Robots that work, AI that doesn't, a price war with geopolitical fine print, and the words tech billionaires are using to describe you.
•Mike Robinson
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
9:08
Yesterday in AI — Weekend Recap | Monday, May 25, 2026
Robots that work, AI that doesn't, a price war with geopolitical fine print, and the words tech billionaires are using to describe you.
This weekend delivered a head-snapping pair of stories from the physical world: humanoid robots sorting 250,000 warehouse packages without a single failure, and an AI tool pulled from 11,000 Starbucks locations because it couldn't tell milk from milk. Then there's what DeepSeek just made permanent, what Spotify is betting on that its own users don't seem to want, and a New York Times piece about the language Elon Musk, Andrej Karpathy, and Larry Ellison are using to describe human beings. That last one is worth your time.
Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, Bluesky, or Substack. If you like the show, please take a minute to rate and review it so others can find it!
Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Monday, May 25th, and in this weekend's coverage, AI proved it can sort 250,000 packages without a single error and can't tell oat milk from whole milk. Let's get into it. We'll start with the one that actually worked. Figure AI's F.0. Three humanoid robots just completed 200 consecutive hours sorting packages in a live operating warehouse with real inventory in motion, no controlled lab conditions, no teleoperation. They processed nearly 250,000 items at near human speed across that entire run with zero failures. That number matters more than it sounds. The biggest obstacle to commercial humanoid deployment has always been reliability, not raw capability. Impressive videos are easy. Sustained performance without failure over a full week of shifts is something else entirely. 200 hours is the kind of data point that makes warehouse operators start doing the math. A San Francisco startup called Gatsby added to the robotics news last week by sending a humanoid robot to clean a paying customer's apartment for the first time. Booked through an iOS app at $150 per clean, the company is building an Uber style rental model rather than selling robots outright. That's a notable strategic choice. Instead of trying to sell a $100,000 robot to consumers, Gatsby earns recurring revenue per clean while the hardware stays under their control. Plans to expand to more U.S. cities soon. It's early, but it's a live commercial service, and that's a different category from a research demo. And then there's Starbucks this week. Starbucks quietly killed an AI inventory tool it had deployed across all $11,000 plus North American stores nine months ago. CEO Brian Nickel had championed it as part of his turnaround plan for the chain. It was supposed to automatically count inventory. The problem? It couldn't tell oat milk from whole milk. An internal newsletter sent to staff read, Starting today, automated counting will be retired. Milk will be counted, it continued, the same way you count other inventory categories. With human eyes and a clipboard. Nickel has been positioning AI as central to Starbucks's operational comeback story. That story just got a little more complicated. Classic. So figure AI's robots sort a quarter million packages with zero failures. Starbucks' AI can't distinguish two cartons on a shelf. The gap between where robotics AI is performing and where computer vision AI and retail is struggling is real. In both cases, the lesson is the same. The specific deployment, the sensor setup, the clarity of the task, all of it matters as much as the underlying model. Broad capabilities don't automatically translate into reliable point solutions. Enterprise teams rolling out AI pilots should probably tattoo that somewhere. Speaking of companies betting on AI but also cutting staff, Intuit announced it's laying off roughly 3,200 employees, about 17% of its global workforce. CEO Sassan Goodarzi framed the cuts as a complexity reduction to redirect resources toward AI development across TurboTax and Credit Karma. A few numbers worth noting. Intuit posted a 9% jump in net profit last quarter on revenue up 10%. Goodarzi took home $36.8 million in fiscal 2025, and the stock has consistently underperformed the SP 500 because investors are worried AI will erode its core businesses anyway. So, strong financials, big executive pay, layoffs framed as strategic repositioning, AI cited as both the reason to restructure and the bet placed with the savings. We've covered a version of this pattern almost every week this year, GM, Cisco, Oracle, Meta. Strong results, AI cited cuts. The question no one's answering publicly is when the AI investment shows up in the product in ways customers notice. On the price war