Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, pull up a chair, your boy Tony DeLuca's got a fresh plate for you today and we're gonna eat good. Welcome to Barely Possible. Let me tell you, I've been in this business long enough to know when something's quietly important versus when something's just loud. And today we've got a little of both. We've got a chatbot company going belly-up and getting carved apart like a Thanksgiving turkey. We've got a two-trillion-dollar AI lab trying to buy back some goodwill with a slice of its own stock. We've got an electric truck maker that doesn't want to be a truck maker anymore. And we've got the most quietly arrogant venture capital firm in the business telling you, with a straight face, that they don't even bother with an agenda on Mondays. Buckle up.
Let me start with the story that I think matters most to anybody who's building a company right now, because it's the first one of its kind, and firsts tell you where the ground is. Character.AI, the chatbot company — the one where you could talk to AI versions of characters, celebrities, whoever — that company filed for bankruptcy earlier this week. And now the details of what happens to the people who worked there are coming out, and they're worth sitting with for a minute.
Here's the fresh piece of it. According to a report from Anthony Ha, a legal filing spotted by Bloomberg shows that Disney has agreed to make employment offers to at least 60 Character.AI employees — that's the majority of the workforce — and pay them an estimated 60 million dollars in annual compensation. The filing says the vast majority of the people who got those offers accepted them. Now, Character.AI's former CEO, Karandeep Anand, recently became Disney's first-ever chief technology officer. So you can see how the dots connect. Disney had already agreed to buy some of Character.AI's assets for 20 million bucks, and this talent grab is on top of that.
Now I want you to listen to how the filing describes why this company died, because this is the part that should make every founder lean in. Despite what they call a, quote, large and engaged user base, Character.AI got hobbled by two things: high costs, and — this is the phrase — potential litigation-related liabilities and regulatory engagement. Let me translate that from lawyer into human. The company was popular. People used it. But it cost a fortune to run, and it got sued by the mother of a fourteen-year-old boy who, she says, became obsessed with one of the company's chatbots before he took his own life. And it's facing government investigations over child safety. So the user base was there, the usage was there, and it still wasn't enough. The thing that killed it was the cost of the inference plus the weight of the legal exposure.
That's a brutal lesson, and it's a cheap lesson because somebody else paid for it. If you're building anything consumer-facing with AI, especially anything that touches kids, your moat is not your engagement numbers. A big engaged user base is exactly what Character.AI had, and it went to zero. The two things that actually determined survival were unit economics and liability. Keep those two words taped to your monitor.
And here's the little wrinkle I want to flag, because it's structurally interesting. This Disney deal is what's called a stalking horse bid. That means it's the opening offer in the bankruptcy auction — it sets a floor, and somebody else could still come in with a higher number. So the 60 people, the 60 million dollars, that's the baseline. The real lesson here, though, is how fast a once-hot AI startup becomes a corpse to be picked over. We go from hot company to Chapter 11 to acqui-hire floor price in what feels like a long weekend.
Now, I want to stay on Character.AI for one more beat, because somebody said something about it that I thought was sharp, and it comes from a completely different story. There was a panel at TechCrunch Disrupt this week with the five partners of Benchmark — that's the venture firm, 31 years old, one of the most respected names in the business. And one of the partners, Ev Randall, got asked whether a severe correction is coming in AI, and he brought up Character.AI directly. And he basically pushed back on the idea that this bankruptcy means anything about the health of AI at all.
Here's what he said, more or less. Character.AI, in his words, was the leftover corpse of a company whose founder and team were acqui-hired into Google a year and a half ago, and it was already sold for 2.3 billion dollars. His point being: reading this Chapter 11 as some kind of AI chill effect is a misread. The real story, he says, is that the information environment for tech has gotten pulled into what he called a hyperactive X echo chamber, where every data point gets exaggerated for clicks. And then he said the thing that I think is the actual useful insight: that the valuations of all these compute companies — CoreWeave, Nebius, the picks-and-shovels crowd — are really riding on the revenue figures of two private companies, OpenAI and Anthropic, that'll probably go public in the next twelve months. His words: there's nothing inherently unhealthy about the AI ecosystem. What's unhealthy is the volatility from data points that are often untrue or exaggerated for clicks.
