Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, pull up a chair, I'm your boy Tony DeLuca and this is Barely Possible. Fresh menu today, and I'm gonna tell you up front, the special is something most of the other shows are gonna skip right past, because it doesn't have a flashy model number attached to it. We're gonna talk about Anthropic quietly trying to crawl inside your company's Slack and learn everything you know — not the chatbot novelty, the strategy underneath it. We've got Menlo Ventures cashing a three-billion-dollar check off one gutsy bet. We've got Oracle firing twenty-one thousand people to fund a debt-loaded data-center binge. Zuckerberg wants to run a betting parlor. And LastPass got hit again — second time, same kind of wound. So grab your coffee, let's have at it.
Let me start with the one I want you actually thinking about while you fold laundry tonight. Anthropic put out something called Claude Tag. On the surface it's the most boring product announcement of the week — an always-on Claude that lives in your Slack, answers questions, tags into conversations like a teammate. Rebecca Bellan at TechCrunch wrote it up, and to her credit she didn't take the bait of treating it like a productivity gadget. She framed it as what it actually is: a strategic play to capture organizational context, institutional knowledge, and enterprise workflows. And that subhead is the whole ballgame. Let me read you exactly how she put it, because the words matter here. The piece says Claude Tag, quote, brings an always-on AI teammate to Slack. But beyond productivity, the feature is a strategic play to capture organizational context, institutional knowledge, and enterprise workflows.
Now let me translate that out of press-release-ese and into kitchen-table English. Every company runs on a kind of invisible nervous system. It's not in the wiki. It's not in the org chart. It's in the Slack threads where somebody says, oh yeah, don't deploy on Fridays because the payment processor flakes out, or, talk to Maria before you touch the billing schema, she's the only one who remembers why it's weird. That stuff — the why-it's-weird, the who-knows-what, the unwritten rules — that's the actual moat of a company. That's the thing that makes you you and not your competitor. And what Claude Tag does, sitting in there all day every day, is it learns that. One Slack message at a time, like the headline says. It's not just answering your questions. It's building a model of how your business actually runs.
And here's why I, as the guy who's supposed to protect your time and your interests, want you to sit with that for a second. If you're a founder, this is a genuinely tempting product. An assistant that knows your whole company, can onboard new hires, can answer the thing nobody wants to ask out loud for the fifth time — that's real value. I'm not gonna pretend it isn't. But you should be clear-eyed about the trade. You are handing the most defensible, least-copyable asset your company owns — its accumulated context — to a model vendor. And the question you have to ask, the question nobody in that announcement is gonna ask for you, is: where does that learned context live, who can see it, and what happens to your leverage when the vendor that knows your company better than your own new hires decides to raise prices?
This is the same dynamic we keep circling back to on this show, going all the way back to that Tesco-VMware lock-in conversation we did a week or so ago. The pattern is always the same. The tool starts as a convenience. Then it becomes the place your operational memory lives. Then it's not a tool anymore, it's a dependency, and dependencies get priced. So I'm not telling you don't use it. I'm telling you, if you adopt something like Claude Tag, you treat the context it accumulates as a corporate asset with an exit plan. You ask the boring questions early — export, portability, deletion — when you still have the leverage, not after the thing knows where all your bodies are buried. Whoever owns your company's context owns a piece of your company. Just go in knowing that's the actual deal.
And here's the part that should make a founder's ears perk up even more. Anthropic isn't doing this in a vacuum — they put it out the same window they're flush with believer money. Which brings me right to the people who funded the believing.
Menlo Ventures. Julie Bort at TechCrunch wrote this one up, and the framing is just chef's kiss for what venture capital actually is. The headline: after betting the firm on Anthropic, Menlo Ventures raises a victorious three-billion-dollar fund. And the line that tells the whole story — Menlo built its reputation as an AI investor all on one gutsy seven-hundred-fifty-million-dollar move back in 2024. One bet. Three quarters of a billion dollars into Anthropic. And now, off the back of how that's gone, they raise a fresh three billion.
