Apify, check your email for a notice, reconnect integrations when asked, reset two-factor authentication if told to, and replace any Apify token stored in other tools like n8n or Make.

Sourcesx.com post ↗ · status.apify.com ↗

Keenable 1.0: a web search API built for AI agents

Keenable launched version 1.0 of its search API for agents on 12 November, adding 100 million pre-built summary "knowledge cards" and snippets that pull only the relevant passage from a page. It plugs in through MCP, and a free key covers 100,000 searches a month, then $4 per 1,000. Keenable says it's #1 on the Artificial Analysis search benchmark and on NEEDLE; NEEDLE is Keenable's own benchmark. Shared by @dr_cintas.

Why it mattersIf you're building a research or support agent that needs live web results, the free allowance is enough to test it properly against the search tool you use now.

Sourcesx.com post ↗ · keenable.ai ↗

InstaBox gives each AI agent its own cloud computer

InstaBox opened to everyone on 12 November. One command starts a cloud Linux machine where you can run Claude Code, Codex or OpenCode, close your laptop and let it keep working, or copy the machine to run several agents in parallel; Android, Mac and Windows machines are available too. There's a free plan with one machine; paid plans start at $20 a month plus usage. Shared by @dr_cintas.

Why it mattersLong agent jobs no longer need to tie up your laptop. Useful if you run coding agents for hours or want a throwaway machine to test something safely.

Sourcesx.com post ↗ · instabox.sh ↗

box by Simular: one shared cloud computer for all your AI assistants

Simular launched box on 12 November: a cloud desktop with a browser and terminal that you add to Claude, ChatGPT, Gemini or coding agents as a single connector. Every assistant you connect shares the same computer, logins and files, so one can start a job and another can finish it, and you can watch or take over live. The box itself is free; the AI it runs uses Simular credits. Simular's "10x faster and cheaper" line is its own claim. Shared by @dr_cintas.

Why it mattersIt's a low-effort way to let the chat assistant you already pay for fill in web forms or work in sites with no integration. Keep sensitive logins off it until you've tried it on low-risk tasks.

Sourcesx.com post ↗ · simular.ai ↗

Vercel's UI SDK: interface building blocks made for coding agents

Vercel announced the UI SDK on 12 November, an open-source library of more than 50 interface behaviours, such as popups, selection, typeahead and focus handling, with components for React, Vue, Svelte, Solid and Web Components. It's designed so coding agents build interfaces from tested parts instead of improvising them. Shared by @ctatedev; @steipete pointed out Vercel's pitch that you won't write this code yourself, your agents will.

Why it mattersIf agents are building screens for your product, giving them a standard library like this tends to make the result more consistent and accessible. Ask your developer whether it fits your stack.

Sourcesx.com post ↗ · github.com ↗

Be wary of glowing reviews of a big agent launch, warns @steipete

@steipete says a new agent is launching later this week and marketers are offering a lot of money for positive posts about it on X, so look out for honest reviews.

Why it mattersWhen a launch floods your feed with praise, check whether posts are labelled as paid partnerships and wait for hands-on reviews before you switch tools or sign a contract.

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.