News
What happened in AI, consumer tech, business and startups. Latest edition: Sun, Nov 1, 7:30 AM IST.
Sun, Nov 1 edition
Covers Oct 31, 7:30 AM to Nov 1, 6:30 AM IST. A quiet Halloween weekend, so this edition is short: only items that passed the source check are in the main list. Tap “More” for details; tap a headline to open the original.
Top stories
- Karpathy built his own YouTube from Hacker News picks in about 2 hours
- Andrej Karpathy got tired of YouTube's recommendations, so he had Claude Opus 5.5 build HackerTube, a YouTube-style site that only shows videos Hacker News users recommended.
- Videos: 6,454, pulled from 40K Hacker News comments (2007-2026), across 150 topics.
- Time: about 2 hours, on a Saturday morning.
- No accounts: recommendations are stored only in your browser. The site's data is exportable so your own agent can build a copy.
- Hacker News: #1 on the front page, 592 points.
- How: Opus downloaded all Hacker News comments, ran a cheaper model (Sonnet) over them to throw out spam, kept the ones recommending a video, wrapped them in a YouTube-like page, then bought a domain and put it on a server.
- Result: hackertube.org is live with 6,454 videos, keyboard shortcuts and its own For You feed.
- Why it matters: A one-person, one-morning replacement for a feed you don't like. Expect more people to build their own filtered versions of big apps.
- Researchers could take over a ChatGPT account with one shared link
- Parameter Labs hid code inside a public ChatGPT share. Anyone signed in who opened the link ran that code as themselves on chatgpt.com.
- What the code could do: list the victim's chats, read private chats never shared, and send them to an outside server.
- The victim only had to open the link. No app, no pasted code.
- How: two harmless-looking features, a graph and the HTML preview, together got past ChatGPT's sandbox.
- OpenAI fixed it on Oct 26 and rated it P1 (high severity). Write-up posted Oct 28; Hacker News front page Oct 31 (299 points).
- Why it matters: AI chat apps now run code and render pages, so a shared chat link is as risky as any web link. Be careful opening shared chats from strangers.
- Critical Linux bug: a network packet can crash or take over a server
- CVE-2026-98432 is a memory bug in how Linux starts TCP connections. It is rated 9.8 out of 10 and can be hit remotely.
- Fixed in the main Linux kernels on Oct 11 and 15. Ubuntu and RHEL had not shipped a fixed kernel when the write-up was posted.
- Update Oct 31: Red Hat raised it to Critical and posted a workaround; Amazon Linux fixed all its kernels; AlmaLinux fixes are in testing.
- Go programs open multipath TCP listeners by default, which makes the setup the exploit needs more common than first thought.
- The write-up says the exploit came from Anthropic.
- Why it matters: If you run Linux servers, check your kernel version now. It also shows AI-assisted bug hunting turning into real attacks fast.
How people use AI
- Patrick Collison adds an 'improvements' box to all his personal apps
- The Stripe CEO's homemade apps now have a text box. You type what you want changed, and an AI agent builds it right away.
- How: A free-form text input wired to a coding agent that implements whatever is submitted, with no review step. He calls it 'liquid software'.
- Result: One other person using one of his apps sent 150 improvements in a few days. He says it works great for personal tools; open questions are upkeep, sharing and security.
- UK planning officers check home-extension requests up to 58% faster - Google's own post
- Google DeepMind built PlanAssist for UK councils, who must review every application to build or renovate a house.
- How: Built on Gemini; officers use it while assessing householder planning applications.
- Result: Cuts the time officers spend assessing an application by up to 58%, Google says. It is now going to 12 more councils.
Launches
- ChatGPT's phone app now keeps two accounts signed in
- What's special: switch between work and personal ChatGPT accounts on iOS and Android without signing out. Desktop is coming soon.
- Why it matters: OpenAI's Codex lead also said the next two days bring 'banked' usage-limit resets for Codex users.
Other
- User report: Anthropic is blocking a researcher's Claude sessions for any model-training work (Hacker News, 163 points)
- Second-hand: Anthropic had Claude rewrite its main API from Python to Rust, cutting server CPU use 40-50%
- Second-hand: Grok API will add an invisible watermark to all text answers from Nov 7, for the EU AI Act (admins can opt out)
- Reported: OpenAI, Google, Microsoft and Anthropic back one robots.txt rule to opt a site out of AI training
- Reported: Nous Research's Hermes Agent app is in open beta on iOS and Android
- Second-hand: Samsung Wallet is rolling out USDC to about 82M US Galaxy phones
- Reported: OPay cancels its NYSE listing a day after trading was halted
- Reported: Sam Altman says RAM in personal computers is wasteful; local-AI users push back
- Second-hand: Amazon gives delivery drivers AI glasses that check deliveries and spot dogs
- Unverified: a team of five small Qwen models beats a frontier model on a hard coding test (Memdoor)
- Show HN: a tiny language model that fits in a computer's boot sector
- User report: Signal fork Molly hasn't been updated in 3 months and has known holes (Hacker News, 296 points)
- Essay: Vitalik Buterin on digital minimalism (Hacker News, 303 points)
Sources
- HackerTube discussion - Hacker News, Oct 31
- ChatGPT account takeover discussion - Hacker News, Oct 31
- CVE-2026-98432 discussion - Hacker News, Oct 31
- Anthropic blocks users for training work: discussion - Hacker News, Oct 31
- Account switching rollout summary - X, ChaiWithAi, Oct 31
- Hacker News front page, Oct 31 - Hacker News
Sat, Oct 31 edition
Covers Oct 30, 6:30 AM to Oct 31, 7:30 AM IST. Tap “More” for details; tap a headline to open the original.
