Yesterday in AI - Amodei Goes to Washington

Yesterday in AI | Saturday, April 18, 2026

Anthropic's CEO walked into the White House on Friday, and the person who showed up to meet him wasn't a junior official. A 187% IPO surge in Hong Kong says more about where the AI race is really being fought than any benchmark. The talent pipeline feeding Silicon Valley is in freefall, and a new Stanford data point puts a hard number on how fast. An open-source AI client just gave enterprise IT a third option nobody saw coming. And new research out of Nature has something to say about who actually wins when AI goes up against a PhD.

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

Transcript

SPEAKER_00 0:05

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 Saturday, April 18th, and the most important AI meeting of the week didn't happen at a conference, it happened in the White House. Let's start with Anthropic, because Friday's news is the most significant development in the Pentagon dispute since the blacklist dropped. Dario Amade met with the White House Chief of Staff, Susie Wiles, and Treasury Secretary Scott Basent in Washington. The meeting was confirmed by Axios, which broke the story, and the framing from sources is cautious but unmistakable. This looks like the beginning of a negotiated off-ramp. Here's the backstory for anyone who's been following from a distance. Earlier this year, Defense Secretary Pete Hegseth demanded Anthropic grant the Pentagon unrestricted access to its AI models, including for potential use in autonomous weapon systems and domestic surveillance. Amade said no. The Pentagon designated Anthropic as a supply chain risk. A federal appeals court declined to block that designation while litigation moved forward. Anthropic was effectively cut out of the fastest growing government AI market in the world. The fact that Wiles and Bissent both showed up Friday signals something. The chief of staff doesn't sit down for routine tech disputes, and neither does the Treasury Secretary. When two of the most senior people in the administration clear calendars for a single AI company, someone in the building has decided this situation needs to be resolved, not litigated. The technology at the center of it is mythos. Parts of the U.S. intelligence community in CISA have already been quietly testing it for defensive cybersecurity. The UK's Bank of England, Financial Conduct Authority, and HM Treasury have treated it as a present-tense risk to financial infrastructure. The model has real weight, and the government knows it. The unresolved question is whether Anthropic can find a path back into government contracting without agreeing to strip the ethical guardrails off its most powerful systems. Amade has said publicly and repeatedly that those lines don't move. We'll know soon whether the administration is willing to work within that. Now to a financial story that validates everything about where AI infrastructure investment is heading. TSMC reported a 58% jump in Q1 profit Thursday, beating estimates and setting a new record. The company raised its 2026 capital expenditure guidance from$40.5 billion to between$52 and$56 billion. ASML, which makes the lithography machines that TSMC and other chipmakers depend on, lifted its own 2026 revenue forecast and said demand will outstrip supply for the foreseeable future. ASML CEO Christophe Fouquet said the constraints span AI, smartphones, and personal computers simultaneously. On the analyst call, TSMC CEO Ceci Wei stated, AI demand is so strong our customers and our customers of customers continue to provide us with very strong signal and positive outlook. If you've been tracking the macro uncertainty argument, this is the counter data. Tariffs, trade tensions, geopolitical friction, none of it has dented the capital commitments from major cloud providers. They're still building, still ordering, still pulling the semiconductor supply chain forward at a pace ASML can't fully keep up with. When the company that makes the machines that makes the chips says they can't fill all their orders, the build out has legs. Let's stay with the geopolitics, because Friday brought a data point from the Stanford AI Index that deserves more attention than it got. Fortune published a detailed look at the report's China findings, and the numbers are sharper than I expected. In May 2023, the top U.S. model, OpenAI's GPT-4, led China's best model by more than 300 arena benchmark points. By March 2026, that gap has compressed to 39 points. Anthropics Claude Opus 4.6 leads China's Dola seed 2.0 by 2.7%. U.S. and Chinese models have traded places at the top of the performance rankings multiple times since early 2025. The AI index covered the broad Stanford report when it dropped Tuesday. The China angle is worth its own moment. The U.S. still produces more top-tier models and higher impact patents. China leads in publication volume, patent output, and industrial robot installations. On raw capability benchmarks, it's close enough that nearly erased is the phrase Stanford's researchers chose, not a media interpretation. The talent pipeline is the detail I keep coming back to. The number of AI researchers and developers moving to the U.S. has dropped 89% since 2017. 