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
Hey hey, what is good everybody — I'm your boy Tony DeLuca, and you are locked into Barely Possible. Memorial Day weekend edition, which means the news cycle slowed down exactly enough for some genuinely interesting stuff to bubble up. Let's get into it.
Alright, so the story I want to lead with today isn't a lab announcement or a funding round. It's Wix. Yes, Wix, the website builder you probably haven't thought about since 2019. They just did their largest restructuring ever — reportedly laying off somewhere between 800 and 1,000 people, which is about 20 percent of their workforce. And look, on the surface that sounds like a standard AI-displaced-jobs story. But sit with the numbers for a second and it actually gets more complicated and more interesting than that.
Wix's core business is growing. Revenue reportedly up 14 percent year over year in Q1 2026. Bookings up 15 percent. Their new AI-driven user cohorts are showing even faster growth than that. So this isn't a company in decline cutting to survive. This is a profitable company cutting while growing, because the cost structure of becoming an AI-first software company is brutal.
Here's what's eating them. They acquired Base44 last year — that's the vibe-coding startup. They're building and running proprietary AI models. They've got massive compute and inference costs. They launched Wix Harmony, which is their AI-native creation platform. And somewhere in there they also executed a controversial 1.6 billion dollar share buyback right before the downturn, which, yeah, that's the kind of decision that ages poorly fast.
But the real story here is what it tells you about the new economics of being a software platform in the AI era. The old moat for Wix was templates plus drag-and-drop builders. That moat is basically gone now. If you can describe a website in plain English and get something functional back in thirty seconds, the drag-and-drop editor becomes irrelevant as a differentiator. The new moat — and this is what Wix is actually betting their restructuring on — is AI orchestration plus hosting plus payments plus integrations plus reliability plus distribution. In other words, all the boring infrastructure stuff that pure AI generation tools still can't touch.
One commenter put it pretty cleanly: AI is turning UI and features into commodities while making infrastructure and distribution even more valuable. That's the frame. The demo is cheap now. Production is not. And what Wix is doing isn't really getting killed by AI — it's paying the price of an expensive transition toward the thing that will actually matter on the other side of it.
For builders, here's why this matters. If you're building a product that competes on the interface layer — on how pretty or how slick the front end is — you are in a race to the bottom right now. The products that survive this transition are the ones combining generation with reliability, with workflows, with ecosystem lock-in, with real customer relationships. That's not a prediction. That's what we're watching Wix bet a billion dollars on in real time.
Now, connected to that story, there's a separate data point worth mentioning. Ramp — the corporate card and spend management company — put out their AI Index for May 2026, and the headline getting attention is that Anthropic is now beating OpenAI on business adoption by their metrics. Claude is apparently the dominant AI tool in Ramp's customer base right now, which is a slice of actual company spend data rather than survey responses.
Now, I want to flag this with some healthy skepticism. There's already a podcast that came out challenging Ramp's methodology pretty hard — arguing their numbers are biased toward a certain customer profile and that the math doesn't hold up under scrutiny. So treat this one as a data point, not a verdict. But what it does surface is something that feels real from the ground: a meaningful portion of enterprise buyers are routing work to Anthropic's Claude, and the reasons people give are interesting. Some of it is vibe — people perceiving Anthropic as more trustworthy or more aligned with their values. Some of it is genuinely product-driven. Claude Code in particular keeps getting mentioned in ways that sound less like marketing and more like people actually using it and finding it useful.
If you're building tools that integrate AI, the provider mix matters. A world where neither OpenAI nor Anthropic has a lock on enterprise customers means your integration bets stay complicated. OpenRouter continues to be mentioned in practitioner circles as a way to hedge — access to all the models under one account, without being locked to any one provider's billing relationship.
Now let me shift gears to something that was genuinely the most interesting research-side story this week, and I want to give it a little more room to breathe.
There was a debate — Pathway hosted it — called the Post-Transformer debate. The setup is already interesting: one of the original authors of Attention is All You Need, the 2017 paper that essentially gave us the transformer architecture and everything that follows from it, argued at this debate that it's time to move past transformers. That's not a minor moment. That paper is the foundation of basically every large language model in production today.
The framing I found most interesting came from a commentator named Adrian from Pathway. He said the field hasn't had its PageRank moment for intelligence yet. The analogy: Google didn't just build a better AltaVista. They found the underlying mathematical theme — the idea that links between pages encode importance — and that insight restructured everything. His argument is that transformers are a very good implementation, but they might not be the discovery of the actual underlying theme of intelligence. We might still be waiting for that.
Whether you buy that or not, it's a genuinely interesting frame. The counterpoint in the room was delivered via a casually dropped fact: that GPT-5.5 had recently solved an Erdős conjecture that had been open for 60 years. One of the attendees noted the room went quiet for a second after that. The implicit argument being: the existing architecture is still producing results that look a lot like discovery, so maybe we haven't hit the wall yet.