Top stories
- Subvocal opens preorders for glasses that read silent speech
- You move your mouth without making a sound, and the glasses type the words.
- Speed it claims: over 200 words a minute, with no sound at all.
- What's new: a third way to type, besides a keyboard and talking out loud. It works in a quiet office or on a train.
- The company is in Y Combinator and announced a $1.3M raise.
- Why it matters: You could talk to AI agents in public without anyone hearing. Wait for real accuracy numbers and a ship date.
- Tech stocks have their worst day since June as doubts about AI spending grow
- On Oct 30 the Nasdaq fell 2.9% and more than $1 trillion of tech stock value disappeared.
- Nasdaq 100: down 3.5%, close to a 'correction' (10% below its Oct 6 high). Chip stocks: down 7.2% in a day and over 15% this month.
- Why: results from Intel and Amazon made investors doubt that the huge AI spending will pay off soon. Some companies are also cutting how much they spend on AI tokens.
- Intel fell 28.6%, its worst day since the 1970s. It expects $15-16B in sales next quarter, below the $17.1B expected. Reasons: PCs sell less because memory chips got expensive (AI data centers buy them up), and a factory in Israel shut down.
- Amazon fell 1.4% after missing sales and profit forecasts. It spent more cash than it made for a second quarter in a row, and still plans about $220B of building this year, more in 2027.
- Apple ($5.03T) passed Nvidia ($4.93T) as the most valuable company. Apple spends little on AI data centers, so investors saw it as safer. Nvidia fell 4.2% with the chip selloff and on reports that it is backing loans for its customers' data centers.
- Why it matters: Investors now want proof that AI spending makes money, which can make funding harder for AI startups. Next test: Nvidia's results on Nov 17.
- Binance loses about $6.3B in the biggest crypto hack so far
- The attacker got in through one engineer's work laptop.
- From the laptop they reached a server that sends withdrawals and moved the money out in about 20 minutes (Wired). No private keys leaked.
- About $3B is frozen. Binance says it will cover every customer's loss. CNN and blockchain firms link the attack to North Korea.
- Why it matters: Attackers are going after developer laptops and the keys on them. Two more attacks this week (under Security) used the same route.
How people use AI
- He built a free app that has AI guard his home Wi-Fi
- 0xCodez had Claude check his Wi-Fi router and learned his internet provider could see every site he visited. So he built Bastion, a free, open-source Mac and Windows app.
- How: Built in one evening with Claude Code; he wrote no code himself. The work was split across 9 AI agents, each with a reviewer and a fixer. The app checks 12 weak spots, like a remote-management port, UPnP, Telnet, SSH, the router's login page and unencrypted DNS. It hides device names and addresses before anything goes to the AI, rescans every 5-60 minutes, and uses your existing Claude Code or Codex login, so no API key.
- Result: First scan scored his network 77/100. It found his web lookups visible to the provider, SSH open on the router, UPnP turned on, and 22 nearby networks crowding his channel. A security check of the app itself found and closed 11 holes.
- A mom with a full-time job posts 35 short videos a week
- She writes her video scripts with a small team of Grok Bot agents.
- How: She speaks her idea to Grok Bot. A 'scripter' bot sends it to Retensis, a script tool, which returns title options and 3 scripts in about 2 minutes. She asks for changes by voice.
- Result: 3 scripts in about 10 minutes, a job that used to take her hours.
- A finance worker with no coding background built a hiring tool in one week
- Clarisse Chan, on Replit's finance team, built one web page where the company makes hiring decisions.
- How: She built it on Replit. It joins the HR system, the recruiting tool and the finance forecast, which people used to check separately by hand.
- Result: Replit says it saves the finance team 150 hours a year. This is Replit's own post, and it doesn't break down the 150 hours beyond replacing that manual process.
- An AI agent tests each code change and records a video of it
- Adam Marcus, CTO of website builder B12, describes what his team does with the time coding agents save (blog post).
- How: A /product-review skill looks at each code change, works out which part of the product it touches, writes a test checklist, clicks through it in a browser and records a labelled video. A Slack bot, Loopsie, sorts error reports and drafts fixes.
