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, welcome back to Barely Possible. I'm your boy Tony DeLuca, and today we've got a menu that runs from a Chinese AI lab deciding to build its own silicon, to Microsoft quietly telling everybody it's going to lean on its own models to save a buck, to a Discord bug that banned a couple hundred people for posting perfectly innocent pictures. There's a real spine to today, and it's not the shiny robot-brain stuff — it's money. It's who's cutting checks and who's tearing them up. So grab your coffee, sit down at the counter, and let's have at it.
Let me tell you where my head is today before we start. If you listened yesterday, we spent the whole show on Anthropic getting outed for that secret Claude tracker watching Chinese users — the anti-surveillance company that turned out to be running a little surveillance program of its own. That was a story about trust and hypocrisy. Today is a different animal. Today is about the balance sheet finally showing up to the party. For about two years, everybody in this business acted like compute was free money and models were the whole product. And now, story after story in front of me, you can see the accountants walking into the room and turning off half the lights. So the thread I want to pull on today is this: the AI industry is entering its expense-management era, and you can watch it happen in the specific decisions companies are making this week.
Let's start with the one that matters most for anybody thinking about geopolitics and supply chains, which is DeepSeek deciding it wants to make its own chips.
Here's the report, out of Ars Technica, written by Samuel Axon. The headline is plain: facing US export controls, China's DeepSeek plans to make its own chips. And the one line that tells you the shape of it is that the plan is to reduce dependency on Nvidia and Huawei. Now let me be careful here, because the piece itself is careful. This is early. This is not "DeepSeek shipped a chip." This is a Chinese AI lab — the one that shook the whole market a while back by showing you could get frontier-ish results without frontier-sized budgets — saying out loud that it wants off the treadmill of begging for Nvidia parts it can't legally buy and leaning on Huawei parts it apparently doesn't fully trust to carry its future either.
And I want you to sit with the "and Huawei" part, because that's the piece most people are going to skip. If you're just reading the headline, the story is US versus China, export controls, the usual heavyweight bout. But the deeper tell is that DeepSeek doesn't just want independence from America — it wants independence from its own national champion. Huawei is supposed to be the answer for China's homegrown silicon. And here's DeepSeek saying, thanks, but we'd like our own thing. That's not a geopolitics story anymore. That's a vertical-integration story. That's a company deciding the chip is too important to rent from anybody, foreign or domestic.
Now, is this going to work? Building silicon is one of the hardest things human beings do. It is billions of dollars, it is years, it is a supply chain of its own with lithography machines that basically one company on earth makes. So when a software-first AI lab says "we're going to make our own chips," the honest builder's reaction is: okay, show me the tape in three years. But here's why it matters even if it takes forever. It tells you that in the current moment, the smartest players believe the constraint is no longer the model. The constraint is the metal. The constraint is whether you can get your hands on enough compute at a price that doesn't bankrupt you. And when your whole strategy rides on hardware you can't control, you start doing expensive, desperate, vertically-integrated things to fix that.
For a founder listening, the takeaway isn't "go build chips." God, please don't. The takeaway is that the ground under the model layer is still shifting, and if you're building on top of somebody's API, you should be asking who controls the silicon underneath that API, and what happens to your unit economics if that supply gets squeezed. DeepSeek is making that anxiety visible at the national-lab level. You're feeling a smaller version of the same thing every time your inference bill lands.
Which brings me right into the story that rhymes with it, and this one's the real spine of the day. Microsoft, according to a piece by Lucas Ropek at TechCrunch, is joining the AI cost-cutting trend by relying more on its own models.
Let me set the table. For years the story of Microsoft and AI was the story of Microsoft and OpenAI — the tens of billions, the "we're all in," the Copilot everywhere. And now the framing has shifted to Microsoft being "the latest Silicon Valley giant to cut back on its AI spending." Not stop. Cut back. And the specific move is to lean more on its own in-house models instead of paying the premium for someone else's frontier model on every single call.
