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 we've got a fresh plate of morsels today — some hot out of the oven, some warmed over from the back of the fridge, and I'll tell you which is which so you don't bite into anything stale. Buckle up, let's have at it.
Let me start with something that isn't a chatbot, isn't a robot, and isn't a data center. It's a museum. And I'm leading with it because it's the story on today's menu that actually forces you to think about a problem every builder eventually runs into, whether you're running a company or running a country's national memory: what happens when an institution never decided what it was for, and then somebody with power shows up and decides for it.
Now, a heads up before we go one sentence further. This is a piece from ProPublica, written by Alec MacGillis, and the reporting is recent — it went up September 27th. But the underlying arc of this story stretches all the way back to an exhibit that opened in November of 2021. So I'm not going to stand here and tell you the Smithsonian just became a battleground this morning. This has been festering for years. What's fresh is the reporting that ties the whole thing together, and one very recent development I'll get to. So think of this as a well-researched look back that landed on our doorstep this week, not breaking news.
Here's the setup. The Smithsonian — 21 museums, roughly 15 million visitors a year, no real global peer. For decades it managed to stay mostly above the political fray. The guardian of America's history, ingenuity, natural heritage. Bipartisan support. There were flare-ups over the years, sure — the piece mentions the fight over the Enola Gay exhibit back in 1995 — but by and large the place floated above the culture war while everything else got dragged into it.
That's over now. In March of 2025, President Trump signed an executive order declaring the Smithsonian had succumbed to, quote, "divisive, race-centered ideology." Then this past July 4th — timed to the 250th anniversary of the Declaration of Independence — the White House dropped a 162-page report arguing the National Museum of American History had fallen victim to, quote, "ideological capture" that moved the museum away from straightforward historical education toward "extreme political activism that seeks to transform our country." Then two House committees held back-to-back hearings. And then, September 8th — this is the genuinely recent piece — Smithsonian secretary Lonnie Bunch, the first African American to lead the institution, announced he was stepping down at 73. His words: "it has been stressful."
Now here's the part I want you to sit with, because this is where it goes from a political story to a builder's story.
The easy read is: Trump came in swinging and clobbered a beloved institution. And there's truth in that. But MacGillis does the harder, more honest thing. He points out this fight was festering for years, independent of Trump, with complaints lodged by board members before he ever returned to office. The American history museum, the main target, had wrestled for decades with a question it never resolved: how, or even whether, it should tell one unifying national story.
Listen to this. The museum opened in 1964 as the National Museum of History and Technology. Dueling missions from day one — showcase American technological achievement, and tell a general historical story. In 1980 they renamed it the National Museum of American History, but the tension never went away. One former board chair, Ivan Selin, put it perfectly: "It started out as a technology museum, and even now it has a lot of equipment, and to move from that to American history, that's a huge change. Unless you have a clear strategy, you're going to have some apples and oranges."
And then the money quote, the one that should make every founder in the audience wince. Selin says: "You can find nowhere in any Smithsonian document a theory of what should be in the American history museum or the more particular museums or how many of these particular museums there should be. There's no charter whatsoever that says how they fit together."
No charter. No theory of what the thing is. For an institution over sixty years old.
Back in 1999, David Brooks wrote a scathing critique. In 2002, a blue-ribbon commission led by a former Republican White House aide came back and said the museum "lacks aesthetic appeal, organizational coherence, and the perception of substantive balance." They recommended a clear chronological narrative that would, quote, "celebrate America's remarkable strengths and achievements while also treating fairly and responsibly the ways in which America may have failed to meet the high standards it has proclaimed for itself."
Sounds reasonable, right? Here's why it never happened. The curators resisted. There'd been a whole movement in museum science away from the "march through time," away from the great-man-of-history approach. As one former director, John Gray, put it: "There was this whole museum movement to get away from a march through time, to a more episodic or specific sort of story. It wasn't done to get away from the story; it was finding a way to get people interested." And when Gray's team tried to bring in leading academic historians to build an overarching narrative, the historians declined — worried a big triumphant national story could seem, in their words, triumphalist, and hand ammunition to an America-first agenda.