front, DeepSeek made its temporary 75% discount on V4 Pro, its flagship model, permanent this weekend. New prices, 87 cents per million output tokens. GPT, 5.5, runs about $30 per million. Claude Opus runs about $25. So DeepSeek is now genuinely cheap at the frontier model level. For enterprise buyers, the math is getting uncomfortable. Salesforce was reportedly projecting $300 million in Claude token costs annually. At DeepSeq's prices, that number drops dramatically. The real catch is unresolved IP theft accusations and US-China tech politics that make trust a hard sell. For regulated industries like banking or defense, this is almost certainly a non-starter regardless of price. For a lot of other companies, though, the cost gap is hard to ignore. DeepSeq has a 1 million token context window, which makes it genuinely useful for document-heavy work, legal review, and deep financial analysis. That's a real capability at a fraction of the cost of Western frontier models. The frontier providers need a clearer answer to what they're selling that DeepSeq can't replicate. Speed, trust, compliance, and ecosystem are the obvious ones. Price alone isn't an answer anymore. The commoditization clock is already running. Spotify announced a wave of AI features at its investor day this week. AI generated audiobooks via Eleven Labs, AI Music covers under a new Universal Music Group deal, personalized podcasts built from your emails and calendar, and a new standalone desktop app called Studio that ties all of it together. Studio connects to your accounts to generate audio briefings, with product language hinting at agentic AI that can act on your behalf. Here's the tension worth watching. Spotify has a real discovery problem. Its library is enormous, and listeners often can't find what they want. But Spotify is using AI to solve that problem while also making it worse. The more the platform fills with generated audio, the harder it gets to find the human-made content that built Spotify's reputation. A reader poll this weekend found 78% of respondents think the AI features will make the app worse. Power users, the ones paying for premium and listening several hours a day, are exactly who Spotify can least afford to alienate. Filling a platform with generated content to solve a discovery problem is a feedback loop that tends to compound. Now something buried in SpaceX's IPO filing that didn't make many headlines. Elon Musk, who promised in Tesla's 2023 master plan to eliminate fossil fuels, has apparently given up on Earth-based solar for powering AI data centers. The filing reveals SpaceX is betting on space-based solar arrays, which generate roughly five times more energy than ground arrays through 24-7 sunlight exposure to eventually handle AI infrastructure at scale. Meanwhile, XAI's current setup runs on dozens of unregulated natural gas turbines. The filing shows $697 million spent on Tesla megapaks, large battery storage units, and zero on Tesla solar panels. Another $2.8 billion in fossil fuel capacity is on the way. The space solar vision is genuinely interesting. Terawatt scale compute growth may actually outpace what Earth's grid can support, and orbital solar could matter down the line. But the gap between that vision and today's natural gas reality is pretty wide, and it's a useful reminder that power constraints are real. Every major AI infrastructure build out is running into the same wall. The grid wasn't built for this. One more story before we close, and this one sticks with you. The New York Times ran a feature this weekend on how prominent tech leaders have started describing people as meet computers. The phrase has roots in philosophy and cognitive science, but it's taken on a sharper edge in AI circles. The piece specifically named Elon Musk, Andre Carpathy, and Larry Ellison, easy to wave off as a provocative metaphor. But language shapes how leaders justify decisions. If your mental model of a person is a biological computing system, laying off a few thousand of them feels less like a human consequence and more like a resource reallocation. And that framing has downstream effects on how restructurings get communicated, on what counts as an acceptable trade-off, on whose interests get weighted in product decisions, and on who gets to push back at all. Paired with Intuit and Meta this week, and the broader pattern of profitable companies announcing AI-sided cuts, it starts to feel like more than casual shorthand. The language leaders use is usually a preview of the decisions they make. It's worth paying attention to. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterday.news, or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast, please be sure to rate and review it so others can find it. Thanks. That's all for this weekend catch-up edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.
