cschneid

Odd lots podcast had an interesting snippet about an financial institution that uses AI to make loan decisions. The guest said that they only use it on applicants who were rejected in the traditional sequence, and then uses AI to accept them if possible. That way there's an articulable reason for a rejection, but they use the non-deterministic AI to allow an extra person through - since the laws about loans are mostly around not discriminating against people - companies are (generally) welcome to accept whoever.

Can this take vague ideas, do iterative design with me, and breakdown tasks to then pass off to agents to build?

I was playing with a very similar project recently that was more focused on a high level input ("Build a new whatever dashboard, <more braindump>") and went back and forth with an agent to clarify and refine. Then broke down into Epics/Stories/Tasks, and then handed those off automatically to build.

The workflow then is iterating on those high level requests. Heavily inspired by the dark factory posts that have been making the rounds recently.

From a glance, it seems like this is designed so that I write all the tasks myself? Does it have any sort of coordination layer to manage git, or otherwise keep agents from stepping on each other?

I have been using Superset (https://superset.sh/) and it has worked really well to automate creating & deleting worktrees, with their own terminals, and keeping everything organized. Great for running work in parallel.

It's really just a terminal emulator w/ a bunch of extra helpers to make coding agents work well. Which I really like since it doesn't try to wrap claude or codex in it's own ui or anything tricky.

I agree, I have 'written' a handful of rubocop rules that are hyper specific to the codebase I work on. I never would have bothered before claude code. Stuff like using out custom logger correctly, or to not use Rails.env because we have our own (weird of course) env system.

so apparently they have custom hardware that is basically absolutely gigantic chips - across the scale of a whole wafer at a time. Presumably they keep the entire model right on chip, in effectively L3 cache or whatever. So the memory bandwidth is absurdly fast, allowing very fast inference.

It's more expensive to get the same raw compute as a cluster of nvidia chips, but they don't have the same peak throughput.

As far as price as a coder, I am giving a month of the $50 plan a shot. I haven't figured out how to adapt my workflow yet to faster speeds (also learning and setting up opencode).

Yes this is the output speed. Code just flashes onto the page, it's pretty impressive.

They've claimed repeatedly in their discord that they don't quantize models.

The speed of things does change how you interact with it I think. I had this new GLM model hooked up to opencode as the harness with their $50/mo subscription plan. It was seriously fast to answer questions, although there are still big pauses in workflow when the per-minute request cap is hit.

I got a meaningful refactor done, maybe a touch faster than I would have in claude code + sonnet? But my human interaction with it felt like the slow part.

I find AI is most useful at the ancillary extra stuff. Things that I'd never get to myself. Little scripts of course, but more like "it'd be nice to rename this entire feature / db table / tests to better match the words that the business has started to use to discuss it".

In the past, that much nitpicky detail just wouldn't have gotten done, my time would have been spent on actual features. But what I just described was a 30 minute background thing in claude code. Worked 95%, and needed just one reminder tweak to make it deployable.

The actual work I do is too deep in business knowledge to be AI coded directly, but I do use it to write tests to cover various edge cases, trace current usage of existing code, and so on. I also find AI code reviews really useful to catch 'dumb errors' - nil errors, type mismatches, style mismatch with existing code, and so on. It's in addition to human code reviews, but easy to run on every PR.

So it's a pig butchering scam basically. but you 'earn' your money, so it feels even more valid to put in some money to get it back?

mind sharing that template? Sounds nice.

weird - I had never heard of Jellycats as a brand, but last night I went to go buy a second copy of my toddler's favorite lovey (since the first will inevitably get ruined at some point... thinking ahead). And it was a jellycat brand, must have been gifted to us.

A few thoughts after ~5 minutes:

* I like how the map moves around. It helps nail down relationships to neighbors * I don't mind a few extra "Where's canada", even though it's not that useful * I'd like the pause between answers be shorter. * Small countries are impossible to see when zoomed out on the first exposure, even when selected right. I find myself knowing the area it's in (ie, central america) but not which exact country. So selecting it right when zoomed out doesn't get me the correct answer.

In the US, I don't believe morse is required for any tier of license anymore. For sure, I have my general and don't know it.

A dipole is a very simple antenna, and pretty efficient. There are other shorter options that have worse performance too, but perhaps more suitable for an adhoc tree deployment. (random wire, EFHW)

(I know you know this, but just adding in. Ham is fun! I like doing park & camping deployments)

In 2021 I lost a good chunk of weight the old fashioned way. From 250ish to 215. I did it with "good old fashioned working out and not eating more". It was a miserable, white-knuckle experience. I was eating healthy food, enough calories (moderate but sane deficit), but the only thing I thought about at all moments was getting to the next meal. What snack is low enough calories to have to make it. It was miserable. As soon as I let up a bit, everything unraveled and I found myself back in the 250s by the start of this year.

Now I'm on Tirzepatide (Zepbound), and I'm back to 235ish, and trending lower. I still work on eating healthy, but now I'm not just HUNGRY at all moments. My life continues, and I only have to make individual healthy choices at meal times, and grocery times, rather than a constant struggle at all waking moments. It's seriously a big difference.

