Generative AIProduct
Shared memory is not a brief you would send
Claude's chat and Cowork now share one memory. I packed the same topic files into three agent tasks and checked which remembered asides left with the draft.
23 min readSplit
Writing
Ship the Paper takes one research result into a production decision. Generative AI is for everyone else who has to use the models this week.
Archive
- 42 min
Ship the PaperLong-Horizon Planning & Reliability
A compaction verifier hears only the loud failures. The quiet ones decide the task
The paper scores a context summary by the blocked or repeated tool calls it causes over the next few steps. I rebuilt the scorer to ask how much of a summary's damage ever makes that noise.
Wait - 21 min
Generative AIInfrastructure
The model is not the assistant
I ran one scripted model inside four different apps and asked the same questions about one expense sheet. The app decided what came back right, and how the failures sounded.
Split - 20 min
Generative AIApplications
The prompt is not the application
I used to treat a good prompt as the whole application. Then I built the same launch email four ways and counted which numbers from the notes ended up in front of the customer.
Split - 20 min
Generative AIData
Dumping the folder is not using your data
The upload box treats every file as current. I packed the same HR folder of old and new files into a prompt three different ways and checked which of two conflicting leave policies won.
Split - 20 min
Generative AIModels
A two-point index lead is not a reason to switch models this week
Vendor leaderboards this month look like a race. I read the boards the way a regular user should, then gave the leading models two pieces of my own everyday work.
Split - 30 min
Ship the PaperSafety & Security
Attack success rate charges every jailbreak the same. Compute-aware evaluation fixes that, then breaks in CI
Ehghaghi and colleagues price a jailbreak attack by the compute it burns, not the queries it sends. I rebuilt both of their summary metrics to see which one a nightly build job could trust.
Wait - 37 min
Ship the PaperAgents
The harness decides what a coding agent looks capable of. The prompt layer does not
Swap the harness around a fixed model and a coding agent's score moves almost as far as swapping the model. I rebuilt the step behind most of that gap: how the agent's work gets collected.
Ship - 31 min
Ship the PaperAgents
Browser agents fail quietly. Falsifiable commitments make them fail loudly
FCPAgent makes a browser agent write down, before each step, what would prove that step wrong. I rebuilt the checking part to see what that buys, and what it can never catch.
Ship - 24 min
Ship the PaperAgents
A smaller interface makes browser agents better. It doesn't make them safe
AgentOccam shrinks what a browser agent sees and what it is allowed to do. I rebuilt that shrinking step to find out what gets lost, and where a cleaner interface stops helping.
Wait
Sections
Ship the Paper
5 essaysStartups
Nothing yetGenerative AI
5 essays