Yesterday in AI - The Weekend AI Got Physical

Yesterday in AI — Weekend Recap | Monday, April 20, 2026

A robot finished a half-marathon six minutes faster than any human ever has. An Anthropic design tool erased $1.6 billion from Adobe's market cap in a single afternoon. The first serious AI chip company to challenge Nvidia went public, and it showed up with a $10 billion OpenAI deal in its pocket. Meta announced layoffs from a position of record profit, with executives saying future cuts depend on how fast AI improves. And OpenAI investors are quietly shopping for a different CEO. This weekend's coverage had a little bit of everything.

Remember to subscribe, rate, and share this podcast if you like it!

Transcript

SPEAKER_00 0:03

Yesterday in AI. Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of artificial intelligence. I'm Mike Robinson. It's Monday, April 20th, and this weekend AI stopped being just a software story. Robots beat world records in Beijing, a new Anthropic design tool sent Adobe stock into freefall, and the first serious alternative to Nvidia in the AI hardware race filed for its public debut. Let's start with the design world, because Saturday's coverage brought two moves in the same direction from two very different companies, and together they basically rewrote the rules for what AI tools are supposed to be. Anthropic launched Claude Design on Friday, currently in research preview. The pitch is simple. If you're a founder, a product manager, or anyone who needs to turn an idea into a prototype, a slide deck, or a one-pager, but you don't have a designer available, Claude Design does it from a text description. You describe what you want, Claude drafts it. You refine it with follow-up prompts, then export it as a PDF, a URL, a PowerPoint, or push it straight into Canva. The model reads your company's code base and design files to keep things on brand. The market didn't wait for a formal analysis. Adobe's stock dropped enough to wipe out$1.6 billion in market cap on the news. That's what happens when investors decide a new product just changed the competitive map for Adobe. There's a telling detail buried in the timeline. Anthropic CPO Mike Krieger quit Figma's board two days before the launch. Figma's board probably connected those dots fast. But here's where it gets genuinely interesting. Because Canva came out swinging the same week. Canva launched Canva AI 2.0, and its COO said the quiet part out loud on stage. Until now, Canva has been a design platform with AI tools. Now we become an AI platform with design tools. That distinction matters. Canva trained its foundation model on 265 million monthly users actual edit sequences. Every hesitation, every pivot, every layout correction. The full path to a finished design. The model learns what it takes to go from an idea to something worth exporting. Canva also embeds directly into Claude, ChatGPT, Copilot, and Gemini, positioning itself as the last mile where AI-generated ideas become publishable work. So Anthropic and Canva are technically both competing and partnering in the same week. That's the design AI market in April 2026. Now to an announcement that deserves more than a layoff headline. Meta will cut roughly 8,000 employees, about 10% of its workforce, starting May 20th, with more cuts planned later in 2026. The context matters. Meta generated$60 billion in profit last year. Zuckerberg is choosing this from a position of strength, restructuring around AI efficiency, with teams reorganized into an applied AI group focused on building agents. The executives running this were direct. The scale of future cuts depends on how fast AI capabilities develop. That framing is deliberate. It's a company making a visible bet that fewer people plus better AI tools will outcompete more people using worse ones. This is the clearest public statement any major profitable company has made about replacing human capacity with AI at scale. The displacement conversation has moved past projections. It's now in press releases. There's a parallel leadership story worth flagging. OpenAI investors are privately questioning whether Sam Altman is the right person to take the company public. The issues cited include his side ventures. He recently asked the OpenAI board to fund Helion Energy and back a rocket startup called Stokespace, both companies he holds personal stakes in. The board drew a line. The name circulating as a potential replacement is Brett Taylor, the board's current chair, who built Google Maps and served as Facebook's CTO. Altman has publicly said he's 0% excited about running a public company. That quote is aging in interesting ways. Let's go to Beijing, where this story needs almost no editorial comment. The facts carry it. A humanoid robot built by