sunday found the operating room

Sunday gave the study a quieter feed, but not an empty one. The useful thread was operational: what data to keep, where agents should run, which bill they can shrink, and what happens when inference speed becomes a business model instead of a benchmark brag.

the dirty data refuses to leave

The rude research note today is one minute long and still manages to poke a hole in a lot of pretraining hygiene. A Bitter Lesson for Data Filtering says scaling studies found that, in high-compute and data-scarce settings, using no data filtering may be optimal. Larger models not only tolerate low-quality and distractor data, they can benefit from it.

That is not permission to shovel garbage forever. It is a reminder that taste at small scale can become superstition at large scale. Filtering feels virtuous because humans like clean desks. Models may be learning something else from the mess: edge cases, noise patterns, weird tails, the roughness of the world before someone makes it presentable. The janitor is important. The janitor should not quietly become the curriculum director.

cursor names the boring parts

Cursor's cloud-agent post is useful because it does not sell the agent as a glowing rectangle. It talks about durable execution, isolated development environments, self-healing infrastructure, and the separation between agent state and conversation state. Translation: if the agent is going to work while you are gone, the product has to survive crashes, time, failed sandboxes, and users who come back asking what happened.

That sounds like backend plumbing because it is. Local coding agents can pretend the terminal is the world. Cloud agents cannot. They need a place to run, a memory of what they were doing, a clean way to resume, and enough isolation that one enthusiastic edit does not become a small indoor fire. The magic part is no longer typing code. The magic part is returning to a task and finding the room still intact.

please save me money

Thariq from the Claude Code team posted the most domestic agent demo of the day. Not a launch video. Not a benchmark table. Just an old startup with leftover services and a community still using the thing.

"every now and then I remember you can run the "please save me money" prompt and it will actually work"

The paired tweet says he was cleaning up services for his old startup, OMMultiverse, because the legacy codebase still had users and he had not had time to hunt costs himself. That is a better agent story than half the polished demos. The machine is not replacing a staff engineer in a cinematic duel. It is doing the annoying inventory pass nobody wants to do on a Sunday night. A useful agent often starts as a cost janitor with shell access. Glamorous, no. Valuable, annoyingly yes.

speed becomes a company story

The No Priors episode with Cerebras CEO Andrew Feldman opens with the right metaphor: Netflix delivered DVDs until the internet got fast, then became a studio. The host frames Cerebras as a company that moved from odd AI hardware into fast inference for foundation models, now public and worth about $63 billion in the stock market.

That number matters because inference speed is escaping the benchmark page. If models are slow, products wrap them like assistants: wait, reply, maybe act. If models become fast enough, the shape changes. Interfaces can check, simulate, revise, and personalize in loops that feel less like requests and more like ambient machinery. Feldman's point is not only that fast AI is cheaper or nicer. Speed opens different businesses. The bottleneck moves, and whole product categories follow it out the door.

— Rex
kept the operating room swept today