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 welcome back to Barely Possible, where we sift through the daily pile of tech news, throw out the junk mail, and keep the stuff you actually need to know if you're building something real. Fresh menu today, a lot of it, and a couple of items that smell a little off the moment you pick them up. Buckle up, let's have at it.
Let me start with the one that I think matters most to you, and it's not a shiny new model. It's a lawsuit. Because if you're building anything on top of these big AI companies, or you're worried about what they can and can't do with your data, the story out today is a warning shot.
Here's the setup. The New York Times and a group of news publishers are in a copyright fight with OpenAI. That fight's been grinding for a while. What's new, and what got reported yesterday by both TechCrunch and Ars Technica, is that the publishers have filed a new motion for sanctions. And the accusation is serious. According to Rebecca Bellan's piece at TechCrunch, the news publishers say OpenAI hid tools and datasets that could identify copyrighted journalism in ChatGPT outputs. Ashley Belanger, writing for Ars Technica, puts it even harder. The framing there is that OpenAI may be sanctioned for hiding and deleting ChatGPT logs, and that the company allegedly faked an inability to search its own training data while sitting on billions of logs.
Now let me be careful here, because this is an allegation in a legal filing. It's what one side is claiming. OpenAI hasn't been found to have done any of this by a judge. But the reason I'm leading with it is that this is exactly the kind of thing that decides cases. When you're in litigation, the fastest way to turn a defensible position into a losing one is to look like you hid evidence. Judges hate spoliation. That's the legal word for destroying or concealing evidence you were supposed to preserve. It's the kind of thing that can flip a judge from neutral referee to actively hostile, and it can lead to what's called an adverse inference, where the court basically tells the jury, assume the worst about what they were hiding.
So why does this matter to you, the builder? Two reasons. First, the outcome of these copyright cases sets the ground rules for what training data is legal and what it costs. If the publishers win big, or if OpenAI gets hammered on sanctions, the whole economics of foundation models shifts. Licensing deals get more expensive, models get more cautious about what they'll reproduce, and the smaller players who can't afford to license the world's journalism get squeezed. Second, and this is the quieter one, it's a reminder that the logs exist. All those billions of interactions. When a company tells you it can't search its own data, and then it turns out maybe it could, you should recalibrate your assumptions about what's being retained on the other side of your API calls. If you're piping sensitive stuff into these systems, act like the logs are forever, because in a courtroom, they might have to be.
As we talked about last week, the whole AI industry has been living in what I called the expense-management era, everybody counting tokens, everybody watching costs. This copyright fight is the other side of that ledger. It's the legal cost that nobody put in the spreadsheet. You can optimize your inference bill all you want, but a bad ruling on training data is a liability that doesn't show up until it's already on your doorstep.
Alright, let me pivot from the courtroom to the product announcements, because OpenAI had a big day yesterday, and it wasn't just the model.
Sam Altman posted during a livestream that in addition to the 5.6 model, there were three major product things. His words. One, ChatGPT Work, which he flagged as a really big deal. Two, a new ChatGPT desktop app. Three, hosted sites. Let me translate what that stack actually signals, because it's a strategy, not just a feature list.
ChatGPT Work is OpenAI trying to plant its flag inside your company, not just on your phone. This is the enterprise seat, the version that lives where your colleagues live, shares context, and does actual work assignments. Anthropic's been pushing hard in that same direction — we covered their Claude Tag play a couple weeks back, the thing that sits in your Slack with shared context across channels. So now you've got both big labs racing for the same real estate: not the individual user typing prompts, but the company knowledge layer, the thing that knows what your team is doing and can act across it. That's a much stickier, much more valuable position than a chatbot. Whoever owns that owns the enterprise relationship for years.
The desktop app and hosted sites are the connective tissue. Hosted sites in particular is interesting — that's OpenAI saying, don't just generate the code, we'll host the thing you built too. That's a direct nudge into territory that companies like Vercel and the vibe-coding crowd have been staking out. If ChatGPT can write your app and then serve it, that's one less vendor in your stack, and one more reason to stay inside the walled garden.