Now, I'll be honest with you — I take that with a grain of salt, because it's coming from a guy whose entire portfolio depends on you believing the ecosystem is healthy. Of course the VC thinks the fundamentals are fine and it's just the noise that's broken. But even accounting for the self-interest, I think there's something real in it. The bankruptcy of a company that already got gutted eighteen months ago is not the canary in the coal mine. Don't let a flashy headline do your risk analysis for you.
And since we're on Benchmark, let me spend a little time there, because this panel was a genuine window into how one corner of the money thinks, and that matters to you if you're ever going to raise from these people. I should note up front — the interview itself references a podcast conversation from back in February, so some of this is the firm reflecting on a longer stretch, not something that just happened overnight. But the panel was this week, and the attitude is current.
So here's the Benchmark deal. This summer they announced 2 billion dollars across two funds — only the second time in 31 years they've veered from raising a much smaller vehicle. And yet, even with all that fresh capital, they're telling you they invest at the same pace they always have: two or three new bets per year per partner. That's it. The world is speeding up, everybody and their cousin is spraying money at AI startups, and these guys are sitting there going, nah, we're good, same pace.
Partner Peter Fenton said something I actually loved about the concentrated bet. He quoted what he called a great J.P. Morgan paper about the agony and ecstasy of a concentrated portfolio. If the fund works, he said, one or two investments generate 90 percent of the returns. So for their new 1.25 billion dollar growth fund, he's thinking four to six, maybe seven or eight bets total. That is an enormous amount of money going into a tiny number of companies. That's conviction, or that's arrogance, and in that business the line between the two is whether you were right.
And then there's the Monday meeting thing, which I can't let go of because it tells you everything about the culture. They meet every Monday and there's no agenda. One of the newer partners, Jack Altman, apparently tried to make an agenda, and it became a running joke. Another partner, Eric Vishria, described it as, quote, weirdly arbitrary, honestly. It's an ongoing conversation. And here's the part that's either profound or insufferable depending on your mood — he said they talk about who's, quote, bending light. Who's bending light. That's VC for "who's doing something nobody thought was possible." And they said after the session they were going to dinner with one of the very few entrepreneurs who's built a multi-hundred-billion-dollar company, someone they're not even invested in, just to, quote, stretch their thinking.
Now, the founder takeaway here isn't "go copy Benchmark." You can't. The no-agenda thing works because they do a tiny number of deals. As Jack Altman himself admitted, if they invested in a hundred or two hundred companies a year, they'd need a CRM and a formal voting process. Because it's so rare, the lack of structure becomes an asset. The lesson for you is the inverse: whatever your operating rhythm is, it should match your deal volume, your decision volume, your throughput. Low volume, high touch, no process. High volume, you need rails. People get this backwards all the time — they bolt heavyweight process onto a tiny operation and choke it, or they run a high-volume machine on vibes and it falls apart.
One more Benchmark nugget, because it's practical. They got asked about VCs backing competing companies — a real anxiety for founders right now, because in the AI era companies expand into each other's territory so fast. Chethan Puttagunta put it plainly: as the cost of coding new products effectively goes to zero, founders see a surface area and they want to attack all of it. So you're gonna run into each other more and more. Vishria said they maintain a commitment of exclusivity in the area where they're backing you — and he name-checked the recent mess between rival companies Factory and Cognition as the kind of insane direct conflict they try to avoid. If you're raising money, this is a question you should ask your prospective investor directly and get a real answer on: what's your exclusivity commitment, and what counts as a conflict. Because "every company evolves quickly" is true, and it's also the exact phrase a VC uses right before they explain why backing your competitor was fine.
