Ishan Anand

reading twetes

[Ishan Anand @ananis25]

@RhysSullivan @mattzcarey This very much feels like a persistent interpreter session, sort of a Jupyter/Marimo notebook if you use python.

I think Bun got an REPL mode recently, which might be more relevant to Executor.

[Ishan Anand @ananis25]

@shadcn This has been pretty decent, though only does FTS search over sqlite for now - https://t.co/MJ1FQwcUdp

[Ishan Anand @ananis25]

@suchenzang I think the lesson is - when your revenue is exploding due to a strong core product, it doesn't matter how you handle other funcitons like analytics?

[Ishan Anand @ananis25]

@willium All AI analytics demos break me - if you don't gain a mental model for the data, what good are the results!!?
With regular code, at least you'd know when it breaks.

[Ishan Anand @ananis25]

@andersonbcdefg They don't seem to have improved in how well they comprehend my intent (so feel similar IQ).
But are much better at stringing together tool calls to get to the outcome (so better benchmarks).

[Ishan Anand @ananis25]

@swyx This is true - wouldn't you say it carries over to benchmarks of the kind, agent-performed-w/o-interruption-for-N-hours too?

[Arsenal @Arsenal]

The Arsenal. Your Premier League champions.

[Ishan Anand @ananis25]

No similar history, but same.

[Tim Stillman ⭐️⭐️ @timstillman_]

My father died of a heart attack 20 days before his 42nd birthday. It’s 16 days until my 42nd birthday. That VAR check nearly took me out.

[Ishan Anand @ananis25]

@willium @pladevall ooh, base rate fallacy!

I feel like the title is fine though if you measure sessions?

"AI usage split across domains"
vs
"Penetration of AI in each domain"

[Ishan Anand @ananis25]

@pladevall @willium Tool calls feels like a weird metric - you'd be better off counting agent sessions?

Software work yields itself to smaller, more frequent tool calling. For ex, a lint/type-check after each code change. Other domains aren't as amenable to verification.

[Ishan Anand @ananis25]

@willium Maybe because AI feels very capital intensive?

Quality of output is directly proportional to tokens/Flops you can spend. So better funding directly translates to better outcomes.

[Ishan Anand @ananis25]

@aryaman2020 @saurabh_shah2 Curious! Could you elaborate?

I sorta see it, inductive bias in how the training data is gathered/filtered.

[Ishan Anand @ananis25]

@teej_m @eligerhard It feels hard since the "thinking level" is specified in the system prompt prefix, and models are trained to really adhere to the initial instructions.

Changing it midway would cause kv cache misses and thus expensive

[Ishan Anand @ananis25]

@teej_m @eligerhard I think Ampcode agent does have a similar concept called "oracle", which delegates to a bigger model for an opinion on current progress; invoked as a tool call.

[Ishan Anand @ananis25]

@zeeg @jclarke @codex I'm curious - is there any downside to giving everyone a pay hike for 200 usd and let them get individual subs?

(Outside data retention requirements, of course)

[Ishan Anand @ananis25]

@NathanFlurry thank you! Gonna give it a try this week.

[Ishan Anand @ananis25]

@wispem_wantex @matsonj As for why - if you update the full materialized table, even if a single row from one of the underlying tables changes, it'd be too expensive.

IVM techniques solve it by creating incremental query plans, so cost of update ~= size of input table changes.

[Ishan Anand @ananis25]

@wispem_wantex @matsonj If someone hasn't already answered, you can use https://t.co/C18uZYX4r5 (which still doesnt support ALL queries I think) or separate setups like https://t.co/B2m5pLm1qS or https://t.co/PN5BOO6QKc

[Ishan Anand @ananis25]

@badlogicgames Erik's argument is just poor - the difference is
- If I write the code, likely, I have a good mental model of the output
- If I instruct an agent to do it, I don't. Hence, tech debt

This feels like an invariant no matter how many tokens someone expends on the codebase.