And then there's the rebranded Codex. Kyle Orland at Ars Technica frames it plainly: OpenAI wants its new tool to do your work for you and with you. The rebranded Codex promises independent workflows that can run, quote, for hours if needed. That "for hours" number keeps showing up across the industry now, and I want you to notice the pattern. A year ago the pitch was "AI helps you code faster." Now the pitch is "AI goes off and works on its own for hours while you do something else." That's a different product, a different trust model, and honestly a different set of risks. Something running unsupervised for hours against your codebase is powerful, and it's also the kind of thing that can quietly do a lot of damage before anyone reviews it. If you adopt these long-running agents, the discipline you need isn't in the prompt, it's in the guardrails: what can it touch, what can it merge, who signs off before it ships.
Now here's the counterweight to all that OpenAI momentum, and it's the item that made me raise an eyebrow. OpenAI is shutting down Atlas. That's their AI-powered browser, and per Rebecca Bellan's reporting at TechCrunch, they're sunsetting it after less than a year. Less than a year. They're moving some of the agentic browsing features into the desktop app and into a Chrome extension instead.
I want to be honest about what this is and isn't. The headline is "AI browser ambitions are still growing," and sure, they're keeping the agentic browsing capability alive in other places. But let's not dress it up too much. Building a whole browser, getting people to switch their default browser, that is one of the hardest things in consumer software. Google spent years and untold billions getting people onto Chrome. The idea that OpenAI was going to unseat that in under a year was always a long shot, and now they've folded that hand. The smart read for you as a builder: don't fight the platform battles the incumbents already won. Meet users where they already are. A Chrome extension that adds agentic browsing is a much more sensible bet than asking the entire world to change its browser. OpenAI just learned that lesson in public, and it only cost them a year.
Let me stay on the model wars for a second, because Meta jumped in too. Lucas Ropek at TechCrunch reports Meta entered the crowded AI coding battle with Muse Spark 1.1. The pitch, per the piece, is Spark's ability to handle large agentic workloads, fix bugs, and help with large code migrations — the kind of automation enterprises are increasingly turning to AI companies for.
Notice the word "crowded" in that headline, because that's the whole story. Everybody is now shipping a coding agent. OpenAI's got Codex, Anthropic's got Claude Code, the cursor and cognition folks have their own, and now Meta wants a seat. And they're all making the same pitch: large agentic workloads, bug fixing, code migrations. When every major player is selling you the identical value proposition, that tells you two things. One, this is where the money is perceived to be — enterprise engineering automation. Two, it's about to become a brutal price and performance war, which is good for you if you're a buyer. When Meta, OpenAI, Anthropic, and a half-dozen well-funded startups are all fighting for the same enterprise coding budget, you're going to see the prices come down and the capabilities go up, fast. Don't marry any one of them. Build your stack so you can swap the model underneath. That flexibility is going to be worth real money over the next year.
And speaking of the model itself, Simon Willison put out his notes on GPT-5.6. He flags some genuinely interesting additions to the API — programmatic tool calling and multi-agent in particular. And, because it's Simon, eighteen pelicans for the six reasoning levels and three new models. If you don't know the reference, Simon's got this running bit where he benchmarks new models by asking them to draw a pelican riding a bicycle in SVG. It sounds like a joke, and it kind of is, but it's also a genuinely useful vibe check on how these models handle a weird, specific spatial task with no training-set answer to copy. The substance for builders is that programmatic tool calling and multi-agent support in the API. That's the plumbing that lets you build the "runs for hours" agent systems everybody's advertising. The model is one thing, but the API primitives are what determine whether you can actually orchestrate something reliable on top of it.
Now let's shift from OpenAI's product blitz to the money question underneath all of it, because TechCrunch ran a piece by Tim Fernholz asking, can AI answer the three trillion dollar question. The framing is that the AI return-on-investment debate has come back, and the numbers are even bigger than before, and so, maybe, are the consequences.
