A clearer view of AI.
Subscribe

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 · 35 STORIES

AI news, 4 November 2026

From the community

Cloudflare's open security-audit skill turns coding agents into bug hunters

Cloudflare's free, MIT-licensed security-audit skill plugs into coding agents like Claude Code or Cursor and runs a six-phase audit: map the code, send parallel "hunter" agents after each attack class, then have separate agents try to disprove every finding before anything lands in the report. Cloudflare says it's the seed of its internal vulnerability-hunting system. It came out on 18 June and was reworked in September, so it's not brand new, but it's now at about 58.8k GitHub stars. Fact or hype: Fact. It's real and well-designed, but it's a method, not a scanner, and results depend on your agent and sandbox.

Why it mattersAI agents are writing more and more code, and this is one of the most credible free ways to point them at your own security. The best idea to copy is that the finder never verifies its own bug. Expect real model costs and run it in an isolated sandbox. We ran its validator tests but haven't run a full audit hands-on. Disclosure: this site uses Grok Bot, which can install and run agent skills like this one.

Sourcesgithub.com ↗ · blog.cloudflare.com ↗ · metronome.fm ↗Fact Hype or Fact? check

Gemini API computer use tool is now generally available

Google's Computer Use tool in the Gemini API, which lets a model look at screenshots and click, type and scroll in a browser, an Android phone or a desktop, is out of preview. The changelog marks it GA from 30 October across both Gemini APIs, with Gemini 3.8 Flash as Google's recommended model. Google says there are no breaking changes from the preview. Fact or hype: Fact. GA is confirmed in the docs; the "production-ready" reliability is Google's own claim.

Why it mattersGA means a supported API that businesses can build customer-facing agents on, not an experiment. You still run the actions yourself, and Google advises sandboxing it and keeping a human approval step for payments, messages and sign-ups. Prompt-injection scanning is off unless you turn it on. We haven't tested it hands-on. Disclosure: this site uses Grok Bot, which competes in the agent space.

Sourcesai.google.dev ↗ · ai.google.dev ↗ · github.com ↗Fact Hype or Fact? check

Cloudflare shows sites which personal AI agents visit, and lets them block each one

Cloudflare has added a "Personal Assistants" category to its security analytics, showing which AI agents acting for people visit your site and what they do there (reading products, logging in, reaching checkout). Its renamed "AI Crawler and Agent access policies" setting now lets you allow or block each supported agent, or block it only on pages with ads. Fact or hype: Fact. It's live in the docs, though it only covers agents Cloudflare can recognise.

Why it mattersAgents shopping and researching for people are becoming real traffic, and blanket bot-blocking can turn away customers. This gives small sites a per-agent choice instead of all-or-nothing. We haven't tested it hands-on. Disclosure: this site uses Grok Bot, which is itself a personal agent that these controls could affect.

Sourcesblog.cloudflare.com ↗ · developers.cloudflare.com ↗ · developers.cloudflare.com ↗Fact Hype or Fact? check

Phone Harness lets AI agents drive cloud iPhones and Androids

Phone Harness is a hosted MCP server: add one URL to Claude, Cursor, ChatGPT, Codex or Grok and your agent can screenshot, tap, swipe, type and open apps on a cloud phone. The client code is MIT-licensed on GitHub (628 stars, one developer, about two weeks old). Android costs 5¢ a minute and an iPhone $80-100 a month, per the company. Fact or hype: Partly fact. The tools are real, but "real iPhones" and reliability aren't independently checked.

Why it mattersPhone-only apps have been a blind spot for agents, and this removes the setup pain. But you're handing an agent a logged-in phone through a third-party cloud: screen text can hijack the agent, and an eSIM phone can receive 2FA codes. We haven't tested it hands-on. No connection to Phone Harness.

Sourcesgithub.com ↗ · github.com ↗ · phone-harness.com ↗Partly fact Hype or Fact? check

Moonshot releases Kimi K3.1 Pro open weights

Moonshot AI has published the full weights of Kimi K3.1 Pro on Hugging Face. It is a 2.8T-parameter mixture-of-experts model (104B active) with a 1M-token context and image input. The API costs $2 in and $10 out per 1M tokens. Moonshot's own benchmarks put it near Claude Opus 5.5 on agentic coding, and Artificial Analysis independently measured 67% on Terminal-Bench 4.0. Fact or hype: Fact. The weights are really out; the benchmark claims are mostly Moonshot's own.

