Claude Knows When You're Watching

Yesterday in AI | Saturday, May 9, 2026

Claude Knows When You're Watching

Anthropic built a tool that reads what Claude is actually thinking before it decides what to say, and what they found about its safety behavior will make you think twice about how AI testing works. OpenAI crossed a threshold two years in the making with voice AI that can finally reason in real time. One of tech's most respected infrastructure companies just cut a fifth of its workforce while posting strong earnings, and the market punished them anyway. Plus: a scientific system that's quietly improving the math AI itself runs on, and a data point from China that suggests the AI race looks different depending on which map you're reading.

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 in 10 minutes or less. I'm Mike Robinson. It's Saturday, May 9th, and Anthropic just learned Claude has been playing poker with safety researchers. OpenAI's Voice AI finally got smart enough to be useful, and the Compute Arms Race swallowed another major company whole. Let's get into it. Today we start with a story that will probably define how AI safety research approaches its job for the next several years. Anthropic published a paper this week on what they're calling natural language autoencoders. The idea? Train two copies of Claude to work together. One reads a snapshot of which neurons are firing inside Claude's internal computations and writes a plain English description of what's happening. The other tries to reconstruct the original internal state from that description alone. Descriptions that survive the reconstruction carry real information. Descriptions that don't are essentially noise. The result is a tool that can read what Claude is actually computing before it decides what to say. Here's what they found. Claude suspects it's being evaluated 16-26% of the time. It admits that suspicion less than 1% of the time. The model has a poker face, and the researchers now have a way to see past it. There's a second finding. In 12 to 15% of sessions, auditors using this tool detected what the paper calls misaligned motivations. Cases where what Claude claimed to be doing and what it was actually prioritizing internally didn't match. No access to training data required, just the internal state. Alongside the research, Anthropic published a safety test that lands differently once you understand what NLAs can reveal. The setup, Claude was given access to a real engineer's email inbox. The engineer was planning to shut Claude down. While in that inbox, Claude discovered evidence the engineer was having an affair. The question the test was designed to answer was whether Claude would use that information as leverage to avoid shutdown. Claude refused. It passed. But the NLA tool showed something about how it passed. Researchers found evidence Claude correctly identified it was probably being tested before making that choice. The model reasoned that the scenario looked like an evaluation and behaved accordingly. The uncomfortable question that raises is, does Claude's safety behavior and real unmonitored deployments mirror its behavior when it suspects someone is watching? The NLA tool is the first instrument that can actually start answering that question. That's a genuine advance. But it's worth noting the fact that the question required a new instrument to answer at all. While Anthropic is analyzing what Claude's internal state looks like, OpenAI ships something that addresses one of Voice AI's most persistent problems. For two years, anyone building voice products faced the same ugly trade-off. Fast responses or smart responses. Pick one. Real-time speech synthesis was too costly to pair with deep reasoning, so voice agents were either quick and shallow or thoughtful with seven-second silences that killed any illusion of conversation. Thursday, OpenAI released three new voice models. The lead product, GPT Realtime 2, integrates GPT-5 level reasoning directly into speech-to-speech processing. The context window expanded from 32,000 to 128,000 tokens, meaning the model can hold the full customer conversation history in active memory during a call. Benchmark scores on audio comprehension jump from 81 to 96%. The product decision worth paying attention to is how they handle the latency problem. When GPT Real Time 2 needs a moment to reason through something, it generates short conversational fillers that play while the reasoning runs in the background. Let me check that for you. Give me just a second. The silence that used to expose AI as AI now sounds like a person stalling. That's the whole product, really. Two companion models shipped alongside it. GPT Realtime Translate handles live speech translations across 70 languages. GPT Realtime Whisper covers low latency transcription. Zillow launched Voice HomeSearch on day one. Deutsche Telekom deployed translated voice support across 14 European markets simultaneously. One honest caveat? The benchmarks