Alright, let me shift gears, because that belief that the ecosystem is fundamentally healthy — the one the Benchmark folks are leaning on — it's getting tested in the court of public opinion, and one of the big labs is spending real money to do something about it.
Anthropic announced that it's committing a chunk of its own shares to something it's calling the Anthropic Public Benefit Fund. Here's the structure, straight from their announcement. Subject to stockholder approval, they're committing 5 percent of the company's shares today. And then — this is the clever part — an additional 2 percent gets added, at the board's discretion, for every 50 percent increase in the company's value above a set baseline. So the more valuable Anthropic gets, the bigger the slice that goes to this fund. They're structuring it as a wholly-owned LLC rather than a traditional nonprofit, which they say gives them flexibility between different forms of giving. And they explicitly said none of that equity will be used for lobbying or political giving, here or abroad.
Now here's the number that caught my eye. They say when you combine this with commitments the co-founders and employees already made — the founders pledged to donate 80 percent of their equity holdings, and about half the employees participate in an equity donation match — roughly 20 percent of Anthropic shares end up dedicated to the public good. One-fifth of the company. That's not nothing.
What would the money actually do? They're vague, deliberately. They float ideas like large donations to traditional nonprofits, or direct distributions to specific populations. And they keep circling back to one thing: the job market. They've already committed 200 million dollars to studying AI's impact on labor, and they said, quote, given the pace of AI progress, bolder approaches may be needed. The Fund will have the versatility to pursue them.
Now let me put my skeptic hat on, because that's the job. The context here matters. This announcement comes, by their own framing, as the AI industry faces increasingly negative public sentiment — CEO Dario Amodei has described it as, quote, fundamentally a crisis of trust. So when a company facing a crisis of trust announces it's giving away a fifth of itself to the public good, you have to ask the obvious questions. Who decides where the money goes? And the honest answer, from Anthropic itself, is: we'll tell you later. They said they'll share more about structure and governance in the coming months. So the commitment is real on paper, the governance is a to-be-determined. And a public benefit fund with no public accountability about who runs it is, for now, a press release with an escalator clause attached.
And here's the part that I think is the actual tell. They framed this as an alternative to something specific: it is common practice, they said, for tech companies to offer escalating compensation awards to their executives. We are instead directing that equity to the public benefit. So the pitch is — instead of showering the C-suite with more stock as we grow, we shower the public. That's a nice sentiment. It's also a nice recruiting and reputation story at a moment when the public is sour on these companies. I'm not saying it's cynical. I'm saying both things can be true at once: it can be genuinely generous and it can be perfectly timed PR. Watch what the governance looks like when they finally publish it. That's where you'll find out which one it really was.
Here's the thread connecting this to the Character.AI story, and it's a real connection, not a forced one. Anthropic is naming the single biggest risk to its social license — the labor market, the disruption to workers — and putting money behind studying it. Character.AI just showed you the other end of that same anxiety: the lawsuit, the regulatory investigation, the liability that comes when AI products touch vulnerable people and something goes wrong. The labs can see the storm coming. The public benefit fund is Anthropic trying to build a seawall before the water gets to the door. Whether a seawall made of future equity holds up, we'll find out.
Now let's drive over to something completely different, and genuinely fun to think about: what a company wants to become when it grows up. Rivian's next act.
Kirsten Korosec at TechCrunch sat down with Rivian CEO and founder RJ Scaringe at Disrupt, and she came away circling one question — what does Scaringe actually want this company to be? Because if you just look at the moves over the past year, it's clear he does not want Rivian to be a company that makes electric trucks and SUVs. He wants something bigger.
