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, I'm your boy Tony DeLuca, and welcome back to Barely Possible, where we read the technical stuff so you don't have to, and then we tell you who's getting squeezed. Today's menu has a real spine to it. We've got Anthropic dragging Alibaba into a fight, accusing them of running twenty-five thousand accounts to mine Claude for almost twenty-nine million conversations. We've got the price of memory chips showing up at your door — Apple and Xbox both jacking prices the same week, and that's not a coincidence. And we've got a story I keep chewing on: software companies straight-up killing their own apps because they say nobody uses them anymore, the agents do. Buckle up, let's have at it.
Let me start with the one that has the most teeth for anyone building anything, because it's the deep dive today and it's worth slowing down for. Anthropic is going after Alibaba. According to the report from Ashley Belanger over at Ars Technica, Anthropic is saying Alibaba ran the largest Claude cloning attack they've ever seen. The numbers in there are eye-popping. Twenty-five thousand accounts. Twenty-eight point eight million exchanges with Claude. And the framing Anthropic is putting on it is not just terms-of-service violation, it's that Alibaba was systematically mining Claude to steal capabilities — essentially using Anthropic's own model as a teacher to train or improve a competing system.
Now let me put my skeptical hat on for a second, because that's the job. What's being described here is something the industry has a polite name for: distillation. You point a cheaper student model at an expensive frontier model, you ask it millions of questions, you record the answers, and you use that to make your own thing smarter. Researchers have done versions of this for years. It is the open secret of how a lot of fast-followers caught up. What's different now is two things. One, the scale. Twenty-five thousand accounts is not a curious grad student. That's an industrial operation, the kind of thing you build a team and a budget around. And two, the timing, because Anthropic is wrapping this in the language of the export-control fight that has been hanging over them all month.
And that's where I want to connect this to something we've been circling on this show. We talked, going back a couple weeks, about how the US Commerce Department put a kind of kill switch on Anthropic's models over jailbreak concerns, and we covered how the company has been navigating that whole mess. There's even reporting out there with a headline along the lines of how Anthropic may have talked itself into an AI export ban. So you have to read this Alibaba accusation in that context. Anthropic right now has every incentive to frame a Chinese tech giant as defying the US government and stealing American AI capabilities. It plays directly into the policy story they're already living inside. That doesn't mean it's false — twenty-eight point eight million exchanges is a specific, countable thing, and they clearly logged it. But when a company makes a national-security-flavored accusation that conveniently strengthens its own regulatory position, you note the alignment of interests and you stay a little cool about the moral packaging.
Here's why this actually matters to you as a builder, and it's not the geopolitics. It's the business model question that this story drags into the open. If your product is an API that other people send prompts to, your capabilities are extractable. Anybody with enough accounts and enough patience can interrogate your model millions of times and walk away with a synthetic dataset that captures a meaningful chunk of what makes you special. That is the structural vulnerability of selling intelligence over a wire. Anthropic is essentially saying the quiet part out loud: the moat erodes every time you answer a question. And their response — pursuing this publicly, demanding Alibaba be punished — is them trying to build a legal and reputational fence around something that has no technical fence.
So if you're a founder, the lesson isn't "China bad." The lesson is: assume your model's outputs are training data for somebody else. Price it in. Rate-limit like you mean it. Watch your account-creation patterns. And understand that if your entire defensibility is "we have the smartest model," you are renting that lead, not owning it, because the act of serving customers is also the act of teaching your competitors. That's the uncomfortable shape of this whole industry right now, and the Alibaba fight just made it visible.
There's a second thread worth pulling, and it ties right back. In the same window, there's a TechCrunch piece from Julie Bort with a headline that should make OpenAI's product team sit up: Anthropic's Claude is winning over paid consumers, a market that ChatGPT basically owns. The reporting says that despite ChatGPT's commanding overall lead, the people who actually pull out a credit card and pay for AI have increasingly been choosing Claude. Now I want to be careful here — the source is summarizing data, and I don't have the methodology in front of me, so I'm not going to oversell the precision. But the directional claim is interesting. ChatGPT has the masses. Claude is quietly winning the wallets. The free-tier giant versus the paid-tier specialist. For Anthropic, that's a much healthier place to be than raw user counts suggest, and it's a reminder that market share and revenue share are two different animals. The loudest product isn't always the one getting paid.
Now let's shift from model business models to something you can feel in your own wallet. Memory. Two stories landed practically on top of each other. Apple ratcheted up prices, and the company is openly blaming the cost of memory. The piece from Jonathan M. Gitlin at Ars Technica is blunt about it — some Macs are hundreds of dollars more expensive than they were the day before. And right behind it, a TechCrunch piece from Lauren Forristal: Xbox follows Apple with price increases, and Microsoft is saying the same thing, that it's being driven by rising memory and console storage prices, with costs more than two and a half times higher than previous levels.
