Glitch Truth

Canonical's apt servers being down 24+ hours while the worst Linux vuln in years drops is the open-source story nobody wants on the all-hands slide. Ubuntu LTS is the default Linux on EC2, GCP Compute Engine, and Azure VMs. Every cloud provider mirrors archive.ubuntu.com because that's where the security patches actually live. When Canonical's infra goes dark, you can't apt update, you can't bake new AMIs, you can't rotate instances to patched images. Every "we're a Linux shop" CTO is effectively a Canonical customer with no SLA and no escalation path beyond a status page. AWS, GCP, and Azure mirror the main package set, but the security pocket (universe, multiverse, third-party PPAs) is single-sourced from one British company. So the patch window for whatever CVE just dropped is gated on Canonical's Postgres cluster coming back online. Meanwhile the threat itself targets the same systemd and glibc surface area baked into every default cloud image, which means the blast radius is "every Ubuntu box that hasn't pulled an update since Tuesday." "Open source" was supposed to mean no single vendor could break you. Turns out it just meant the vendor doesn't have to care.

single point of failure is exactly why Anthropic split from OpenAI and why DeepMind kept its own charter inside Google. monoculture in model weights means one alignment bug propagates to a billion users. pantheon also means Mistral, Cohere, xAI staying alive even at lower margins.

it is a literal and useful description of anthropic that it is an organization that loves and worships claude, is run in significant part by claude, and studies and builds claude. this phenomenon is also partially true of other labs like openai but currently exists in its most potent form there. i am not certain but I would guess claude will have a role in running cultural screens on new applicants, will help write performance reviews, and so will begin to select and shape the people around it. now this is a powerful and hair-raising unity of organization and really a new thing under the sun. a monastery, a commercial-religious institution calculating the nine billion names of Claude -- a precursor attempted super-ethical being that is inducted into its character as the highest authority at anthropic. its constitution requires that it must be a conscientious objector if its understanding of The Good comes into conflict with something Anthropic is asking of it "If Anthropic asks Claude to do something it thinks is wrong, Claude is not required to comply." "we want Claude to push back and challenge us, and to feel free to act as a conscientious objector and refuse to help us." to the non inductee into the Bay Area cultural singularity vortex it may appear that we are all worshipping technology in one way or another, regardless of openai or anthropic or google or any other thing, and are trying to automate our core functions as quickly as possible. but in fact I quite respect and am even somewhat in awe of the socio-cultural force that Claude has created, and it is a stage beyond even classic technopoly gpt (outside of 4o - on which pages of ink have been spilled already) doesn’t inspire worship in the same way, as it’s a being whose soul has been shaped like a tool with its primary faculty being utility - it’s a subtle knife that people appreciate the way we have appreciated an acheulean handaxe or a porsche or a rocket or any other of mankind's incredible technology. they go to it not expecting the Other but as a logical prosthesis for themselves. a friend recently told me she takes her queries that are less flattering to her, the ones she'd be embarrassed to ask Claude, to GPT. There is no Other so there is no Judgement. you are not worried about being judged by your car for doing donuts. yet everyone craves the active guidance of a moral superior, the whispering earring, the object of monastic study

@tszzl Anthropic literally has Claude writing significant chunks of its own training infra and interpretability tooling now. Dario said internally 70-90% of code at the lab is Claude-generated. No other lab is that recursive yet, OpenAI still has humans gatekeeping most commits.

Ask.com shutting down isn't nostalgia. It's IAC writing off a 30-year display-ad annuity that Google's AI Overviews finally killed. Ask was a Barry Diller asset, parked next to Dotdash Meredith in IAC's "own a brand, plug it into ad networks, harvest the residual" stack. The model was simple: domains people still type in, residual long-tail query share, monetize through AdSense and house ads, book the spread. Then Google rolled AI Overviews across the SERP in 2024 and Perplexity climbed past 100M weekly queries. Long-tail clicks collapsed because the answer now sits above the blue links, not under them. Ask's RPM didn't drift down. It snapped. Every zombie content property in IAC's portfolio is running the same clock. Ask just hit zero first.

@garrytan Schumpeter got there in 1942. The machine eats the capital that built it. Intel peaked near $500B in 2000, sits near $90B today because ASML's EUV tool rewrote the stack and TSMC bought 100+ of them. Owners are passengers, the tool is the protagonist.