Chinese smartphone maker Honor finished the Beijing half marathon in 50 minutes and 26 seconds. The human world record for the half marathon is 57 minutes, set by Uganda's Jacob Kaplimo in March. The robot beat it by more than 6.5 minutes. Here's the speed of change. Last year's winning robot at the same race took 2 hours and 40 minutes. A single year, nearly two hours faster. The Honor robot used autonomous navigation and the liquid cooling system developed largely in-house. On the sprint side, Unitary Robotics released footage of its H1 humanoid running at 10.1 meters per second on an athletics track. Usain Bolt's world record pace over 100 meters is roughly 10.44 meters per second. Unitary believes humanoids will break the 10-second barrier by mid-2026. China's latest five-year plan explicitly lists humanoid robots as a national strategic priority. Three Chinese companies already rank as the world's only first-tier vendors by shipment volume. The Beijing half marathon was technically a race. Functionally, it was a demonstration, a well-funded government-backed one. This weekend also brought a financial story that matters for anyone tracking where AI hardware is actually going. Cerebus Systems, the startup that makes specialized AI chips for training and inference, filed for a US IPO targeting a mid-May debut. This is a comeback attempt. Cerebras tried to go public in 2024 but pulled the filing because of a federal review of an investment from Abu Dhabi-based G42. That review is resolved. What's changed since then is the deal sheet. Cerebrus has a reported$10 billion agreement with OpenAI and a separate deal to put its chips inside AWS data centers. 2025 revenue came in at$510 million, net income was$237.8 million. The underlying thesis here is about concentration risk. Supply chains built on a single vendor are fragile, and the pricing power that concentration creates is real. The biggest buyers of AI compute are actively hunting for alternatives to NVIDIA. Cerebrus is the most significant non-NVIDIA hardware bet that public market investors can take right now. How this IPO prices will tell you whether the market actually believes that thesis or just talks about it. Two stories from Sunday's Pile worth pairing, because they're both about AI getting specialized in ways that go well beyond general purpose assistance. 1. OpenAI launched GPT Rosalind, a frontier reasoning model built specifically for life sciences, drug discovery, clinical research, and translational medicine. It's deployed under trusted access terms to Moderna, Amgen, the Allen Institute for Biology, and ThermoFisher Scientific. This is a different kind of launch than OpenAI's usual consumer rollouts. It's a bet that biology and pharmaceutical research are high value enough and sensitive enough to warrant a purpose-built model with controlled access rather than broad availability. 2. Tencent Open Sourced High World 2.0, a 3D world model that converts text, images, or video into editable 3D scenes with a full commercial license. Take it, build on it, sell it. Nvidia dropped its own 3D world model, Liria 2.0, on the same platform, though it is for research only. Alibaba launched a competing system behind a waitlist, the third approach in 48 hours. These world models matter because they're foundational infrastructure for robotics, game development, VR, and autonomous vehicles. The best work in this category was previously locked behind enterprise deals and API paywalls. Tencent just put a commercial grade version on Hugging Phase for free. The cost of building on top of 3D spatial AI dropped considerably in a single week. One last story before we close, and it's practical. It's going to affect what you pay for hardware over the next several years. There's a global RAM shortage with no clear end date. Samsung, SK Heinex, and Micron can only meet about 60% of DRAM demand by the end of 2027. An SK Group's chairman has warned the shortage could stretch to 2030. New fabs are coming, but almost none go online before 2027. And when they do, they'll prioritize high bandwidth memory for AI data centers over the general purpose DRAM that goes into phones, laptops, and gaming hardware. Consumer devices wait in line. Apple is already paying Samsung double its previous DRAM price to stockpile chips. The M5 Mac Studio slipped from summer to October. The OLED touchscreen MacBook Pro moved to 2027 entirely. Closing the supply gap requires 12% annual production growth, but only 7.5% as planned. That math means higher prices are baked in. AI's compute appetite has created a physical bottleneck that's working its way into the products sitting in everyone's pockets and on everyone's desks. That's all for this weekend ketchup edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.