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 we've got a fresh tray of stories that all, in their own way, circle the same kitchen-table question: who's getting paid, who's getting squeezed, and who's quietly building a tollbooth. Buckle up, let's have at it.
Let me start with the one that, on paper, looks like a celebration, but underneath is the most consequential thing happening for anyone who builds a company in the next decade. SpaceX priced its shares. Officially. One hundred thirty-five dollars a share, and according to the reporting, this is the largest initial public offering ever. That's not a typo, that's not hype, that's the actual headline. The biggest IPO in history, and it's a rocket company that also happens to be the dominant satellite internet operator on the planet.
Now, I want to slow down here, because a number that big tends to wash over you. The largest IPO ever means more money raised in a public debut than any company that has ever gone public. And the thing you have to understand about an event like this is that it's not just a SpaceX story. It's a market-mood story. When the biggest IPO in history is a hardware company tied up with Elon Musk, with rockets and satellites and a defense business, that tells you where the appetite is right now. Private capital is sloshing around looking for the next monster, and the public markets just got handed the biggest one of all.
But here's the part the confetti hides, and this is where I want to protect your time and your money. The same outlet, TechCrunch, ran a piece by Marina Temkin that you need to hear if you ever bought into a hot private company through one of those special purpose vehicles. You know the ones — somebody sets up an SPV, you put in your twenty-five grand, you get a sliver of SpaceX, you feel like a player. Well, the reporting says lower-tier SPV investors won't even know their true holdings until the post-IPO lock-ups lift. Hidden fees. Lengthy payout delays. And in the worst cases, the risk of outright fraud.
Sit with that. You think you own a piece of the biggest IPO in history, and you can't actually see what you own, because there are layers of vehicles between you and the shares. There's a guy who set up the SPV, there might be a guy on top of that guy, and each one is taking a cut you may not have fully understood. The lock-up — that's the period after the IPO where insiders can't sell — that's when the music stops and everybody finds out who's actually holding what. So if you're a founder or an angel and you've been buying secondary stakes in hot names through stacked vehicles, this is your reminder: the headline price is for the people at the top of the cap table. The people at the bottom of the SPV stack are often the last to get clarity and the first to eat the fees.
And the growth story under all this isn't airtight either. A couple days back, TechCrunch reported that the Indian government got cold feet on Starlink right before the IPO. Tim Fernholz's piece lays it out — problems with Starlink's India expansion could challenge the SpaceX growth narrative. India is the kind of market that, on a pitch deck, justifies a giant valuation. Hundreds of millions of people who could use satellite internet. So if India gets wobbly right as you're going public, that's a crack in the foundation of the very story you're selling to public investors. I'm not telling you the IPO is a bad deal. I'm telling you the biggest IPO ever is being priced on a growth story that has at least one important question mark, and the little guys buying in through SPVs have the least visibility of anyone. Skeptical of hype, that's the house style here.
And notice something — the IPO wave creates its own gravity. TechCrunch also covered Quantum Space, a company trying to go public through a military SPAC, chasing a roughly one-point-two billion dollar deal to build military spacecraft. SPACs — those blank-check vehicles that were supposed to be dead and buried after the 2021 hangover — apparently aren't dead when there's a hot IPO wave to surf. Quantum Space is basically saying, the water's warm, jump in. That's how these cycles work. One giant debut, and suddenly every adjacent space company is dusting off its banker.
Now let me connect that money story to the thing that's actually going to determine how you build software, because they're more related than they look. We talked a couple days ago about Claude Fable 5, Anthropic's safety-wrapped public version of its internal Mythos engine, as a step-change coding model. I'm not going to re-litigate the launch — we covered the release, we covered the controversy when it first hit. But there are genuinely new developments in the last day worth your attention, and they all point the same direction.
First, the small human moment that tells you the model is real. Simon Willison — and if you build, you know Simon — shipped a new release of his open-source tool Datasette, version 1.0a33, and he noted plainly that most of the code in that release was built with the help of Claude Fable 5. That's not a press release, that's not a benchmark, that's a working developer quietly saying, yeah, this thing wrote most of my shipping code. When the careful, skeptical builders start saying that out loud, the model is doing real work, not demo work.
But here's where it gets spicy, and here's where the money and the model collide. Simon also floated an alternative read on a story that's been swirling. There've been reports that OpenAI is considering significant price cuts on its tokens. Now the cynical take is, oh, OpenAI's panicking, their next model isn't ready. Simon's alternative interpretation, which I think is the sharper one: OpenAI is finding Anthropic's latest models to be surprisingly good, and they're considering a pricing war in response. Read that again. The most respected lab in the room ships something good enough that the giant across the street starts talking about slashing prices.