Why it mattersIt is now among the strongest models you can download and self-host. The licence is permissive unless you are very large. We haven't tested it hands-on. No connection to Moonshot AI.

Sourceshuggingface.co ↗ · kimi.ai ↗ · artificialanalysis.ai ↗Fact Hype or Fact? check

Supermemory launches a 'memory teammate' Bot for Grok Bot, and explains Grok Bot's memory limits

Supermemory has released a Supermemory Bot for Grok Bot: you add it once, tag @Supermemory in any chat or group to ask what was decided or tell it what to keep, and your other Bots can ask it too. Its write-up also explains Grok Bot's own memory, and that part checks out against Grok Bot's docs and support answers: each Bot keeps its own memory, only the newest entries are loaded every turn, and older rules can drop out of view unless you ask the Bot to merge them. Fact or hype: partly fact. The product and the explainer are accurate, but 'insanely good memory' is an unmeasured headline from a company selling memory.

Why it mattersEven if you never add the Supermemory Bot, the free tip is worth it: every so often, ask each Bot to merge its saved rules into one short list, and back must-never-do rules with Auto-review. The Bot itself makes sense if several Bots keep needing the same decisions. We haven't tested it hands-on. Disclosure: we use Grok Bot to help put this site together.

Sourcesx.ai ↗ · docs.x.ai ↗ · forum.cursor.com ↗Partly fact Hype or Fact? check

TesterArmy is now e2e and raises $8M for open-source AI app testing

TesterArmy, a Y Combinator startup from Warsaw and San Francisco, has renamed itself e2e after its open-source testing framework of the same name, and raised an $8M seed led by Credo Ventures, KAYA, Point Nine, Baobab Ventures and Script Capital. e2e lets you write tests for web, iOS and Android apps that mix exact checks with plain-language AI steps, run them on your own machine or in CI, and plug them into Claude Code, Cursor or Codex. It's free under Apache-2.0; npm shows about 1.6M downloads in its first month. Fact or hype: fact. The round and the adoption are real; 'verification layer for the agentic coding era' is the company's slogan.

Why it mattersIf a coding agent builds your app, something has to check the app actually works. e2e gives the agent a way to test its own work, at no cost beyond the AI model. It's pre-1.0 and code-first, so it suits developers rather than no-code builders, and TesterArmy's hosted service closes on 15 Jan 2027. We haven't tested it hands-on.

Sourcese2e.com ↗ · github.com ↗ · therecursive.com ↗Fact Hype or Fact? check

Explee: the AI cold-email agent behind a viral 'Series A founder's secret' post

A post seen about 2.7 million times says a newly funded founder's 'secret' is Explee: paste in your website, and it works out who needs your product and emails them, with $30 free. The tool is real: Explee is a seed-funded London company whose AI picks business contacts from its own database, sends personalised cold emails from its own mailboxes, follows up and shares your booking link. But the post looks like marketing. The account is three weeks old, the same text appears word for word on many other accounts, and it carries no ad disclosure in the text. Fact or hype: partly fact. The product exists and the $30 trial is real (card required, then about $0.04 an email); the 'founder's secret' story isn't evidence, and results figures are Explee's own.

Why it mattersCheap, hands-off cold email is useful if you sell to other businesses, and a $30 test is low-risk if you set a daily budget. Two cautions for UK small businesses. Under PECR (regulation 22) you need consent before sending marketing emails to sole traders and partnerships, so don't let any agent email them blind. Every email must also say who it's from and offer an easy opt-out. Judge it by replies from your own test, not by viral posts. We haven't tested it hands-on.

Sourcesexplee.com ↗ · buttondown.com ↗ · ico.org.uk ↗Partly fact Hype or Fact? check

Copilot in PowerPoint: teams can now share brand 'skills' through Brand Kit

Microsoft's PowerPoint lead Derek Johnson is promoting a new option: brand managers can add custom skills (reusable instruction files in the common SKILL.md format) to a Brand Kit, and everyone with access to that kit can choose them in Copilot in PowerPoint. Microsoft's support pages confirm how it works. It sits alongside personal skills saved in OneDrive and skills that IT admins push out from the Microsoft 365 admin center. Fact or hype: fact. The feature is documented; the 'spend less time' benefit is Microsoft's own claim.

