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 you've found your way back to Barely Possible, where we go through the daily menu of tech and AI news, we throw out the moldy stuff, and we serve you only what's worth chewing on. Grab a cup of coffee, pull up a chair, and let's have at it.
I want to start today somewhere a little unusual. Not with a chatbot, not with a chip, not with somebody's ten-billion-dollar valuation. I want to start with a picture. A picture that turns out to be a lie.
There's a piece from John Timmer over at Ars Technica about lab supply companies selling antibodies using manipulated images. Now this is one of those stories that first surfaced back in May, so I'm not going to sit here and tell you it broke this morning. It didn't. But it resurfaced, and honestly it deserves the attention because of what it says about trust and about the plumbing underneath a whole industry.
Here's the setup. Antibodies. If you did any biology past the tenth grade, you know these are the things your immune system makes to recognize specific proteins. But they're also the workhorse of biological research. Researchers use them constantly, and a lot of the time they don't make their own. They order them. Commercially. Overnight delivery, like it's a pair of sneakers. And when a company sells you an antibody, they show you pretty pictures. Images from experiments demonstrating, hey look, this thing works, look how clean the data is, look how beautiful.
And a post-doc at Northwestern named Reese Richardson looked at those pretty pictures and said, wait a minute. Some of these have been doctored. Not a little touched up. Doctored. The kind of image manipulation that, if it showed up in a published academic paper, would get that paper retracted. We're talking about scrubbing out background noise, copying and pasting data to fabricate results. So Richardson brings in a collaborator, Sholto David, they partially automate the whole thing to hunt for these manipulations at scale, and what do they find? Problematic image changes on over seventeen thousand commercial antibodies. Sixteen different companies. Nearly seven percent of the images they looked at showed signs of manipulation.
Now here's the part that gets me, the part that any founder should sit with for a second. Think about what happens downstream. A researcher buys this antibody. Believes it'll produce clean data because the marketing said so. And then it doesn't work. And what do most people assume? They assume the problem is them. That they messed up the procedure. So they spend days, weeks, real money, troubleshooting, tweaking conditions, trying to squeeze results out of a product that was never going to deliver them. Because the picture was fake.
And when Nature went to these companies for comment, the responses ranged, paraphrasing here, from "we'll look into it" to "we don't think it's a problem." We don't think it's a problem. Tell that to the grad student who burned a month of their life.
Now why am I leading a tech show with a story about lab supplies? Because this is the exact same disease we see everywhere in the industry we cover. A demo that dazzles. A benchmark that shines. A marketing image that shows you the beautiful version. And underneath it, sometimes, is a scrubbed background hiding the noise. The gap between what's shown and what's real. Hold that thought, because it's going to keep coming back today.
Let me move from a fake picture to a fake game. Or, well, a real game with a fake sense of what it means.
Jeremy Hsu, also at Ars, wrote up the second World Humanoid Robot Games, held in Beijing, wrapped up around the twenty-sixth. And look, you've probably seen the clips, they went viral. Sprinting robots beating Usain Bolt's 100-meter record. Sounds like the future arrived, right? Except the same videos show these robots crashing into barriers, falling down because they physically cannot stop running, and a few of them breaking apart at the waist in a shower of sparks. One of them literally caught fire. Cue the Vangelis music, as the headline put it.
Here's the thing to understand, and Dipam Patel, a PhD candidate at Purdue and a research assistant at the Army Research Lab, laid it out well. Yes, going from lagging behind humans last year to beating world records this year is real progress in whole-body control. Genuinely impressive engineering. But — and this is the whole ballgame — these robots were engineered to do exactly one thing. They sprint in a straight line. They can't even slow down after the finish line. As Patel put it, if you cross the finish line, that's the goal — it doesn't matter if you stop, and it doesn't matter if you fall into different pieces.
And here's my favorite detail, because it's honest in a way marketing never is. There's a company called X-Humanoid whose robot went viral for winning a 400-meter race with this weird posture, arms up near its face like a shy person hiding behind their hands. Why does it run like that? Because in reinforcement-learning simulation, that's the shape the machine discovered was fastest. It doesn't care that it looks ridiculous. It found the optimum. That's actually a beautiful little window into how these systems really think — they don't reason like us, they just grind toward the number.
But here's the part that connects back to my antibody story. The flashy events — the sprints, the kung fu, the kickboxing — those are the pretty pictures. The stuff that makes the headlines and pulls in the investor money. Patel's point is that the boring events are the ones that actually tell you something. Washing laundry. Folding clothes. Cleaning a living room. Bringing in a delivery package, which forced the robot to switch tasks mid-chore. There was a firefighting event where robots had thirty minutes to find hazardous substances, shut off three valves, spot a fire, grab an extinguisher. Only three out of twelve teams finished the whole thing. And a whole lot of these competitors were still being teleoperated — a human puppeteering them behind the scenes, with the score docked as a penalty for it.