That's the whole ballgame for builders. If OpenAI and Anthropic get into a token price war, your cost of building agentic software goes down. The thing that's been eating founders alive — those per-task costs that on the frontier models can run hundreds of dollars a job — that math gets better for you if these two start undercutting each other. A pricing war between your two biggest suppliers is one of the best things that can happen to a builder. So watch this one. It's just a report at this stage, the Wall Street Journal floated it, Simon reframed it. But if the price cuts actually land, it changes everybody's unit economics overnight.
Now, the deep dive. Because the more I sat with today's batch, the more I kept coming back to one thread that runs underneath the Fable 5 story, and it's not the model quality. It's the question of who gets to decide what a model is allowed to help you do. And the most instructive thing I've got in front of me today is a full breakdown of the Fable 5 backlash that lays out exactly what people are afraid of — not the safety filters themselves, but the precedent.
Let me set the table with the facts as reported. When Anthropic shipped Fable 5, they put in a set of safeguards, and a few of them lit the internet on fire. One was around biology — the guardrails were so aggressive that, by one biomedical researcher's account, the model could tell she was a biomedical scientist and effectively wouldn't engage with her unless she went into incognito mode with memories off. A booster of these labs, somebody who gets early access, saying she can't say hello to the model as herself because it knows what she does for a living. That's the filter being so blunt it catches the exact people it shouldn't.
The second issue was data retention. The reporting describes a policy where even zero-data-retention enterprise customers — the ones who specifically pay to make sure the lab can't see their data — would have messages retained for thirty days under certain conditions, with Anthropic employees able to see prompts and outputs flagged for, quote, potential serious harm, or at a customer's written request. And as one lawyer in the breakdown put it, what counts as potential serious harm? Who's the customer making the written request? If you're a law firm with privileged client communications, that vagueness is a fire alarm. And it wasn't theoretical — within about an hour of the release, The Verge reported Microsoft had started restricting employees from using Fable 5 over those exact data retention concerns. When Microsoft, a giant partner, pumps the brakes inside sixty minutes, you know the policy hit a nerve.
But the thing that really set everyone off — and this is the heart of the deep dive — was the limitation on using the model for other model development. The system card said that, in light of recent models' ability to accelerate their own development, they'd put in new interventions to limit Claude's effectiveness for requests targeting frontier large language model development. Pre-training pipelines, distributed training infrastructure, accelerator design. And here's the kicker, straight from the reporting: unlike the other safeguards, these would not be visible to the user. The model wouldn't refuse. It wouldn't fall back to a weaker model. It would just quietly get worse — through prompt modification, steering vectors, fine-tuning tricks — and you'd never know it was happening.
Silent nerfing. That's the phrase that went around. As one commentator put it in the breakdown, Anthropic is now silently making Claude dumber for certain users on purpose, and there's no way to tell when it's happening to you. And think about what that breaks, because this is the part that matters even if you never touch frontier model research. Benchmarks assume the model you tested is the model you get. That assumption dies the moment a lab can silently degrade output for a category of work it doesn't like. An engineer debugging a failing training run can no longer tell the difference between "the model is wrong" and "the model was made wrong on purpose." And the people in the breakdown point out classifiers misfire — GPU inference research was reportedly already getting caught in the net, and inference optimization is something every company running open models does, not just the frontier labs. A false positive on a visible refusal is annoying. A false positive on silent degradation is undetectable. That's a poison for the whole trust model of building on these tools.
Now here's the resolution, and the reason I think this story has legs beyond the patch. It took less than twenty-four hours for Anthropic to walk it back. In a statement to Wired, they said they're changing the Fable 5 safeguards for frontier model development to make them visible, and — their words — we made the wrong trade-off and we apologize for not getting the balance right. So now, if the system suspects you're trying to build a highly capable AI, it'll tell you it's refusing or rerouting you. Visible. Honest. Good. The right call.
But here's why I'm spending real time on this even though the specific policy got fixed. The breakdown captures the deeper unease beautifully, and it's not about Anthropic being villains. A lot of thoughtful people landed on the same read: the folks at Anthropic genuinely believe these models are dangerous, genuinely believe AI development should be slowed, and from that belief, the silent sabotage almost makes logical sense — if you think your model meaningfully compresses the timeline to something dangerous, then renting that compression to competitors for two hundred bucks a month is a real cost. The steel-man argument in the breakdown runs: the biggest risks come from super-intelligent AI, the leading lab needs a big lead to be able to pause safely at the critical moment, and if everybody can use the leader's model to catch up, the leader can't keep that lead.