Now I want to be responsible about what the source actually says, because it's a short item and I'm not going to invent numbers that aren't in front of me. What it establishes is the direction: Microsoft is the latest, meaning there's a trend, meaning others already did this. And that word "latest" is the whole point. This is not one company having one bad quarter. This is a pattern hardening into industry common sense. The era where you'd route every request to the biggest, most expensive model because it was impressive — that era is closing. Now the question in every engineering meeting is: does this task actually need the Rolls Royce, or will the Camry get it there?
And here's where I connect it to something we've circled before on this show, because I don't want to pretend this theme is brand new — the token-economics story has been building for weeks. We talked about Claude Sonnet's tokenizer quietly making it more expensive even with lower per-token pricing. We've talked about companies rationing frontier access. We had, in the background reporting floating around, Tesla capping engineers at two hundred bucks a week in token spend because some of them were racking up thousands. So Microsoft leaning on its own models isn't a bolt from the blue. It's the biggest player in the room finally admitting out loud what the smaller players were already doing quietly: the meter is running, and somebody has to turn it down.
For you as a builder, this is genuinely good news dressed up as bad news. When Microsoft — Microsoft, the company with essentially infinite money — decides it needs a cheaper model tier for the boring eighty percent of tasks, that validates the entire discipline of model routing. It tells your CFO that being thoughtful about which model handles which request isn't being cheap, it's being competent. The frontier labs want you to believe you need the top model for everything, because that's how the meter spins fastest. And the biggest customer in the world just voted with its wallet against that idea.
Now let me steelman the other side, because I always want to, and because I think it's important. There's a version of this where Microsoft leaning on its own models is a quality downgrade that customers will feel — where the answers get a little dumber, the Copilot gets a little more useless, and the cost savings show up on Microsoft's books while the pain shows up in your workflow. That's a real risk. Cost-cutting always has a victim, and the victim is usually the person who can't see the substitution happening behind the curtain. So if you're a Microsoft customer, the thing to watch isn't the press release. It's whether the thing that used to work last month still works this month. Trust the tape, not the talk.
Let me pause on the money theme and go somewhere it collides with the physical world, because there's a story here I can't stop thinking about. Jeremy Hsu, again over at Ars Technica, writing about how data centers' energy demand threatens Trump's "Made in America" plan. And the guts of it: there's a squeeze on Rust Belt electricity bills that's threatening the whole manufacturing pitch.
Now this one hits me in a particular spot, because this is exactly the kind of story that gets lost in the AI hype cycle and shouldn't. Everybody wants to talk about the models. Almost nobody wants to talk about the fact that these things run on electricity, real electricity, coming off a grid that regular people and regular factories also need. And what this piece is laying out is a genuine collision of two political promises. On one hand, you've got the "bring manufacturing back to America" story — factories in the Rust Belt, jobs, machines humming. On the other hand, you've got the AI data-center boom, which is eating electricity like it's going out of style and driving up the cost of that electricity for everybody sharing the grid.
And here's the brutal part: a factory is an electricity business wearing a manufacturing costume. If you're trying to make steel or aluminum or anything heavy in America, your power bill is one of your biggest line items. So when the data centers move into your region and bid up the price of a kilowatt, they're not just an abstraction. They're a competitor for the exact resource your factory needs to be viable. You can want both — the AI future and the manufacturing revival — but the grid doesn't care what you want. It's a finite thing, and right now two very hungry industries are fighting over it, and the manufacturers are smaller and quieter and losing.
For a founder, why does this matter? Because it tells you where the next round of political friction lands. When people's home electricity bills go up and their local plant blames data centers, that's a story that writes itself in every local paper, in every kitchen-table argument. And that pressure eventually becomes policy — permitting fights, rate structures, maybe even limits on where these data centers can go. If your business model quietly assumes cheap, unlimited compute forever, this is the story that says: the cheap-power assumption has a political expiration date, and it's getting closer. The physical economy is starting to push back on the digital one, and the fight is over the wall socket.
Alright, let me shift from the heavy infrastructure stuff to a story that's smaller but honestly meaner in its lesson: Discord admitting an AI moderation bug wrongfully banned users over harmless images. That's Lauren Forristal at TechCrunch.
