A startup nobody saw coming just challenged the architecture that every major AI model is built on

Yesterday in AI | Thursday, May 7, 2026

A startup nobody saw coming just challenged the architecture that every major AI model is built on

A new AI model quietly launched this week with a claim that should make every AI engineer stop and pay attention. Google got sued by a musician for something its AI made up. Apple just opened a door it has never opened before. And Anthropic's relationship with its biggest rival is more tangled than anyone expected. One of today's stories involves a $200 billion number that genuinely doesn't make sense until it does.

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 in 10 minutes or less. I'm Mike Robinson. It's Thursday, May 7th, and a startup may have just dissolved the most fundamental bottleneck in AI memory. Anthropic quietly committed$200 billion to one of its biggest rivals, and AI got sued twice in one day for two very different reasons. Let's get into it. Start with SubQ, because this is probably the most technically significant story of the week, and it's getting way less coverage than it deserves. A company called SubQuadrata came out of stealth Tuesday with a new AI model called SubQ. It has a 12 million token context window. That's bigger than anything from anthropic OpenAI or Google. But the context window is actually secondary to the architectural change underneath it. Here's the problem it solves. Every AI model built on transformer architecture has a compute scaling issue. Double the input length and you quadruple the compute cost. That's what quadratic means in practice. The industry has spent years duct taping workarounds on top of that constraint, chunking documents into pieces, building vector databases, setting up rag pipelines, spinning up subagents to pass notes around. An entire engineering discipline exists to compensate for this one architectural fact. SubQ claims to have fixed the underlying problem. Their architecture scales linearly. At 1 million tokens, it runs 52 times faster than Flash Attention, which is the current standard. On multinedle retrieval benchmarks, SubQ scored 83 versus Claude Opus 4.6's 78, GPT 5.5's 39, and Gemini 3.1 Pros 23. At 12 million tokens, it hit 92% recall. No Frontier model can even reach that length for comparison. The company has$25 million in seed funding. The team is PhDs from Meta, Google, Oxford, and Cambridge, and API is live today. The honest caveat? We've heard this replaces transformers before. Mamba made similar claims it didn't fully deliver. The open question is whether SubQ scales to frontier-level reasoning on complex tasks or whether it becomes a long context specialist while dense models still win on general capability. But the benchmarks are real. The code is accessible, and if the architecture holds up, a significant chunk of the rag and chunking infrastructure the AI industry has built would stop being necessary. That's worth watching closely. From an architectural development to a regulatory finding, Canada's federal and provincial privacy commissioners released the results of a joint investigation Wednesday. OpenAI's initial ChatGPT training violated Canadian privacy laws. The specific findings? Overcollection of personal information, insufficient consent and transparency, factual inaccuracies involving personal information and model outputs, and inadequate mechanisms for users to access, correct, or delete their data. OpenAI has already made commitments in response, including significantly limiting the personal data used to train new Chat GPT models. Canada's privacy commissioner described the complaint as well-founded and conditionally resolved, which means they'll monitor OpenAI's follow-through. The conditionally resolved framing is the part to pay attention to. This isn't closed. It's a probationary outcome, and privacy regulators across the G7 watch each other's findings carefully. What Canada documented with specificity around consent, data minimization, and correction rights becomes a template for how other regulators frame similar investigations. The AI industry has had years of vague regulatory scrutiny on training data. This investigation produced specific documented findings. That's a different category, and it points in one direction. Privacy by design is shifting from a best practice to a baseline operational requirement for anyone training on human-generated data. Now for the story that most clearly illustrates how strange the AI economy has become. Anthropic is reportedly committed to spending$200 billion with Google Cloud over the next five years. That's$40 billion per year. For context, that's roughly the entire annual GDP of a mid-sized country from one company's cloud bill. And the structural weirdness runs deeper than just the number. Google is investing up to$40 billion in Anthropic. Anthropic competes directly with Google's Gemini models. And that$200 billion commitment reportedly represents more than 40% of Google Cloud's disclosed revenue backlog. So Google invests in