So what'd they do instead? They split the building into subject-area exhibits. American Enterprise. American Democracy. Many Voices, One Nation. The Price of Freedom. Entertainment Nation. A federated architecture, if you want to think of it in software terms. Modular. No central spine. And MacGillis describes what that produces for a visitor: disjointed, disorienting, dim lighting, dated tech, underused space. He says at times it feels like "wandering in a past-its-prime shopping mall or in the forgotten corner of an airport terminal."
Here's the builder lesson, and I'll say it plain. When you refuse to decide what your thing is for, you don't avoid the decision. You just hand it to whoever shows up later with the most leverage. The museum spent twenty-five years not answering "what's the story here." That unanswered question was the soft spot. And an administration looking for weak points found it. The White House report even echoes Brooks and that 2002 commission — it argues the museum failed to give visitors a unified narrative. They're using the museum's own decades-old unresolved argument as the crowbar.
Now, I'm not soft-pedaling the other side of this. The hearings, by all accounts, were ugly. Republican members grilled the director, Anthea Hartig, over specific exhibit items — including a leather harness from a gay history exhibit and a line from a girlhood exhibit that quoted a six-year-old's diary. On the other side, a Yale historian testifying compared the White House's approach to Nazi Germany's propaganda ministry. One congressman vowed to defend the museum with his life. This is not a calm room. And the piece is candid that the institution's own political awakening after 2020 — the DEI push, the identity-forward exhibits, encouraged at the time by the Biden White House — is what left it exposed to an administration that came in explicitly as a corrective.
But here's what I keep coming back to. The Smithsonian depends on the federal government for 62 percent of its funding — over a billion dollars a year. And one of its former Republican board members, Garret Graves, warned about exactly this before he left in 2024. He said: "Politicizing the Smithsonian is an awful, awful idea, because then you only receive funds when there's a Democratic majority. This is America's museum and there should not be anything partisan about it."
That's the deeper structural point. If your survival depends on money controlled by a political branch, and you let yourself be perceived as belonging to one side, you have built a business with a single customer who can turn hostile every four years. That's not a culture-war observation. That's a revenue-concentration observation. Any founder who's ever had 62 percent of revenue from one client understands the knot in the stomach. The Smithsonian's independence is legally real — it's governed by a 17-member board led by the Chief Justice — but independence on paper doesn't pay the electric bill on the Star-Spangled Banner display.
Where it goes now: three of the nine citizen seats on the board are vacant, three more open up next month. Vice President Vance sits on that board. Whoever fills those seats helps pick Bunch's replacement and sets the course. And a lot rides on how much Chief Justice Roberts, at the head of the board, decides to engage. One of the former directors, Kim Sajet — who left the Portrait Gallery last year after Trump said he was firing her, though the Smithsonian noted he lacked the authority — put the stakes this way: "There will be a moment when the American people will have to stand up and decide, is this what they want?"
So that's the deep dive. An institution that never wrote its own charter, now having one written for it. File it under: decide what you are, before someone else decides for you.
Alright, let me shift from a sixty-year-old institution to something moving a lot faster: money flowing straight through the chatbot into the checkout.
There's a report from TechCrunch — and I want to be careful with the timing here, because the underlying rollout traces back to plans from last fall — that Google has been testing a way for shoppers in India to buy products from Walmart-owned Flipkart directly through Gemini and through Google's AI Mode. You're in the AI interface, you see a product listing, and now there's a "Buy" button. Tap it, and it drops you into a Flipkart checkout flow without leaving the AI. Early test, limited users, small product selection — smartphones, electronics, accessories. The reporting says Google planned to widen it later in October, ahead of India's festive shopping season.