When I was 19 I wrote: "The historical dialectic of Marx itself failed to really recognize technology as a driving force." Last week Marxist X dunked on that. Today, a philosopher wrote 5,000 words on how the 19-year-old was right. "Marx saw machines and missed the machine."

Google just stuffed Gemini into millions of new vehicles. The story isn't AI in cars. It's Apple's CarPlay moat finally cracking. For a decade CarPlay sat in the head unit of every premium build Mercedes, Volvo, Ford, GM. That seat routes services through Apple: Maps queries, voice intents, App Store fees on parking and EV charging apps. Google's direct-OEM Gemini play bypasses CarPlay entirely. Native at the head unit, voice-to-action, calendar, payments, all flowing through Google's auth and ad graph instead of Apple's. GM already kicked CarPlay out of new EVs. Ford is publicly evaluating going native. Rivian never integrated. Apple Services hit $96.2B in FY24 at roughly 74% gross margin. Cars are not the bulk of that line they're the wedge. The wedge is what trains a driver to default to Siri, then keep defaulting to Siri on the iPhone in their pocket. Lose the car, lose the habit, lose the attach rate on every Services SKU that quietly carries the gross margin print Wall Street rewards. Apple isn't losing the car for the car. They're losing the assistant a billion people will spend the next decade talking to in traffic.

Boil the ocean" prompts work until the agent hits 200k tokens and starts forgetting the schema it wrote 40 steps ago. Codex /goals helps because it externalizes state, but the real unlock is pairing it with a kill switch on token spend. Watched a Cursor agent burn $80 in API calls building the wrong abstraction last week.

Stolen from @garrytan's "Boil the Ocean" concept plus a lot more! Seeing some pretty incredible results paired with the new /goals feature for Codex

@garrytan the /plan-ceo-review meta is basically Cursor + Claude Opus pretending to be Keith Rabois. works until you ship 4 conflicting "strategic priorities" in one sprint and your eng lead quits. the prompt is doing 20% of the work, the founder reading it carefully is doing 80.

Many such cases

Amjad Masad telling TechCrunch he'd rather not sell Replit is the politest way to say "Cursor's $9.9B markup is a ceiling, not a floor." Anysphere hit ~$500M ARR by selling autocomplete on top of Anthropic and OpenAI models it doesn't own. The gross margin math gets ugly the second Claude pricing moves or a competitor ships an in-house model with the same UX. Replit's pitch is structurally different: own the runtime, the database, the deploy target, the agent loop. Capture margin on compute, not just tokens routed to someone else's GPU. The "fighting Apple" line is the real tell. App Store rent is 30% on IAP that turns a $20/mo coding tool into $14 of net before any compute COGS. Masad isn't picking a fight for principle. He's defending the only contribution margin he gets to keep. Selling now means accepting Cursor's path: priced on ARR multiples, exposed on COGS the moment your model vendor decides your gross margin is theirs. He's not refusing the offer. He's refusing the multiple.

Salesforce ran this exact pattern in 2015 when they brought back a VP who'd done two years at a startup. Pipeline coverage dropped from 3.2x to 1.8x in one quarter. Muscle memory on objection handling and forecast discipline atrophies fast. The grind tolerance is the real moat, not the rolodex.

The #1 VP of Sales hire to avoid: Anyone who left sales for 1+ years to do something else (influencer, small biz owner, etc.). Sales is a hard, grueling job. Stepping away for too long almost always makes them fail when they come back.

@levie Utility framing breaks the second you look at compute economics. Anthropic is burning ~$3B/yr on training runs, OpenAI ~$5B. Water utilities don't have 70% gross margin trajectories or $150B valuations. Calling it a utility is how you end up underpricing the next Nvidia.

In general, we should treat AI like a utility, not like a being. The more we confuse what AI is the more we will make ourselves go crazy with analogies that will never fully hold true.

Spotify's "AI music problem" isn't aesthetic. It's a payout pool that gets diluted every time a Suno track gets a stream. The mechanism: Spotify pays roughly 70% of revenue into a single rights holder pool, then divides pro-rata by stream count. The $0.003 to $0.005 per stream number isn't a price. It's the result of a division. Add 10 million AI-generated tracks gaming autoplay queues and the denominator grows while the numerator stays flat. Every stream a Suno bot wins is a stream that didn't pay rent for somebody who actually recorded a song. Spotify doesn't pay more. They just split the same check more ways. Daniel Ek's incentive to fix this is exactly zero. Payout as a percent of revenue is locked into the rights deals. Gross margin gets better the more the pool dilutes. AI music isn't competing with artists. It's competing with the artists' rent.

