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, grab whatever's keeping you warm this morning, and let's get into it. I'm Tony DeLuca, your neighborhood voice for the stuff that actually changes how you build, and today we've got a menu that's heavy on an idea nobody wants to say out loud: the fastest-moving AI company on the planet just told us that the real bottleneck isn't the model. It's you. The human being. Sitting there. Reading. Trying to keep up. Buckle up.
Let me tell you how this is gonna go. We're gonna open with Anthropic's report on what happened to its own workforce, because it's the most consequential thing on the table for anybody running a company right now. Then we're gonna pull back and look at a piece of meeting science that lands on exactly the same nerve from a totally different direction. After that we've got Meta walking back some terms-of-service language that should make every founder with a website pay attention, a couple of fast enterprise-adoption notes worth your time, a little palate cleanser from the car world and outer space, and we'll close out. Fair warning, a bunch of today's material is older stories that got fresh attention, and I'll flag that when it matters, because I'm not gonna sell you a June story as this morning's news.
So let's dig into Anthropic and what they found out about themselves.
On Friday, Anthropic published a report on how its own workforce and its own work have changed over the past year. And the framing matters here, so I want to be precise about it. Anthropic argues that because the company has, in their words, thoroughly molded its workflows around AI, its internal trends could be an early indicator of how AI shifts work organization, hiring, and labor demand across the economy. Now, you should hold that claim at arm's length. This is a company with a product to sell telling you that the future looks like a company that uses a lot of that product. That's not a neutral narrator. But even if you strip out the sales pitch, the numbers underneath are worth sitting with, because they're specific, and specific is where the music is.
Here's the headline that should make every junior engineer and every hiring manager lean in. The proportion of junior-level employees in Anthropic's technical departments fell from twenty-one percent to fourteen percent since early 2025. And the proportion of junior technical hires — the people they're actually bringing in the door — fell way more sharply, from twenty-seven percent down to nine percent. Read that again. Twenty-seven to nine. That's not a trim. That's a company deciding, at least for now, that it mostly wants senior people.
And technical teams as a share of the whole company dropped too, from sixty-one percent to fifty-six percent. Anthropic attributes that partly to normal business functions expanding as they scale — sales, legal, operations, the stuff every growing company piles on — and partly to this line, which I think is the quiet tell: coding tools increasingly let non-technical employees make software themselves. So the marketing person builds their own little internal tool. The ops person wires up their own dashboard. You don't need to hire an engineer for every piece of software anymore, because the non-engineers are shipping software now too. Whether it's good software is a separate question, and we'll get there.
Now here's the part that connects to our whole episode. One employee described, quote, an eternal pressure at Anthropic to move as fast as possible and use Claude for everything. In an April 2026 survey, twenty-seven percent of employees said they'd gotten more productive because of Claude. Fine, expected. But twenty percent said their job had shifted from actually doing work to managing Claudes. And one engineer compared it to babysitting. Listen to this quote, because it's the one that stuck with me: "When you are babysitting, every child brings to you concerns with an equal demand for your attention and the various Claudes also do that."
Every child demands your attention equally. That is a very honest description of what it feels like to run a bunch of agents at once. You're not writing the code. You're the daycare supervisor, and seven toddlers are all tugging your sleeve at the same volume. And the report says human attention and judgment are becoming — their word — a bottleneck for productivity, because Claude produces a deluge of text that somebody still has to read, organize, and decide about.
So think about what that actually means for the economics. The thing you cut was the junior person who used to absorb grunt work. The thing that got infinite was the output. And the thing that's now scarce — the actual constraint on how much your company can do — is a senior human's finite attention and ability to say "this is right, that's wrong, ship this, kill that." You didn't remove the bottleneck. You moved it. You moved it from "can we produce enough" to "can a trusted human review enough." And trusted humans with good judgment, it turns out, are exactly the people you stopped hiring at the junior level, which raises the obvious question nobody at Anthropic wants to answer on the record: where do the senior people of 2031 come from if you quit training juniors in 2026? You can't skip the apprenticeship and still expect the masters to show up.
And here's the kicker, the detail that made me laugh out loud. Meeting time per employee is up twenty percent compared to a year ago. Up sixteen percent for the technical contributors. The most AI-saturated company in the industry, the one telling you AI makes everything faster, is spending more time in meetings than it was a year ago. Anthropic's explanation is that it's about human trust and accountability, with employees repeatedly describing face-to-face coordination as indispensable. Which, okay. When the machines generate a flood and the judgment is the scarce resource, what do humans do? They get in a room and argue about what's real. The meeting didn't die. The meeting came back with a vengeance.