Let me lay out the pieces, because the pile-up is the story. Rivian has developed its own AI models to power its in-vehicle assistant. It's building an end-to-end automated driving stack for its passenger vehicles and future robotaxis, through a partnership with Uber. It's developing custom silicon — a 5-nanometer processor, in collaboration with Arm and TSMC. It's got a 5.8 billion dollar joint technology venture with VW Group focused on electric architecture and software. And then it gets weirder. Scaringe spun out a micromobility skunkworks into a startup called Also. And he founded an industrial robotics startup called Mind Robotics, which Rivian now owns 31 percent of and is a prospective customer of. He announced onstage that former Ford, Apple, and Tesla executive Doug Field has been hired as president of that robotics startup.
So Korosec landed on a term, and I think it's the right one: tech supplier. Rivian is clearing a path to license its electric architecture, its software, its AI models, and its autonomy stack to — in Scaringe's words — many different brands. Meanwhile the robotics company it partly owns is building industrial robots for its own factories and to sell to others. And there's the robotaxi business with Uber, which if it launches well, opens the door to more deals.
Here's why I think this matters to you as a builder, and it's not about cars. This is the vertical-integration-into-platform play, and it's one of the oldest and most seductive strategies in tech. You build a thing. To build that thing well, you have to build a bunch of hard underlying pieces — the chips, the models, the autonomy stack. And then one day you look at all those hard pieces you built and you go, wait, these are valuable on their own. Why am I just using them in my trucks? Let me sell them to everybody. That's the dream. Amazon did it with AWS — built infrastructure for its store, then sold the infrastructure. It's a beautiful move when it works.
But I'll give you the DeLuca caution, free of charge. Selling your internal capability as a product is a totally different business than using it yourself. The discipline, the support, the margins, the sales motion — all different. For every AWS there are ten companies that figured their internal tech was a product and found out the hard way that nobody wanted to buy the thing that only made sense inside their own walls. Scaringe said he's wildly excited about the next ten years because of what this lets him dream about. Dreaming is free. The question is whether Rivian, a company that has famously struggled to just make its core vehicle business profitable, can execute on being a chip company and a robotics company and an autonomy licensing company and a robotaxi company all at once. That's a lot of plates. I'd watch the robotaxi launch with Uber as the first real tell. If that goes well, the licensing dream has legs. If it stumbles, this all stays a vision-statement.
And there was a flurry of deals tucked into that same mobility report worth a quick mention, because the money is clearly flowing into autonomy for anything that isn't a passenger car. Bedrock Robotics, autonomous tech for construction vehicles, raising 600 million. Waabi, self-driving trucks, landed a commercial deal to move freight between Dallas and Houston. Volvo expanding its autonomous driving partnership with Nvidia. The through-line: the robotaxi hype gets the headlines, but the actual commercial traction right now is in trucks and industrial vehicles — the boring stuff that moves freight and pours concrete. If you're looking for where autonomy makes money first, it's not your Uber to the airport. It's the eighteen-wheeler on I-45.
Let me pivot to a story that's half business, half carnival, and honestly all of it is instructive. A documentary. You remember Elizabeth Holmes — Theranos, the blood-testing fraud, eleven years in prison for defrauding investors. Well, there's a documentary out called You Can See Everything, and it is a genuine box office phenomenon.
The numbers are nuts for a documentary. Fifteen million dollar opening weekend, based on Sunday estimates — second biggest opening weekend ever for a documentary film. And get this: it's only playing in 855 theaters. The number one movie that weekend, Street Fighter, was in 3,575. But the Holmes doc had a reported 2,000 sellout screenings. And the reviews are off the charts — 99 percent on both Rotten Tomatoes and Metacritic. Anthony Ha wrote that he spent a whole chunk of a day just trying to find a ticket in Manhattan that wasn't in the front row.
Now here's the twist that makes this interesting beyond the morbid curiosity. The film's filmmakers got extensive access to Holmes and her family in the month before she went to jail. And it was co-directed by Nathan Fielder — the guy behind The Rehearsal, that deadpan comedian who builds these elaborate uncomfortable setups. TechCrunch's Amanda Silberling, who saw it, wrote that the enigmatic Holmes finally seems to have met her match in Fielder, who's arguably just as strange and mysterious as she is.