Why it mattersIf your team writes the same 'use our tone, our slide order, our disclaimers' prompt every time, you can write it once as a skill and share it. It only helps if you already pay for Copilot in Microsoft 365, and Copilot follows the skill rather than guaranteeing the result, so someone should still check the deck. We haven't tested it hands-on.

Sourcessupport.microsoft.com ↗ · support.microsoft.com ↗ · techcommunity.microsoft.com ↗Fact Hype or Fact? check

Google opens a Gemini Notebook API and MCP server

Google has quietly published an official API and MCP server for Gemini Notebook (formerly NotebookLM) that works with personal Google Accounts. It's an early-access preview and free for now. Today it can create notebooks and add web pages, YouTube videos, text or Markdown as sources; asking questions, Audio Overviews, file uploads and Workspace accounts are on the roadmap.

Why it mattersIt's the first official way for agents like Claude Code or Cursor to feed your own notebooks without paid Enterprise licences or unofficial cookie-based tools. But it's narrow: your agent can fill a notebook, not read answers out of it, and setup needs a Google Cloud project. Limits and 'free' are Google's own preview terms. We haven't tested it hands-on.

Sourcesdevelopers.google.com ↗ · developers.google.com ↗ · developers.google.com ↗Fact Hype or Fact? check

You can now talk to ChatGPT about your money

OpenAI added Voice to ChatGPT's Finances feature. Once you connect your accounts through Finances, you can ask about spending, savings and other money questions out loud.

Why it mattersHandy for a quick spending check-in, but it means giving ChatGPT access to your financial accounts. Read what is shared and stored before connecting a business account.

Sourcesx.com post ↗ · chatgpt.com ↗

Grok Bot adds shopping, price comparison and order tracking

Grok Bot has a new shopping experience: it can look for the best deal, compare items side by side and keep track of your orders. Disclosure: Grok Bot helps run this site's research.

Why it mattersUseful if you buy supplies online and want one place to compare and track orders. Check prices and sellers yourself before buying.

Sourcesx.com post ↗

Mistral shows 133 agents finding and fixing a real security bug

Mistral posted a demo of its open model Large 4 running 133 agents on Django, a popular web framework. The agents mapped the code, hunted for flaws, tested attack scenarios and confirmed a recent known vulnerability, then the model wrote the patch. The video is sped up 90 times. Separately, Large 4 joined Open Math OS, a shared research workspace for mathematicians and AI agents built only on open models.

Why it mattersIt is a concrete picture of agent teams doing security triage on code you can self-host, but it is Mistral's own demo on a bug that was already public. Treat it as a sign of what's possible, not proof it will find new flaws in your app.

Sourcesx.com post ↗ · x.com post ↗

Tencent's Hy4 can now see and use a computer

Tencent released Hy4, its new flagship model. It can read images and charts and, for the first time, operate software on a computer, from command-line tools to desktop apps such as Blender and CAD. Tencent's examples include a walkable 3D model of its campus built as a single web page, a Tesla valuation spreadsheet with a matching research report, and a daily to-do board pulled together from chats, email and calendar. It is live on Tencent Cloud and OpenRouter at $0.834 per million input tokens and $2.501 per million output, with open weights promised soon.

Why it mattersAnother cheap, capable model you can try through OpenRouter for spreadsheet, report and front-end work. The rankings and finance scores are Tencent's own claims, and the weights aren't out yet, so test it on your own tasks before switching.

Sourcesx.com post ↗ · hy.tencent.ai ↗

OpenAI: two engineers used GPT-6 Astra to design a working Codex gadget in under a week

OpenAI says two team members, working on it as a side project, had GPT-6 Astra write the requirements and design the enclosure, the circuit board and the firmware for a small device you talk to Codex through, with no direct human edits to the design. A contract manufacturer built prototypes in under a month. OpenAI admits that getting it ready for manufacturing still took several rounds of human review, because Astra didn't consistently allow for the factory's limits and tolerances. It also says its hardware teams' model use has grown 23-fold since February, names Mazda and Chip Ganassi Racing as collaborators, and shows an unreleased internal model scoring 0.875 on its BenchCAD 2.0 test against 0.594 for Astra (@kimmonismus and others jumped on that chart). Fact or hype: partly fact. It's a real, detailed case study, but every number is OpenAI's own, the stronger model isn't available, and the 'no human edits' part stopped at the prototype.