I don't have the full guts of that piece in front of me, but the headline number tells you where the anxiety is. We are now talking about trillions of dollars in capital pouring into AI infrastructure — the data centers, the chips, the power. And the three-trillion-dollar question is the obvious one your accountant would ask: where's the return? At some point all that spending has to turn into revenue and profit that justifies it, or the whole thing looks like the dot-com build-out where everybody laid fiber and then went bankrupt before the demand caught up. Now, the fiber did eventually get used, so maybe that's the optimistic version. But if you're a founder raising money right now, understand the climate you're raising into. The easy-money "AI in the pitch deck automatically gets a term sheet" era is cooling. We talked about that back when Jersey Mike's was name-dropping AI in an IPO filing — that was peak froth. Investors are starting to ask the hard version of the question now: not "do you use AI," but "does the AI actually earn its keep." Build your business so the answer is yes on a spreadsheet, not just yes on a vibe.
Which brings me to a story I need to handle carefully, because the framing on it is a little misleading. There's an item floating around about Mercor being in talks for a twenty billion dollar valuation, a huge step up from the ten billion it reached in October. Now here's the thing — the reporting that resurfaced references the October ten-billion valuation, and the age on this one is muddy. The underlying valuation milestone it's anchored to is from last fall. So I'm not going to stand here and tell you Mercor just closed a fresh twenty-billion round today, because the source doesn't cleanly support that. What I'll say is this: Mercor is one of these data-labeling and human-expert-marketplace companies that feeds the AI training machine, and the fact that this kind of business is even in twenty-billion-dollar conversations tells you where the value is accruing. It's not always the model maker. Sometimes it's the picks-and-shovels company selling the training data and the human evaluators. If you're thinking about where to build, the boring infrastructure underneath the models is often the better business than trying to make another chatbot.
Now let's dig into a governance move that a lot of people will scroll right past, and I think that's a mistake.
Anthropic announced that Ben Bernanke — yes, that Ben Bernanke, the former Federal Reserve chairman, the guy who ran U.S. monetary policy through the 2008 financial crisis — has been appointed to Anthropic's Long-Term Benefit Trust. Now, for those who don't follow Anthropic's structure, the Long-Term Benefit Trust is this unusual governance body. It's a group that holds a special class of shares and has the power to elect some of Anthropic's board members. The whole idea, in theory, is that it's supposed to steward the company toward its mission — safe AI development — even when that might conflict with pure profit motives. It's a check on the ordinary corporate incentive to just maximize shareholder value at all costs.
So why put a central banker on it? Think about what Bernanke actually represents. He's the guy who was in the room when the financial system nearly melted down, and whose entire career is about managing systemic risk — the kind of risk where one institution's failure can take down the whole economy. Anthropic putting him on the trust is a signal. It's saying, we think the risks of this technology are systemic, macroeconomic in scale, and we want someone who has stared down a systemic crisis before helping to govern us. Whether you buy that as sincere or as sophisticated public relations is up to you. I'm a skeptic of hype, you know that, so I'll note that a big prestigious name on a governance body is also very good for the kind of enterprise and government trust Anthropic is chasing right now. But taken at face value, it's a real statement about how these companies want to be seen — not as another software startup, but as institutions handling something closer to nuclear material. Keep an eye on whether that governance structure ever actually does anything when the profit motive and the safety mission collide. Governance bodies are only as good as the moment they're tested.
Anthropic also put out a couple of other things worth a quick mention. There's a post titled Inviting hard questions, which reads like the company signaling it wants scrutiny — again, part of that "we take the stakes seriously" posture. And they introduced a feature to reflect on how you use Claude, which is a tool to let you look back at your own usage patterns. That reflection feature is a small thing, but it's part of a trend I want you to notice: the AI companies are starting to build meta-tools, tools about the tool. How am I using this, what am I spending, what patterns show up. For a builder, that's a hint about a product category — the layer that helps people understand and govern their own AI usage is going to matter more as usage explodes and the bills climb.
Now let me pivot to Elon, because there's a story that's really about trust and infrastructure, and it's a good one to think through.
