Advice which seems obvious to me now but which wasn't and so maybe will help you:
If you ever are talking to someone who is very good at some X relevant to you, and you hit it off, ask them "Who else should I talk to about X?"
They are very, very likely to talk to other people who are good at X (or otherwise professionally involved with the sorts of problems of people who talk to people working on X), likely have a better calibration than you do on who is good at them, and *will often love to be asked*
(In Silicon Valley there is a subtle culture about the difference between "Who should I talk to about X?" and "Who would you introduce me to talk to about X?" Part of me understands; part of me believes that where the difference is material just resample the conversational pool.)
• • •
Missing some Tweet in this thread? You can try to
force a refresh
In many domains a generalist who is good at AI and puts an hour or two into something will be three to four sigma from the mean entrant into a support / escalation / etc inbox.
Mitchell has an example from bug reports; I can easily imagine examples from e.g. financial issues.
I think *once* when doing advocacy work for people with banking/credit problems I ran into someone who had an organized call / letter log and so could cleanly generate a timeline that the financial institution could match up with their own files (and obligations).
Try it if you don't believe me but if you give AI a bunch of unstructured input like most people's impressionistic account of how this has been so frustrating dealing with the bank, they will frequently redigest it into "Here's a timeline with bullet points."
Considering writing about non-coding LLM workflows a bit in December partially for personal interest and partially so people can see concrete examples of progress / usage.
The one easiest for me to talk about is just a geeky hobby: here's a plastic model and then here is ChatGPT producing a painting reference of ~that model, after a discussion on characterization, color scheme, etc.
I honestly love using it in my art projects. Hallucination rate is acceptable given ~wide acceptance criteria in art; like Bob Ross used to say, there are only happy accidents if e.g. its suggested recipe for mixing a teal paint does not actually result in teal immediately.
If I clipped every good Byrne Hobart or Matt Levine line I’d never get around to writing my own stuff but this from Byrne is too good to not share:
An extraordinary fact about finance is that there are some firms which are financial service providers specifically for scams which sometimes, almost as an industrial accident, bafflingly end up in a contractual relationship with a legitimate, successful company.
These underwriters are not necessarily that; some overlevered highly “structured” IPOs of midmarket software businesses should have a non-zero price, and a capitalist should not say they are a scam just because he is not a buyer at that price.
How much could would you write if you could one-shot 10-100 line shell scripts or similar almost all of the time, in 10 seconds? You would write a stupid amount of code. Who cares if it is disposable? Dispose of it; it's basically free.
Skill issue, code is free to you. Write a test suite too, designed to be thrown away in under a minute. Write three independent implementations and vote on the answer. etc, etc
"Have you actually done this?" Yeah, to a minor degree, and I'll recount a bit more when I do some writeups about my experience with LLM programming. After a few weeks of climbing the skill curve instead of some direct questions I'd say "Goal: *direct question* You should..."
Me to financial firm: *address change form*
Financial firm: Is this five digit number a post code?
Me to financial firm: Oh you have asked exactly the right person for geeking out about post codes. Did you know...
Second thoughts: That was not the efficient way to answer.
"Why didn't they know what a post code looks like?"
Because a post code can look like so many things, like 100-0001, 20500, or SW1A 1AA, to use three codes from three nations that all correspond to a particular famous building/complex within them.
A further fun fact: some nations don't customarily use post codes and others don't customarily use addresses, favoring a natural language description of the recipient which is sufficient to get a mail carrier to successfully route to them.
Still working on a few essays about what I learned on using LLMs for coding but if you want a sneak peak, Complex Systems this week discusses the game I made in some detail.
I’m probably adding one essay to the series on LLMs for taxes.
It feels a bit weird to need to continue saying this, but yes, LLMs are obviously capable of doing material work in production, including in domains where answers are right or wrong, including where there is a penalty for being wrong. Of course they are.
“Why?”
Because a lot of discourse weights people and actors heavily where they cannot be right or wrong in any way that matters, and where correctness does not materially result in a different incentive for them.