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, pour yourself the big coffee, because today we're starting with the one number that should make every founder in my audience sit up straight. Zero. That's the number of dollars of financial payoff that AI has delivered to most of corporate America, according to a guy whose whole job is counting dollars. I'm your boy Tony DeLuca, and this is Barely Possible, the show where we read the technical stuff so you don't have to, and then we tell you who's getting squeezed. Let's have at it.
So let's start with that zero, because it's the story that quietly matters more than anything flashier that crossed my desk today.
Torsten Slok is the chief economist at Apollo Global Management. Apollo is a private equity giant, and if you want the fun footnote, Apollo used to own TechCrunch, which is where this note surfaced. So this is not some anti-tech blogger throwing rocks. This is a money guy, at a firm whose entire reason for existing is to find returns, looking at the biggest companies in America and saying: I don't see it.
Here's how he did the math, and I like this because it's clean. He takes the S&P 500 and he strips out the seven names everybody already knows are riding the AI wave: Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla. What's left he calls the S&P 493. That's the rest of corporate America. The banks, the retailers, the manufacturers, the healthcare companies, the insurers. The actual economy. And when he looks at their profit margins and their revenue per worker, he finds it flat. No bump. No evidence that all this AI spending is turning into productivity.
His words, and I'll quote him because he said it plainly: "Despite massive investments and widespread hype around artificial intelligence, the technology has delivered zero financial payoff to corporate America outside of tech."
Now the obvious pushback is, well, it's early, these companies just lag the tech leaders. And Slok heard that one coming. He's skeptical of it. He argues it's going to be genuinely hard for regular corporate America to turn, as he puts it, increasingly cheap intelligence into higher revenue and fatter margins. And then he drops the line that I think every builder should tape to their monitor. "AI may be a great calculator, but in business, intelligence is mostly judgment." He goes on: AI does not understand accountability, trust, culture, unwritten rules, ethical trade-offs, or fairness. All of the stuff, in other words, that actually drives whether a customer trusts you, whether a deal closes, whether your reputation survives a bad quarter.
Now, I want to be careful here, because this is where a lot of coverage goes wrong. This is not me telling you AI doesn't work. It clearly works for a lot of specific tasks. But there's a gap between "this model is impressive in a demo" and "this changed our income statement," and Slok is standing in that gap waving his arms.
And here's why it matters for you specifically, the founder and the builder. If you're selling AI into enterprises, the economic argument you've been leaning on — the productivity story, the do-more-with-less story — the data isn't backing it up yet at the aggregate level. Your buyer's CFO may be reading the exact same Apollo note. So the pitch that wins in 2027 is probably not "AI makes everything faster." It's "here is the specific, measurable judgment-shaped problem we take off your plate, and here is the dollar figure attached." Specificity is your friend. Hand-waving about intelligence is going to get you shown the door, because the guy across the table has now seen a chart that says the hand-waving didn't pay off.
And this connects to something we keep circling on this show. Back when we covered Anthropic's own report, the one arguing that AI's real bottleneck is human cognition and not the model itself — that's the research-lab version of the same truth. The model is not the hard part anymore. Getting an organization, with its culture and its unwritten rules and its accountability chains, to actually absorb the thing and change how it operates — that's the hard part. Slok is just pricing that difficulty in dollars instead of describing it in theory.
Now hold that thought, because I want to put it right next to a completely different story that, I swear, is the same story wearing a different hat.
Grok released seven minutes of its AI-generated version of The Odyssey. Now, the underlying saga here isn't brand new — Elon Musk announced back in July he was going to make a full-length, quote, historically accurate Odyssey, partly because he was feuding with Christopher Nolan's version. Nolan, for his part, has called himself a techno-skeptic and said he'd never seen a technology, meaning AI, so completely rejected by the public. And Nolan's Odyssey, the real one, the one shot across six countries with a cast and crew of thousands, came out this past summer and became an enormous box office hit, selling out Imax screenings for months.