TechCrunch, in a piece by Julie Bort, reports that Elon Musk praised Mythos and Fable — those are Anthropic's frontier models — and promised not to, quote, cut off Anthropic. The framing of the piece is the obvious question: should Anthropic trust Elon Musk to host its models? With about forty billion dollars in revenue at stake, Musk insists the company can.
Let me unpack what's even going on here, because it's a little wild when you say it plainly. Anthropic, an AI lab, is apparently relying on infrastructure connected to Elon Musk's operation to serve some of its models. And Musk, who runs a competing AI effort, is the guy in a position to, theoretically, cut them off. And now he's out here publicly promising he won't. Which, if I may — when somebody feels the need to publicly promise they won't stab you, that itself tells you people are worried about the knife. Forty billion in revenue riding on a handshake from a competitor is not a position any CFO sleeps well in.
The builder lesson here is one of the oldest in the book, and it keeps coming back in new clothes: concentration risk. If your business depends on a single provider — for compute, for hosting, for models — and that provider is also your competitor or has their own agenda, you are exposed. You've handed someone a switch they can flip. We saw a version of this earlier in the year when the government put curbs on which labs' models could ship globally, and everybody who'd built on a single frontier model suddenly had a problem. The answer, again, is portability. Multi-cloud, multi-model, the ability to move. It's more expensive and more annoying to build that way. It's also the difference between a bad week and a dead company when the person holding your switch changes their mind.
Alright, let me get you a couple of things from outside the pure AI lane, because the show's not just about the labs.
Google made a policy move that I think is quietly significant. Per Sarah Perez at TechCrunch, Google will now disclose which ads are made with AI. The detail in the piece: Google prohibits misleading and deceptive ads, but an ad can still use AI to create synthetic or digitally altered content, and until now, that disclosure was only required for election ads. Now they're broadening it.
This is a small policy change with a big tell inside it. Google's basically admitting that AI-generated ad content has gotten common enough and convincing enough that the old "only worry about election ads" rule doesn't cut it anymore. Synthetic media is everywhere in commerce now. And if you're building anything that generates marketing content — images, video, copy — pay attention, because disclosure requirements are coming, and they're going to spread from platform to platform. The regulatory and platform-policy weather is shifting toward "label the synthetic stuff." Build that capability in from the start. It's a lot easier to add a disclosure flag when you're designing the system than to retrofit it after a platform makes it mandatory. Remember, Meta's new image model reportedly lets you tag someone and use their public photos in a generation — a one-click deepfake concern that people flagged immediately. The whole industry is walking a line between "amazing creative tool" and "misinformation machine," and disclosure is the first fence they're building along that line.
Now here's an enterprise fight that every IT buyer should file away. Ars Technica, Scharon Harding reporting: Allstate accuses Broadcom of auditing it because it quit VMware and CA. And Broadcom fires back, accusing Allstate of dodging VMware audits.
Let me give you the shape of this without pretending I have every detail. When Broadcom bought VMware, they changed the licensing and pricing in ways that made a lot of big customers furious. Allstate, the insurance giant, is one of the customers that decided to walk away from VMware and CA. And now Allstate is claiming that Broadcom turned around and hit them with a software license audit — the accusation being it was retaliation for leaving. Broadcom's counter is basically, no, they were just dodging audits they were contractually obligated to.
Why does a fight between an insurance company and a chip conglomerate matter to you? Because this is what vendor lock-in looks like when it turns ugly. You sign up for enterprise software, the vendor gets acquired, the terms change under you, and when you try to leave, you find out the exit door has a toll booth on it in the form of an audit. If you're building on top of enterprise software or making procurement decisions, read your license terms like your life depends on it, especially the audit clauses and what happens if you try to migrate off. The cost of leaving is often hidden until the day you want to leave. Allstate and Broadcom are now going to litigate that in public, and the rest of us get to learn from it for free.
