Greetings!
This week: Microsoft quantifies how we fragment ourselves across devices and hours, Stanford reveals that one in twenty AI benchmarks is fundamentally broken, OpenAI maps the widening gulf between AI power users and everyone else, and researchers race to build frameworks that can actually measure what we've created. The hype cycle has cooled, and the age of accountability has begun. But now that we've stopped selling the idea that there might be a ghost in the machine, can benchmarks and frameworks alone carry the weight of what comes next?
Microsoft's 37.5 Million Conversations Confirm What We Already Suspected. In a study that scientifically validates the obvious, Microsoft's new Copilot Usage Report analyzed data from January through September 2025 to discover that—hold your gasps—we behave differently depending on which rectangle we're staring at. Desktop You asks about work and technology during business hours, while Phone You turns to questions of health, relationships, and philosophy late into the night. Religion and philosophy queries peak in the predawn hours, suggesting either profound existential awakening or insomnia-induced desperation. Relationship queries spiked on Valentine's Day. The researchers poetically note that users have agreed to "weave AI into the fabric of their daily existence, turning to it for code reviews at 10 a.m. and existential clarity at 2 a.m."
Stanford Finds 5% of AI Benchmarks Are Fundamentally Broken. When we measure AI capability, we use benchmarks—standardized tests with questions and known correct answers. Researchers at Stanford HAI mathematically scoured thousands of these benchmarks and found that one in twenty contains serious flaws they playfully call "fantastic bugs." These aren't minor quibbles: errors include formatting issues that mark correct answers wrong (grading "$5.00" as incorrect when the answer key says "$5"), culturally biased questions, logical inconsistencies, and mismatched labels. The consequences cascade badly—in one example, the model DeepSeek-R1 jumped from third-lowest to second place after benchmark corrections. Since benchmark scores drive funding decisions, research priorities, and resource allocation, we've essentially been judging AI capabilities with broken rulers. The researchers are now working with benchmark developers to fix these flaws and advocate for ongoing maintenance rather than treating benchmarks as finished products once published.
OpenAI's Enterprise Report Reveals a Widening Capability Gap. OpenAI's State of Enterprise AI report goes beyond headline productivity stats to reveal something more interesting: dramatic variation in how deeply organizations and individuals have integrated AI into their work. "Frontier workers" send vastly more messages than median workers, with the gap widening for advanced features like coding and data analysis. Frontier firms show similarly deeper integration. But perhaps most striking: three-quarters of workers report completing tasks they previously couldn't perform—coding, data analysis, spreadsheet automation. Non-technical teams now increasingly perform technical work that was previously confined to specialist roles. AI isn't just accelerating existing workflows; it's reshaping what workers are capable of doing in the first place.
Google Opens Willow Chip to UK Researchers. Google and the UK government announced a partnership giving researchers access to Google's Willow processor—the 105-qubit chip that last year achieved below-threshold quantum error correction and ran a benchmark in five minutes that would take classical supercomputers 10 septillion years. Scientists can now submit proposals and work with Google and the UK's National Quantum Computing Centre to design experiments. The goal: find practical applications for a technology still searching for real-world traction. Some experts believe quantum computers capable of meaningful performance could arrive within a decade, with potential applications in chemistry, medicine, and materials science. For identity security, the timeline matters: quantum computers powerful enough to break current encryption would also be powerful enough to do useful work. The race is on.
Quantum Technology Reaches "Transistor Moment." A new Science paper authored by researchers from UChicago, Stanford, MIT, and others argues quantum technology now stands at a turning point similar to the early computing age before transistors transformed everything. The authors surveyed six leading hardware platforms and—in a delightfully meta touch—used ChatGPT and Gemini to assess technology readiness levels. Their verdict: advanced prototypes exist, but meaningful applications like large-scale quantum chemistry simulations could require millions of physical qubits with error performance far beyond current capabilities. As one researcher notes, a high readiness level today doesn't mean the science is done—just as 1970s semiconductor chips couldn't do much compared to modern integrated circuits.
MIT Discovers LLMs Learn the Wrong Lessons. MIT researchers found that large language models sometimes learn to associate grammatical patterns with specific domains rather than actually understanding content. An LLM might answer "France" to the nonsense question "Quickly sit Paris clouded?" simply because it sounds like a geography question. Worse, this shortcoming can be exploited: by phrasing harmful requests using syntactic templates the model associates with "safe" datasets, attackers can trick safety-trained models into generating harmful content. The researchers developed a benchmarking procedure to evaluate this vulnerability before deployment—but the finding raises uncomfortable questions about whether AI "understanding" is understanding at all, or merely sophisticated pattern matching wearing a convincing mask.