So three months after the big promise, what did Grok actually ship? Seven minutes. They're now calling it "Episode 1 of our Odyssey series," which, if you've ever worked in software, you recognize immediately. That's the sound of a full movie quietly becoming a "series" because the full movie wasn't happening.
And here's the review, from Anthony Ha at TechCrunch, and I think it's fair. He said it's not as spectacularly ugly as he expected. There were moments he might've thought these were real actors on a real sea. But what it lacks, in his words, is any sense of artistry, inspiration, or human interest. The backstory voiceover goes on forever, you get one bearded face after another shouting stilted lines, and the whole thing feels airless. No momentum. No reason to care about anyone on screen.
Now why am I putting this next to the Apollo note? Because it's the exact same gap. The technology can produce a passable simulacrum of a thing. Individual shots look real. But the part that makes the Odyssey worth three hours of your attention — the judgment, the taste, the reason to care — that's the part the machine doesn't have. Slok said AI is a great calculator but business is judgment. Well, filmmaking is judgment too. You can render the waves. You can't render the reason anybody should keep watching. Same lesson, two very different conference rooms.
And for a founder, there's a practical takeaway buried in the embarrassment. "Historically accurate full-length movie" became "seven-minute Episode 1." When you're setting expectations — with your users, with your investors, with the press — the gap between what you promise and what you ship is the thing people remember. Musk could've shipped seven minutes and called it a tech demo and looked fine. Instead he called it a movie and it reads as a retreat. Scope your promises to what the technology can actually carry.
Alright, let me shift from what AI can't do to the people who've been warning us about who's selling it, because there's a genuinely interesting piece out this week.
There's a book from the year 2000 called Cyberselfish, by a writer named Paulina Borsook, and it's coming back into print in a new edition from Tin House. And the reason it's a story is that a book about technology from a quarter century ago has somehow become relevant again, to the point where used copies were getting scooped up and European academics were cold-emailing the author. The revival got a push from Gil Durán, who wrote The Nerd Reich, and he's writing the introduction to the new edition.
The book was about what Borsook called techno-libertarianism — a strain running through Silicon Valley culture that she describes as, her phrase, violently anti-regulation. And her central jab, which she repeats in the interview, is this: no sector of society has benefited more and suffered less from the government than these guys. The whole industry was built on government-funded R&D, on public infrastructure, on decades of the state doing the expensive, unglamorous work. And then a chunk of the people who got rich off it turned around and decided government was the enemy.
She's got a line about this that stuck with me. She talks about the attitude that survived the dot-com crash, "percolating around, contaminating the aquifers. It was uncorrected, and it just spread out more and more." Like Giardia in the water system, she says. Not the most flattering metaphor, but you get it.
Now, I'm not here to turn Barely Possible into a political show, and I'd push back on some of her broad strokes — she calls AI right now a big Ponzi scheme, which, I think the Apollo economist we just covered would call it something more precise than that. But there's a piece of this interview that I think is genuinely useful for my audience, the builders, and it's the part about young people.
She says the kids come to her, the ones in CS and engineering who don't buy the toxic stuff, and they ask, what do we do? And her honest answer is rough: AI is sucking up all the investment money right now. If you want to do almost anything that isn't AI-related, or defense-related, the capital isn't there. Government's broke. Nonprofits are struggling. And then she says the thing that actually lands for me. She remembers when you could start a small company, just you and a friend or two, bootstrap it, let the revenue support the growth, and not have to — her word — elephantize into some grotesque growth machine. And she's not sure how much of that world still exists.
That, right there, is worth sitting with. Because the two stories I've told you so far — AI not paying off for most companies, and the hype outrunning the product — both point in a direction she's pointing too. If the easy-money, growth-at-all-costs AI narrative cools off, the bootstrapped, revenue-funded, actually-solves-a-problem company might not be the sad consolation prize. It might be the thing that survives. She's cynical about whether it's possible anymore. I'm a little more hopeful, because I've watched the pendulum swing a few times from the cheap seats. But she's naming a real tension, and if you're a founder deciding whether to chase a mega-round or build something that pays for itself, this is the conversation to have with yourself.