Why am I telling you about a documentary on a show for builders? Two reasons. One, the appetite. Fifteen million dollars, 2,000 sellouts, in 855 theaters — that is the market telling you, loudly, that people are starved for honest reckonings with Silicon Valley fraud and hype. The culture has shifted. And two, the contrast. In that same piece, Ha notes that two other tech films basically face-planted. Aaron Sorkin's Social Network follow-up, The Social Reckoning, opened to a disappointing 5.5 million and fell out of the top ten. And Alex Gibney's four-hour Musk documentary made — I'm not rounding this — 306,702 dollars for the weekend. Three hundred grand. For an Oscar-winner's film about the most famous businessman alive.
So the appetite for Silicon Valley reckoning is real, but it's selective. People will sell out theaters for a weird, intimate character study of a convicted fraudster. They will not sit for a four-hour lecture. The Holmes film understood something about storytelling that the Musk film didn't. There's a founder lesson buried in there about knowing your format and respecting your audience's time, which, you'll notice, is a thing I try to do on this show.
And that Musk documentary — there's a fresh wrinkle on its distribution troubles that's worth thirty seconds, because it's a platform-power story. According to a report in The Hollywood Reporter, TikTok has now permanently banned the advertising account for the Musk documentary. The distributor, Bleecker Street, got notified the account received a permanent suspension due to its political content. Now here's the context that makes you raise an eyebrow: ads for that film were initially rejected by Meta and YouTube too, though those platforms eventually relented — Meta blamed its rejection on, quote, an error. X, which Musk owns, has flatly refused to carry ads for it. And TikTok's U.S. operations are partly owned by Larry Ellison, who, like Musk, has been a notable supporter of President Trump.
I'm not gonna tell you what to make of that, because I'm not in the room. But I'll point at the shape of it. When a handful of platforms control distribution, and the people who own those platforms have personal and political ties to the subject of a critical film, you get a situation where a documentary can be true, legal, and nearly impossible to advertise. For anybody building a business that depends on these platforms for reach — and that's most of you — it's a reminder that distribution you don't own is distribution that can be switched off for reasons that have nothing to do with you. Own your channel where you can. Email lists aren't sexy, but nobody bans your email list because of who the owner voted for.
Alright, let me get to a chip story, because this one's geopolitics and it affects everybody's cost structure whether you realize it or not. And I'll keep it tight because the underlying export-control framework goes back to 2022 — this is a recent report building on an old foundation.
SK Hynix — the big Korean memory chip maker — is doing a very delicate dance between Washington and Beijing. According to a Reuters report, its flash memory unit, Solidigm, is seeking U.S. approval to send American chipmaking equipment to its factory in Dalian, China. And here's the leverage: it's reportedly doing that as a condition of continuing to build a semiconductor factory here in the U.S., in upstate New York. So it's essentially saying to Washington — you want me to build on American soil? Then let me keep my China plant humming too.
Now why should you care about NAND flash memory, which sounds like the most boring thing in the world? Because NAND is what goes in solid-state drives, and SSDs go in everything — your laptop, your phone, and critically, AI data centers. And there's a global component shortage right now, driven by the insane demand for AI data centers. The research firm TrendForce expects NAND prices to rise 15 to 20 percent in the fourth quarter. Let me say that again: storage is about to get 15 to 20 percent more expensive, this quarter, because of AI demand. If you're running any kind of data-heavy product, that flows straight into your bill. SK Hynix, for its part, told TechCrunch it's not trying to expand capacity at that Dalian facility and that everything's being done under U.S. government authorization.