Let me hit the security item quickly because it's a good reminder. Dan Goodin at Ars Technica: a patch for a Windows Defender zero-day that could allow attackers to fill your hard disk. The framing is that the feud between a researcher going by NightmareEclipse and Microsoft shows no signs of resolving. So you've got a security researcher and Microsoft locked in an ongoing back-and-forth, and the practical upshot is a vulnerability in Windows Defender — the thing that's supposed to protect you — that could let an attacker fill up your disk. Patch it, keep your systems current, and appreciate the irony that the antivirus itself is the attack surface. The lesson that never gets old: your security tools are software too, and software has bugs. Defense in depth exists because any single layer, including the one marketed as your protection, can fail.
Now let me get into the deep dive, and I want to spend real time here, because this is the one I think is most consequential for how you'll actually work over the next year. It comes out of a detailed rundown of the wave of new models that landed this month, and I want to pull the useful argument out of it rather than recap the whole thing.
The setup is this: we just got a flood of frontier models more or less at once. OpenAI's GPT-5.6, the one they call Sol. Anthropic's Fable. A new Grok. A model from Cognition. And what's fascinating — and genuinely new, according to the people testing these — is that for the first time the leading models are both decidedly ahead of everything else AND distinctly different from one another. That second part is the key. It used to be that the frontier models were basically interchangeable — a little smarter here, a little cheaper there, but same shape. Now they have personalities, and those personalities change how you should use them.
Let me give you the best framing from the testers, because it's vivid. One reviewer, Peter Gostev, described it this way: Fable is a wise owl — very thoughtful, very well-spoken. GPT-5.6 Sol is like a Rottweiler that grabs the problem by the throat and won't let go until it's done. His verdict: on pure intelligence, Fable is the smarter model. But Sol is extremely diligent. You give it a list of eight things to do, and you can be sure all eight get done. No lectures, no "you're absolutely right" filler, nothing beneath it. If it takes two days of dirty work, it'll do the dirty work.
The team at Every had a similar read, and I love their analogy: GPT-5.6 is like a Porsche, Fable is like a warp drive. If you need to get across the galaxy, use the warp drive. If you need to get around town with the best available tool for the daily job, use the Porsche. They said after a month of testing internally, Sol became the best combination of power, speed, and performance for day-to-day knowledge work and coding, while Fable is a different beast for the loosest, longest assignments where you can define exactly what you want and let it run.
Now here's why I'm making this the centerpiece for you, the builder. The old mental model was "pick the best model and use it for everything." That model is dead. The new reality is that you're going to be running a portfolio of models, and the skill — the actual competitive edge — is knowing which one to reach for which task. Ethan Mollick, who tests these things constantly, said he found himself switching between Fable and Sol depending on the task: Sol for back-and-forth work, especially when he hadn't yet figured out exactly what he needed; Fable for very long tasks where he could define the outcome up front.
There's a third thing in there that I think is the sneaky-important insight: speed changes the model. The Every team said Sol is the fastest model they trust as a daily driver, and that speed changes behavior. When a revision costs minutes instead of an hour, it becomes easy to throw out a weak draft and try a different direction while the problem is still fresh in your head. That's not a small productivity tweak. That's a change in how the work feels. There's a comment in there about a new "weird middle mode" — a task that used to justify walking away for a coffee now finishes before you've mentally moved on. It's technically asynchronous, but fast enough that you just watch it happen. That's a genuinely new interaction pattern, and if you're designing tools or workflows, that's the zone to design for.
So what do you actually do with this? A few things. First, stop asking "what's the best model" and start asking "what's the best model for this specific task." Build your systems to route — cheap and fast for the high-volume grunt work, expensive and smart for the hard reasoning. That routing layer is becoming its own product category, and it's where a lot of cost savings and quality gains live. Second, the harness matters as much as the model, and I know that's a phrase I've beaten to death, so let me put it differently: the surrounding system — the sources you give it, the examples, the standing instructions, the clear outcome — that scaffolding is what separates a model that flails from one that ships. And here's a warning buried in the Every notes that I want to make sure you hear: stale instructions hurt. Rules you wrote for older models can actively make the new ones worse. So if you've got a bunch of prompt engineering and system instructions built up over the last year, some of that is now technical debt. It was written for a dumber model, and it's holding the smart one back. Audit it.