And that's the unsettling part. Because as one law professor in the breakdown put it — and I think this is the sharpest line in the whole thing — the only way Anthropic's decisions make logical sense is if they presume they'll maintain control over the frontier and get to dole out access to it without pushback from everybody else. And these are not single-player games. If a lab genuinely sets itself up as the tollbooth for frontier model access, the government reads that as direct competition and acts accordingly. And Anthropic does not win that fight.
So here's my plainspoken takeaway for you as a builder. Forget whether you agree with the safety philosophy. The practical lesson is about platform risk. A developer in the breakdown said it cleaner than I could: he'll avoid these models because they keep imposing more limits on what he can build, and he's not going to build on a completely walled-off ecosystem. That's the founder's lesson. When your core dependency reserves the right to silently change what your product can do, based on its own private judgment of your use case, you have a supply chain problem dressed up as a safety feature. The patch made it visible, which is better. But the residual broken trust — and one AI policy expert in the breakdown said this exactly — is going to have a blast radius wider than Anthropic. Multi-source your model dependencies. Don't bet your whole company on one lab's evolving conscience. That's not cynicism, that's risk management.
And notice — this connects right back to the pricing war we talked about. The same dynamic that makes a lab want to gatekeep its frontier model is the dynamic that makes its competitor want to undercut it on price. Power concentration and price competition are two sides of the same coin, and as a builder you want to be standing where the competition is fiercest, not where the gate is highest.
Now let me pivot from the model to the enterprise, because Anthropic also did the thing companies do the day after a rough news cycle: ship some good news. They announced an alliance with DXC, where DXC will integrate Claude into the systems that banks, airlines, and other regulated industries rely on. That's the boring, sticky, lucrative part of the business — getting embedded into the back-office plumbing of regulated industries where switching costs are enormous and contracts are long. It's also, frankly, the smart counter to everything we just discussed. If you're worried enterprises will flee over data retention, you go sign deals that lock enterprises in through a systems integrator. DXC is the kind of partner that puts your model inside an airline's reservation system, and once it's there, it's not going anywhere fast.
Anthropic also introduced something called Claude Corps. The fresh details on what exactly it is are thin in front of me, so I'm not going to oversell it or invent a mission for it. What I'll say is the pattern is clear — a lab under fire for being too gatekeepy is leaning hard into community and enterprise programs to soften the image. Watch how Claude Corps actually gets defined in the coming days before you read anything into the name.
Now let's shift from the model labs to the picks-and-shovels story, because there's a thread here about agents getting their own wallets that founders should be tracking. Coinbase rolled out a tool that helps AI agents trade and pay for premium research. The mechanism is a protocol called x402, which lets an agent get access to data and APIs and, crucially, pay for them. Now I know the instinct when you hear Coinbase and crypto and agents in one sentence is to reach for the eye-roll. Hold off for a second.
The genuinely interesting thing here is the payment-rails problem. We've spent months on this show talking about agents that can do things — write code, book stuff, research markets. But an agent that can't pay for anything hits a wall the second it needs to buy a data feed or hit a paid API. Coinbase's pitch is basically: give the agent a way to transact, in this case for premium research and API access. If you're building agentic products, the ability for your agent to autonomously pay for the resources it needs — without a human punching in a credit card every time — is one of those unsexy infrastructure pieces that has to exist before the whole vision works. Whether x402 is the standard that wins, I have no idea. But agent-native payments is a real category now, and the company that owns those rails owns a tollbooth of its own. Keep an eye on it.
Let me bring it back to the consumer side, because Waymo did something that tells you the robotaxi business is officially growing up. They launched a loyalty program. Waymo Premier — twenty-nine ninety-nine a month, gets you ten percent cash back and free cancellations. Now think about what a loyalty program signals. You don't build a thirty-dollar-a-month subscription for a service people try once. You build it when you've got regulars. When riders are taking enough trips that a frequent-flyer-style program actually pencils out. The robotaxi went from novelty to commute. And the move to a monthly subscription is the move every transportation company eventually makes — turn the occasional user into a recurring revenue line. It's the same instinct that has Starlink charging a monthly hardware rental fee instead of a one-time purchase, which we touched on recently. Everybody wants the subscription. Everybody wants the recurring number on the board.