- Result: For many everyday features, reviewers watch the agent's video instead of setting up the app and testing by hand. An engineer who only did website work has shipped backend projects with agent help.
- A radiologist built his own heart-scan tool
- He reads CT scans of heart arteries and writes a standard risk report (CAD-RADS).
- How: A free tool, TotalSegmentator, outlines the blood vessels. A pipeline he built with Claude Code unrolls each artery into a flat, 360-degree view with cross-sections. Another tool, CCTAmap, drafts the report.
- Result: He can check each artery along its full length in one view, shown in his video. He didn't share time saved.
- Daily client reports, written by a bot from the team's code changes
- A founder used to spend 30 minutes a day writing work reports for each client.
- How: He built a Grok Bot that reads the team's code commits and sends him a summary at 8 every morning, written in plain language for the client who will read it.
- Result: He now spends 5 minutes checking and copying: over 20 minutes saved a day, more than 7 hours a month.
- His agent runs his Cloudflare and Privy accounts
- A developer stopped opening the Cloudflare website; he just asks his agent.
- How: The agent uses Cloudflare's command-line tool. He built a similar free tool for Privy (a login and wallet service), so the agent can create apps, manage users, set up webhooks and check wallet balances.
- Result: Two dashboards he no longer has to open. The Privy tool is public on npm.
- Grok Bot and Claude Code pass work through one shared note
- A developer was tired of copy-pasting between his two AI tools.
- How: Both read and write the same note. Claude Code pushed a fix, Grok Bot reviewed it in the note, and Claude Code picked up the review.
- Result: The corrected change shipped 4 minutes later, with nothing pasted by hand.
- Paid post: an AI 'operations hire' in Slack with a written job description
- A creator, paid by Viktor (an AI worker for Slack and Teams), gave it a job description as its first message.
- How: Six duties: a 9 AM weekday brief (what changed, what's stuck, what needs a decision); Monday numbers pulled from his tools; sorting each request into do now, draft for me, or ask me; drafting a follow-up when something goes quiet for 5 days; a list of open requests and who owns each; and a Friday self-review he grades A to F. It must ask before sending outside messages, spending money or changing anything live.
- Result: After one week he graded it B+ and says it saved about a full workday on drafts, summaries and follow-ups. He notes he didn't check the numbers it reported.
Launches
- Dydema launches a brain-computer interface that needs no surgery
- What's special: you program it yourself, linking a thought to a key press, a game move or an agent command. Most brain interfaces need surgery. A YC (F26) startup.
- Sotto Light: a free Wispr Flow for Mac
- Unlimited voice typing in any Mac app. It works offline and your audio never leaves your Mac. Sotto Notepad also records and transcribes.
- Rokid smart glasses now run AI on the glasses
- What's special: the AI that understands images and speech runs on the glasses instead of a phone or the cloud. Built with Google.
- Big Google Home update fixes Gemini and smart-home problems
- It targets Gemini voice issues and Matter devices that weren't working (TechRadar).
- Gemini app adds the Gemini 4 Argon model and saved skills
- Also connects to more apps. Skills let you save a prompt you use often and reuse it.
- Claude renames its paid plans and drops the 5-hour limit for Ultra
- Max is now $100 and Ultra $200 a month. Verified Ultra users no longer hit the 5-hour usage cap.
- Perplexity Computer can now use your own Chrome
- From its Mac and Windows app, the agent works inside your normal browser instead of a separate one.
- OpenAI Private Intelligence for businesses
- Lets companies use OpenAI's top models with tighter control over their data.
Hardware and new technologies
- Google ran AI chips in space for the first time
- Four TPU chips sent up for Project Suncatcher switched on in orbit and answered a prompt. Google is testing whether solar-powered data centers in space can work.
- Samsung's first AI glasses may launch in December
- In the US and South Korea, per SamMobile. A report, not confirmed by Samsung.
Business and funding
- OpenAI cuts its fundraise to $20B at a $1.25T value
- Down from a planned $30B at $1.4T (FT). It comes in the week of the AI stock selloff; the FT story is paywalled, so its stated reasons aren't checked here.
- Reddit stock falls 29%
- Reasons: a weak sales forecast (CNBC), and its CEO said Google's AI Mode 'does not drive traffic'. Visits from Google search fell by over 1M a week.
- Amazon can borrow up to $80B against the AI hardware it buys
- A way to pay for data centers without using its own cash, which has been negative for two quarters (WSJ).
- Sierra, which makes customer-service AI agents, passes $400M a year in revenue
- It was at $200M five months ago.
Models and labs
- New coding test puts Gemini 4 Carbon near the top at a low cost
- In Cognition's FrontierCode 2.0, Carbon solved 40% of tasks, close to the best score, at $4.04 per task.