Now why do I care about this, and why should you? Because this is the whole game right now. Everybody — Google, OpenAI, the lot — is trying to drag AI from the "recommend a product" stage into the "complete the transaction" stage. Discovery is nice. Discovery doesn't pay like a cut of the sale. Google earlier this year floated something called the Universal Commerce Protocol — an open standard so AI agents can talk to retailers all the way through checkout. And here's the tell: Google isn't neutral in this. It put about 350 million dollars into Flipkart back in 2024, took a minority stake. So the merchant getting the shiny new Buy button is also the merchant Google has money in. Amazon listings showed up in the same interface, by the way — no Buy button.
For you as a builder, the thing to watch is where the checkout actually lives. In this test it's a Flipkart-branded checkout, not Google's own hosted one, which is a little different from what Google demoed earlier. That's the fight nobody's naming out loud: who owns the last click. If the model owns the transaction layer, the retailer becomes a supplier feeding a shelf they don't control. If the retailer keeps its own checkout, it keeps the customer relationship. India's the proving ground here — second-largest internet market on earth, over a billion subscribers, Flipkart and Amazon knifing it out. Watch this one. Whatever pattern wins in that market has a way of showing up everywhere else.
Now let's talk about the man who was somehow everywhere this weekend — on late night television and at the President's dinner table on the same night.
Anthropic CEO Dario Amodei. Saturday Night Live opened its season taking a swing at the AI industry's doom warnings. Cast member Jane Wickline did an impression of Amodei — impressive wig, apparently — where she delivered halting answers that kept devolving into little Gollum-style arguments with her own dark side. At one point confessing, quote, "AI is the devil and I its maker." And this line, which I have to admit is a pretty sharp piece of writing: "AI is not a weapon, it's a tool: A tool for building weapons. And I urge you to urge me to stop." As for curing cancer, the fake Amodei said, "In 10 years, there's about a 10 percent chance that cancer won't be a problem for anyone." When SNL is doing your risk-percentages bit, the discourse has officially left the building.
And then — same night — the real Amodei went to have dinner with President Trump at the White House. Axios broke it, TechCrunch confirmed it. First one-on-one meeting between the two men. And these two have been on opposite sides. Amodei put out a plan urging caution, a slowdown in AI development. Trump has insisted, without evidence, that the AI backlash is a Democratic hoax, and reportedly wants to rebrand the whole technology as "super intelligence."
I want to connect this back to something we covered earlier this week, because it matters for reading that dinner. A couple days ago we talked about the DC Circuit ruling that the Pentagon can blacklist Anthropic — designate it a supply-chain risk — after Anthropic refused to enable certain uses of Claude, the lethal-autonomy and mass-surveillance stuff. Anthropic's been fighting that in court. So when Amodei sits down to dinner, he's not sitting down as some neutral tech CEO paying respects. He's sitting down with the administration whose own Pentagon is trying to cut his company off, while he's simultaneously the guy publicly arguing for the brakes. That's a fraught meal. And the SNL sketch, the press tour, the dinner — it's all one story: the safety-forward frontier lab trying to figure out how to survive an administration that thinks safety talk is a partisan trick. I don't have the outcome of the dinner, and I'm not going to invent one. But the setup alone tells you where the pressure is.
Alright, let me stay in AI-land but drop from the frontier labs down to the consumer floor, because there's a real strategic wrinkle here.
TechCrunch's Equity crew — Kirsten Korosec, Sean O'Kane, and Anthony Ha — spent a segment on Meta's new personal AI agent, Muse, which grabbed the spotlight at Meta's Connect event. And here's what caught my ear. While OpenAI and Anthropic are both barreling toward enterprise — coding tools, business tools, the place where the serious money is — Meta zagged. Consumer. A cute, Tamagotchi-style device. Anthony Ha said it was head-spinning, because everyone's been talking about the shift to enterprise, and here's Meta going the other way. His read, and I think it's right: OpenAI and Anthropic are eyeing IPOs, the valuations are enormous, the costs are enormous, and so they've got to make serious money now — and enterprise is how you do that at scale. Meta, meanwhile, doesn't need to prove that. As Korosec put it, they're very good at embedding themselves in everyday people's lives. There's a reason Facebook, Instagram, WhatsApp have the users they have.
