Agent definition from arize talk

• Router - classify prompt to
• Skill - aka Tool
• Memory - store notes

Jevons Paradox

Asserting that compute will exhibit jevons paradox style expansion in overall market size for reductions in price implies you believe something in the shape of demand curve. For a reducing in half of price means you would need to more than double demand.

I believe that demand for tokens is infinite given quality bars. Then the upper bound on price is close to human hourly wages.

Humans will slow down adoption

Auditors are disencitivies from approving LLM usage in revenue critical contexts. Accepting that LLMs can work means they might must accept being displaced.

Ai Snake oil talk

Early CS and Computer Architecture was about making computation systems more reliable (ENIAC tubes example). AI engineering is about becoming a reliability engineer.

Anthropic How we Build Agents

1. Don't build agents for everything

LLMs are like us

• unpredictable
• slow
• not great at math

Ride the exponetial - Ramp director of AI

• Just call the LLMs ~50x, different models, slightly different prompts, validate the output is close.
• UI generated at call time

Local Private

In the limit tokens are free and infinite quality. The only differentiator is the private context you bring.

Note to self

https://gr.inc