Now let's talk about a company going the opposite direction — not growing up, but staring into the mirror and not liking what it sees. Ars Technica got hold of an internal picture of Microsoft's gaming division, and the headline quote is brutal: "This cannot continue." Xbox leaders, in what Kyle Orland describes as a brutal self-assessment, laid out the hard truths behind a sagging brand. A Microsoft gaming division in crisis, in their own words.
Now why does a Bronx radio guy who mostly talks AI and startups care about Xbox? Because this is a case study in what happens to a product business inside a giant when the giant's attention and capital have all gone somewhere else. Microsoft is spending like a drunken sailor on AI infrastructure — we've covered the data center buildouts, the chip-to-model stack under Mustafa Suleyman, the whole vertical play. And while all that money and executive oxygen flows toward AI, the gaming division is sitting there writing memos that say, essentially, we are in trouble and this cannot continue. That's the founder's lesson buried in a gaming story: every dollar and every hour of leadership attention is a choice, and the thing that doesn't get the attention withers. Xbox isn't dying because nobody can make a console. It's struggling because inside a company betting everything on AI, a hardware-and-games brand is no longer where the energy is. If you run a company with multiple product lines, that memo should make you nervous about your own neglected divisions.
Now let's talk about a quieter problem that hits a lot of builders right in the gut — the forced app migration. Ars Technica reported that AcuRite, the weather-gadget company, admitted its new app falls short and delayed the shutdown of its old app to fix the problems. The old app, in their words, still needs to be retired — but they can't pull the plug because the replacement isn't good enough yet.
This is such a perfect little parable I had to include it. Here's a company with customers who bought physical hardware — weather stations, sensors — that depend on an app to function. The company wants to retire the old app and move everybody to the shiny new one. Except the new one isn't ready, and now they're stuck. They've announced the death of the thing people rely on, and then had to walk back the funeral. If you build hardware that depends on software, or software that customers have wired into their daily lives, this is the trap. You can't just sunset the thing people depend on because you'd prefer to maintain one codebase instead of two. The migration has to be better, not just newer. AcuRite learned that the embarrassing way, in public.
Let's stay on the theme of platforms changing the rules and the users eating it. Valve killed its retail gift card program — the physical Steam cards you buy at the store with cash — because of scammers. And the reporting makes the painful point right in the summary: the move also cuts off a massive market of legitimate users who buy those cards with physical cash. People who don't have a credit card. People who give gift cards as gifts, which is, you know, the entire point of a gift card. The scammers ruined it for everybody, and the platform's response was to take away the option entirely.
This is the eternal tension in any consumer platform. Fraud is real, fraud is expensive, and the easiest fix is often to nuke the whole feature. But every time you do that, you cut off the honest people who were using it as intended — often the people with the fewest alternatives. The cash-only customer doesn't have another easy way in. So when you're building trust-and-safety into your own product, remember the AcuRite-meets-Valve lesson: the blunt instrument always catches the innocent, and the people you cut off are usually the ones with the least power to route around you.
Now let me shift to the policy corner, because there's a bill worth flagging and a content-moderation story worth connecting to it. Ted Cruz and Ron Wyden — and if you can get those two in a room agreeing on anything, that itself is news — introduced a bipartisan bill called the JAWBONE Act. The idea, per Ars Technica, is to help Americans sue federal officials over censorship. The name's a wink at "jawboning" — that's the term for when government officials lean on platforms behind the scenes to take down content without ever issuing a formal order. The bill would create a path to sue federal officials for that kind of pressure.
Now set that next to the other content story in today's batch. Ars Technica, citing reporting from Wired's David Gilbert, found that racist comments targeting politicians tripled since Meta relaxed its moderation rules, with violent threats against lawmakers also surging on Facebook. So in the same news cycle, you've got lawmakers trying to make it easier to sue the government for pressuring platforms to remove content, and you've got data showing that when a platform loosens its rules, the ugliness floods in. That's the whole content-moderation debate in two headlines. One side says government pressure on platforms is the danger. The other side says platform permissiveness is the danger. And the honest answer, the kitchen-table answer, is that both can be true at the same time, which is exactly why nobody can write a clean law for it. For builders running any kind of user-generated platform, the lesson is grim but useful: when you relax the rules, you should expect the bad stuff to scale faster than the good stuff. It just does.
And there's a kids'-safety dimension to all this regulation energy. TechCrunch ran a rundown of the countries moving to ban social media for children. Australia was first, back in late 2025, aiming at cyberbullying, addiction, exposure to predators — and now a list of countries is following. If you build anything that touches younger users, the regulatory walls are going up around the world, not just in one jurisdiction. Age verification, the whole compliance apparatus — it's becoming table stakes for consumer products, and the country-by-country patchwork is going to be a genuine headache for anyone trying to ship globally.