Google's Titans Architecture Gives AI Genuine Long-Term Memory. Current AI models face a fundamental limitation: they can only "remember" what fits in their context window—everything else vanishes between conversations. Google Research's Titans and MIRAS frameworks attempt to solve this by giving AI a neural long-term memory module that learns and updates while the model is running, without requiring retraining. The key innovation is a "surprise metric": the model detects when new information contradicts its existing memory (high surprise = prioritize for storage; low surprise = safe to skip). The architecture can scale to over 2 million tokens of context while outperforming GPT-4 on long-document reasoning tasks. If it works at scale, AI identity could become genuinely persistent rather than perpetually amnesiac.
Vibe Coding's Dirty Secret: 80% of "Working" Code Is Insecure. Carnegie Mellon's SUSVIBES benchmark delivers sobering news for the "vibe coding" movement. Across 200 real-world tasks, even the best-performing agent (SWE-Agent with Claude 4 Sonnet) achieved only 10.5% security pass rates despite 61% functional correctness. Translation: four out of five "working" AI-generated solutions contain exploitable vulnerabilities. The researchers note that beginner programmers are much more likely to be vibe coding optimists. Frontier AI companies admit to using vibe coding in production. The non-human developer is prolific but dangerously naive.
Google's Differential Privacy Framework: Watching AI Without Seeing You. AI companies face a dilemma: understanding how people use chatbots helps improve safety and service, but those conversations often contain sensitive information. Existing approaches—like Anthropic's CLIO framework—use LLMs to summarize conversations while prompting them to strip out personally identifiable information. But this relies on heuristic privacy protections that are difficult to audit and may not hold as models evolve. Google Research's Urania framework takes a different approach: mathematical privacy guarantees. The system clusters conversations, extracts keywords with added noise, and generates summaries without ever showing the LLM actual conversations. Even if keywords accidentally contain PII, the framework's differential privacy guarantees prevent that information from surfacing. When tested against membership inference attacks—attempts to determine whether a specific conversation was in the dataset—the system performed no better than random guessing.
Agentic AI Wants to Do Your Shopping. Bain's analysis finds 17% of holiday shoppers started their journey with ChatGPT or similar assistants, with AI referrals to retailers growing 7× year-over-year in the US. Around half of consumers remain uncomfortable letting AI complete transactions without involvement—but that number is shrinking. The agents are learning your preferences, remembering your sizes, and optimizing for your wallet. The question isn't whether non-human shoppers will emerge, but whose interests they'll ultimately serve.
Tim O'Reilly on AI Hype and What Actually Matters. Tech visionary Tim O'Reilly isn't buying the current AI narrative. "The idea that previously software was a tool, now it's a worker—that's been true for a long time," he tells Big Think. "Software was doing work long before AI. It's all in service of the narrative that this technology is different from anything before it." He's equally dismissive of imminent AGI claims: "AI is not singular, it's 'normal technology.' There are still laws of physics, things to be built, and many constraints." His advice for businesses: AI is transformative and worth investing in, but figure out how it actually impacts your business rather than buying the hype. Plan scenarios, develop robust strategies that survive multiple circumstances, and focus on genuinely improving customers' lives. "When the bubble bursts," he says, "you want to be making things that genuinely improve your customers' lives. The market will eventually reward people doing the real stuff."
The hype cycle is cooling, and everyone's reaching for their rulers.
What stands out this week isn't any single breakthrough—it's the collective pivot toward measurement. Stanford discovers that one in twenty AI benchmarks contains serious flaws, meaning we've been grading these systems with broken instruments. OpenAI publishes enterprise metrics revealing dramatic gaps between power users and everyone else. Microsoft analyzes 37.5 million conversations to quantify patterns we could have guessed. Carnegie Mellon builds a benchmark showing that most AI-generated code is insecure. Google develops mathematically provable privacy frameworks because the existing heuristic approaches aren't rigorous enough to audit.
This is what the post-hype era looks like. After years of breathless capability announcements and trillion-dollar valuations built on vibes, the industry is now obsessed with validation. The investors who wrote the checks are asking what they bought. The enterprises experimenting with AI are asking whether it actually works. The researchers are asking whether their measurements even mean anything.
There's something almost poignant about this moment. We rushed to build systems we don't fully understand, and now we're frantically constructing frameworks to figure out what we've created. The benchmarks themselves turn out to be flawed. The privacy guarantees turn out to be heuristic. The "working" code turns out to be riddled with vulnerabilities. Even our measurements of how people use AI reveal that we're different selves depending on the time of day—professional and task-oriented at our desks, introspective and philosophical in the late-night hours.
There is irony in trying to use math to settle our nerves. Descartes believed he could build a world on the bedrock of logic, yet here we are, surrounding ourselves with sophisticated mirrors that we mistake for foundations. We are no longer looking for the ghost in the machine; we are frantically trying to measure the shadow it casts, hoping that a precise enough number will finally pass for the truth.
