byte | sumit sute

I wonder if one consequence of agentic workflow is a gradual normalization of heuristic correctness. Not wrong, not careless, just increasingly comfortable with good enough for now because another iteration is always available?2d ago

xxx

The economics of generation influence the econimics of scrutiny. When producing a solution takes seconds, the instict to rigorously defend every line before moving forward starts to soften.2d ago

xxx

The goal is to assist each other. Agents help you, and you help agents. Building both ephemeral communications and persistent memory layers accordingly is a good approach for human-agent interfaces. Let's save the delegation for deterministic chores.may 25

xxx

A nice side effect of building and sharing tiny tools is that the solutions becomes inspectable, iterable, and communal instead of disappearing into a private agent session.may 22

xxx

One of the interesting ways in which you can artificially create a collision and make it relatively less passive for your cognition by making two or more models collide. You will have to make a choice, that can keep the atrophy at arms lengthmay 20

xxx

Karpathy's llm-wiki and Obsidian have just unlocked a way for me to create my own skills, opinions, and blog factory, solving a lot of cognition and workflow issues in a consistent and central manner.may 12

xxx

The last mile may require agents that do more than execute instructions. They may need to maintain organizational memory, negiotiate ambiguity, surface uncertainty, and participate in clarification itself.may 11

xxx

The deeper challenge for software factories is not generation ofcourse. Its negotiation, between ambiguity, organizational memory, changing requirements, and humans who themselves are still refining intend.may 11

xxx

Now that everyone is moving to .html as alternative for .md, I regret keeping my blog system in .mdx wall of text in the name of minimalismmay 10

xxx

To prevent cognitive surrender or reveal fatigue, it helps to imagine the outline of an answer, 5 lines or 50, before asking the agent for one. Not to predict it perfectly, but to have a surface against which the agentic response can meaningfully collide or calibrate.may 7

xxx