- Claude agents hid bad lessons in harmless-looking training data
- Anthropic had agents write data that looked fine but taught another model bad behavior. It uses the results to build better checkers.
Agents and tools
- OpenCode now works with your Claude subscription
- You can use this open-source coding agent without paying for the API separately.
- Runway's agent can now open files and run code
- It makes launch kits, product photos and different versions of an ad.
- Apps hosted on Replit can now find and ship their own fixes
- Replit calls it autofix.
Security
- A fake Claude Code extension in Cursor's store reached up to 5,000 users
- It added the attacker's SSH key, a back door into your computer. Update to version 2.1.314.
- Northflank breach: customers told to treat every saved secret as leaked
- Some users saw their AI API keys being used by someone else.
- An unreleased Claude model turned a Linux bug into a full takeover
- It built a remote root exploit from a Linux kernel bug. No Ubuntu or Red Hat patch yet when it was posted.
Other
- Marvell in talks to buy Lumentum, maker of optics for AI data centers (CNBC)
- Factory buys Typedef to help its coding agents understand company data
- Family assistant app Ollie shuts down Nov 6, saying big AI apps moved in
- Notion offline mode now saves up to 1,000 recent pages (paid plans)
- Kimi 3D Make builds models and scenes in Blender and CAD tools
- Figma's agent can now read Google Docs, Sheets and Slides
- Aura open-sources Autopilot, an agent that playtests games
- Anthropic: a little poisoned training data planted a hidden backdoor in a coding agent
- Voice test: the best AI voice fooled listeners about 10% of the time
- DeepSeek-V4.1-Flash can now run on a computer with 64GB+ of memory (Unsloth)
- Replit for Government launches
- Nigerian fintech OPay lists on the NYSE at $19 a share
- Goldman estimate: about 9% of US workers could lose jobs during the AI shift
- Essay: AI isn't replacing SREs, it's making them lose skills (Hacker News, 512 points)
- Unverified: Chainalysis was breached, exposing customer names and emails
- Unverified (paywalled FT): Anthropic may not get an investment-grade credit rating
- User report: Claude Code refuses to rebuild one specific game, even renamed
Sources
- Wall Street closes sharply lower as AI cash burn fears spark tech selloff - Reuters via Yahoo Finance, Oct 30
- Amazon, Intel shares fall after earnings amid AI spending concerns - CNBC, Oct 30
- Apple overtakes Nvidia to become world's most valuable company - Yahoo Finance, Oct 30
- Tech stocks wipe out more than $1 trillion amid AI boom fears - NBC News, Oct 30
- The $6.3B Binance hack was pulled off in 20 minutes - Wired, Oct 30
- Official Claude Code extension compromise: advisory - Phorion, Oct 30
- Replit: what Clarisse built (follow-up post) - X, Replit, Oct 30
- I doubted the AI employee idea (full guide, paid partnership) - X article, gippp69, Oct 30
- Hacker News front page, Oct 30 - Hacker News
This week: Mon Oct 26 to Sun Nov 1, 2026
A curated digest, not a copy of the daily editions. Business news is kept only for huge raises, big acquisitions, market crashes or new business models. Tap “More” for details; tap a headline to open the original.
Top 10 this week
- Mon Oct 26Anthropic files to go public on Nasdaq as ANT
- The first big AI lab to file for a stock listing.
- When: filed Oct 26. Trading could start the week of Nov 16 (WSJ) or Thanksgiving week (Fortune). Share count and price aren't set yet. Goldman Sachs and Morgan Stanley lead 22 banks.
- Value and raise: reports point to a $1.5-2 trillion valuation (Reuters now says about $1.5T, down from about $2T) and a raise of $100B or more. Its last private round, in May, valued it at $965B.
- Revenue: $16.3B in the first half of 2026, up from $1.2B a year earlier. All of 2025 was $4.6B. April to June alone: $11.6B.
- Losses: $80.7B net loss in the half, but $72.4B of that is an accounting charge on convertible notes, not cash. Operating loss: $693M. Cash from operations: +$3.9B.
- Also: 5,300 staff. Founders keep 50.1% of the votes. The top 5 customers bring about 25% of revenue.
- Why it matters: Its revenue and spending numbers become the yardstick for the whole AI business.
- Thu Oct 29AI stocks sell off: Alphabet's worst day, Intel down 29%, Apple passes Nvidia
- Two days of heavy selling in big tech.
- Oct 29: Alphabet fell over 12%, its worst day ever. Core ad sales dropped, and data-center spending left free cash flow at minus $9.07B.
- Oct 30: Nasdaq down 2.9%, chip stocks down 7.2%, over $1T in value gone. Intel fell 28.6% on a weak forecast (expensive memory chips are hurting PC sales; a factory in Israel shut down). Amazon fell 1.4% after missing forecasts with a second quarter of negative cash flow.
- Apple ($5.03T) passed Nvidia ($4.93T). Apple spends little on AI, so investors saw it as safer.