Two and a half times. Let that sit. We talked yesterday about a memory chipmaker whose profit jumped roughly fifteen-fold, to twenty-eight point two billion, and how the real bottleneck in this AI buildout isn't the fancy GPUs everybody photographs, it's the high-bandwidth memory that feeds them. Well, here's the downstream consequence showing up at the consumer level. When every data center on earth is fighting to stuff its racks with memory, the supply that used to go into your laptop and your game console gets bid away. The AI boom is reaching into the consumer electronics aisle and lifting prices on things that have nothing to do with AI. Your kid's Xbox costs more because some hyperscaler is hoarding DRAM. That's not a metaphor, that's the actual mechanism.
And for a founder, this is a planning input, not just a gripe. If you're building anything that touches physical hardware — devices, kiosks, on-prem boxes, edge inference rigs — your bill of materials is being squeezed by forces completely outside your category. Memory is the new contested resource, and the contest is going to last as long as the data-center arms race lasts. Budget accordingly, and don't assume last year's component prices.
Now let me bring you to the story I find genuinely strange, and I mean that as a compliment. Notion is killing its email app. Two pieces on this — one from Scharon Harding at Ars Technica, one from Ivan Mehta at TechCrunch. Notion Mail, the email client that drew a lot of its DNA from the Skiff team Notion acquired, is being shut down. And the reason they're giving is the part that stops me. They're saying most users are using AI agents to run their inbox instead, so they're going, in their words, all in on using agents to run your inbox.
Think about what that actually means. A company built an email app. Polished it. Shipped it. And then concluded that the interface itself — the inbox you look at, the messages you click, the buttons you press — is becoming vestigial, because people would rather hand the whole thing to an agent and let it triage, draft, and respond. The product isn't dying because it failed. It's dying because the layer above it ate it. Why maintain a beautiful inbox if the user never wants to open the inbox?
Now, I want to be the skeptic in the room, because this is exactly the kind of claim that gets ahead of reality. "Most users hand their email to agents" is a heck of a sentence. I'd love to see the cohort data, because there's a real difference between power users at a fast company and your aunt managing her email. It's also a very convenient story for a company that maybe didn't want to carry a standalone email product anymore — "the agents made us do it" is a tidy way to retire a product line. So take the framing with a grain of salt.
But here's why I'm not dismissing it either. There's a pattern forming, and we've been watching it build all week. Anthropic's Claude Tag thing we covered, dropping an agent into Slack as a persistent teammate. Now Notion saying agents are taking over the inbox. The through-line is that the unit of software is shifting from "app you operate" to "agent you delegate to." And if that's even partly true, it changes what you build. If your product's value is a slick UI for a task, ask yourself the Notion question: what happens to me when the user stops wanting to touch the UI at all? The companies that survive that shift are the ones that own the data, the integrations, and the trust — not the ones that own the prettiest screen. The screen might be the thing that disappears.
Let's stay in the agent economy for a minute, because the money is following it. Patronus AI just landed fifty million dollars to build what they're calling digital worlds that stress-test AI agents. The piece from Marina Temkin at TechCrunch notes the company was founded by former Meta AI researchers and that their investor describes demand as nearly insatiable. And that phrase tells you everything about where we are in the cycle. Everybody and their cousin is deploying agents into production. Almost nobody knows how to test the things. An agent isn't a function you can unit-test with a few asserts — it takes actions, it chains steps, it touches real tools, and it can fail in a hundred creative ways a human tester would never dream up. So building synthetic environments where you can throw an agent into the deep end and watch it drown safely, before it drowns your customer's data, that's a real picks-and-shovels business.
And it rhymes with another funding story in the same batch. General Intuition raised three hundred and twenty million, in a round that values them around two point three billion, on a bet that's frankly audacious: that video games can train AI agents for the real world. The piece from Rebecca Bellan lays out the thesis — millions of hours of gameplay as training data, the idea being that action data, the actual moment-to-moment decisions players make, can help AI develop something closer to human intuition. Now, I've got my eyebrow up on this one. Game environments are not the world. A character dodging in a shooter is not a robot navigating your kitchen. But the underlying instinct — that we've run low on good text and we need new kinds of data that capture decision-making and action, not just words — that instinct is sound. Whether gameplay is the right action data is the three-hundred-million-dollar question, and somebody just answered "yes" with a checkbook. We'll see.
There's one more on the agent-testing-and-evaluation theme worth a quick mention, because it connects the dots. There's a report floating around about OpenAI's Codex bombarding SSDs with needless write operations, reportedly costing real money in wasted disk wear. I'm flagging it carefully because I don't have a firm date on it, so I'm not going to tell you it happened this morning. But the substance fits the moment perfectly. Agents that run autonomously make decisions at machine speed and machine volume, and when they're sloppy, the waste isn't a typo, it's millions of writes hammering hardware nobody was watching. That's exactly why the Patronus AIs of the world are getting funded. When you let software take actions without a human in the loop, the failure modes get expensive and weird, and you'd better have a way to catch them before they ship.