Now let's get into our deep dive, and I'm picking this one because it's the single most consequential thing today for anybody building actual software products. OpenAI just made its Decisions API widely available, and this is the start of a real category fight that's going to change how you architect AI into your product.
Let me set the table, because the backstory here is a good one. There's a startup called TypeSafe, founded by a former researcher from one of the frontier labs, and their model is called Jev. And what Jev did was create a whole new category. These things are called decision models, or System One models. Think of them as super-classifiers. You give the model a set of options, and instead of writing you a paragraph, it just picks one, fast, and gives you a confidence rating on its choice. That's it. Given these choices, pick, and tell me how sure you are.
And here's the thing developers figured out quick: a huge number of workflows that people were using big language models for don't actually need a big language model. They need a fast, cheap decision. Is this email spam or not. Is this support ticket urgent or not. Does this image contain a product or not. Should this agent route left or right. You don't need a model writing you an essay for that. You need a snap judgment at a tenth the cost and a fraction of the latency. TypeSafe built a nearly eight-billion-dollar company on that insight.
So OpenAI's product lead for their APIs, a guy named Nikunj Handa, tells TechCrunch that his colleagues saw Jev and said, quote, "wow, we have like a million use cases for this." And here's the part I love, because it's honest in a way these announcements usually aren't. OpenAI didn't go build some specially calibrated decision model from scratch. They took one of their existing ChatGPT models, something they call Luna, and they reconfigured it to speed up the inference and drop the cost. Handa's words: "it's just plain old Luna that gets the job done."
And the numbers are where this gets real for builders. Handa says the big unlock is speed — shrinking responses from one-to-three seconds down to a hundred, a hundred-twenty, a hundred-eighty milliseconds. That's not an incremental improvement. That's a category change. One of their early customers, a company called Speak that builds AI language tutors, their CTO says the Decisions API let them cut the time between actions from 1.4 seconds down to 0.13 seconds. Think about what that does to a conversation with a tutor. 1.4 seconds of dead air is awkward. A tenth of a second is natural. Same with OpenAI's own voice assistant — they used this to make it more responsive. Other early trial customers: Box, Datadog, Ramp. Serious engineering shops.
Now here's where it gets interesting as a competitive story, and this is the part I want builders to really hear. The market already has a bunch of players, and OpenAI is not winning on the benchmarks. There are public benchmarks — the AIM System One Index, the Artificial Analysis Decision Index — and the leaders on those tend to be a model called D1 Ultra, which a company called Liquid released, and something called Elephant-1 from Base Lab. The latest Jev model is competitive, sitting seventh and third on those two indices. And OpenAI's Decision Model? Eighth and fourteenth. So on the scoreboards, OpenAI is in the middle of the pack, even behind.
But — and this is the lesson — TypeSafe's CEO, Diogo Almeida, distrusts benchmarks entirely. He tells people to trial the models in their own actual work, because what makes a good decision model is almost impossible to pin down in the abstract. The piece says it plainly: nobody really knows what makes a good decision model other than "it works for you." The right balance of intelligence, cost, calibration, and speed is different for every single workload. The thing that's perfect for routing support tickets is not the thing that's perfect for classifying video frames.
And my favorite detail, because it's just delightful: there's a benchmark called JevMan where these decision models play Pac-Man. The guy who built it, Felix Wunderlich at Opper AI, explains that Pac-Man is a great test because it's a nonstop stream of quick, high-stakes decisions. At every junction the model's got a few hundred milliseconds to decide where to go. The latest Jev model tops that one. So somewhere out there, the state of the art in machine decision-making is being measured by how well it eats dots and dodges ghosts. God bless this industry.
So let me pull the builder lessons out of this, because there are three and they all matter.
First: stop reaching for the biggest model by default. If you've got workflows in your product where you're paying for a full language model to essentially make a yes-or-no or an A-B-C choice, you are overpaying in both money and latency. The decision-model category exists specifically to eat that use case. Audit your own stack for places where you're using a sledgehammer to pick between two options.