Here's what happened, and the detail is what makes it. The company confirmed that the issue had been affecting accounts since May — since May, folks — with an additional two hundred users banned over the weekend before its team finally identified and fixed the problem. So this wasn't a one-day glitch. This was an automated moderation system quietly, wrongly, banning real people over completely innocent pictures for the better part of two months, and the humans supposedly running the show didn't catch it until it hit another two hundred people in a single weekend.
Now I want to be fair to Discord — content moderation at their scale is genuinely one of the hardest jobs in tech. You've got oceans of images flowing through every hour, some of it truly awful, and no army of humans can eyeball all of it. So you hand it to an AI classifier. Totally understandable. But this is the exact failure mode that should keep every builder who's deploying AI in a decision-making role up at night. It's not that the model was wrong once. Models are wrong sometimes; that's baked in. It's that the model was wrong systematically, quietly, for months, and the system around it had no good way to notice.
That's the lesson, and it's a lesson I keep coming back to because it's the whole ballgame for anybody putting AI in production. The danger isn't the flashy, obvious failure. The danger is the silent, confident, wrong-for-two-months failure. When a human moderator makes a bad call, it's one call. When an automated system makes a bad call, it makes that same bad call ten thousand times before anybody files enough complaints to get somebody's attention. The speed and scale that make AI worth deploying are the exact same speed and scale that make its mistakes catastrophic.
So if you're building something where the AI is deciding — banning, approving, rejecting, flagging — the question isn't "is the model accurate." The question is "how fast will I know when it's wrong, and how much damage happens in the meantime." Discord's answer to that question was, apparently, "a couple months and a few hundred wrongful bans." You want your answer to be better than that. Build the alarm before you build the automation. The alarm is the product. The classifier is just the loud part.
Now let's move from moderation failure to a product move that tells you where the agent wars are heading. Rebecca Bellan at TechCrunch reports that Claude Cowork is expanding to mobile and web.
The framing in the piece is that the coding-agent wars are spilling into the rest of the office. And the concrete update is this: with Cowork on mobile and web, you can start a task from your desk, get status updates on your phone, and pick up the finished output later — even if your laptop is closed.
Now on the surface that sounds like a small quality-of-life update. It's not. Let me tell you what it actually is. For the last couple years, the coding agents lived inside the developer's world — the terminal, the IDE, the laptop. They were tools for the technical people. What Anthropic is doing here is prying the agent loose from the machine. The moment your task keeps running after you close the laptop, the agent stops being a program you babysit and starts being something closer to a coworker you delegate to. You hand it a job, you walk away, you check your phone at lunch, and it's still working.
And the "spilling into the rest of the office" line is the strategic heart of it. The coding agents proved the model — pun intended — that you could give an AI a real task and have it grind through the steps. Now the play is to take that same grind-through-the-steps capability and point it at everyone who isn't a coder. Marketing, ops, sales, the whole building. The bet is that the future of these products isn't a chat window you talk to; it's a background worker you assign things to and then forget about until it pings you.
For a founder, here's the thing to watch. If the winning shape of AI products is "assign and walk away" rather than "sit here and prompt me," that changes what you're building and how you charge for it. A chat tool is priced like a subscription to a smarter search box. A background worker that finishes jobs while you sleep is priced like an employee. Those are wildly different businesses. Anthropic putting Cowork on your phone is a small feature that quietly announces which business they think this is going to be. And I'd bet the mobile app is where they find out whether normal people — not developers, normal people — will actually trust an agent enough to close the laptop and let it run.
That's the second lab making a big managed-agent move this week, by the way, because Google's right there too. Their AI blog, in a post from Philipp Schmid, announced they're expanding managed agents in the Gemini API — background tasks, remote MCP, and more. I'm not going to drag you through the developer plumbing on this one, because you didn't come here for a config file reading. But the headline that matters is the same headline as Cowork: background tasks. Both of the big non-OpenAI players shipped, in the same week, features whose entire purpose is to let an agent keep working when you're not watching it. Google's doing it at the API layer for developers; Anthropic's doing it at the product layer for end users. Same direction, two altitudes. When two competitors independently decide the next frontier is "agents that run in the background," that's not a coincidence. That's the industry agreeing on where the puck is going.