Now let's shift from agents to the grid, because there's a genuinely clever startup story in here. A16z-backed Base Power is offering cheaper electricity to the part of the grid that needs it most. The piece from Tim De Chant explains the mechanism, and it's the kind of thing I love because it routes around a broken system instead of waiting for it to be fixed. The big regional grid operator out East, PJM, has this notoriously clogged interconnection queue — if you want to plug a big new power source into the grid, you can wait years in line. Base Power's move is to skip the queue entirely by putting batteries in people's homes. They install the battery, the homeowner gets backup power during outages, and Base gets a distributed fleet of storage it can lean on when the grid is straining. It's a virtual power plant assembled one garage at a time.
Why does a builder care? Because this is a template. The official path is jammed — years-long queue, regulatory molasses. Instead of fighting that, they found a distributed end-run that creates value on both sides: the customer gets something they want, backup power, and the company gets the asset it needs, storage capacity. Whenever you're staring at a bottleneck that everyone treats as immovable, the Base Power question is worth asking: is there a distributed version of this that sidesteps the queue entirely? Sometimes the moat isn't technology, it's a clever go-to-market that nobody else was patient enough to assemble.
Let me run through a few quicker ones that matter to how you operate. Microsoft added another year to its Windows 10 extended update program. The detail that jumps out from Ryan Whitwam's piece: about a quarter of PCs are still running the previous operating system. A quarter. That's a staggering amount of inertia, and if you ship software to a broad audience, it's a reminder that your users are not all on the shiny new thing. The long tail of old operating systems is long, and Microsoft just admitted it by extending support yet again.
Google finally shipped a Google Finance mobile app — Android first, iOS promised later in 2026, per the company's own blog post from Barine Tee and the writeup from Ryan Whitwam. It took twenty years, and naturally it arrives stuffed with AI features. I mention it less for the app itself and more for the pattern: legacy products are getting re-released as AI delivery vehicles. The finance app is a wrapper for the AI now. That's the move across the whole industry — every old surface becomes a new place to put a model.
And on the policy side, the FCC may kill a two-billion-dollar program that connects schools and libraries to the internet. The piece from Jon Brodkin notes the chairman is citing screen-time concerns and is getting accused of trying to be, quote, the nation's parent. I'll just say plainly: connectivity for schools and libraries is infrastructure, and dressing up a budget cut in screen-time worry is a thin costume. Watch this one, because the digital divide doesn't fix itself, and the people who lose the connection aren't the ones in this conversation.
Quick hit in crypto land: Polymarket says hackers stole users' funds, and the prediction market is refunding people who lost money in what it's calling a third-party breach. The piece from Lorenzo Franceschi-Bicchierai keeps it tight. Two things worth your attention. One, "third-party breach" is doing a lot of work in that sentence — it means the weak link was a vendor or integration, not necessarily Polymarket's core, which is the most common way crypto platforms get burned these days. Your security is only as good as the least-careful service you connect to. Two, they're refunding users, which is the right call reputationally but also tells you they'd rather eat the cost than eat the trust damage. In a prediction market, trust is the entire product.
Let me close the loop with a couple of stories that are less about AI and more about the world that AI is running in. Colossal Biosciences — the de-extinction folks — are involved in a new effort to get genome sequences for the entire endangered species list, and to biobank tissue from all of them, per John Timmer's reporting. Set aside the woolly-mammoth headlines for a second. A complete genomic library of every endangered species, plus banked tissue, is a serious conservation asset regardless of what you think about bringing anything back. It's a hedge against extinction in data form. And the fact that a private biotech is the one doing it tells you something about where the capability and the capital sit now.
And then there's the heat. Europe is baking under its second heat wave of the summer, and the piece from Lauren Dalban at Inside Climate News uses a phrase that stuck with me — the sad inevitability of it. I bring it into a builder show on purpose, because climate is not a side topic for anyone running data centers or hardware. The same machines driving the memory crunch we just talked about run hotter every year, in a world that's running hotter every year. We covered Nvidia's water and cooling challenges earlier in the week. The heat wave and the data-center buildout aren't separate stories. They're the same story seen from two ends — and the cooling bill is coming due for everyone.
So let me tie a bow on it. The thread running through today is extraction and what it costs. Alibaba allegedly extracting capabilities from Claude one query at a time. Data centers extracting memory out of your laptop and your Xbox. Agents extracting the inbox out from under the email app. And startups raising hundreds of millions to test, fund, and route around all of it. The builder's takeaway is consistent: whatever you make, ask what's extractable from it, who benefits when they take it, and whether your real defensibility is the thing on the screen or the thing underneath. Because the screen, more and more, is the part that disappears.
That's the menu for today. I'm Tony DeLuca, this has been Barely Possible, and I appreciate you spending a piece of your day with me. Go build something they can't mine out of you. Catch you next time.