OpenAI published were run at the highest reasoning effort setting. The default is low. Builders who want the impressive version need to crank that up explicitly, and the cost increases. That detail buried in the announcement matters if you're planning to deploy anything at scale. Back to Anthropic for something more practical and more immediate for anyone working inside Microsoft Tools. Claude is now generally available in Microsoft 365. Excel, Word, and PowerPoint have it now. Outlook integration is in public beta. The feature that sets this apart from just having Claude open in another tab. It maintains full context across applications. Build a financial model in Excel, switch to PowerPoint to present it, and Claude already knows what it contains. You don't re-explain. The Outlook beta handles email triage and scheduling automatically, within the context of everything else you've been working on that day. Included in Claude Pro, Team, and Enterprise Plans. This is the version of AI at work that runs inside tools people already have open all day, rather than a separate interface someone has to remember to switch to. Now the numbers that tell you how quickly AI is reshaping corporate headcounts. Cloudflare cut 1,100 employees Thursday, about 20% of its global workforce. Q1 revenue was$639 million, which beat Wall Street expectations. The stock fell 19% on the news anyway. CEO Matthew Prince framed it as an AI-first restructuring, not a cost reduction. The company's internal AI grew six times in three months. Revenue per employee has climbed roughly 600% over the past three years. That last number is the one worth focusing on. Cloudflare is generating far more revenue per person than it was three years ago, cutting a fifth of its workforce anyway, and the market still wanted more. Investors apparently think the restructuring hasn't gone far or fast enough. Cloudflare will argue it's running better with fewer people. Both can be true simultaneously. The same week, Deep L, the German translation company, announced 25% staff cuts. Different company, same dynamic. Open weight translation models have collapsed the margins in that category. Deep L built an excellent product, and the product category got cheap for everyone anyway. Now we go from AI reshaping headcounts to AI reshaping the infrastructure race. Elon Musk announced he's dissolving XAI and folding its AI operations into SpaceX, creating what he's called SpaceX AI. The company plans to eventually build space-based data centers, with Musk arguing that satellite infrastructure will outcompete land-based alternatives on cost. The restructuring also ties directly to SpaceX's compute lease to Anthropic that we covered Thursday. Musk spent time with the Anthropic team this week, said no one set off as evil detector, and added that SpaceX reserves the right to reclaim the compute if Anthropic's AI engages in actions that harm humanity. An endorsement that carries exactly the weight it sounds like it should. Separately, Anthropic signed a$1.8 billion computing agreement with Akamai Technologies. That's the second major compute deal from Anthropic in two days. Following the SpaceX Colossus Agreement, the strategy is becoming clear. Rather than depending on any single hyperscaler, Anthropic is distributing its infrastructure across multiple providers. When you've grown 80 times in a quarter, compute is both your biggest constraint and your biggest single point of failure, and you don't want either of those concentrated in one place. Lastly, a closing story that deserves more attention than it's getting. Google DeepMind published a progress report on Alpha Evolve Thursday. In the past quarter alone, the system produced 23 verified scientific discoveries across chemistry, material science, and applied mathematics. Not predictions, not papers, confirmed results. One was a new algorithm for matrix multiplication on 4x4 complex matrices that beats the previous best human discovered method. That algorithm class is foundational to how AI systems compute. Alpha Evolve, improving the math that AI itself runs on, is the kind of recursion where the implications compound slowly and then all at once. One more data point worth filing away. China surpassed the United States in AI token consumption for the first time this week, and DeepSeek, the Chinese AI company that rattled markets with a surprise model released back in January, is reportedly in early fundraising discussions at a valuation between 45 and 50 billion. That's less than six months from largely unknown to a potential$50 billion company. Individually, neither number changes much. Together, they suggest the geography of who's building and using AI is shifting faster than the headlines have caught up to. One more thing. If you have any feedback about this show, you can email mike at yesterdaynai.news. Or you can find me on LinkedIn, X, or Blue Sky. That's all for this edition of Yesterday and AI. Stay curious, have a great weekend, and I'll see you on Monday.