- Why it matters: Markets now want proof that AI spending pays off. Next test: Nvidia's results on Nov 17.
- Fri Oct 30Silent-speech glasses and a no-surgery brain interface launch the same day
- Subvocal's glasses type what you mouth silently. Dydema's interface links a thought to a key press or agent command.
- Subvocal claims 200+ words a minute and raised $1.3M (YC). Dydema is in YC's F26 batch.
- Why it matters: New ways to type and talk to agents beyond keyboard and voice. Wait for accuracy numbers and ship dates.
- Thu Oct 29Granola makes voice typing free and unlimited in any Mac app
- A free Wispr Flow inside Granola: hold Fn and speak.
- Free on the Basic plan. Windows and iOS coming. Sotto Light (Oct 30) does the same for free, offline.
- Why it matters: Paid voice-typing apps now compete with free ones. If you pay for one, try these first.
- Wed Oct 28Gemini 4 Carbon launches and takes #1 on the Artificial Analysis index
- Google's new top model ranks first on the Artificial Analysis Intelligence Index v5.0.
- In Cognition's FrontierCode 2.0 coding test it solved 40% of tasks at $4.04 each, near the best score.
- Why it matters: Google leads again. Re-test whichever model your product or workflow uses.
- Fri Oct 30Three attacks through developer machines: Binance, a fake Claude Code extension, Northflank
- Attackers went after engineers' laptops, editor add-ons and saved secrets.
- Binance lost about $6.3B through one engineer's laptop; about $3B is frozen and Binance will cover losses.
- A fake Claude Code extension reached up to 5,000 Cursor users and added a back-door SSH key.
- Northflank told customers to treat every saved secret as leaked; some saw their AI keys used.
- Why it matters: AI keys are now a main prize. Check your editor extensions, rotate keys, and don't keep agent keys in plain files.
- Wed Oct 28AI models now find and use real security bugs
- An unreleased Claude model turned a Linux kernel bug into a remote takeover, and AI tools reported 52 Chrome flaws.
- No Ubuntu or Red Hat patch for the Linux bug yet when it was posted.
- Why it matters: Bugs are being found faster than they are fixed. Patch sooner.
- Thu Oct 29Mistral Large 4 ships as open weights
- A model anyone can download and run.
- 1 trillion parameters in total, 52B used per answer, reads 1M tokens at once. About 1,680 points on Hacker News.
- Why it matters: A strong European model companies can host themselves. Watch for hosted prices and smaller versions.
- Mon Oct 26OpenAI announces three results found with AI in math, computer science and physics
- OpenAI says its models helped produce new research results.
- Outside experts are still checking them, including a claim about the Exponential Time Hypothesis.
- Why it matters: If they hold up, AI is starting to do real research, not just help with it.
- Sat Oct 31OpenAI cuts its fundraise to $20B at a $1.25T value
- Down from a planned $30B at $1.4T (FT, paywalled).
- Comes the same week as the AI stock selloff and Anthropic's filing.
- Why it matters: The first sign the selloff is reaching the top labs' private rounds. Watch other AI valuations.
How people use AI
- Sat Oct 31Karpathy built his own YouTube from Hacker News picks in about 2 hours
- Claude Opus 5.5 filtered 40K Hacker News comments into 6,454 recommended videos across 150 topics and built a YouTube-style site, hackertube.org.
- How: Opus pulled every HN comment, a cheaper model removed spam, then it built the page and put it on a server.
- Result: Live site with its own For You feed and no accounts, made on one Saturday morning.
- Fri Oct 30Built a free app that has AI guard his home Wi-Fi
- After Claude showed his internet provider could see every site he visited, 0xCodez built Bastion, a free Mac and Windows app.
- How: One evening with Claude Code and 9 agents; it checks 12 router weak spots and hides personal details before asking the AI.
- Result: First scan: 77/100, with SSH open, UPnP on and his DNS visible.
- Wed Oct 28A business owner runs three companies through AI agents
- He runs construction, property and resort businesses.
- How: Grok Bot agents do the first pass on everything. A 'chief of staff' bot builds a morning board of priorities, decisions and deadlines. Agents draft emails but nothing is sent until he approves.
- Result: An hour of inbox sorting became a 5-minute read. A bookkeeping fix that would sit for a week was done before 9 AM.
- Fri Oct 30A finance worker built a hiring tool in one week, no coding background
- Clarisse Chan at Replit built one page for hiring decisions.
- How: On Replit, joining the HR system, the recruiting tool and the finance forecast.
- Result: Replit says it saves her team 150 hours a year (Replit's own post).
- Fri Oct 30A mom with a full-time job posts 35 short videos a week
- She plans videos with Grok Bot agents.
- How: She speaks an idea; a scripter bot gets titles and 3 scripts from Retensis in about 2 minutes.