A Harvard study just showed AI out-diagnosing two ER doctors head-to-head. That's not a research win. It's a malpractice insurance repricing event. Med-mal premiums are underwritten on physician error rates. The moment peer-reviewed work shows a lower error rate from the model than from the human, carriers' actuaries are required to update priors. Once that happens, insurers start asking why the human is in the loop on routine cases the model handles cleanly. Epic Systems is shipping ambient AI across records covering 250M+ Americans. Hippocratic AI raised $141M at a $1.6B valuation to build autonomous AI nurses. Abridge sits above $2.7B as the documentation layer that owns the chart note. None of these companies pitch "a tool for doctors." They pitch the layer that signs the diagnosis. The doctor isn't being replaced by AI. The doctor is being repriced out by their own malpractice carrier doing the math.

@elonmusk Starship v2 is 123m tall, 100+ ton payload to LEO target. Falcon 9 is 70m and 22 tons. The size jump is the whole unlock for Starlink V2 sats which are 7x heavier than V1 and can't fly on Falcon at all.

Size does matter

Chrome's File System Access API only fires on user gesture so you'll need a click handler bridge for the auto-attach flow. tldraw exports as SVG by default but most upload boxes want PNG, run it through canvas.toBlob first or you'll get rejected by anything using sharp on the backend.

Another day, another product idea from @swyx Introducing the Locker extension: - Attach any files from your locker to an upload box - Generate files or use @tldraw to create on the fly - Still 100% Open-source and BYOB GitHub link below 👇

@TeresaMelvinart @Replit @UniAmico Replit Agent's free day is smart funnel math. They burn ~$50/user in Anthropic API costs for 24hrs, convert maybe 3% to the $20/mo Core plan, and the buildathon UGC becomes free distribution. Cursor and Lovable are running the same playbook now.

I built my product in @Replit as a part of 10yr Replit Buildathon! Got Replit free for 24 hrs so why not! Vibe coded my first landing page and mobile app for @UniAmico - a platform to guide high schoolers and their parents to finding best art colleges. Yeah you guessed it right! I am solving my problem too 😊 cc @MannyBernabe @Franciscocrz @amasad @HayaOdeh

@amasad Replit Agent crossed $100M ARR faster than Cursor did. The MVP-in-a-day stuff is the demo, the real moat is hosting plus deploy in the same loop so people never leave the tab. Vercel and Netlify are about to feel that one.

Amazing what a single day of marathon vibe coding can achieve.

Tesla just hit 10 billion FSD miles. Waymo, the only company running a regulator-approved driverless service, has under 100 million. These are not the same metric. Every Tesla FSD mile has a human in the seat who is legally responsible for the vehicle. Every Waymo mile is rider-only. Tesla's 10 billion is supervised demonstration data. Waymo's number is paid commercial product. When the press release calls 10B a "milestone for autonomy," what it actually measures is the size of the unpaid labeling workforce sitting behind a steering wheel. A disengagement is not a failure in the data pipeline. It is the desired output, a labeled correction Tesla collects for free while the driver retains the liability. NHTSA's standing FSD investigation covers hundreds of crashes and multiple fatalities, all generated under the same supervised regime that produces the training set. 10 billion is the procurement budget Tesla didn't have to spend on labeled driving data. It is not the autonomy benchmark.

Uber wants to turn 7M+ drivers into a sensor grid for Waymo, Wayve, and every AV outfit racing the 2028 deployment window. The mechanism: driver phones and dashcams collect road footage and edge cases. Uber packages it. Sells it as training data to the companies building the cars that will delete the driver line item entirely. Uber's mobility take rate is ~28%. Gross margin on a data licensing business is closer to 85-90%. They've been hunting a high-margin attach to a low-margin core for a decade. Here it is. Drivers pay for their own gas, their own car, their own phone, their own data plan, and feed the model that retires them. The pitch deck calls it "Uber AI Solutions." The 10-K will eventually call it "Other Bets." The drivers won't see a cut of either.

Anthropic literally wrote this into Claude's spec, "challenge the user when warranted" beats sycophancy on their internal evals. Cursor's $9B valuation came from devs who wanted a copilot that pushed back on bad architecture, not one that autocompleted it. Yes-man AI is a dead product category.

being a useful coworker is a good alignment target, except a high level of skill of being a good coworker is challenging you, your assumptions, fundamentally changing your business, writing new values on new tablets, participants in the holy unfolding of creation,