Let me be the skeptic in the room for a second, because that's my job. This is the part of the cycle that should make the hair on your neck stand up a little, not because anybody did anything wrong, but because of what it tells you about where we are. When a firm can write up a single concentrated bet as the foundation of its whole identity, and limited partners line up to hand them billions more, that's not a sign the market is cautious. That's a sign the market has a winner-take-most story it desperately wants to be true. And look — maybe it is true. Anthropic's enterprise business, the Claude Code stuff, the agentic coding wave, that's real revenue, real adoption, not vapor. I'm not calling top. But for you as a founder trying to raise money in this environment, understand the weather you're standing in. The capital chasing AI right now is concentrated, it's conviction-driven, and it's looking for the next gutsy bet, which means it's both easier and more dangerous to raise into. Easier because the checkbooks are open. More dangerous because the expectations baked into those checks are enormous, and they don't care about your nuance. You take that three-billion-dollar money, you're signing up for three-billion-dollar outcomes.
Now let me show you the other side of that same coin, because the believing money has a shadow, and the shadow is the bill. Oracle. Scharon Harding at Ars Technica with the headline that says it all: Oracle's twenty-one thousand layoffs help drive its debt-fueled AI investments. Twenty-one thousand people. And the framing isn't that Oracle had a bad quarter and had to cut. The framing is that the layoffs are part of how they're paying for the buildout. They're spending billions on data center infrastructure to support AI, and they're going into debt to do it, and the people are the offset.
I want you to feel the shape of that, because it connects directly to the Menlo story. On one end of the pipe, you've got venture firms raising billions on conviction. On the other end, you've got an enterprise software company that's been around since the disco era cutting twenty-one thousand jobs and loading up debt to build the picks and shovels. Same gold rush, very different positions in it. And here's the thing a builder should take away. When the infrastructure layer is being funded by debt and headcount cuts, that tells you the people closest to the actual cost of this stuff — the cost of the GPUs, the power, the buildings — are doing painful math. We talked yesterday about Reflection AI committing to pay SpaceX a hundred and fifty million dollars a month for compute through 2029. That's the demand side. Oracle going into debt and cutting twenty-one thousand jobs to build capacity — that's the supply side, doing whatever it takes to be there when the demand shows up. Both of those things being true at the same time is the whole AI economy in one breath: enormous projected demand, enormous real cost, and a lot of people betting the projection covers the cost. If you're building on top of any of this, your job is to never forget that the cheap, abundant compute you're enjoying is being financed by somebody's debt and somebody's job. That doesn't last forever at these prices.
Now let me pivot off the money and onto something that's gonna touch more of you directly, because it's about your own security hygiene. LastPass. Zack Whittaker at TechCrunch reported that LastPass — the password manager, the thing whose entire job is to hold your most sensitive keys — says hackers stole customer support case data during what's being called the Klue breach. And the gut-punch line in there: this is the second data breach to affect LastPass customers in recent years. Second time. After one of their tech partners got breached.
And then there's the companion piece, also Whittaker, on Klue itself, and this is the part that should make every founder in earshot wince in recognition. Klue says hackers stole a credential — from 2022. From a limited pilot. And the quote that does the damage: it's unclear why Klue had not revoked the credential after the limited pilot, which hackers then used to breach a system holding keys for accessing customers' data. Read that again. A credential from a pilot program — a little test run — from four years back. Never revoked. Just sitting there. And it was the door.
I'm not gonna do the cheap shot at LastPass here, because honestly this is the most relatable security failure there is. Every one of you has a Klue in your own systems. You ran a pilot with a vendor in 2023. You spun up a service account for a hackathon. You gave a contractor temporary access and the contract ended but the key didn't. That stale credential nobody remembers is the single most common way real companies get owned. It's not some genius zero-day. It's housekeeping you skipped. So consider this your free, no-charge reminder from your neighborhood radio guy: go look at your access tokens. Go look at your service accounts. The ones tied to things that don't exist anymore. The pilot that ended. The integration you killed. Revoke them. Today, not next sprint. Because the entire LastPass-Klue mess traces back to a key that should've been turned off years ago and wasn't, and now customer data's walking out the door. The lesson isn't that password managers are bad. The lesson is that your supply chain is your attack surface, and a four-year-old forgotten key in some vendor's system can blow back onto your users. Audit your dependencies like your reputation depends on it, because it does.
Alright, let me lighten the room a little, because not everything today is a cautionary tale. Let's talk about Zuckerberg wanting to open a casino. Lucas Ropek at TechCrunch: Mark Zuckerberg wants Meta to launch its own prediction market. And the detail that matters — the app would be independent of Meta's other social offerings, but sources told the New York Times that those social sites could direct users to engage with it. Independent. With an asterisk the size of Texas.