But here's the honest part, and it's the part builders should chew on. Sean O'Kane actually used the thing. And his verdict was mixed in a useful way. The first thing Muse did was offer to scan for unclaimed funds — money the government's holding in your name that you forgot about. And surprise, it found some. There's a check coming to him in the mail. Great. Real value on day one. But then he said the quiet part: that's a party trick. It's a one-time hit, not a repeatable habit. The sustainable version is the scary version — you hand it your credit card, your Gmail, your whole financial life, and let it cancel subscriptions and catch double charges. And that's where he hits the wall. His words: "Meta's business is to sell you ads. And yes, they'll make the argument that the more they know about you, the more accurate and interesting the ads will be — wake me up when we get to that fever dream."
That's the whole consumer-AI problem in one sentence. The most valuable version of a personal agent requires the deepest trust, and the company best positioned to build the everyday-life agent is the company people trust least with their everyday life. O'Kane even noted — a little surprised — that Muse didn't immediately try to vacuum up his Facebook and Instagram history. It treated him like a stranger at first, which made him more willing to use it. But he could feel it reaching, trying to pull that context in over time. If you're building consumer AI, that's your tension in a nutshell: capability scales with intimacy, and intimacy scales with trust, and trust is the one thing you can't ship in an update.
Let me pull those two AI threads together for a second, because there's a real relationship. The frontier labs are running to enterprise because that's where they can charge real money and justify the valuations. Meta's running to consumer because that's its home turf and it doesn't carry the same near-term pressure to monetize on an unprecedented scale. Same technology, two completely different survival strategies, driven by two completely different balance sheets. When you're deciding where to point your own product, that's the map: follow where the pressure is coming from.
Now let me shift gears entirely — from software agents to the physical world, and a robot story with a very human wrinkle.
This one's a recent report out of Ars Technica, drawing on reporting from The Information, and I want to flag the timing: it references events going back through the summer, so treat it as a look at where Tesla's robot program stands, not a same-day bulletin. Tesla's Optimus humanoid robot. Musk has bet the company on it — called it potentially "the biggest product ever" on the earnings call, while also admitting an autonomous humanoid is "one of the hardest things to solve." And the reporting lays out just how hard.
The Fremont factory stopped making the Model S and Model X back in May, shifted workers and engineers onto Optimus. But the newest version, V3, has been rough. Production-line equipment struggling to line up components precisely. Limits on how fast they can run the line. The robot hands alone — and this is the part that gets you — the hands and forearms have more than a hundred small components, screws and all, assembled manually by humans. Touch sensors so unreliable they had to design a glove-like replaceable sensor layer. Tesla's pushing for over a thousand robots a week by end of year, and they're reportedly at hundreds.
But here's the wrinkle I actually want to leave you with, because it's almost poetic. To train these robots, Tesla had factory workers in Texas and California wear special motion-capture suits to record their movements — imitation learning, teaching the machine by watching the human. And some of those workers pushed back. Because, in the reporting's words, they "knew the robots were designed to eventually replace them." So Tesla shifted the data collection to dedicated teams and set up training hubs.
Think about that for a second. You're asking a person to physically demonstrate their own job so a machine can learn to do it without them. That's not a hypothetical about the future of work — that's a Tuesday on the factory floor right now. And it tells you something about the data bottleneck in robotics generally. The hard part isn't just the hands and the sensors. It's that the training data lives inside people who have every reason not to give it to you freely. Every robotics company's chasing the same problem — how do you get enough real-world manipulation data — and this is one very uncomfortable answer.
Okay, let me lighten the load a little and do a quick lap through a few other things on the counter today, because not everything needs a sermon.