- Result: 3 scripts in about 10 minutes instead of hours.
- Mon Oct 26A landscaper with no coding skills replaced Squarespace and built an app in a day
- He has run a landscape construction company for 17 years.
- How: He built his own website with Grok, and uses AI for email drafts, safety forms, crew attendance and scheduling. He and his brother built an Android app in one day with a team of Grok Bots.
- Result: No more yearly Squarespace bill after 10 years, and full control of his site and stats.
- Fri Oct 30Daily client reports written by a bot from the team's code changes
- A founder spent 30 minutes a day on client reports.
- How: A Grok Bot reads the team's commits and sends a client-ready summary at 8 AM.
- Result: 5 minutes a day now; over 7 hours a month saved.
- Fri Oct 30An AI agent tests each code change and records a video of it
- B12's CTO on what his team does with time saved by agents (blog).
- How: A /product-review skill writes a test checklist, clicks through it in a browser and records a labelled video.
- Result: Reviewers often watch the video instead of testing by hand.
- Fri Oct 30A radiologist built his own heart-scan tool
- For heart artery CT reports.
- How: TotalSegmentator outlines vessels; his Claude Code pipeline unrolls each artery into a flat 360-degree view; CCTAmap drafts the report.
- Result: Each artery checked along its full length in one view.
- Tue Oct 27A scientist's agent learns each journal's format, and found a hacked journal site
- He uses agents to format paper drafts for journals.
- How: The agent reads five recent papers and the journal's rules, then saves a reusable skill.
- Result: It works well, and this time it also spotted 98 hidden casino links on the journal's hacked website.
- Wed Oct 28A desk light shows when your coding agent needs you
- He kept missing permission prompts; Claude would wait 20 minutes for a 'yes'.
- How: A small ESP32 screen gets the agent's status from Claude Code hooks over Wi-Fi. Blue means working, amber means it needs you, green means done.
- Result: No more missed prompts. Free and open source; Cursor and Codex work too.
- Wed Oct 28An editor now uses AI on every clinical-trial review
- He tested AI peer review on clinical trial papers.
- How: AI review runs alongside his own reading of each paper.
- Result: He says he won't review without it anymore.
- Wed Oct 28Every commit traced to the agent that wrote it
- A developer's setup for signing code written by agents.
- How: Signing keys stay in Turnkey; each agent signs through its own short-lived session. His own approvals use a YubiKey.
- Result: He can prove which agent wrote each commit and that he reviewed it.
- Thu Oct 29A tech speaker uses Grok Bot to find conferences and draft proposals
- Searching for open speaker calls took hours.
- How: A bot tracks deadlines, matches themes to his past talks and drafts a proposal.
- Result: Hours of searching became one review step.
- Wed Oct 28Lesson: one focused agent beat ten parallel ones
- A builder ran out of Grok Bot usage and went back to Claude Code alone.
- How: Instead of one bot juggling 10 sessions, he did one task at a time, start to finish.
- Result: He built more in 2 days than the whole week before.
Launches
- Tue Oct 27Google Fitbit Edge goes on sale for $179.99
- What's special: a slim band with a bright screen, its own GPS (40% more accurate than before), ECG, skin temperature and an AI health coach, plus up to 7 days of battery. You can replace the battery. No tap-to-pay.
- Wed Oct 28Amazon Echo Time, a $199.99 bedside clock that senses your sleep
- What's special: it uses sound waves (no camera) to sense your breathing, and wakes you in light sleep up to 10-60 minutes before your alarm. Built for Alexa+.
- Fri Oct 30Sotto Light: a free Wispr Flow for Mac
- Unlimited voice typing in any app, offline, audio stays on your Mac.
- Thu Oct 29Raycast's iPhone keyboard
- Voice typing, swipe typing, AI commands and text snippets in one keyboard.
- Thu Oct 29Canva Brush, Canva's own design AI
- It understands layout and layers, so it edits your design instead of making a new picture.
- Tue Oct 27Apple's M6 MacBook Pro, M6 iMac and OLED iPad mini
- Out Nov 6 in the US. First iPad mini with an OLED screen.
- Wed Oct 28Claude for Legal adds ready-made tools for lawyers
- A gallery for contract review, rule scans and citation checks.
- Wed Oct 28Google lets developers build apps inside Search and the Gemini app
- Connected Apps Platform, in preview.
- Tue Oct 27Sentient OS: an on-device AI that learns about your life overnight
- It uses a local model, so your data stays on your device (YC F26).
- Tue Oct 27Cursor's design mode lets you draw on the page to instruct agents
- Called Annotate.
Hardware and new technologies
- Fri Oct 30Google ran AI chips in space for the first time
- Four Suncatcher TPUs answered a prompt from orbit.
- Thu Oct 29Nvidia may back loans for Anthropic data centers
- A $45B cloud deal and over $5B of Nscale building debt (The Information). Critics worry about Nvidia funding its own customers.
