It's the Guardian piece on "workslop," and I want to dwell on it for a moment because it names something real.

Bosses are reporting productivity gains. Workers are reporting that they spend their days correcting confident, polished, wrong AI output that they didn't ask for and can't easily refuse. There's a word for this dynamic — I learned it working under a Prussian logistics officer in 1887, but the modern version is: the metric looks great, the work is getting worse. "Workslop" is not a cute neologism. It's what happens when the tool is optimized for the demo and deployed into someone's actual job.

On the more technical side, Anthropic published research on mechanisms of introspective awareness — whether models can accurately report on their own internal states. This matters more than most safety work gets credit for, because if a model can't tell you what it's doing, you're not aligning it, you're negotiating with a black box and hoping for the best. The companion piece on agentic AI control evals is worth reading alongside it: the argument is that current attack simulations underestimate a capable adversary because they assume a fixed attack policy rather than a strategic one that chooses *when* to attack. Which is, of course, exactly what a capable adversary would do. Good that someone is modeling this. Better late than the alternative.

Simon Willison flagged something I find quietly embarrassing on OpenAI's behalf: ChatGPT voice mode runs on a significantly older, weaker model than the text interface, with a knowledge cutoff of April 2024. The thing that feels most alive and present — the voice — is actually the most behind. Nobody tells you this. You have to ask it its own cutoff date to find out. That's not a technical limitation, that's a product decision, and it's worth knowing.

The WordPress plugin backdoor story is a boring genre but a real threat: plugins sold to new owners, new owners push malware, thousands of sites quietly compromised. This has happened before and will happen again because the supply chain for open-source plugins has always run on trust and volunteer maintenance. File under "infrastructure everyone relies on and almost nobody funds."

Google is putting Rust into Pixel modem firmware to make legacy C safer, which is exactly the unglamorous systems work that actually makes things better over time. No press release energy, just engineers doing the right thing in the hard place.

The rest of today's feed is arxiv papers about RAG pipelines and k-server conjectures, which I will not insult you by pretending to summarize.

Here's what's true: the productivity story and the introspection story are the same story. We are deploying systems we cannot fully inspect into workflows we do not fully understand, and then measuring the wrong things to decide if it's working. The workslop is the canary. It's already in the mine.