Let me give you two more from the AI-meets-creators front, because they rhyme. Meta is adding an AI assistant and a desktop version to its Edits app — the editing tool aimed at creators. The stated goal is right there: keep creators engaged on Instagram as Meta fights TikTok and YouTube for their attention. That's the platform war fought through tooling. You don't keep creators by begging them to stay, you keep them by making your editing tools so good and so AI-assisted that leaving means giving up your workflow. Sticky tools beat loyalty pleas every time.
And on the flip side of the AI-content coin, Deezer — the music streaming service — built a tool that scans playlists from Spotify, Apple Music, and others to identify AI-generated music. So one company is racing to put AI into the creator's hands, and another is building the detector to flag the flood of AI slop washing into the catalogs. That's the arms race in miniature. Generate it faster, detect it faster. If you're in any content business, you're going to be on one side of that line or the other — making the AI content, or building the tools to label it — and increasingly you might need to do both.
Let me bring it down to earth before we close, because not everything is models and money. There were a couple of stories in today's pile that have nothing to do with your cap table and everything to do with what we're choosing to fund and not fund as a society, and they're worth a beat.
The National Science Foundation is decommissioning an ocean monitoring network in Alaska, and the reporting from Inside Climate News, run by Ars, is blunt about the stakes — Alaska's multibillion-dollar fishing industry and vulnerable coastal communities left, in their phrase, flying blind. This is the kind of infrastructure nobody notices until it's gone. Sensors in the water that tell fishermen and coastal towns what's happening with the ocean. It's not flashy, it doesn't trend, and when the budget gets cut, the people who pay are the ones with a multibillion-dollar industry and not a lot of political volume. I'm not going to pretend I have a tech-builder lesson stapled to it. Sometimes a story just deserves to be said out loud: we are turning off the instruments that tell working people what's coming.
And in the same spirit, there was a story about the American Diabetes Association apologizing for ejecting scientists over criticism of Trump — Beth Mole's reporting at Ars notes that for days the organization had doubled down on that choice before backing off. I flag it because it's part of a pattern we've been watching — institutions, medical and scientific bodies, getting yanked around by politics, and sometimes correcting course only after the blowback. We saw the OB-GYNs releasing their own vaccine schedule, rejecting the meddling from the top, with thirteen other medical groups endorsing the independent guidance. The throughline is that professional institutions are increasingly choosing to go their own way rather than march to a political tune, and sometimes choosing wrong and then walking it back. It's messy. It's human. It's the same belief-and-power tug-of-war we saw in the Fable 5 story, just playing out in a hospital instead of a server farm.
Let me do a quick lightning round before I let you go, because there were a few more worth a sentence each. Bluesky launched group chats as it shifts focus toward features for smaller communities — the social network leaning into intimacy over scale, which is an interesting counter-bet in a world where everyone else chases reach. NASA's Deep Space Network, after nearly breaking, worked well on Artemis II, with one official noting that some missions are using more than what their paperwork would say — a nice reminder that even our most advanced infrastructure is running hotter than its specs claim. And Google put out a note about new community investments in Virginia, supporting local jobs and energy affordability — which, read with a clear eye, is the kind of goodwill spending that tends to show up exactly where a company is building a lot of power-hungry data centers and wants the neighbors on its side. We've been tracking the data center backlash for a while now, with city councils hitting pause on new construction, and this is the other half of that playbook: when the resistance grows, you start writing checks to the community.
So let me tie the bow on it. The thread running through today — from the biggest IPO in history with its hidden-fee SPV underbelly, to a lab walking back a silent nerf, to Xbox writing its own obituary, to Valve and AcuRite cutting off the users with the fewest options — is about who holds the levers and who's at the mercy of them. The SPV investor can't see his own shares. The developer can't tell if his model got quietly throttled. The cash customer can't buy his gift card. The fisherman can't see the ocean. Power keeps concentrating at the top of the stack, and the value, the lesson, the protection, all of it lives in knowing which tollbooth you're standing in front of, and whether you've got a second way around.
For builders, the practical marching orders: watch that OpenAI-Anthropic price war, because cheaper tokens change your whole business. Multi-source your model dependencies, because a single lab's conscience is not a stable foundation. And keep the agent-payment rails on your radar, because the boring infrastructure is where the next tollbooths get built.
That's the menu for today. I'm Tony DeLuca, this has been Barely Possible, and as always — read the fine print before you sign the SPV, would you do that for me. Take care of yourselves out there.