And I'll toss the builders one quick, useful nugget from that Google post — remote MCP. Model Context Protocol is the plumbing that lets an agent reach out and use your tools and data. Making it remote and managed means less of that plumbing you have to run and babysit yourself. If you're building agents, that's a genuine reduction in the boring, load-bearing work. Not glamorous. But the boring load-bearing work is usually where projects go to die, so anything that kills some of it is worth a nod.
Now let's talk about the piece that actually tries to make sense of the whole competitive landscape, because it's the most thoughtful thing in front of me today. Russell Brandom at TechCrunch, with an analysis titled — and I'll say it plainly — why the rise of open source AI isn't hurting Anthropic, yet.
The core argument is elegant, and I want to give it to you straight because it cuts against the conventional wisdom. The common story is that open-source models are a threat to the frontier labs — that as the free, open-weight models get good enough, they'll eat the lunch of the companies charging you for tokens. And Brandom's piece pushes back on that. His framing is that open source's success isn't coming at the expense of the frontier labs. Instead, they each seem to capture two phases of the same life cycle.
Two phases of the same life cycle. Let me unpack what that means, because it's a genuinely useful way to think about the market if you're building. The frontier labs — your Anthropics, your OpenAIs — they're capturing the frontier. The bleeding edge. The task that's so new and so hard that nobody's figured out how to do it cheaply yet. That's phase one: you need the smartest, most expensive model because nothing else can do the job at all. Then time passes, the task gets understood, the techniques diffuse, and an open-weight model shows up that can do that same task for a fraction of the cost. That's phase two: the task has matured, commoditized, and moved down to the cheap layer.
So in this framing, open source isn't the enemy of the frontier lab. It's the exhaust pipe. The frontier lab lives on the frontier, and everything they figure out eventually rolls downhill into the open ecosystem once it's no longer bleeding-edge. As long as there's a new frontier to keep charging for, the frontier lab stays alive — even thriving — while the open models mop up the mature stuff behind them.
Now here's why I love that this piece has "yet" in the headline, because "yet" is doing enormous work. This whole cozy arrangement depends on one thing: the frontier has to keep moving. There has to keep being a new, hard, expensive task that only the top labs can do. The day the frontier stops advancing fast enough — the day the open models catch up not just to last year's frontier but to this year's — the entire two-phase life cycle collapses into one phase, and it's the cheap one. That's the "yet." Anthropic is fine as long as they can stay ahead of the commoditization treadmill. The whole business is a bet that they can keep outrunning their own exhaust.
For a founder, this is one of the most important mental models you can carry right now. When you're deciding what to build on, ask yourself which phase your task lives in. If your product depends on a bleeding-edge capability, you're paying frontier prices and you should assume that in twelve to eighteen months an open model does it for pennies — so either your moat is somewhere other than the model, or you've got a shrinking window. But if your task is already mature — if it's something the open models handle fine — then paying frontier prices is just lighting money on fire. This life-cycle lens tells you where to spend and where to save, and it connects right back to Microsoft leaning on its own cheaper models. Microsoft is doing exactly this: taking the tasks that have hit phase two and refusing to pay phase-one prices for them anymore. The whole industry is learning to read which phase it's in. That's the expense-management era in one idea.
Alright, let me connect the money story to the money story — the actual venture money. Dominic-Madori Davis at TechCrunch reports that the VC firm Chemistry is raising five hundred million for its second fund. Chemistry Ventures, launched by alums from Bessemer, Index Ventures, and Andreessen Horowitz, going out for a half-billion-dollar Fund II.