Second: the speed unlock opens doors you didn't know were closed. Handa floated getting this down to 60 milliseconds, at which point, he says, you could imagine an API call for every single frame of a video. That's a different kind of product. Real-time classification on a live video stream. Decisions fast enough to sit inside a natural voice conversation. When latency drops by an order of magnitude, the thing you build changes shape. It's not the same product running faster. It's products that weren't feasible before.
Third, and this is the strategic one: in this category, benchmarks lie, or at least they don't tell you what you need to know. OpenAI is eighth and fourteenth on the public indices and it almost doesn't matter, because the real question is whether the model is calibrated for your specific workload. Where OpenAI might actually have an edge, per Handa, is multimodal stuff — image classification, where their Luna model has been showing better calibration. So the move for you is not to read a leaderboard and pick the winner. The move is to run two or three of these against your own data, on your own task, and measure your own numbers. TypeSafe's own CEO is telling you that. When the vendor says "don't trust the benchmarks, just try it," believe him.
And the bigger picture for the industry: this is the commoditization and specialization of AI happening in real time. The frontier models get all the headlines, but the actual money in production software is increasingly going to flow to fast, cheap, specialized models doing narrow jobs billions of times a day. OpenAI reconfiguring an existing model and shipping it into this fight tells you they see where the volume is. It's not in the three-hour conversation. It's in the hundred-millisecond decision, repeated forever.
Alright, let me come up for air and give you a few shorter ones, because there's good stuff in the margins today.
Let's talk cars, briefly, because there's a regulatory detail here that's actually kind of charming. Federal safety regulators — the National Highway Traffic Safety Administration — have proposed that autonomous vehicles use cyan-colored lights, blue-green, to signal when they're driving themselves with no human monitoring the road. It's not a mandate, more of a strong suggestion, steady light, not flashing, visible from all sides. The idea being that you, the pedestrian, the other driver, should be able to tell at a glance that the car next to you has nobody minding it.
And buried in the same proposal is something more consequential: they'd remove the requirement for manually operated controls for turn signals, high beams, and hazard lights. Why does that matter? Because the current federal motor vehicle safety standards were written for cars that have steering wheels and pedals and mirrors. And vehicles like the Tesla Cybercab and the Zoox robotaxi don't have those things. So the rulebook is slowly being rewritten to allow for cars that were never designed for a human to touch. That's the real story under the pretty blue-green lights — the regulatory scaffolding for a steering-wheel-free world is quietly going up.
Same roundup had GM laying out its plan — overhauling the computing guts of its future vehicles, a conversational AI assistant replacing the thirty-year-old OnStar, an eyes-off highway driving feature coming in 2028 with the Cadillac Escalade IQ. GM says it's used enterprise AI internally to cut its development timeline — normally four to five years — roughly in half. Now, you'll notice, after the Apollo note, I file "company says AI cut our timeline in half" under claims I'd like to see on the income statement before I believe it. But directionally, the legacy automaker is clearly preparing for a bigger fight than just chasing Tesla.
Now let me shift to something on the science and research front, because two of the labs made moves worth a builder's attention.
Anthropic put out an open call for guest investigators in the life sciences. Here's the deal: principal investigators and their research groups — at universities, nonprofits, government labs — can apply to work directly with Anthropic's researchers for eight to twelve weeks, using Claude to go after real questions in human biology, molecular biology, drug discovery, clinical science. Anthropic runs the models and the harnesses at no cost to the scientists. And here's the part that matters: the investigators keep the rights to their own data, and the rights to any follow-on research. Anthropic's not asking them to hand over their datasets.
Why do I flag this for founders? Two reasons. One, it's the latest signal that the labs see the life sciences as a serious frontier, not a side quest — this sits alongside their Life Sciences Verification Program and their AI for Science credits. And two, the structure is smart and worth stealing. Direct access, time-boxed, keep your own IP. If you're building a developer platform or a research tool, that model — we'll give you real hands-on help for a fixed window, and you keep what you make — is a much better way to earn trust than "sign here and give us your data." Trust is the currency, and Anthropic is spending on it deliberately.