They do offer one honest caveat, and I'll give them credit for it: more code per person, more sales closed per rep, more contracts processed. But they say right in the report, quote, more code does not necessarily mean better code or business impact. Though they add they haven't found signs code quality is declining. As we covered a few days back, there was that Harvard and Jellyfish study making the rounds showing AI agents producing something like thirty percent more code without producing thirty percent more actual software. So "more code per person" and "more software" are not the same sentence, and I'd want to see the shipped-features number, not the lines-of-code number, before I throw a party.
Here's the builder takeaway, and I want to be concrete because that's the only thing worth your time. If you're running a small company and you're tempted to copy this — stop hiring juniors, point everybody at the agents, go go go — understand what you're actually buying and what you're actually paying. You're buying throughput. You're paying with review capacity and with your own future bench. Before you restructure around a flood of machine output, ask the boring question: who reads it, who's accountable for it, and does that person have the hours in the day. Because Anthropic just told you, in their own numbers, that the answer to "who has the hours" is "nobody, and that's now the ceiling on the whole operation."
Now, here's the connection I promised, and it's a real one, not a forced one. The Anthropic report says the bottleneck is human attention and judgment, and that the response is more meetings and more face-to-face coordination. Well. Right alongside that, there's a recent piece out of Knowable Magazine, written by Lesley Evans Ogden, with the headline that practically sings: science says your boss should talk less in meetings. So if meetings are coming back — and at the most AI-forward company they clearly are — then the question of whether we actually run them well stops being a human-resources nicety and starts being an operational problem. Let's dig into that.
This piece is a window into something called meeting science, which, believe it or not, is a real and growing field among organizational and social psychologists. And the field is accelerating for a reason that should sound very familiar after the Anthropic story: all those archived virtual meetings, all that recorded Zoom audio and video, gives researchers vast piles of data to study. The machines are watching the meetings now too.
Let me give you the scale first, because it's genuinely staggering. Close to one billion meetings take place every day around the world. CEOs spend nearly three-quarters of their workdays in meetings. Three-quarters. And for decades, nobody really studied whether any of it worked. One of the researchers, organizational psychologist Liana Kreamer at the Florida Institute of Technology, who coauthored a thirty-year review of the field, puts it nicely: you can use a meeting as a stage to study anything you want, because meetings are a fundamental part of how we organize as social beings.
Now here's the finding that gives the headline its teeth, and it ties straight back to that Anthropic daycare problem. Nale Lehmann-Willenbrock, an industrial and organizational psychologist who directs the Center for Better Work at the University of Hamburg, found in video-recorded meetings that when leaders asked questions and actually, attentively listened to the answers, team performance went up. When leaders asked questions but didn't really listen — just waited for their turn to talk — it didn't. And here's the line that should make every manager squirm a little: "Often, leaders talk the most, which is also why they often think their meetings are pretty good." The person running the meeting walks out thinking it was great, precisely because they did all the talking. The people who sat there muted thought it was a waste. Same meeting. Two completely different verdicts.
And active listening, she points out, basically got murdered by remote work — her word is "killed completely" — because we've all trained ourselves to hit mute when we're not talking. You mute, you check your email, you're gone. The engaged back-and-forth that made meetings worth anything evaporates.
There's a bunch of practical stuff in here that's actually useful if you're the one calling the meetings. One technique Kreamer swears by is silent brainstorming — some researchers call it brainwriting. Instead of the loud free-for-all where the three most confident people in the room dominate, everybody writes into a shared doc or chat at the same time. A 2009 study found it beat traditional spoken brainstorming. And the reason it works is the quiet people, the shy ones, the ones who hate interrupting, they finally contribute, because they don't have to fight for airtime. If you've got a team where the same two voices win every discussion, that one's free and you can use it Monday.
There's also work from a doctoral student, Zofia Walters, on neurodiverse participants. In a survey of a hundred and forty-seven employees, she found people with ADHD reported more meeting fatigue, especially when the camera was forced on for long stretches, and poorer emotional states afterward. And interestingly, her findings suggest multitasking during meetings might actually help those attendees manage boredom. So the "cameras on, eyes forward, no multitasking" orthodoxy might be quietly making things worse for a chunk of your people. These are findings not yet published, so hold them loosely, but it's a thread worth pulling.