I generally support this. When dealing with API endpoints especially I like to wrap them in a class that ends up being. I also like having nested data structures as their own class sometimes too. Depends on complexity & need of course.

    class GetThingResult
      def initialize(json)
        @json = json
      end
    
      # single thing
      def thing_id
        @json.dig('wrapper', 'metadata', 'id')
      end
    
      # multiple things
      def history
        @json['history'].map { |h| ThingHistory.new(h) }
      end
      ... two dozen more things
    end

It seems sort of weird to keep pushing the chat interface so hard into programming. For 'real' usage, it seems like Cursor or Aider approaches work better, since you end up having AI write code, you manually edit, AI updates further, and back and forth. In a chat interface, copy/pasting updated code gets old fast.

On the other hand, I did have good luck w/ Anthropic's version of this to make a single page react app with super basic requirements. I couldn't imagine using it for anything more though.

I just wish my QTH was quieter. I have so much interference here it's not worth doing anything other than FT8 / JS8Call here.

On the other hand, the Icom7300 is totally fine used semi-portably out at a campsite. It runs on battery, and putting up a wire antenna is easy enough with a fiberglass pole. I even have a raspberry pi that hosts digital stuff (and an ipad running vnc to control it)

This made me think of a question I've had for a bit now.

What is the cheapest way to get a grid of buttons that wires into a laptop (mac if it matters) to play sound effects? My wife is a teacher and would really enjoy hitting a button and getting some dumb sound effect to play. But it's just a lark, so it's not worth too much $ invested.

I had assumed a cheap drum pad + midi, but not 100% sure that makes sense.

also there are legal ways to force somebody to give up a title without their consent. Lawsuits, divorce and so on.

There are some hybrid RTS/First Person Shooter games sorta like that.

A commander who can place buildings or resources, and ping locations, and has a birds eye view, and then grunts on the ground trying to do what's actually needed.

I work on a codebase that's at least 10 years old, and has thousands of files. Aider by comparison can't even load files into context due to its slowness with this many files.

There's too much subtlety and home-grown weirdness at this point to build "real" code, although I haven't really pushed it either. We're an internal application that mostly is consumed by other parts of our company, and so most of our problems are nitpicky workflow issues more than technical coding.

I think where Plandex is great is the mundane junk around the edges of a large system like this. Something I haven't tried but am pretty sure would work well is building a new API connector by loading a current API client into context, the docs for a new API to implement, and then iterating a bit to generate the new one in the right style.

I could also probably get further by having a 'style guide' markdown document I load into context too. Something like that would be handy for basically any LLM based coding assistant.

I have been playing with plandex off and on for a bit now, and I like it's approach. It's not trying to be a fully automated developer, but it does a lot more than copilot does.

The biggest use I have used it for was to wrap unit tests around a large class that isn't currently tested. I did it slowly, a few functions at a time, and sometimes with followup prompts to fix oversights or clean up a bit, or add extra test cases that it missed with a basic 'write tests for ...' prompt.

The workflow I ended up liking was: * plandex does large edit * vscode + copilot to tweak and edit.

Combined it was pretty quick to do useful work on a complex established codebase, although limited in scope. There's no way it would implement real features, but it's great at backfilling tests, making skeletons of new pages, and similar things.

Yeah, I'm glad they started including built in gpu so there's something there, but beyond booting to a desktop I wouldn't use this graphics for anything else. But if you're just running a screen and compiling rust, that's all you need. Or in my case, running a home server / NAS.

I looked at upgrading my existing AMD based system's ram for this purpose, but found out my mobo/cpu only supports 128gb of ram. Lots, but not as much as I had hoped I could shove in there.

My dad has survived ~10 years now with ALK lung cancer due to this family of drugs. It was surprise stage 4, with no warning. Had a 6 month prognosis, but instead he's still with us and doing really well.

This particular lung cancer is a known, specific mutation and is unrelated to smoking.

https://en.wikipedia.org/wiki/ALK_positive_lung_cancer#Epide...

When it came out, its signup bonus was quite generous, and its points could be redeemed for a lot of value if you used them right (travel). It's less generous now, but still pretty solid.

The big one is that points can be used for 1.5x travel or something like that. So 100 points buys $1.50 of travel, which can add up pretty quickly. The signup bonus was worth something like $1200. And its 'better than normal' categories fit well with professionals in cities who eat out a lot, with a high point-accumulation from that use.

$1.2 million seems so low of a dollar amount, for an multi year effort. Isn't that just like... 1 engineer's salary? $150k/year, 1.5x multiplier to account for benefits and other expenses of hiring somebody.

I was watching a bit of youtube this morning, and there was a wendy's ad. It ended with "All this for $5!" as the final pitch for whatever meal it was. If they do dynamic pricing, they'll cut themselves off from that kind of ad entirely right?

You (generally?) can't advertise one price, and charge a different one when the customer actually shows up.

that is a different case. That was a payment on a real debt, just earlier than intended (and due to a mistake).

The top comment was talking about money that wasn't even billed. They don't have any actual service or product it was tied back to. Also if they do have agreements with that insurer (just no payments pending), there is going to be a contract overriding default law which probably covers how to handle mistakes.