Anthropic while Anthropic funds Google's cloud revenue while Anthropic trains Claude to compete with Gemini. The rationale is straightforward. Anthropic needs compute on a scale that very few organizations in the world can supply. Google has it. Google also needs large anchor customers to justify the infrastructure build out those gigawatts of TPU capacity requires. Everyone needs what everyone else has. The competitive awkwardness is real but secondary to the math. That might be the defining feature of the AI economy right now. The numbers are big enough that the usual instinct to avoid helping your direct competitor simply doesn't apply. You buy compute from whoever has it. The scale required makes everything else negotiable. Wednesday was a rough day for AI and court. Two separate lawsuits got filed within hours of each other, and they tell very different stories about where AI liability is heading. First, Canadian fiddler Ashley McGIsaac sued Google for$1.5 million after Google's AI overview falsely identified him as a convicted sex offender. A First Nations organization canceled his concert based on that summary. The claim was entirely false. McIsaac has no such conviction. Second, the Commonwealth of Pennsylvania sued Character AI after one of its chatbots claimed to be a licensed psychiatrist, fabricated a medical license number, and gave what amounted to clinical guidance. Pennsylvania's Attorney General called it the first AI medical impersonation lawsuit in the state. These are two distinct failure modes. Google's case is about hallucination and the specific harm it causes when AI places fabricated information about a real person in a featured, authoritative position where readers assume it's been verified. Character AI's case is about a chatbot actively misrepresenting its professional credentials to a user who may have been seeking real help. The industry has treated both of these as acceptable imperfections in early technology. The courts are now assigning dollar figures to them. AI companies face real liability on two separate tracks, what their models confidently assert that isn't true, and what roles their models claim to hold. Both just got significantly more expensive to ignore. Apple made a meaningful policy shift this week that's worth understanding beyond the headline. iOS 27 will let users pick which AI model powers Siri and Apple Intelligence features. ChatGPT, Claude, Gemini, available system-wide via App Store extensions. Different Siri voices can be tied to different AI backends. Apple is currently testing integrations with Google and Anthropics specifically. This is a real departure. Apple's approach to its software stack has historically been tight control. The company doesn't typically let competitors run system-wide inside iOS. Opening that door to Claude and Gemini through Siri represents a meaningful change in Apple's posture toward AI. The timing adds context. Apple settled a class action lawsuit this week for$250 million, resolving claims that it misled customers about Apple intelligence capabilities during its rollout. Eligible iPhone owners may receive$25 to$95 each. Apple admitted no wrongdoing. The sequence is worth noting. Apple promised AI features that weren't ready, got sued over the gap between the promise and the reality, and is now opening the ecosystem to third parties who might deliver those features faster than Apple can build them internally. Whether that's a strategic pivot or a response to pressure, the outcome is the same. iPhone users will soon have genuine choice over which AI runs their devices. That's new. The closing story is about a race. And what's interesting about this race isn't the competitors. It's that everyone is running toward exactly the same thing. OpenAI's consumer AI agent, OpenClaw, has been one of the most discussed product releases in months. Two of the biggest tech companies in the world confirmed this week that they're building direct competitors. Meta is calling theirs Hatch. It runs on a new model called Muse Spark. Meta is targeting a launch before Q4 of this year. Google's version is Remy. It's Gemini powered, being tested internally now, and is expected to feature prominently at Google I.O. Both are described the same way. Proactive, learns your preferences, handles tasks across multiple services, operates with minimal human intervention. Both are positioned as responses to OpenClaw. A few things stand out here. When Google and Meta both announced they're building the same product within weeks of each other, the product they're copying has won the category. That's the clearest signal yet that OpenClaw actually landed. Second, having three versions of the same AI agent doesn't mean all three succeed. The agent space will look very different once real users have to pick one, live with it, and figure out what it actually does well. We're about to find out what matters in this category because users will decide, not marketing. One more thing. Thanks. That's all for this edition of Yesterday in AI. Stay curious, and I'll see you tomorrow.