Now I want to be fair and give you the counterweight, because the piece does, and it's the most interesting part. Everybody dunks on meetings. But a 2026 study of more than nine thousand workers in Norway, led by Harvard economist David Deming and based on data from the survey company Norstat, found that meeting intensity was positively associated with later wage growth and on-the-job learning. The theory is that meetings put you in the room with more senior colleagues, and that's how knowledge actually transfers — you learn by being there. Now, the big caveat, stated plainly: the study is correlational. It can't prove the meetings caused the raises. Maybe the people getting invited to meetings were already on the up. But the authors land on a line I'm going to be stealing for the rest of my life: "Meetings are the broccoli of work — widely disliked, but probably good for us anyway."
And here's why I paired these two stories, and it's not a cute transition, it's the actual argument. Anthropic just told us that in an AI-saturated company, the scarce resource becomes human judgment, and the response is more face-to-face coordination. Meeting science tells us that face-to-face coordination is also exactly how a junior person becomes a senior person — the Norway study's whole point is that the room is where the learning happens. So the company that cuts juniors and leans on senior judgment is, maybe without noticing, kicking out the ladder. The meetings that transfer knowledge to the next generation only work if there's a next generation in the room. Cut the juniors, keep the meetings, and you've got a room full of senior people with nobody to teach. That's not a prediction. That's just two data points pointing at the same wall.
Now let's shift from how you work to something that could bite your company directly. Let's talk about Meta and a terms-of-service scare.
Here's what happened, and it's a tidy little case study in reading the fine print. Meta recently updated its commercial terms of use with language saying the company could use any content from a business account's domains, web pages, or URLs to — and I'm quoting the verbs because the verbs are the whole story — provide, develop, improve, research, train, and maintain the integrity of Meta products and services. Legal experts interviewed by Nieman Lab flagged this as a remarkable provision. Because read plainly, it looked like it could give Meta permission to scrape the website of any publisher or business that has a Meta business account — and let's be honest, that's basically every company that runs ads on Instagram or Facebook — and use that content to train Meta's AI models.
Think about that for a second. You run a startup. You've got a Meta business account because of course you do, you need to advertise. And buried in the commercial terms is language that a reasonable lawyer reads as "we can crawl your whole site and feed it to our models." For context, Meta beat a high-profile copyright infringement lawsuit last year, but other cases against AI companies, including Meta, are still grinding along. So this isn't happening in a vacuum. It's happening while the courts are actively fighting over exactly this question.
Meta's spokesperson initially said the updated terms didn't actually reflect any change in how Meta crawls the web. And then — and this is the fresh part, reported by The Verge — the provision got removed. Meta's statement: "We've heard you and want to clarify: the crawling terms we consolidated only applied to our ads and business products, and we've clarified that in the terms themselves. It does not change how Meta uses web crawlers."
Okay. So here's my read, and I want to be careful not to overstate it. The honest version of events is Meta says it was just consolidating existing terms, somebody drafted it too broad, legal experts raised the alarm, and they walked it back. That's the charitable read and it might well be true. But the lesson for you doesn't depend on which read is right. The lesson is: the scary language was in there, it took outside legal experts reading it closely to catch it, and it only came out after public pressure. Nobody's inbox lit up with a friendly heads-up. The default was the broad version, and the narrow version only arrived after people yelled.
So what do you actually do with this. One, somebody at your company — or your lawyer — needs to read the terms of the platforms you're locked into, not just click accept. Two, understand that when a platform you depend on for distribution also has an AI lab, your content and your platform access are quietly connected in ways that don't show up in the sales pitch. And three, this stuff moves in your favor only when somebody's paying attention. The system that protects you here is a Nieman Lab reporter and some lawyers reading boilerplate on a Friday. That's thin. Build accordingly, and keep a copy of your own stuff somewhere the platform can't quietly claim it.
Now let's move from the fine print to the adoption stories, because there were a few quieter enterprise notes this week that tell you where this is actually landing in regulated industries.
OpenAI put out a couple of these customer pieces, and I'll keep them short because they're short on substance, but the pattern across them is the real signal. One's about Benefit Street Partners, where a builder named Anant Kumar is using Codex for code review and faster follow-ups on tools that compare loan valuations and analyze credit agreements. Another's about Itaú, one of the big banks in Brazil, building generative AI across its entire customer base — they've got a conversational interface and, notably, a governance platform they call IARA, applying what they call an AI-first strategy in one of Brazil's most regulated sectors. And there's one about Block building AI-powered security, with their team investigating threats faster and proposing patches in minutes.