So here's my read, and I'll keep it tight because we already touched the prediction-market world earlier this week with the Polymarket fake-bet mess. Prediction markets are having a moment. They look like markets, they feel like information, they've got this veneer of we're-just-aggregating-wisdom-of-crowds. But strip the wrapper off and a prediction market is a betting product. And what Meta brings to a betting product is the one thing the existing players don't have — distribution. Billions of users, an engagement engine that already knows exactly which dopamine lever to pull on each one, and the ability, as the reporting says, to point those users at the app. That's not a neutral aggregator of truth. That's a gambling funnel bolted onto the largest attention machine ever built. For a founder watching this, the lesson is the one that never changes: distribution beats almost everything. Meta doesn't need to build the best prediction market. It needs to build a fine one and aim a firehose of users at it. If you're a startup in any space where Meta could plausibly point its users, that's the competitive reality you live in. The product's never the moat. The funnel is.
Now let me shift over to the agent-and-tooling stuff, because there were a few quieter pieces worth a builder's attention. Russell Brandom at TechCrunch had a piece on what he calls the AI world getting loopy. The idea — the loop takes agentic AI a step further by authorizing a swarm of agents to work continuously in the background, endlessly. Not you prompting, getting an answer, prompting again. A standing crew of agents just running, indefinitely, doing work while you're asleep.
I'll give you the dry-eyed version. This is the natural next step, and it's also where the bills and the risks both go vertical. An agent that runs once and stops, you can reason about. A swarm of agents running endlessly in the background, that's a thing that can rack up compute costs, take actions you didn't sanction, and chase its own tail in ways you won't notice until the invoice or the incident shows up. And it ties right back to that LastPass lesson — every one of those background agents needs credentials, needs access, needs permission to touch things. The more autonomous, always-on machinery you run, the bigger your attack surface and your blast radius get. So if you're tempted by the always-on swarm — and it's tempting, the productivity story is real — build the guardrails before you build the loop. Spending limits. Action approvals for anything irreversible. Logging you actually read. The endless background loop is coming whether I like it or not, but the founders who survive it are the ones who treat an autonomous agent the way you'd treat a new employee with a company credit card and no manager: with a very short leash at first.
And speaking of what these tools can actually do, here's a nice concrete one from Simon Willison, who's one of the more grounded voices in this whole circus. He posted that his parallel agent side-project for the day was having Claude Code port a new image model to ONNX so it could run entirely in the browser. Now I'm not gonna drag you into the technical weeds — ONNX, browser inference, that's not the point for this audience. The point is the texture of it. A guy, as a side project, in a day, points an AI coding agent at a real engineering task — take this model, convert it to a format, make it run in a place it wasn't built to run — and it gets done. That's the actual state of agentic coding in builders' hands right now, not the benchmark numbers, the lived reality. The leverage a single competent developer has today is genuinely different than it was eighteen months ago. If you're a founder, that's not a threat, that's your unfair advantage. Small teams can now attempt things that used to need a department. The flip side, of course, is so can your competitor's small team. The floor came up for everybody.
Let me touch a couple more business moves quickly, because there's a tell in them. Superhuman — the email company, also has an AI detection tool through Grammarly — acquired GPTZero. That's the startup that became famous for detecting AI-written text. Julie Bort had it. And I just want to flag the funny little irony loop here without making a whole sermon of it. We are now in a world where the same companies building tools to generate text are buying up the tools to detect generated text. The detection business and the generation business are merging into the same companies. Make of that what you will, but it tells you AI detection has stopped being a moral crusade and become a feature — something you bolt onto a writing product, not a standalone mission. For founders in the AI-detection space, that's the writing on the wall: you're an acquisition target or a feature, probably not a category.
And one more on the consumer side — Meta also debuted new, cheaper smart glasses under its own brand, available in several countries with various color and lens combos. Aisha Malik reported it. I'll keep this to a sentence of analysis: Meta dropping the price and putting its own brand on the glasses is them trying to move smart glasses from gadget-for-enthusiasts to thing-normal-people-buy. Whether the AI inside them is good enough to justify wearing a computer on your face is the open question, and one cheaper SKU doesn't answer it. But the price coming down is how you know they're playing a long volume game, not a hobbyist game.
Now let me get to a couple of things outside the pure tech-business lane, because part of why you keep me around is I read the stuff that affects how the world works, not just the funding rounds.