80% of that decline happened in the last year alone. Researchers don't move to places where the political environment is hostile, the visa system is unreliable, or the welcome feels conditional. Whatever your view on the AI competition, the human capital side of it is moving in the wrong direction for the U.S. Here's a market story from Hong Kong that connects directly to that China picture. ManyCore Tech listed on the Hong Kong Stock Exchange Friday, and its shares rose 187% on their first day of trading. The company raised 156 million in its IPO. The Hong Kong public offering was oversubscribed 1,591 times. ManyCorps started as an interior design platform. Over 15 years, it accumulated one of the world's largest libraries of 3D spatial data, rooms, environments, physical spaces that real users built inside its software. Now it's selling that data as training sets for robot makers. The company is calling itself the world's first publicly listed spatial intelligence company, and investors appear to believe the framing. The oversubscription number is the TEL. Physical AI, meaning systems that understand and navigate real world environments, need training data that mirrors the physical world. Most of that data either doesn't exist in usable form or is locked inside proprietary systems. Companies that have it at scale have something genuinely scarce. Whether many core specific library is as valuable as Friday's trading suggests, is a fair question. But the investor appetite for robotics data plays is clearly not hypothetical, and China producing the world's first publicly listed company in this category is not a coincidence given where Beijing is directing national resources. Away from the geopolitical drama, there's a quieter story from Thursday that I think the enterprise IT crowd should know about. Mozilla Ship Thunderbolt, an open source self-hosted AI client built by MZLA Technologies, the for-profit arm behind Thunderbird. The pitch is direct, everything Microsoft Copilot or Cloud Enterprise does, running entirely on your own infrastructure with no data leaving your network. Thunderbolt supports chat, search, research, and automation workflows. It connects to any OpenAI compatible API which covers Claude, Codecs, DeepSeek, and most of the major models. They have native apps for Windows, Mac, Linux, iOS, and Android. It's currently under a security audit and working toward enterprise production readiness, so it's not a day one production deployment for most organizations. But the architecture exists and the code is on GitHub. The market for this is bigger than most AI coverage suggests. Law firms, hospitals, financial institutions, government contractors, any organization where data sensitivity is a genuine compliance concern can't simply route their work through third-party cloud AI. They either build something custom at significant cost and time or they wait. Thunderbolt is offering a third path. An open source client you can vet, audit, and run yourself. Mozilla isn't going to outfeature the hyperscalers. They don't have to. They need to be the answer for the CIO who's been told by Legal that company data can't touch an external server. One more before we close, and it's a useful corrective to the AI is replacing scientists conversation. Nature published a piece Thursday drawing on the Stanford AI Index research, and the headline is one that doesn't get as much traction as the hype pieces. Human scientists with PhDs still outperform the best AI agents on complex scientific tasks by roughly two to one. The best agents perform at about half the level of the domain experts when the work is genuinely hard. Researchers are adopting AI tools widely. The Stanford data is clear on that. What's also clear is that there's not much evidence yet that adoption is actually improving scientific productivity in measurable ways. The tools are everywhere. Demonstrated impact on research output is still murky. This is the honest place to be on the question. The math proof that GPT-5.4 produced this week was real and remarkable, but that result came from a problem where you can verify the answer automatically. Most research doesn't have that property. Most science requires judgment calls that can't be scored by a program. The two to one human advantage on hard problems is a reality check worth keeping on the table. That's all for this edition of Yesterday and AI. Stay curious, have a great weekend, and I'll see you on Monday.