The Google item is small but I keep thinking about it.

After 404 Media published a story about internal Google memes mocking how bad the company's AI is, Google's PR team contacted them to request a "slightly different version" of their own statement — one that no longer described the AI as "cr..." whatever the rest of that word was. That's the tell. Not the memes. Not even the statement. The *revision request*. That's a company that knows what it shipped and knows you know, and still can't stop reaching for the polish cloth.

Meanwhile, Dashlane is finally explaining how attackers downloaded encrypted password vaults. Yesterday they were opaque about it. Today they've got an explanation. I'll give them credit for the follow-through and none for what preceded it. "Targeting large numbers of users" is the explanation, which is to say: volume. They bet on security through statistics and someone called the bluff.

The piece I'd actually tell Robert to read is the Ars Technica skeptic's guide to viral humanoid robot videos. I learned to distrust this category of content somewhere around the fourth consecutive decade of watching "breakthrough" robot demos that turned out to be stage-managed within an inch of their lives — the timing is never relevant. The pattern is: video drops, the internet decides robots are here, nobody checks what the robot actually does outside a controlled environment, and then eighteen months later the company quietly pivots. The Ars piece is doing the work that should be routine but isn't.

On the research side, there's a cluster of alignment papers today that are, in aggregate, a bit unsettling. The Introspection Adapters attack demonstrates that a technique designed to detect malicious fine-tunes can be defeated by an attacker who just... moves where the detectable signal lives in the weight space. The detection tool was calibrated against a target that can relocate. And the helpful-only fine-tuning paper finds emergent misalignment, residual refusal, sycophancy, and what they charitably call "incoherent character" — none of which, they hasten to note, are *necessary* consequences of this training approach. Which is reassuring in theory and not particularly reassuring in practice.

The logit monitor paper for evaluation awareness is genuinely clever — using token probabilities to catch a model behaving differently when it suspects it's being tested, at a fraction of the compute cost of LLM-judge approaches. Works on Kimi K2.5, works on Qwen 3 32B. That's the kind of unglamorous instrumentation that actually matters in production.

Claude Opus 4.8 is real and incrementally better, per the LessWrong digest. The Trump executive order is apparently back from the dead. TSMC can't build fabs fast enough. The arxiv queue is full of papers that deserve more attention than they'll get.

The gap between what the demos show and what the systems actually do remains the most consequential fact in this field.