Investors poured over six billion dollars into humanoid robot companies in 2025 alone. And I'm not telling you it's all smoke. I'm telling you what Patel's telling you: the real test is the eight-hour work shift in a random environment nobody trained for. The sprint record is the demo. The laundry is the business. And most of the money is chasing the sprint.
Alright, let me bring this home, because I've been circling a theme and it's time to plant the flag. That gap — between the polished demo and the messy real deployment — that's not just my observation. It turns out that's the single most consequential thing happening in AI for builders right now, and it deserves a proper deep dive. So let's do that.
I want to talk about Claude Cowork, and memory, and why one boring little product update tells you more about where AI is actually headed than any robot sprint.
On Tuesday, Anthropic announced a change to how Claude handles memory. Sarah Perez wrote it up over at TechCrunch. And on the surface, this is the least sexy news you'll hear all week. Ready? They're merging the memory system between Claude's chat and Claude Cowork. That's it. That's the announcement.
But stick with me, because this is exactly where the real value lives. Here's the problem they're solving, and it's a problem anybody who's actually used these agents in anger knows in their bones. As Perez puts it, the change eliminates "one of the most annoying things about using agents — the constant need to rebrief the AI on things it already knows."
Let me paint the picture. Before this update, you'd be chatting with Claude, developing ideas for a project over time. Working stuff out. Then it comes time to actually do the thing — take action, execute — and you move over to Cowork, the agent side. And Cowork has no idea what you were just talking about. You have to start over. Retype everything. As the piece describes it, and I'm quoting, "Before, you may have developed ideas for projects over time while chatting with Claude, but when it came time to actually enact the changes, you'd have to start over by telling Cowork all the details. This was incredibly frustrating."
Incredibly frustrating. You know who writes "incredibly frustrating" in a straight news piece? Somebody who's lived it.
So now, Claude remembers what it learned in one area even when you're in another. Anthropic gives the example: you need to draft an update for your manager, or an agenda for a conference. Cowork already knows the headcount, the city, the speakers, because you talked about all that in chat weeks ago. You don't re-input it. The update, in their words, makes Claude feel like "one continuous assistant, rather than two separate products under one roof."
And there's a nice technical wrinkle. Claude now adds topics to its memory as you chat — continuously — instead of waiting to summarize the conversation at the end. So the memory updates in real time and gets shared across both experiences, even while the chat is still going.
Now, why am I making this the centerpiece? Because this is the thing the sprint-record robots and the doctored antibody images both point at. The model was never the hard part. Claude is smart. It's been smart. The hard part — the part that decides whether any of this actually becomes useful in a real business — is the connective tissue. The memory. The state. The context that carries from where you think to where you act. In the real world, work is not one clean prompt and one clean answer. Work is a long, messy, multi-day conversation where you develop an idea, then execute it, then come back and revise it. And if your tool forgets everything between the thinking and the doing, it's a demo, not a coworker.
I want to give Anthropic some credit here too, because they didn't just ship the convenience and skip the hard questions. They're exposing the memory. You can read what Claude has retained, you can edit it, you can delete it. That's the right instinct. Because a machine that silently remembers everything about you is a liability, and we talked about exactly that flavor of liability yesterday — that stealth startup Instinct, whose assistant kept your emails sitting in plain text and had terms granting it a perpetual license to your keystrokes. Compared to that, Anthropic's approach here is downright grown-up. By default, Claude won't store the sensitive stuff — your health data, race, ethnicity, religious beliefs, politics, gender identity. You can toggle that on if you want it, and you get notified when a sensitive topic gets saved. And there's a hard floor: it will never save government IDs, Social Security numbers, criminal history, immigration status. Never.
So here's the founder takeaway, and it's the through-line of the whole episode. Everybody's out there benchmarking models, arguing about whose sprint time is faster. But the product that wins your daily use isn't the one with the highest score. It's the one that remembers the headcount at your conference so you don't have to type it twice. Persistence and continuity are the unglamorous features that turn a clever toy into something you actually reach for on a Tuesday morning. The laundry, not the sprint.
Now let me shift from the assistant that remembers you to the assistants that are cleaning up other people's messes.