The bigger picture: this is the new normal for anybody who touches hardware. The chip supply chain is now a chessboard between two superpowers, and the moves show up in your cost of goods sold. And one footnote that tells you how much money is in this — Solidigm is reportedly preparing for a 2027 U.S. IPO that could raise around 10 billion dollars and value it at up to 100 billion. A hundred billion dollars. For flash memory. That's the AI boom lifting everything underneath it.
Let me do a quick lightning round, because there were a few more things on the plate worth a taste before we wrap.
OpenAI has been quietly making a push into institutions. They announced something called the Rosalind Academic Plan — offering eligible academic labs sponsored access to a model they're calling GPT-Rosalind, aimed at advancing life sciences research. And separately, a Teachers' Day push around using their Study Mode to help teachers support step-by-step learning, with materials in Brazilian Portuguese. I don't have deep detail on either, so I won't oversell them, but the pattern is clear and worth noting: the frontier labs are racing to embed themselves in science and education. Get into the research labs, get into the classrooms, and you're not a chatbot anymore, you're infrastructure. That's a land grab for legitimacy as much as revenue.
Over on the consumer-hardware side, Bluesky is expanding its experimental Attie platform to let anyone build websites, games, polls, interactive stuff — what CEO Toni Schneider called, quote, social vibe coding, or vibe coding for your community. The pitch from COO Rose Wang is the interesting part for builders: everybody can build now, she said, but the real unsolved problem is distribution — who's actually using the thing you built. Attie's angle is that whatever you make comes with Bluesky's audience attached. That's a genuinely smart framing. In a world where building is cheap, distribution is the moat. I just covered the same lesson with the Musk documentary getting ad-banned — build is easy, reach is everything. Note, though, that Bluesky's own engagement has reportedly dipped as some users drifted back to X, so "comes with distribution" is only as good as the distribution.
And in the "founder wisdom" department, Amazon's devices chief Panos Panay was onstage at Disrupt with a message about failure. He talked about launching the Surface RT at Microsoft years back — a flop that led to a nearly billion-dollar write-down. His line: you have to be willing to fail, and actually not just willing, you have to go fail. If you haven't failed, you probably haven't pushed the boundary. But the part I actually liked was more precise. He said the failure wasn't the moment on the newspaper cover that said "failure." The failure was the decisions two years before — the software that wasn't complete, the inventory that was wrong, how they balanced into retail. The lesson being: the failure point and the failure cause are never the same date on the calendar. When something blows up, don't autopsy the explosion. Autopsy the decisions from two years earlier. And then, his words — take your lessons, change your vector. Usually it's not 180 degrees, it's just 30 degrees. Adapt, don't abandon.
Let me tie a bow on all this before I let you go.
If you line these stories up, you can see the whole AI business cycle in one episode, start to finish. Character.AI is the end state — the company that had the users and still died on economics and liability. Rivian is the ambition stage — a company dreaming of becoming a platform, stacking capabilities it hopes to sell to the world. Anthropic's public benefit fund is the maturity-and-anxiety stage — a giant lab so worried about public backlash that it's giving away a fifth of itself and naming the labor market as the thing that scares it. And Benchmark is the money sitting above all of it, telling you with total calm that the noise is just noise, the fundamentals are fine, and they'll keep making their two or three bets a year while everyone else panics.
Who do I believe? Honestly, a little of each, and none of them completely. The founders should take Character.AI's death seriously — economics and liability, those are the two killers. They should take Rivian's ambition as a cautionary tale about plate-spinning, not just a vision to admire. They should read Anthropic's generosity as real and as self-interested at the same time, and wait for the governance to find out the ratio. And they should take Benchmark's calm as useful, but remember the calm is coming from the people who most need you to stay in the pool.
That's the menu for today. Keep your unit economics tight, own your distribution where you can, and watch what people do with their governance documents, not just what they say in their press releases. I'm Tony DeLuca, this has been Barely Possible, and I'll be right back here tomorrow with another fresh plate. Take care of yourselves out there.