OpenAI, meanwhile, had a cluster of quieter posts — a safety and security partnership with Poland's NASK, a piece on how they've hardened their frontier model training environment with offline environments and high-compute red-teaming, and an engineering post on scaling their storage to serve over a billion ChatGPT users, which apparently involved squeezing twelve-times-faster reads out of their system. I'm not going to deep-dive the storage internals on you, because that's exactly the kind of plumbing story I promised I'd spare you. But the one-line takeaway is: a billion users is a number, and the infrastructure work to hold that up without falling over is real, unglamorous, and the kind of thing that quietly becomes a moat. Also note the theme — training-environment containment, red-teaming, a government safety partnership. The labs are spending a lot of visible effort right now looking responsible. Read into that what you will.
A quick one on the hardware wars, because it's a good reminder about defensibility. Oura, the smart ring company, has sued its rival Ultrahuman again, trying to block US sales of Ultrahuman's Ring Pro. And the detail I can't get past: one of the two patents Oura is asserting was issued to them on October 20th — two days before they filed the lawsuit. The ink was barely dry. Ultrahuman's CEO points that out pretty pointedly, says there's no court order affecting the ring and it's still on sale. This is at least their third round of fighting since 2023, and Oura's gone after Samsung, RingConn, Circular, and others too — some of whom just folded and agreed to pay royalties.
The builder lesson here is cold but true: in a hardware category, patents aren't a footnote, they're a weapon, and the incumbent will use them to make your life expensive right when you're trying to break into the US market. Oura shelved a 2.2 billion dollar IPO recently, citing market uncertainty — so part of this is a market leader protecting its position while it waits out a rough window. If you're building physical product, budget for the legal fight the way you budget for tooling. It's not optional.
And one more, because it made me laugh and then it made me wince. The New York Knicks — the actual basketball team — sent their employees a phishing test. And the bait was a fake offer of a free championship ring. "Congratulations, your Employee Championship Ring is here," click to pick your design. This is a team that just won its first title in over half a century, so a ring is not an implausible thing to be offered. Folks who clicked got sent to a mandatory cybersecurity class. Madison Square Garden later called using the rings as the lure an "unfortunate mistake," and noted they were actually gifting hundreds of real rings worth millions to staff.
Now, phishing tests are normal and good — Ars Technica's own owner sends them from a deliberately misspelled domain. But there's a real lesson for anyone who runs a company and is tempted to get clever with internal security training. The whole point is to teach people to be skeptical. If your bait preys on something emotional and genuinely desired — a championship ring after fifty years — you're not just testing them, you're kind of punishing them for hoping. You can train vigilance without being cruel about it. Dangling the dream and then slapping the wrist is how you get employees who trust the IT department a little less, not more. And trust, as our Cyberselfish author said and our Apollo economist implied, is slow to build and quick to lose.
So let me tie the ribbon on this, because there's a through-line today whether I planned it or not. The Apollo economist says AI hasn't paid off outside tech because business is judgment, not calculation. Grok's seven-minute Odyssey shows you a machine that can render the waves but not the reason to care. The decision-model fight shows where AI actually is paying off — not in the grand gesture, but in the narrow, fast, specialized job done a billion times. And the whole Cyberselfish conversation reminds us to keep an eye on who's selling the dream and whether the dream is funded by something real.
If you're building right now, the move isn't to chase the biggest model or the biggest promise. It's to find the specific, judgment-adjacent, measurable problem, attach a dollar figure to it, pick the right-sized tool — maybe a hundred-millisecond decision model, not a three-second essay generator — and ship something that pays for itself. That's not the glamorous version. But glamour, as a couple of today's stories showed us, has a way of shipping seven minutes and calling it a movie.
That's the menu for today. I'm Tony DeLuca, this has been Barely Possible, and I appreciate you spending a little of your one and only life in here with me. Go build something that works for you. Take care of each other.