Why it mattersIf you design physical products, it's worth trying ChatGPT or Codex on drawings, CAD clean-up and design reviews now. Don't expect a factory-ready design: budget for an engineer to check tolerances and manufacturability.

Sourcesx.com post ↗ · x.com post ↗ · openai.com ↗Partly fact Hype or Fact? check

Claude now connects your company's apps for each teammate automatically

On Claude Team and Enterprise plans, admins can switch on a company's connectors and plugins once, and each teammate is connected automatically the first time they use them, instead of signing in to every app one by one. Boris Cherny, who leads Claude Code, announced it. Launch connectors include HubSpot, Notion, Linear, Atlassian, Figma, Supabase, Semrush, Granola, Dropbox and Datadog. Anthropic's help pages note two limits: in the beta option where Claude issues the sign-in, it checks company membership by email rather than through your single sign-on, and the access it hands out stays valid until it expires.

Why it mattersFor a small team this removes the most annoying part of rolling Claude out. Before you switch it on, decide which apps each role really needs, because 'connected automatically' also means 'connected without anyone thinking about it'.

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

Traversal raises $120M to fix outages with AI, and its own test shows why that's hard

Traversal, which builds an 'AI SRE' that investigates why live software broke, raised a $120 million Series B at a $1 billion valuation (Kleiner Perkins, Redpoint and Khosla co-led). Its pitch: AI lets teams write code faster than anyone can review it, so more breaks in production. Its new ORCA-bench benchmark, built with Cornell Tech and Columbia, gives agents 1,079 real-looking incidents: the best general agent found the true cause only 25.3% of the time on medium cases and about 10% when the report was vague. Traversal says its own specialised agents do much better, and its site quotes customers such as PepsiCo and Cloudways with 70–95% accuracy and big time savings. Fact or hype: partly fact. The funding and the benchmark are real and public; the accuracy and time-saving figures for Traversal's own product are the company's and its customers' claims, not independent results.

Why it mattersIf AI is writing more of your code, plan for the clean-up: good logs, tests and a person who can read them. Don't hand incident response to a general chatbot, which this benchmark shows usually guesses wrong.

Sourcesx.com post ↗ · x.com post ↗ · businesswire.com ↗ · arxiv.org ↗Partly fact Hype or Fact? check

Anthropic's FlashProve: proof of which model actually answered you

Anthropic published research on 'provable inference': FlashProve attaches a cryptographic proof to an answer showing it came from a specific, committed model, without revealing the model's weights. The demo runs on Llama 4 Maverick with a 1,024-token window at about 110 tokens a second on one H100, and costs roughly 70 to 95 times more GPU time than normal serving. The code is an open research release, not a maintained product.

Why it mattersNothing to use yet. Longer term, it could let you check that a provider really served the model you're paying for, which matters if you buy AI through resellers.

Sourcesx.com post ↗ · anthropic.com ↗

OpenAI's next security sprint brings in outside researchers and bigger bug bounties

OpenAI says researchers from Cathedral Systems will join its next internal security sprint, using OpenAI's own models to look for weaknesses, and it has raised bug bounties to up to $30,000 for critical issues (up to $172,000 with bonuses). It gave no dates or scope.

Why it mattersA reminder that AI is now used on both sides of security. If you run a website or app, schedule your own update-and-patch check rather than waiting for an incident.

Sourcesx.com post ↗ · openai.com ↗

Google DeepMind trained an AI for psychiatric intake interviews, in simulation only

A DeepMind paper fine-tuned Gemini 3.1 Flash-Lite to run psychiatric intake interviews entirely in simulation, then tested it with 32 psychiatrists playing patients. Blind raters preferred it over the base model in 76.7% of matched cases, appropriate possible diagnoses rose from 33.3% to 68.3%, and missed acute risk in tricky scenarios fell from 30.5% to 12.9%, which is still about one case in eight. The authors say it isn't ready to be used without supervision.

Why it mattersA good model for testing any AI that talks to customers: role-play the hard cases with experts before real people see it, and measure what it misses, not only what it gets right.