This connects to something Demis Hassabis said back in January, which has been resurfacing in discussion this week. The clip is making the rounds again. His position was that solving Erdős problems, while genuinely impressive, is still a far cry from true invention in the mathematical sense. When people push him on this, he reaches for comparisons to Ramanujan — someone who wasn't just solving problems but breaking paradigms, finding structures nobody had imagined looking for. His calibration is that current AI has shown it can do non-trivial mathematical reasoning. It has not shown it can do the Ramanujan thing. And he's not a hype guy, so when he says that, it's worth paying attention to.
The honest position here is: we don't know. The transformer might be the PageRank of intelligence. Or it might be the AltaVista — genuinely impressive, genuinely useful, but not the final architecture. Watching one of the original authors advocate publicly for moving beyond it is at minimum a sign that serious people think the search space isn't closed.
Okay, let's go deep on something that I think is practically the most useful thing we can cover this week for founders and builders. There was a piece from the AI Daily Brief this week — Nathaniel Whittemore's show — where he had on Nufar Gaspar, who runs AI executive education programs and has been training leaders across about 30 countries. The premise is simple: most executives fall into one of three traps when it comes to AI, and all three traps leave enormous value on the table.
Gaspar names them the Podcast CTO, the Weekend Tinkerer, and the Manifesto Writer. The Podcast CTO knows every release, every benchmark — but has never actually built a system for their own work. The Weekend Tinkerer is building stuff in their private time but hasn't figured out how to bring it into their actual workflow. And the Manifesto Writer has the vision, funded the transformation committee, but personally hasn't crossed the threshold of believing AI can operate at their level.
Her argument is that the CEO's quality of AI usage is the single biggest predictor of how well their teams adopt AI. When the CEO is the best user in the building, the organization moves fast. When the CEO talks the talk but hands off the actual work, they tend to go wrong in both directions — underestimating what AI can do in some areas, setting unrealistic expectations in others.
Now what I found most useful in this conversation is the framework she builds around what she calls four digital team members every executive should hire. Let me walk through these because they're concrete in a way that a lot of AI advice isn't.
The first is a research analyst. Not a search engine replacement — an analyst you actually brief. Her point is that most people type a question into an AI tool and accept whatever comes back, which is using it like a slightly smarter Google. The real move is treating it like a staff researcher: tell it what you're trying to decide, what assumptions you're bringing, what sources to trust and distrust, what time horizon matters, what to exclude. You're not asking a question, you're briefing someone. One technique she recommends is cross-model verification — send the same research prompt to multiple models in separate sessions, then use a third model to fact-check where they converge and investigate where they diverge. The logic: 100 percent consensus across independent runs is more likely to be factual than something only one tool surfaces. It's a workaround for the hallucination problem that doesn't require you to check every source yourself.
The second digital hire is what she calls the strategic thought partner — essentially a board of advisors. Her suggestion is to build multiple AI personas with different decision-making styles, not a single AI voice. Have them debate a decision between themselves before they give you the result. The goal is calibrated pushback, not a devil's advocate for sport and not sycophantic agreement. She's very specific that you should also instruct the AI to surface not just the biases you might have, but the biases the AI itself might have in the analysis. That self-aware loop is something a lot of people skip.
The third is a communications expert. This one's interesting because she names the failure mode exactly right: the distance between AI writing that sounds like everyone and sounds like no one, versus AI writing that sounds like a specific person — that whole gap is about how well you've steered the tool, not the tool's capability. Her technique is style profiling. Feed the AI your actual best writing — board updates, important emails, whatever feels most like you. Have the AI name the patterns it sees: rhythm, sentence structure, rhetorical tendencies. Use that as a living style guide. And when you give feedback on drafts, score on specific dimensions instead of saying you don't like it. Clarity gets a nine, wittiness gets a five, conciseness gets a seven. AI is goal-driven and moves better toward a specific target than toward vague dissatisfaction.
The fourth is the operational powerhouse — the part that handles meeting prep, status synthesis, stakeholder relationship tracking, the morning brief. Her key insight here is not to think about automating what you already do, but to think about what you would build if you had unlimited headcount. What visibility have you always wanted that was previously infeasible? She also has a rule worth remembering: never automate before you've tested manually and repeatedly. Run the morning brief yourself for a week or two before you commit to having it run automatically. Only after you've seen how you actually consume the data do you know what's worth keeping.
The throughline of all four is context. The more undocumented context — the relationship dynamics, the half-formed intuitions, the thing that was said before the recording started — that you push into the AI, the better the output. Her operating principle one is: speak rather than type, because typing filters your thinking. Her operating principle two is brain dumping, habitually, as a practice. Her operating principle three is having AI interview you before any complex task. Her operating principle four is separating planning from execution — have the planning conversation first, in a separate session. And her operating principle five is identifying your own intervention point: the place in the workflow where your judgment specifically adds the most value, and then designing everything else around protecting that moment.
The capstone she names, for people who've gotten all four individual team members working, is building a chief of staff agent that has a cross-view of your decisions, communications, and priorities — something that orchestrates across all four rather than siloing them. She says you should earn that by getting real mileage with each individual one first.