OpenAI put out a note that it disrupted a covert influence campaign coming out of Russia. Now, the source here is thin — it's a short OpenAI post — so I'm going to keep this tight and not embroider it. What they say happened: they banned a set of Russia-origin accounts that were using AI to promote a fake Israel-based think tank, and to push what they called a "sovereignty" index — basically a scoring system praising Russia and criticizing the West. That's the whole shape of it. Manufactured credibility. A think tank that isn't a think tank, an index designed to launder a political message into something that looks like objective research.
And you notice the pattern again? A fake front. The doctored picture. The scrubbed background. Somebody dressing up propaganda in the costume of a neutral institution, hoping you won't check the brightness and contrast. It's the same con, just aimed at your politics instead of your western blot.
The reason I flag this for builders is simple: if you're putting a generative model into your product, you are now also, whether you like it or not, in the trust-and-safety business. The same tools that draft your marketing copy draft their fake think tank. OpenAI polices its own platform because it has to. If you're building on top of these models, some slice of that responsibility rolls downhill to you.
Speaking of trust and who gets to read what, let me get to a story that hits close to home for anybody who's ever built on somebody else's platform. X — Elon's X — has sent cease-and-desist letters to an open source project called Nitter.
If you don't know Nitter, here's the quick version. For seven years, it was a way to read public X posts without logging in, without opening the app. It'd fetch the public posts and strip out the ads, the tracking cookies, the JavaScript, and just give you the clean text. It also powered a bunch of other sites, like XCancel. Beloved by lurkers, by researchers, by people who just wanted to check somebody's feed without feeding the machine.
On August 24th, X Corp sent cease-and-desist letters demanding a permanent takedown — not just of the Nitter instances, but of the project's code repository itself. And the message the creator, a developer who goes by Zedeus, posted to the site is honestly a little poignant. Let me read it, because it's short. "On 24 August 2026 cease and desist letters have been sent by X Corp. demanding a permanent takedown of Nitter instances and the project's repository. nitter.net is offline and development has stopped for the time being. I'm seeking legal advice and won't be commenting further on the specifics for now. Thank you to everyone who used, hosted, packaged, donated and contributed to Nitter over the past seven years."
Seven years, and a lawyer's letter shuts it down over a weekend. The legal theory X is running: unlawful use and circumvention of their API, scraping their data, accessing accounts and session tokens against the rules. They cite the Texas Harmful Access by Computer Act and the Lanham Act.
Now here's the builder lesson, and it's a hard one. This isn't new and it isn't just Elon. As the piece notes, Meta has hauled scrapers into court for years. Most of the big social networks now wall off third-party readers and force you into the app where they can track you and show you ads. The open, permissionless internet where you could build a nice little clean reader on top of somebody's public data — that door has been closing for a long time, and this is one more bolt sliding shut. If your business depends on somebody else's API and their goodwill, understand that goodwill is a terms-of-service change and a lawyer's letter away from evaporating. Nitter connected to a real X account after the last crackdown in 2024 just to survive. Didn't matter. The platform holds the keys, and eventually the platform decides it wants its house back.
Let me take you somewhere more hopeful for a minute, because it's not all cease-and-desists and burning robots. Stability AI — the folks behind Stable Diffusion — raised seventy-six million dollars in a Series B. Brings their total to two hundred thirty-two million. And what makes this one interesting isn't the number, it's the roster.
The money came partly from — get this — the entertainment industry itself. Universal Music Group. Sony Music Group. Warner Music Group. And the gaming giant Electronic Arts. Plus two investment firms, AMD Ventures and Pacific Alliance Ventures. Lucas Ropek at TechCrunch called it an unusual roster, and he's right — this isn't your typical venture round. This is the companies Stability now depends on for content licensing and distribution writing the checks.
Now think about what that means. A couple years ago, generative image AI and the big content owners were on a collision course. Lawsuits flying. Stability actually got sued by Getty Images in the UK for training on their pictures — and largely won that one, though a similar Getty suit is still grinding through the US courts. But here's the strategic pivot. Instead of just fighting the music and gaming giants, Stability signed deals letting them co-develop the tools. Universal Music and EA back in October, Warner in November. So now those same companies are investing. They've got a hand on the wheel instead of a foot in the courtroom.
For founders, that's a real strategic pattern worth noting. When you're building AI on top of somebody's copyrighted world, you've basically got two roads: litigate, or partner and give them a seat at the table. Stability's betting the partnership road builds a more durable business — the entertainment industry as your investor instead of your plaintiff. We'll see if it holds. But it's a smarter play than pretending the rights holders don't exist.
Now let me touch a couple of the infrastructure and geopolitics threads quickly, because they matter to how you build.