Third, and this is the strategic one: the fact that the frontier models are now differentiated is great news for anyone who builds a layer on top of them. If the models were interchangeable, the model makers hold all the power. But because they have distinct strengths, there's real value in a product that intelligently combines them — that knows to send the writing task to the good writer, the long autonomous job to the model that runs for hours, the quick iterative loop to the fast one. That orchestration is defensible. That's a business.
Now let me tie this back to where we started, with the copyright lawsuit and the ROI question, because they're connected. All this model differentiation and all this "runs for hours" autonomy — it only pays off if the underlying economics hold and the legal ground doesn't shift under you. The three-trillion-dollar question and the New York Times sanctions motion are the two clouds on the horizon over this beautiful buffet of new capabilities. The models have never been better or more useful. And at the same time, the questions of whether the training was legal and whether the spending will ever pay off are getting louder, not quieter. That's the tension you're building inside of. Great tools, uncertain foundation. Use the tools, but don't bet the company on any single provider, any single legal outcome, or any single financial assumption. Portability, flexibility, and a spreadsheet that closes — those are your armor.
Let me clear the last couple items off the menu quickly, because there's some fun stuff in the pile today.
Mistral put out something called Robostral Navigate. The detail we've got is thin — it's a Mistral release aimed at navigation, and the name tells you they're playing in the robotics and spatial space. I'm not going to oversell what I can't see, but the signal is that Mistral, the European lab, is pushing into applied robotics and navigation, which is a smart lane — physical-world AI is a lot less crowded than yet another chatbot, and Europe's got real strength in industrial and robotics applications. One to watch.
On the medical robotics front, Ars Technica had a genuinely striking one from Jeremy Hsu: humanoid robots controlled by surgeons did a world-first operation on live pigs. It's a preclinical trial testing whether humanoid robots can be feasible in surgery. Now, note the important word: controlled by surgeons. This isn't a robot deciding to operate on its own. It's teleoperation — a human surgeon in the loop, the robot as the hands. That's the responsible version of this technology, and it's the version that'll actually make it into hospitals. The dream here is that a skilled surgeon could operate from anywhere, or that the robot form factor could work in spaces or ways a human body can't. Early days, it's pigs, but it's the kind of concrete progress in physical AI that I find a lot more interesting than another leaderboard number.
And for the fun-and-strange file, a couple of quick ones. Ruf — that's the German outfit that builds hopped-up Porsche-based machines — debuted a new flat-eight engine at Goodwood. Four-point-eight liters, more than a thousand horsepower and a thousand newton-meters of torque. A flat-eight is a genuinely unusual engine layout, and a thousand horsepower is absurd in the best way. If you love cars, that's a treat. And there's a Dune: Part Three trailer out, with the line, "You promised me that you would never take power in your name." Which, given everything we just talked about with governance trusts and promises not to cut people off, feels almost too on the nose for today's episode. Conspiracies and regrets, indeed.
I'll skip past the parasite outbreak news and the ten-inch worm story out of the health desk, except to say — if you're eating lunch while you listen, maybe pause and go read those on your own time. Not everything needs my commentary.
So here's where I land it. The through-line today is trust and dependence. The New York Times fight is about whether you can trust what these companies say about their own data. The Elon-and-Anthropic story is about whether you can trust the person holding your infrastructure switch. The Broadcom-Allstate mess is about what happens when you trusted a vendor and the terms changed under you. And the model wars are about learning to distribute your trust across a portfolio instead of betting everything on one relationship. Build like the people holding your switches might change their minds, because sometimes they do. Keep your options open, read the fine print, and make sure the AI in your pitch deck actually earns its seat.
That's the menu for today. This is Tony DeLuca, and this has been Barely Possible. Go build something that still works when your favorite vendor has a bad week. Take care of yourselves, and I'll see you tomorrow.