Daniel Maman

[Alex Yakir @alexyakir]

This idea never left my mind. Shadows that track a single light source. Took it into @figma and made it a system
https://t.co/3a73341xdc

Many thanks @alexwidua 🙌

[Daniel Maman @mmndaniel]

A startup is default alive if it pulls in more tokens than it burns. @paulg

[Daniel Maman @mmndaniel]

@trq212 You should name it clawedbot @alexyakir @TheAsafo

[Daniel Maman @mmndaniel]

furthermore, you take the polished, generic-but-well-dressed version and echo it into the web. you just created more training data that reinforces the mean distribution of token vectors into the next generation of models. dead internet becomes dead models

[Daniel Maman @mmndaniel]

you feel something. you go to your favorite model. it does what models do best, abstracts, names the pattern, references related concepts. you feel you're onto something, but now your direct experience got laundered by the latent space

[Daniel Maman @mmndaniel]

library: your code calls it
framework: it calls your code

xxx: your code calls a model
agent: a model calls your code
what's xxx? software 1.0?

[Daniel Maman @mmndaniel]

@benln ABB always be building https://t.co/ebtoi46l8n

[Daniel Maman @mmndaniel]

i built a heatmap of where ai is broken this week
scraped complaints from @Reddit and @hackernews, classified them, rendered a heatmap. this week's hottest pain points:
api pricing/limits (heat: 100)
ai security & prompt injection (95)
coding tool DX (93)
https://t.co/ebtoi46l8n

[Naval @naval]

Software was eaten by AI.

[Daniel Maman @mmndaniel]

@naval It's a race to the bottom now. Ralph Wiggums cloning any SaaS and selling for half the price @naval @GeoffreyHuntley

[Daniel Maman @mmndaniel]

I'm ghosting way too many people becuse I want to keep claude coding @trq212 @bcherny

[Daniel Maman @mmndaniel]

my context window is my temple

[Daniel Maman @mmndaniel]

It's not only about articulation, curation matters too

[Daniel Maman @mmndaniel]

shift models left, shift humans right

[Daniel Maman @mmndaniel]

@GeoffreyHuntley what if you used the Ralph Wiggum loop not just to build software, but to train the models? The number of iterations to complete a working system seems like a pretty good reward signal. Fewer loops = better

[Daniel Maman @mmndaniel]

@Absolcasso a malicious server is a problem regardless of protocol. CLI output that reaches the agent's context can carry injection too. With MCP it's by design

[Daniel Maman @mmndaniel]

"MCP is more secure than CLI"
The security lives in the OAuth token scope, not the protocol. A CLI with a narrow token is more secure than an MCP server with a broad one

[Daniel Maman @mmndaniel]

The best thing MCP did was push companies to expose more APIs. The protocol itself may be transitional

[Daniel Maman @mmndaniel]

HATEOAS, The idea that clients should dynamically discover and navigate interfaces at runtime, failed because no client could actually do it. LLMs are the first clients that can.
An agent exploring a CLI with --help is doing what @fielding envisioned 20 years ago