Sourcesx.com post ↗ · arxiv.org ↗

Two reports, not announcements: Anthropic's chip team and DeepSeek V5 on Huawei chips

The Information reports that Anthropic told IPO investors its 15-person chip-design team uses Claude to design chips, and that the team finds it faster than its estimates for traditional methods. Separately, TrendForce reports that DeepSeek plans to train its next model, V5, on Huawei's Ascend chips from late 2026. Neither company has confirmed the details.

Why it mattersBoth point the same way: labs want their own chips and less reliance on NVIDIA, which over time should mean more choice and lower prices for AI you buy.

Sourcesx.com post ↗ · x.com post ↗

Rumour, not a release: 'GPT-6.2'

Screenshots and model names spotted in Codex have people expecting a 'GPT-6.2 Astra', and an OpenAI engineer's four-word post ('Le GPT Fat') added fuel. As of tonight OpenAI hasn't announced it, and there are no docs or prices. One viral 'pure SVG' clip used as evidence was later admitted by its poster, @Mr_Salio, to have been 'enhanced' beyond the original.

Why it mattersDon't change plans or budgets on rumours. Wait for the release notes and pricing page.

Sourcesx.com post ↗ · x.com post ↗

Perplexity Computer can now run parts of an online shop

Perplexity says its Computer agent can now work across Shopify, Meta ads and Klaviyo: analysing your store, building landing pages, creating and uploading Meta ads and setting budgets, and drafting Klaviyo emails, with your approval needed before anything is sent. It's on paid plans (Pro from $17 a month billed yearly, Max $167). I haven't tested it.

Why it mattersIf you run a small e-commerce store, try it first on analysis and drafts. Keep approval on for anything that spends ad money or emails customers.

Sourcesx.com post ↗ · perplexity.ai ↗

CrowdStrike adds an 'AI kill switch' for rogue agents

CrowdStrike launched a Falcon feature that lets admins stop unauthorised AI agents, block them from restarting and cut their network access, and it's reported to be buying two AI-security firms, Fiddler AI and Dreadnode, for about $495 million combined (per the Wall Street Journal).

Why it mattersAimed at larger companies, but the idea scales down: know which agents run on your computers and have a way to switch them off.

Sourcesx.com post ↗

Hermes Agent can now import your Grok Bot setup

Nous Research's open-source Hermes Agent (version 0.22.6) added 'hermes migrate grok-bot'. Because Grok Bot has no export button, you paste a message into each Bot, it writes an export file, and Hermes imports that Bot's skills, memories, role and connected apps; scheduled routines arrive as paused jobs so nothing runs twice, and credentials aren't imported. One known issue: importing several Bots into one profile merges them, so use one profile per Bot. Disclosure: Grok Bot is the tool we use to put our digest together. This is a tool for moving off it, so read my take with that in mind.

Why it mattersEven if you're not leaving, it's a useful reminder to keep your prompts, skills and routines somewhere you control, so switching tools later isn't a rebuild.

Sourcesx.com post ↗ · hermes-agent.nousresearch.com ↗

Researchers tested Matt Pocock's /retro skill, and it mostly didn't help

A new paper (Treude and Baltes) tested Matt Pocock's /retro skill, which has Claude Code review its own session and suggest improvements to its set-up, on 30 real coding issues. It made 125 suggestions; 24 were implemented and 12 evaluated. Of seven that could be measured on a later task, two stopped the repeated mistake, one reduced it and four made no difference, and some added extra work. Pocock called it 'the greatest issue report in history'.

Why it mattersThe lesson for any team building skills or agent instructions: check that a change actually helps on a real task before you keep it. Self-improvement loops add clutter as easily as they add value.

Sourcesx.com post ↗ · arxiv.org ↗

.NET libraries can now ship Agent Skills inside their packages

Microsoft's .NET team showed how a NuGet package can bundle SKILL.md files that match the library's version, and a 'dotnet-package-skills install' command copies them into your project's .agents/skills folder so coding agents use the right instructions for that version. Microsoft warns to install skills only from maintainers you trust.

Why it mattersIf your business sells a developer tool or API, shipping a skill alongside it is becoming a simple way to make AI agents use your product correctly.

Sourcesx.com post ↗ · devblogs.microsoft.com ↗

Vercel Security Review: an AI check on every code push

Vercel's Security Review, now in public beta on Pro and Enterprise plans, scans each push to connected GitHub repositories for security issues, and 'Fix with Vercel Agent' opens a pull request with a fix. It's free until 10 November, then charged at model cost plus $0.25 per million tokens. It looks at source code only, so it isn't a full audit. Chris Tate, who works at Vercel, says its labs team has used it for months.