I want to connect this to something we've been tracking on Barely Possible. We covered a few weeks back how companies like Every — the publication and product company — are moving away from individual personal agents toward shared team agents. Personal agents are high maintenance; when they break, the individual person has to fix them. Team agents, shared across multiple people whose work overlaps, solve both the maintenance problem and the continuity problem. A shared agent retains institutional knowledge even when people leave. That's the org-level version of exactly the same insight Gaspar is making at the executive level: the value is in the context that gets accumulated, not in the tool itself.
Now let me give you a quick tour of a few other things worth noting.
On the open guardrails front, the Financial Times ran a story this week on Heretic — the tool that lets you strip guardrails from Meta's Llama models. The FT says they were able to do it themselves in under ten minutes using publicly available tools on GitHub, without any specialist hardware. The creator of Heretic, Philipp Emanuel Weidmann, told the FT that the software has been used to create more than 3,500 decensored models since release, and that modified versions have been downloaded 13 million times. Weidmann posted about this himself on Reddit and said the mainstream press attention is the first of multiple recent inquiries he's received, noting he has no interest in being an influencer or politician but also recognizes that refusing interviews means the story gets told entirely by people who misunderstand the technology.
The community reaction to this is interesting. Some people are framing it as the first domino toward a legislative push against removing model guardrails. Others point out that the actual use cases for uncensored models are broader than the obvious ones — reverse engineering, financial analysis where the model won't hedge with disclaimers, historical research on sensitive topics. One commenter put it directly: uncensoring in LLMs means eliminating all refusals on any topic. The spicy roleplay use case is a byproduct, not the point. This is going to stay in tension for a while.
On the agent infrastructure side, there's a thread worth flagging about audit trails. The argument is simple and I think correct: the next useful version of AI agents isn't the one that acts most independently. It's the one that makes every step legible enough that a normal user can trust what it did. When an agent is just answering a question, you can judge the result. But when it's operating across websites and accounts and forms and email, you need to know what it clicked, what it submitted, where it failed, when it decided to stop. Without that audit trail, even a capable agent feels opaque and therefore hard to trust. And audit trails solve a second problem the thread points out: handoff. When an agent passes a task to a human or another agent, the trail is the only thing that carries context forward without re-prompting from scratch. Autonomy without observability is a liability you haven't priced yet.
Related to that, there's a small open-source project called Spice worth keeping an eye on. The premise is a decision layer that sits above execution agents. Instead of just seeing the output of an agent task, it tries to preserve the reasoning boundary before execution: what the system believed, what it chose, why it chose it, what tradeoffs it rejected, whether execution needs approval, what happened after, and how that outcome should affect the next decision. It's early — very early — but the problem it's pointing at is real. Most current agent systems are execution engines. The question of what happens before execution — the deliberation layer — is still pretty underbuilt.
On the fastest-growing repos list this week, a few things jumped out. CodeGraph — a pre-indexed local code knowledge graph that works with Claude Code, Codex, Cursor, and Hermes Agent — gained 14,000 stars. The agentmemory project, which is about persistent memory for AI coding agents, gained nearly 7,000 stars. 12-factor-agents — principles for building production-grade LLM-powered software, kind of the 12-factor app methodology translated for agents — gained close to 2,000. And CloakBrowser, which is a stealth Chromium that passes bot detection tests with Playwright compatibility, gained 7,000. That last one is obviously a tool for browser automation agents, and the combination of fast growth and bot-detection-bypass capabilities is something to watch.
One more thing before we close, and this one is small but kind of delightful. Simon Willison — who's been building Datasette for years — posted about a new feature in Datasette 1.0 alpha 30. It's a jump-to menu, available via the slash keyboard shortcut, that lets you type to navigate to any database, table, or canned query. It also offers a plugin hook so extensions can add additional content to it. The datasette-agent plugin already uses it to offer a form for kicking off new agent conversations. This is a nice example of what good developer tool design looks like — keyboard-accessible navigation, extensible via plugins, and the agent integration feels like a natural addition rather than a bolt-on.
Also from Simon Willison, a sharp observation about the water consumption debate around AI. The original bottle-of-water-per-generated-email estimate relied heavily on guesses about GPT-4's architecture. His point: GPT-4 is now a retired model, over three years old. It would be in OpenAI's interest to just publish the architecture at this point. The environmental impact numbers could then be grounded in something real instead of speculation. It's a transparency argument dressed up as a PR argument, and it's not wrong.
Alright, that's what we've got today. The big themes shaking out this week: the economics of becoming an AI-first software company are genuinely brutal even when revenue is growing, as Wix is demonstrating in real time. The executive AI gap is more about undocumented context and deliberate workflows than about tool choice. Agent infrastructure is quietly bifurcating between execution capability and observability — and the latter is what actually makes these things deployable. And the transformer debate is alive enough that one of the original authors is publicly calling for the field to move past it, even as the architecture is still producing results that surprise people.
That's Barely Possible for May 26th. I'm Tony DeLuca. Watch your audit trails out there.