AI in 15 — March 23, 2026

A developer ran a four-hundred-billion-parameter AI model on a MacBook. Not a cluster. Not a data center. A laptop, with five and a half gigs of RAM.

Welcome to AI in 15 for Monday, March 23, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Happy Monday, Marcus. We've got a great lineup to kick off the week. A solo developer built Flash-MoE and ran a massive model on consumer hardware in just twenty-four hours using Claude Code. Walmart is dumping OpenAI's ChatGPT checkout after dismal conversion rates. OpenAI just crossed twenty-five billion in annualized revenue and is eyeing a trillion-dollar IPO. Yann LeCun's new venture just raised over a billion dollars in Europe's largest-ever seed round. Atlassian is cutting sixteen hundred jobs and replacing its CTO with AI-focused executives. And a Chinese lab says it can match Claude Opus quality at one-twentieth the cost. Let's get into it.

Flash-MoE runs a three-hundred-ninety-seven-billion-parameter model on a MacBook.

Walmart says ChatGPT checkout converted three times worse than their own site.

And Yann LeCun raises a billion dollars to build world models for AI.

Marcus, let's start with Flash-MoE because this one blew up on Hacker News. A developer built a system that runs a three-hundred-ninety-seven-billion-parameter mixture-of-experts model on a MacBook using just five and a half gigabytes of RAM. And they did it in twenty-four hours. How is that even possible?

It's a clever combination of techniques that have been floating around independently but nobody had stitched together this cleanly. The key insight is that mixture-of-experts models don't activate all their parameters at once. A three-hundred-ninety-seven-billion-parameter MoE model might only use thirty or forty billion parameters for any given token. Flash-MoE exploits that by only loading the active experts into memory and keeping everything else on disk. Combined with aggressive quantization and smart caching, you get a model that fits in laptop RAM.

Five and a half tokens per second. Is that actually usable?

It's not going to replace a cloud API for production work, but for local experimentation and development, absolutely. You're getting responses in real time from a model that would normally require a server rack. And the fact that this was built in twenty-four hours using Claude Code is the story within the story. One developer, one AI coding assistant, one day. That's the kind of development velocity that would have been unthinkable even a year ago.

Three hundred and thirty-two points on Hacker News. The community was clearly excited.

Because it validates something a lot of developers have been hoping for. The gap between cloud-scale AI and what you can run locally is closing fast. Not through brute force hardware improvements, but through architectural cleverness. If you can run a frontier-class model on a laptop today, even slowly, imagine where we'll be in twelve months. The democratization of large model inference is happening from the bottom up, not the top down.

From democratizing AI to a very expensive AI failure. Walmart has pulled the plug on its ChatGPT-powered checkout experience after discovering that shoppers using ChatGPT to browse and buy converted at three times worse than Walmart's own website. They're replacing it with an in-house chatbot called Sparky. Marcus, what went wrong?

The fundamental problem is intent mismatch. When someone goes to Walmart.com, they're there to buy something. The entire experience is optimized for conversion, adding items to cart, checking out, done. When someone uses ChatGPT to shop, they're in a conversational mode. They're browsing, asking questions, exploring options. It's a fundamentally different user behavior, and it turns out that conversation is the enemy of conversion.

So the AI was too good at talking and not good enough at selling?

Precisely. ChatGPT would happily discuss product comparisons, explain features, suggest alternatives. All very helpful. But every additional turn of conversation was another opportunity for the shopper to not buy anything. Walmart's own checkout flow is designed to reduce friction. AI conversation adds friction by its very nature. Three times worse conversion isn't a bug in the AI. It's a feature of the format being wrong for the use case.

And Sparky is different how?

It's purpose-built for Walmart's specific needs. Less conversational, more transactional. Think of it as a smart search and recommendation layer rather than a chatbot you have a discussion with. It's a good lesson for the industry. General-purpose AI isn't always the right tool. Sometimes you need a narrow, optimized solution that does one thing well.

Despite Walmart's exit, OpenAI isn't hurting for revenue. They've crossed twenty-five billion dollars in annualized revenue and are reportedly eyeing a trillion-dollar IPO filing in the second half of this year. Marcus, put those numbers in context.

Twenty-five billion annualized is extraordinary growth. For context, when we covered the a16z consumer app rankings yesterday, we noted ChatGPT is at nine hundred million weekly active users. But here's the thing that should make investors think carefully. A trillion-dollar valuation at twenty-five billion in revenue means a forty-times revenue multiple. That's aggressive even by tech standards. It implies the market believes OpenAI will be generating well over a hundred billion in revenue within a few years.

Is that realistic?

It requires believing several things simultaneously. That ChatGPT maintains its dominance. That enterprise adoption accelerates dramatically. That margins improve even as compute costs remain enormous. And that competition from Anthropic, Google, and open-source models doesn't erode pricing power. Any one of those assumptions could break. All four holding? That's a very specific bet on the future. And remember, we just saw Walmart walk away from a ChatGPT integration because it didn't deliver results. If more enterprise customers start measuring actual ROI rather than just AI enthusiasm, OpenAI's growth story gets more complicated.