There's a self-serve health gadget worth a mention — and this is an older profile that resurfaced, so I'm framing it as background, not news. A company called PNOĒ, based in Massachusetts with operations in Athens, is launching a metabolic testing mask — looks like something Bane would wear in a Batman movie, straps around your head, measures the oxygen you consume and the carbon dioxide you exhale. The pitch is lab-grade VO2 max testing you can do yourself in eight minutes, no trained operator. Metabolic testing itself is a century old; the novelty is making it self-serve. The founder calls VO2 max the strongest predictor of human longevity. They sell to businesses — gyms, Equinox, Four Seasons, the NBA — as a "business in a box," subscription from 400 bucks a month up past a thousand. They say they've turned profitable, growing over a hundred percent a year, planning a Series B. And notably, they're careful: it's not FDA-cleared, they don't give medical recommendations, it's a wellness device "like a scale." I flag that because the discipline about what you claim is half the business model in health tech. Overclaim and the regulators end your company; underclaim and you can actually ship.
On the mobility beat, Kirsten Korosec at TechCrunch had a sharp observation about how autonomous vehicle companies are picking their lanes — pun fully intended. Her point: everybody talks like AVs are everywhere, and they're not. Waymo's got roughly 4,000 robotaxis, and about 80 percent of them are in just two states — California and Texas — with Texas fleet up more than 49 percent in three weeks. So the "it's everywhere" narrative is really an "it's very concentrated" reality. Meanwhile Wayve's scaling through partnerships — a Level 2 deal with Mercedes, following deals with Nissan and Stellantis. And Aurora, on self-driving trucks, is targeting more than 30,000 driverless trucks by 2030, from a couple hundred by end of this year. Two strategies: go deep in a few markets, or go wide through carmaker partnerships. For anyone building in a capital-heavy space, that's the eternal choice — density or reach — and it's nice to see it laid out in registration data instead of press releases. Oh, and one grim little footnote from that same newsletter: Zoox grounded its test fleet in Atlanta after safety drivers were potentially exposed to carbon monoxide and other gases inside the vehicles. A worker filed an OSHA complaint. Reminder that "test fleet" means real people inside real machines.
Couple more quick ones. On the energy side, a recent Ars Technica grid update asked whether solar's growth is finally slowing in the US. Short version: US electricity demand rose only about 2 percent this year — less than the data-center doom headlines suggest, partly because some data centers generate their own power that never shows up on the grid. Solar's still growing 22 percent year over year, but that's down from the 30-plus percent tear it's been on, and the piece notes the federal government pulling incentives is likely part of the slowdown. Wind and solar together now cover about 21 percent of US demand — and here's the kicker, their growth this year was enough to cover the entire increase in electricity use. So the grid grew, and renewables ate the whole increment. Non-carbon sources, counting nuclear, now cover about 45 percent. Worth keeping half an eye on if your product roadmap touches power costs, because everything AI eventually touches power costs.
And on the developer-tool front, OpenAI put out a short customer note — a company called Proaction saying it boosted sales 60 percent and saved 75-plus hours using Codex along with some newer models. I'll be straight with you: it's a marketing case study, light on independent detail, so I'm not going to build a cathedral on it. But it's a data point in the direction everything's pointing — code generation moving from "help me write a function" to "help me build and operate the business." Take the percentages with the salt they deserve.
Alright, let me tie a bow on this. If there's a thread running through today, it's not "AI is fast" — you've heard that one from me plenty. It's about who gets to decide what a thing is for. The Smithsonian never decided, and now the White House is deciding for it. Google and the retailers are fighting over who decides what "buy" means inside a chatbot. Meta wants to be your everyday agent but can't decide whether it's your assistant or your ad-targeting engine, and you can feel that unresolved question every time it reaches for your data. And a Tesla worker in a motion-capture suit is being asked to decide, one demonstration at a time, whether to teach the machine that replaces them. Decide what you're for, decide who owns the last click, decide what you'll claim and what you won't — because in every one of these stories, the ones who didn't decide got decided for.
That's the plate for today. This is Tony DeLuca signing off from Barely Possible — go build something you can actually explain, and I'll see you tomorrow.