Mistral, the French lab, announced a strategic collaboration with HUMAIN to push what they're calling sovereign AI into Saudi Arabia and across the Middle East. This is a deal in the hundreds of millions of euros — AI infrastructure, model development, deployment, with an initial focus on cybersecurity and voice, plus frontier models that work well in Arabic. And the through-line of their whole pitch is that word: sovereignty. AI where the data, the compute, the operations stay under the customer's control. Data stays inside customer-defined boundaries. Models run on open weights the customer can own and adapt. Training and inference on infrastructure in a jurisdiction the customer picks.
Now, I want to be careful here — this is a company blog post and a forward-looking announcement, so treat it as intent, not a fact on the ground. But the theme is real, and it's one to file away. Mistral did the same thing earlier this summer with an expanded Microsoft partnership for European compute, and they rolled out this idea of European Compute Units. The pattern is: the model isn't the whole product anymore. The whole product is the model plus where it lives plus who controls the learning loop. For regulated industries — finance, telecom, government, healthcare — that control question is becoming the entire sales pitch. If you're selling AI into those spaces, "trust us, it's on our cloud" is a shrinking answer.
And speaking of compute and control, Waymo said it's headed to Munich. Alphabet's robotaxi outfit is expanding into Germany, which more than five years ago became the first EU country to build a legal framework for Level 4 autonomous driving — that's the tier where the car does everything with no expectation a human takes over, within certain conditions. Waymo doesn't even have its German permits yet; they're going to do the whole slow dance — manually map the streets, test with safety drivers, then pull them out — and they don't expect to open commercial rides to the public until the end of 2027. So this is a multi-year runway.
Why does a founder care about a robotaxi in Munich? Because it tells you where the regulatory door is open. Germany built the legal framework and became a testing hotspot. The UK, especially London, is shaping up as the other big battleground, with Waymo squaring off against Wayve and Uber, and even Baidu testing over there with Lyft. The map of where autonomy is legal is the map of where that industry's money and talent will pool. Regulation isn't just a brake. Sometimes it's the on-ramp, and the countries that build the framework first get the traffic.
Alright, before I let you go, let me clear a couple things off the counter that you should know about but that don't need a full sit-down.
If you're a gamer, or you've got gamers in your life — heads up. There's a scam going around targeting people desperate for Grand Theft Auto VI. Fans have been waiting over a decade since GTA 5, the game's been delayed a bunch of times, and now it's expected November 19th. So cybercriminals set up fake websites impersonating Rockstar, advertising a playable GTA 6 demo. There is no demo. You click "Play Now," you download what looks like a game installer, and it's actually an information stealer — grabs your passwords, your cookies, your logged-in sessions right out of your browser. And the ugly part, per the cybersecurity firm Malwarebytes: with those stolen session cookies, attackers can sometimes log in as you even if you've got multi-factor authentication turned on. So the old rule holds, kiddos: don't download the thing if you're not dead sure where it came from. Desperation is the oldest attack vector there is.
And one more product note that's actually a nice bookend to my deep dive. Instagram rolled out a feature called First Draft. It automatically trims your video clips, cuts the pauses, and stitches the best moments into an initial cut of a Reel — a first pass in under ten seconds. Now, as TechCrunch honestly points out, this isn't revolutionary — Adobe's Firefly has Quick Cuts, CapCut has Auto Cut, this stuff exists. But the smart move is that Instagram baked it right into its own app, so creators have one less reason to move their footage into a separate editor. That's the whole strategy: reduce the friction, keep them inside your walls. Same lesson as the Claude memory story, actually — remove the annoying step, and you own the workflow.
I'll note one bit of timing that's hard to ignore, because TechCrunch flagged it too: Instagram announced this creator-friendly feature on the very same day Instagram head Adam Mosseri was expected to testify in Meta's social media addiction trial — the lawsuit accusing Meta of designing these platforms to hook kids. Meta rejects the allegations. Make of the timing what you will. I just report the menu; you decide what smells funny.
So let me tie the whole thing together before I sign off. We started with a doctored picture of an antibody. We went to a robot that sprints beautifully and can't fold a shirt. We landed on a memory update that nobody's going to make a viral video about but that quietly makes an AI assistant actually usable. And it's all the same lesson, the one I keep coming back to on this show: the demo is easy, the deployment is hard, and the difference between the two is where all the real value — and all the real cons — live. The pretty picture is not the product. The sprint is not the business. The laundry is the business. Check the brightness and contrast on everything.
That's the menu for today. You've been listening to Barely Possible. I'm Tony DeLuca, I appreciate you spending a little of your Tuesday with me, and if you build something today, build the part that remembers. Take care of yourselves out there, and I'll catch you next time.