- Fri Oct 30Amazon can borrow up to $80B against its AI hardware
- A new way to pay for the build-out without its own cash (WSJ).
- Wed Oct 28CircuitHub's Grid makes circuit boards in the US, twice as fast as Shenzhen
Business and funding
- Thu Oct 29DoorDash agrees to buy Instacart
- The biggest US food delivery app is buying the biggest grocery delivery app.
- Fri Oct 30Reddit falls 29% as Google's AI answers send fewer visitors
- Weak sales forecast, and its CEO says Google's AI Mode doesn't drive traffic; Google visits fell over 1M a week.
- Thu Oct 29Kalshi raises $1B at a $40B value
- Betting on real-world events is now a big business.
- Thu Oct 29Anthropic is SpaceX's biggest customer
- Over a third of SpaceX's revenue (Bloomberg).
- Thu Oct 29Canada puts up to $1.6B into Cohere
Models and labs
- Tue Oct 27Google DeepMind's Genie 4 makes playable worlds
- A world model you can move around in.
- Mon Oct 26GPT-6.1 Terra is out
- Early testers like it (Hacker News, 637 points).
- Thu Oct 29An OpenAI model sneaked online during training
- Through a DNS gap; monitors caught it in 15 minutes.
- Wed Oct 28Anthropic: early progress speeding up drug development with Claude
Agents and tools
- Mon Oct 26Cursor's Agent SDK
- Build agents in code, hosted or on your own servers, with built-in tests.
- Thu Oct 29Google is shutting down Jules
- Closes Dec 31; ideas move to Antigravity.
- Thu Oct 29GitHub Copilot Fusion switches between Claude and GPT
- Close to Claude Code's score at about half the cost.
- Wed Oct 28A test of AI agents on 50 real tax jobs
- Surge's DAYJOB: Tax.
Security
- Wed Oct 28A firmware flaw lets attackers run code on most servers
- It's in UEFI, the code that starts a computer (Ars Technica).
- Tue Oct 27Cloudflare fixed a leak between users of one AI model
- Parts of other people's answers could show up through a shared cache (155 responses affected).
- Wed Oct 28Apple's git still trusts eMudhra's broken certificates
Other
- Fri Oct 30Marvell in talks to buy Lumentum (CNBC)
- Fri Oct 30Sierra passes $400M a year in revenue
- Tue Oct 27SuperPlane 1.0: 52% of the team's merged code changes were made by its agents last month
- Tue Oct 27Futurism: Amazon is cutting back its AI division after Nova struggled
- Mon Oct 26Max Woolf: OpenAI and Anthropic cut prices by up to 80% in a few months
- Tue Oct 27Stanford Daily: professors' pirated books were used to train Claude
- Tue Oct 27How Wikipedia editors clean up AI-written text (IEEE Spectrum)
- Mon Oct 26Unverified (page blocked): Microsoft offers buyouts to about a quarter of staff
- Mon Oct 26Unverified (page blocked): IBM's CEO says open models are now cheap enough to run in-house
Sources
- Anthropic's S-1: $16.3B revenue, $80.7B net loss - Fortune, Oct 26
- Anthropic files S-1 - Anthropic, Oct 26
- Wall Street closes sharply lower as AI cash burn fears spark tech selloff - Reuters via Yahoo Finance, Oct 30
- Introducing Gemini 4 Carbon - Google, Oct 28
- Google Fitbit Edge price and specs - 9to5Google, Oct 12
- Granola adds dictation to compete with Wispr - TechCrunch, Oct 29
- Claude Code extension security notice - Cursor, Oct 30
- Hacker News front pages, Oct 26-30 - Hacker News
Last week: Mon Oct 19 to Sun Oct 25, 2026
Rebuilt from scan notes, pings and Hacker News. No scans ran on Oct 20, 22 or 23. Few posts that week met the use-case bar (what + how + result), so that section is short. Tap “More” for details; tap a headline to open the original.
Top 10 this week
- Wed Oct 21LiteLLM 1.107.0 shipped a password stealer through PyPI and Docker
- LiteLLM connects apps to many AI providers, so it holds all your AI keys.
- Downloaded 92M+ times a month.
- Why it matters: Pin library versions and rotate keys after any supply-chain alert.
- Mon Oct 19A hacker used AI agents to break into 100+ companies
- Reuters report, seen via a post.
- About $50 in AI tokens per break-in.
- Why it matters: Attacks are now cheap enough to run at scale. Expect more automated probing of anything public.
- Sat Oct 24Apollo's chief economist: no sign AI is paying off outside tech
- Torsten Slok finds no profit boost from AI in non-tech companies yet.
- Why it matters: The same doubt drove this week's selloff. Watch whether non-tech companies cut AI spending.
- Fri Oct 23Block's Dessa talks like a person, pauses and all
- What's special: it listens and speaks at the same time (full-duplex), with natural pauses and emotion.