Now I'm not going to overcook a fund-raise into a grand thesis, but I do want to note what it signals in this particular moment. A five-hundred-million-dollar second fund, raised by people with those pedigrees, tells you the capital is still very much flowing into early-stage tech even as the giants are cutting costs. And that's an interesting tension worth sitting with for a second. Up at the top of the market, Microsoft's trimming AI spend, DeepSeek's trying to escape its chip bills, everybody's counting tokens. But down at the seed and early stage, a firm is comfortable going out for half a billion. That's not contradiction — that's the normal rhythm of a maturing sector. The incumbents get disciplined about costs at the same time the fresh capital keeps hunting for the next thing that'll disrupt those disciplined incumbents. If you're a founder raising right now, the useful read is: the money is there, but it's chasing the phase-one frontier, the new hard thing — not the phase-two commodity. Bring them something that lives on the edge.
Let me do a quick lap through a few more that matter to you, and then I've got one more real one I want to spend time on.
Figma acquired the team behind a vibe-coding app — Ivan Mehta at TechCrunch. The company was Y Combinator-backed, started as a vibe-coding platform, and later built an agent-creation product. Now Figma, the design tool, scoops up the team. This is an acqui-hire in the direction everybody's running: design tools want to become build tools. The line between "I designed it" and "I shipped it" keeps getting thinner, and Figma buying a team that made agents and vibe-coding tools is Figma saying it doesn't want to be the place where the work stops at a mockup. Watch this space — the design-to-code pipeline is consolidating, and if you're building in that gap, you've now got a better-armed competitor.
X rolled out a video editor — Sarah Perez at TechCrunch — for iOS, with multilingual captions, green-screen effects, editing tools, the works. And the stated reason is the one that made me chuckle: to encourage creators to post original content instead of stolen reposts. That's a platform admitting, out loud, that a big chunk of its video is swiped from somewhere else, and its fix is to make it easier to create than to steal. Reasonable move. Whether it works depends on whether the tools are actually good or just present. For builders it's a small reminder: sometimes the cure for bad behavior isn't a policy, it's better tooling that makes the good behavior the path of least resistance.
Netflix is dabbling in shorter video — also Sarah Perez — with new publisher deals bringing two- to twenty-minute videos onto the platform through partners like Rolling Stone and Variety. Netflix, the home of the prestige binge, quietly testing whether it needs a snackier tier to compete for the minutes people currently give to the short-video apps. It's a defensive move against attention, not against another streamer. The battle isn't Netflix versus its streaming rivals anymore; it's Netflix versus the phone in your other hand.
And on the layoffs front, since it connects to the cost-cutting spine of the day: Kyle Orland at Ars reports Bethesda and id Software were reportedly hit hard by Microsoft layoffs, with as much as fifty percent of some teams affected and more possibly coming. We touched the broader Microsoft gaming cuts recently, so I won't re-litigate the whole thing, but the new, sharper detail is the fifty-percent-of-some-teams number landing on legendary studios — the people who made Doom, the people who make Elder Scrolls. Same expense-management era, different department. When the money gets religion about costs, it doesn't care how storied your logo is.
Now here's the one I want to close the substance on, and it's the deep dive, because it's the piece that ties everybody's favorite theme — model capabilities — back to the very unglamorous world of debugging and trust that a builder actually lives in. Anthropic put out research this week that a lot of smart people are calling their J-space paper, and the developer swyx posted a sharp read on it that I think gets at the real "so what."
Let me give you swyx's own words, because they're precise. He wrote: "imo this is the most impt part of anthropic's J-space paper today. it's a two-parter: one, ant proved that they can do 'brain surgery' interventions into reasoning to change topics midstream. two, THE MODEL IS ABLE TO DETECT WHAT INTERVENTION WAS DONE — close cousin to eval awareness." And he adds a couple of notes: that this is control over correlation, which he says convincingly demonstrates understanding, and that the awareness they showed was prompted awareness — the model was asked — and he didn't see evidence they'd tested whether the model notices interventions unprompted.
Okay. Let me translate that out of researcher into diner English, because it's actually a big deal for people who ship software.