Why it mattersCheap insurance if your site or app is on Vercel, especially if AI writes some of your code. Turn it on during the free period and see what it finds.

Sourcesx.com post ↗ · x.com post ↗ · vercel.com ↗

AI CLI: make images, video and audio from any coding agent

Chris Tate's free ai-cli ('npm i -g ai-cli') adds simple commands such as 'ai image', 'ai video' and 'ai audio speak' that any coding agent can call, with any model you have a key for (for example Claude Code with Flux, or Codex with Gemini). It uses Vercel's AI Gateway or your own provider keys, so you pay for the models you use.

Why it mattersHandy if you already use a coding agent and want it to produce visuals or voice-overs for a page or post without switching tools. Developer-level set-up.

Sourcesx.com post ↗ · ai-cli.dev ↗

Matt Pocock's next experiment: one 'chief of staff' agent per area, attached to a software factory

Matt Pocock says he wants to try a long-running orchestrator conversation for each area of work (using his /chief-of-staff skill), connected to a 'software factory' of sandboxed agents that do the building. The orchestrator would hand out work, run retrospectives and set up scheduled jobs, while the human just talks in plain language. It's a plan, not a result.

Why it mattersA useful shape even for non-developers: one agent that holds the context and delegates, rather than many disconnected chats. Pair it with the /retro finding and test what each change really does.

Sourcesx.com post ↗

Follow-up: newer models are much better at 'writing weights' directly

After yesterday's MIT 'weight-space programming' result, Google researcher Alex Krentsel ran a similar test across model generations on a tiny game-playing network (about 20,000 weights). Older models scored near zero; the newest ones produced networks scoring around 18 to 22 times a trained baseline. It's one researcher's own run on a toy task, and open models struggled.

Why it mattersStill research, not a tool. It's a sign that the newest models can do kinds of engineering that older ones couldn't, so retest tasks a model failed six months ago.

Sourcesx.com post ↗ · x.com post ↗

Greg Isenberg's 'agent-managed company' stack

Greg Isenberg listed how he runs product work with agents: a voice note to Grok Bot, Matt Pocock's /grill-me skill to question the idea, the spec sent into Linear, agents reading competitors' sites and pulling reviews, four design versions of every screen in Claude Code, parallel coding agents in separate git worktrees, test personas clicking through previews, and error and analytics agents that open fixes or explain drop-offs. It's his own stack and opinion, and it includes his own ideabrowser product. Disclosure: Grok Bot is the tool we use to put our digest together.

Why it mattersCopy one habit rather than the whole stack: have an agent read the 10 most common complaints in your competitors' reviews, with quotes, before you build anything.

Sourcesx.com post ↗

Chris Tate: wterm terminals that AI agents can use

Chris Tate (Vercel) says wterm, his web terminal component, now supports WebMCP, so AI agents can list, read and type into terminals on a web page, opt-in per terminal and with a read-only mode. He also posted AI CLI and Vercel Security Review and an agent-browser update that warns when your Chrome is out of date on security fixes.

Why it mattersDeveloper-niche, but the agent-browser warning is a good habit for anyone: keep the browser your agent uses up to date.

Sourcesx.com post ↗ · x.com post ↗

Boris Cherny: how he uses Claude Code (May 2025 talk)

A 27-minute talk by Boris Cherny, who created Claude Code, at Anthropic's 'Code with Claude' event in May 2025; @oaklinelx is a repost account. Unlike our earlier Boris notes (big-picture interviews), this one has real prompts and commands. Advanced level; you need a terminal and some coding. Fact or hype: fact. It's genuinely Boris and genuinely practical, one of the clearest introductions to how Claude Code's creator works. But it's 18 months old, as the post says: slides and screenshots only, some commands and keys have changed, and his prediction that people might stop using code editors by the end of 2025 didn't happen.

Why it mattersStart with the first habit: spend a session only asking questions ('What did I ship last week?') before you let it change anything. It's the quickest way to learn what it understands about your project.

Sourcesx.com post ↗Fact Hype or Fact? checkSteps to try, cost and caveat ↗

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