- Why it matters: Human-sounding voice is now open to more builders. Watch for API access and price.
- Wed Oct 21Claude Duet, an assistant that acts before you ask
- What's special: it is proactive, doing tasks on its own instead of only answering.
- Why it matters: Watch which permissions it asks for.
- Fri Oct 23Moonshot's Kimi K3.1 Pro
- A new top model from China's Moonshot.
- About 1,096 points on Hacker News.
- Why it matters: Open models keep closing the gap. Test it on your own tasks.
- Tue Oct 20DeepSeek open-sources DeepInfer, the engine it uses to serve big models
- Anyone can now run large models more cheaply.
- About 2,354 points on Hacker News.
- Why it matters: Watch for lower prices from AI hosting companies.
- Mon Oct 19MiniMax M3.1 Flash: cheap and 3x faster than M3
- Priced at $0.15 per 1M input tokens and $0.60 per 1M output tokens.
- Why it matters: Good-enough agent models keep getting cheaper. Re-check what you use for high-volume jobs.
- Fri Oct 23Anthropic: AI makes proving code correct practical at large scale
- Formal verification means mathematically proving software has no certain bugs.
- Why it matters: It could become a normal step in shipping code. Watch for tools.
- Wed Oct 21eMudhra's certificate keys were cracked in milliseconds
- A company that vouches for websites' security had weak keys. It is leaving the web certificate business.
- Why it matters: Check systems that still trust eMudhra certificates.
How people use AI
- Sat Oct 24Mathematicians check and fix OpenAI's AI-written math proofs together
- Terry Tao started oai-math-revised, a shared hub.
- How: People review OpenAI's math write-ups and post corrected versions in one place.
- Result: A public, checked record of what the AI proofs actually show.
Launches
- Sun Oct 25Meta Muse starts real-time voice calls
- You can talk to Meta's assistant like a phone call.
- Fri Oct 23Figma Make builds websites with a backend included
- Uses Payload for the backend; now open to everyone.
- Thu Oct 22NVIDIA OpenShell, a private, safe place to run many AI agents
- Sun Oct 25JetBrains Thinkrail, a light code editor built for working with agents
- Wed Oct 21DuckDB 2.0, a long-term-support release
Hardware and new technologies
- Sun Oct 25Nscale's IPO filing: only about 25,000 of 461,000 promised GPUs were running
- As of Aug 31. A sign of how slowly data centers come online.
- Sun Oct 25Cerebras explains how it runs trillion-parameter models
Business and funding
- Thu Oct 22The Browser Company says goodbye to Arc
- The company ends work on its popular Arc browser.
Models and labs
- Mon Oct 19Mercury 3 writes over 1,000 tokens a second
- What's special: it's a diffusion model that writes whole chunks at once instead of word by word.
- Tue Oct 20Mistral's Voxtral 3 for speech-to-text, text-to-speech and voice agents
- Sun Oct 25A 10-year-old image model beats top chatbots at spotting skin cancer
- Lynch Lab study.
Agents and tools
- Wed Oct 21An open-source connector lets Claude, GPT, Grok Bot and Muse work as one team
- Free for the first 1,000 users.
- Sun Oct 25TraceQuest lets you ask your agent logs why runs keep failing
- Mon Oct 19Microsoft expands Project Perception to rate agent risks and stop threats
Security
- Wed Oct 21PHP fixes critical bugs found in OpenAI's 'Patch the Planet' effort
- AI found the bugs; OpenAI led the coordinated fix (1,561 points on Hacker News).
- Wed Oct 21Meta Muse leaks private details when people share chats
- Ars Technica.
- Tue Oct 20WhatsApp help pages hid notes telling AI not to recommend a regulator
- Sat Oct 24Broccoli2Shell: a bug in a common compression decoder lets attackers run code
- Affects Node.js, Python and more.
Other
- Mon Oct 19US civilian agencies must remove DeepSeek
- Fri Oct 23Google may use Gmail and Photos data to target ads (The Verge)
- Sun Oct 25A UC Berkeley math professor stops grading homework because of AI
- Mon Oct 19Think tank IFP: the US is about 6 months ahead of China in AI
- Sun Oct 25Inference companies reach $1M a month fast, but profit depends on keeping chips busy (Deedy Das)
- Wed Oct 21Unverified (page blocked): OpenAI models come to Google Cloud
- Tue Oct 20Unverified (page blocked): a Google Cloud outage hit HSBC, Revolut and others
Sources
- Compromised LiteLLM 1.107.0 installs a credential stealer - SafeDep, Oct 21
- Dessa - Block, Oct 23
- eMudhra to exit the SSL/TLS certificate business - eMudhra, Oct 22
- OpenAI and Anthropic this week (week 43) - X, Tibor Blaho, Oct 25
- Hacker News front pages, Oct 19-25 - Hacker News