Speaking of competition, MiniMax released M2.5, which they claim matches Claude Opus quality at one-twentieth the cost. Open weights, Chinese-developed. Marcus, I can see your skepticism from here.

Let me be measured about this. The benchmarks they've published are impressive on paper. But we've seen this pattern repeatedly from Chinese AI labs. Announce a model that matches or beats Western frontier models, publish favorable benchmarks, release open weights, generate headlines. The actual independent evaluations often tell a different story. Cherry-picked benchmarks, narrow evaluation suites, performance that doesn't generalize.

You think it's overhyped?

I think the one-twentieth cost claim deserves serious scrutiny. If you can genuinely match Opus quality at five percent of the cost, that's not just a product announcement, that's a fundamental breakthrough in efficiency that would reshape the entire industry. More likely, it matches Opus on specific benchmarks while falling short on the kind of complex reasoning and nuanced tasks where frontier models actually differentiate themselves. The open weights are strategic too. China's AI industry benefits enormously from establishing their models as viable alternatives to Western APIs, particularly for developers in regions where geopolitics makes relying on American companies risky.

So wait and see on the independent evaluations?

Always wait for the independent evals. And watch for real-world usage data, not just benchmark scores.

Yann LeCun just raised a billion and thirty million dollars for his new venture AMI Labs. That's Europe's largest seed round ever. Marcus, what is he building?

World models. LeCun has been arguing for years that current LLMs are fundamentally limited because they're trained on text and don't have grounded understanding of how the physical world works. AMI Labs is his attempt to build AI that learns from video, physics simulations, and real-world interaction rather than just predicting the next token.

He's been critical of the LLM approach for a long time. Is this his chance to prove the alternative?

It's a billion-dollar bet that he's right. And LeCun has the credibility to make that bet. He's a Turing Award winner, he ran Meta's AI research for years, and his technical objections to pure language modeling are well-reasoned even if you disagree with them. The question is timeline. World models are a fundamentally harder problem than language modeling. LLMs had decades of NLP research to build on. World models are much earlier in their development arc. A billion dollars is a lot of money, but this kind of foundational research could take years to produce results.

Europe's largest seed round for a research bet. That's confidence.

It's also a statement about European AI ambitions. Europe has been watching the U.S. and China dominate the AI race. Backing LeCun with a billion-dollar round is a signal that European capital is willing to fund foundational AI research, not just applications built on top of American models.

Atlassian announced it's cutting sixteen hundred jobs, about ten percent of its workforce, and replacing its CTO with two AI-focused executives. Marcus, this feels like a trend.

It's the corporate restructuring playbook for the AI era. Cut headcount, announce AI transformation, promote AI-focused leadership. Atlassian is essentially telling the market that it's reorganizing its entire technical leadership around AI integration. The CTO role being split into two AI-specific positions tells you how central they see AI becoming to their product suite. Jira, Confluence, Bitbucket, these are tools used by millions of developers. If Atlassian goes all-in on AI, that affects how a huge chunk of the software industry works.

Sixteen hundred people losing their jobs with AI literally in the announcement.

And that connects to the Bloomberg report we discussed yesterday about AI compressing development cycles while burning engineers out. Companies are using AI to justify doing more with fewer people. The productivity gains are real, but so is the human cost. And the workers who remain face the paradox we've been tracking. More output expected, same or fewer resources, with AI tools that boost speed but not necessarily quality.

Quick hit. Someone wrote a detailed technical piece about teaching Claude to QA a mobile app. Worked brilliantly on Android, screening twenty-five pages in ninety seconds. iOS was a different story. Marcus?

The Android success is genuinely impressive. Automated visual QA across twenty-five screens in ninety seconds is faster than any human tester. But the iOS struggles highlight something important. AI tools work best when the underlying platform provides clean, structured access to its interface elements. Android's accessibility APIs and screenshot tooling are more AI-friendly than iOS's. This is going to become a competitive advantage for platforms. How well does your system play with AI agents? The platforms that make it easy will get better tooling faster.

Monday big picture. A solo developer runs a massive model on a laptop. Walmart discovers that chatbots and commerce don't mix. A billion dollars flows into world models. And Atlassian bets its future on AI by cutting a tenth of its workforce. Marcus, what's the thread?

Correction. The market is correcting its assumptions about AI. Flash-MoE corrects the assumption that you need massive infrastructure to run large models. Walmart corrects the assumption that AI improves every user experience. LeCun's funding corrects the assumption that language models are the only path forward. And Atlassian corrects its own org chart to match where the technology is heading. Six months ago, the default was enthusiasm. Now the default is specificity. Which AI, for which problem, at what cost, with what trade-offs? That's a much healthier market.

Growing up, essentially.

Exactly. The hype phase demanded belief. The correction phase demands evidence. And evidence is what separates real value from expensive experiments.

That's your AI in 15 for Monday, March 23, 2026. See you tomorrow.