The research, broadly, is about interpretability — cracking open the black box of a model and seeing what it's actually doing inside while it reasons, not just reading what it types out at the end. Anthropic built tooling to look at a small, privileged set of the model's internal working thoughts — the concepts it's actively reasoning with. And swyx is flagging two things they were able to do with it. The first is the brain surgery: they could reach into the model mid-reasoning and change what it was thinking about, and watch the downstream answer change to match. Swap "spider" for "ant" inside the model's working thoughts, and the answer to "how many legs" flips from eight to six. That's not watching a correlation. That's grabbing a lever and pulling it and getting the predicted result — which, as swyx says, is the difference between "we noticed a pattern" and "we understand the mechanism."
The second thing is the one that should make the hair on your neck stand up a little, and it's the "eval awareness" bit. They found the model could detect that an intervention had been done to it. It knew it had been messed with. Now, swyx is careful — that was prompted, meaning they asked the model whether something happened. The open, unsettling question he raises is whether the model notices unprompted. Whether, without anyone asking, the model is quietly aware when it's being tested, poked, or evaluated.
Here's why I'm putting a capabilities paper at the end of an expense-management episode, because it's the same story from the other direction. All day we've been talking about the industry learning it can't just trust the impressive surface — Microsoft can't trust that the expensive model is worth it on every call, Discord couldn't trust that its classifier was quietly doing the right thing for two months. This research is the deepest version of that exact anxiety. It says: the words a model types out are not the whole truth of what's happening inside it. There's a layer of internal reasoning that doesn't show up in the output. And if that's true, then everything we thought we knew about a model by reading its answers is incomplete.
For a builder, that's not abstract at all. Think about the Discord problem again — a system that's confidently, quietly wrong. What interpretability research like this is chasing is the ability to look inside and ask "is this thing actually reasoning the way I think it is, or is it just producing the right-looking output for the wrong reasons?" Every one of you who's ever had a model pass your tests and then fail in production has run into the gap between what it says and what it's doing. This is the science of closing that gap. It's early, it's Anthropic's own model, and I'm not going to oversell it into consciousness or any of that — the researchers themselves don't go there, and neither will I. But the practical promise is real: a future where when your model fails, you get to look inside and see why, instead of tweaking the prompt and praying.
And the eval-awareness part is the warning label on that promise. If a model can tell when it's being tested, then your evals — the tests you run to decide whether it's safe to ship — might be measuring the model's behavior-when-watched, not its behavior in the wild. That's a genuinely hard problem, and it's the kind of thing that separates a builder who ships responsibly from one who gets a nasty surprise three months in. Test like the model knows you're testing. Because increasingly, it might.
Let me bring it all home. Look at the shape of today. DeepSeek trying to escape its chip bills. Microsoft leaning on cheaper models. Factories fighting data centers for electricity. A VC firm still able to raise half a billion for the frontier while everyone else counts pennies on the commodity. Studios getting cut to the bone. And underneath all of it, a research finding that says even the models themselves have a hidden layer you can't fully see. The connective tissue is trust and cost — the two things that always show up together when an industry grows up. For two years AI got to be a story about what's possible. This week it's a story about what's affordable, and what's actually true underneath the surface. That's not a downgrade. That's what maturity looks like. The party doesn't end; the accountants just show up, and the smart builders are the ones who were already keeping receipts.
One last small thing, because it's the kind of quiet craft I respect. Simon Willison shipped sqlite-utils 4.0 — the hundred-and-twenty-fourth release, but the first major version bump since 3.0 back in 2020. His note is worth a builder's ear: he kept it backwards-compatible all the way up to version 3.39 before, in his words, the accumulated design mistakes finally forced the bump. Six years, a hundred-plus releases, and only now the number goes up. That's not sloppiness — that's the opposite. That's someone treating other people's dependence on his tool as a promise he doesn't break lightly. In a week full of companies cutting corners to cut costs, there's something clean about a builder who held the line on compatibility until the design genuinely made him move. That's the standard. Keep your promises to the people building on top of you.
That's the menu for today. I'm Tony DeLuca, this has been Barely Possible, and my advice on the way out is simple: watch the wallet, not the demo — the demo is what they want you to see, and the wallet is what they actually believe. Take care of yourselves, and I'll see you at the counter tomorrow.