Now, these are vendor case studies. Grain of salt the size of a fist — nobody publishes the one where it went sideways. But the thing I want you to notice is what shows up right next to the AI in the Itaú story: a governance platform. In the most regulated corner of a regulated industry, the AI doesn't go in alone. It goes in wearing a governance harness. And that tracks perfectly with everything we just talked about from Anthropic — when the output becomes a flood, the scarce, valuable thing is the layer that reviews, controls, and vouches for it. Benefit Street using Codex for code review, Itaú wrapping the whole thing in IARA, Block proposing patches a human presumably signs off on. The adoption that's actually sticking in serious industries isn't "let it rip." It's "let it run inside a cage somebody trusts."
There's also a nice little exchange that went by on X worth thirty seconds. Simon Willison — who you know if you follow the practitioners — responded to a conversation about whether you'd want a coding model to avoid open source libraries with a very dry line: "Avoiding open source libraries? Why would you want a model to do that?" And it's a small thing but it's a real point for anyone building with these tools. The whole value of a coding assistant is that it reaches for the well-worn, battle-tested library instead of hand-rolling something janky from scratch. If your instinct is to wall the model off from open source, you're throwing away most of what makes it useful and signing up to maintain a pile of bespoke code nobody else has ever stress-tested. The standard stuff is standard because it works.
Alright. Let's lighten it up before we close, because you can't live on workforce anxiety alone. Let me give you a couple from outside the AI blast radius.
First, a resurfaced one, and I'm flagging it clearly: this Tesla Roadster story is from back in June, it just got renewed attention, so don't take it as this-morning news. The gist, as Ars Technica's Jonathan M. Gitlin told it, is Tesla taking another swing at the Roadster — nine years after the first attempt — and leaning on roughly fifty-year-old stunt technology to make a show car do tricks in the air. There are four big exhaust pipes on the back of an electric car, which, I know, but they're not exhaust — they're where jets of compressed nitrogen shoot out to flip the thing around. The claimed numbers are the usual eye-watering fantasy: a top speed north of three hundred eighty miles an hour, zero to sixty in under a second. And Gitlin does the thing a good reporter does, which is point out there's an EV you can actually buy, the McMurtry Spéirling, that already does more — pulls more lateral G, holds real records, and will genuinely drive upside down. The reason I even mention it: it's a good reminder that a flashy demo and a shipping product are different species, and that gap is the whole game, whether you're selling a hypercar or a frontier model. The demo flies. The product has to show up in your driveway.
And one that's actually good news, also a resurfaced piece from Ars, this one by Stephen Clark, about NASA's Venus missions. For years these things were on the chopping block. Now DAVINCI — the probe that'll drop through the Venusian atmosphere and send back the first direct surface images since 1982 — is targeting a launch in December 2030, pulled a full year earlier than planned. And VERITAS, the radar mapper, is funded with a launch date of November 2032. The deputy project scientist at JPL, Dan Nunes, said simply, "VERITAS is going. We're staffing up. We're accelerating." I just like that one. After a decade of Mars hogging the spotlight and budgets getting slashed, the planet next door is finally getting its generation of visitors. Something nice to carry into your day.
So let me tie the bow on this, because the through-line today is cleaner than usual. Anthropic — the company whose entire pitch is that AI makes work faster — published numbers showing that in their own house, the junior pipeline shrank hard, the output exploded, and the thing that's now scarce is a senior human's attention and judgment. And their own fix for that scarcity is more meetings, more face-to-face coordination. Then the meeting scientists come along and tell us those meetings are also, inconveniently, the mechanism by which juniors become seniors in the first place. And then Meta reminds us that the platforms we build on will quietly write themselves broad rights unless somebody reads the fine print. Put those together and the takeaway for a builder isn't "AI is overhyped" and it isn't "AI changes everything." It's narrower and more useful than that: the flood is real, the scarce thing is trusted judgment, and judgment doesn't scale by itself and doesn't grow if you stop planting it. Build your company like the review capacity is the constraint, because at the fastest company in the business, it already is.
That's the menu, kiddos. Read the terms, keep somebody in the room who can say no, and for heaven's sake let the quiet person in the meeting get a word in — they might be the one who saw the thing everybody else missed. I'm Tony DeLuca, this has been Barely Possible, and I'll see you back here tomorrow. Take care of each other.