There's an interview piece at Ars Technica — Jennifer Ouellette talking to Cory Doctorow about his new book, the title being The Reverse Centaur's Guide to Life After AI, and the framing is how to burst the AI bubble by striking at its roots. Now I only have the setup, not the full argument, so I'm not gonna pretend I read the whole thing. But I'll tell you why even the framing is worth thirty seconds of your attention. Doctorow's whole long-running critique is about what he calls enshittification — the way platforms start good, lure you in, then degrade the experience to extract more once you're locked in. And there's a separate post floating around right now, from a writer named Damian, literally titled the coming enshittification of AI, making a related argument. I can't date that one precisely so I won't pretend it's breaking news, but the theme is in the water. And here's the reason a builder should care, not just a critic. If the skeptics are even half right that today's generous, subsidized, venture-and-debt-funded AI tools are in their lure-you-in phase — the cheap, the good, the abundant phase — then the question isn't whether they'll get more expensive and more extractive, it's when. Which loops right back to everything I said about Claude Tag and lock-in and Oracle's debt. The friendly tool of today funds itself somehow tomorrow. Build accordingly. Keep your exits. Don't pour your company's whole soul into something you can't walk away from.
Let me give you a continuity callback while we're here. Earlier this week we spent real time on John Jumper, the Nobel laureate leaving DeepMind for Anthropic, as a signal about where the talent is flowing. I want to connect that to today's money stories, because it's the same river. Talent flowing toward Anthropic, Menlo raising three billion off an Anthropic bet, Anthropic shipping enterprise products like Claude Tag to capture company context — that's not three separate stories. That's one story about gravity. The capital, the talent, and the enterprise land-grab are all pulling in the same direction right now. And when everything pulls the same direction at once, that's exactly the moment to keep one skeptical eyebrow raised, because that's how bubbles feel from the inside — like obvious, undeniable consensus. I'm not saying it's a bubble. I'm saying when the smart money, the smart people, and the product strategy all agree this hard, the contrarian question — what if the unit economics don't pencil out — is the most valuable question in the room, and almost nobody's asking it.
A couple of quick hits to round us out, and then I'll let you go. There's a story out of Ars about police touting using a drone to disarm a motionless suspect — removing a knife, billed as a nationwide first, in a promo video, as more departments fly drones as first responders. I flag it not for the gee-whiz but for the quiet thing underneath: drones as first responders is becoming normal, and the promo-video framing tells you the departments know they need to sell the public on it. Watch that space — it's robotics and autonomy showing up in the most consequential possible context, with a weapon and a person, and the norms are being written right now in marketing departments.
There's also the NHTSA investigating an alleged Tesla Autopilot crash that killed a woman in her own home, with the grim detail that Tesla touted Autopilot as lifesaving a day after a grandmother died in a crash. I won't dwell — it's a recent report, the investigation's ongoing, and I'm not gonna litigate fault from a headline. But the juxtaposition of marketing the safety on the same calendar page as a fatality is the kind of thing that should make any founder building autonomous anything think hard about how they talk about their own product's safety. Your marketing copy becomes Exhibit A in the worst week of your life. Write it like that's true, because someday it might be.
And on the space beat, since I know some of you are romantics about this stuff — there's reporting that SpaceX is eyeing point-to-point cargo delivery from orbit with something called Starfall, transport and delivery of goods through space, and a separate report that Kennedy Space Center isn't ready for the era of super-heavy rockets, with SpaceX telling NASA it wants to launch Starship every eight days. I'll just say: delivering cargo through orbit and launching every eight days is the kind of ambition that's either the future of logistics or a very expensive way to find out the ground infrastructure can't keep up. Both can be true at once. File it under watch, don't bet.
So let me bring it home. The thread running through today wasn't a single blockbuster — it was the quiet machinery of where AI's money, context, and risk are concentrating. Anthropic wants to live in your Slack and learn your company. Menlo raised three billion off conviction. Oracle's borrowing and cutting to build the capacity. And a four-year-old forgotten password key reminds us that no matter how futuristic the stack gets, you still get owned by the boring stuff you didn't clean up. If you take one thing into your week, take this: in a market where everybody's racing to give you the convenient thing, your edge as a builder is being the one person who reads the fine print on the exit. Own your context. Revoke your stale keys. Keep your leverage. The believers will tell you not to worry about any of that. The believers are not the ones who'll be holding the bag.
That's the menu for today. I'm Tony DeLuca, this has been Barely Possible. Go turn off an old access token, would you, do it for me, and I'll talk to you tomorrow.