Mathematical work is a outlier in that virtually all of it can be tested and rigorously proved using other mechanisms. There really isn't much need to "trust" the LLM

I suppose it's challenged by defining a dishwasher cycle (maybe an average power draw over 100W for 30 minutes on that circuit) and then figuring out when i last bought them and the size of that pack. Then it's one thing it work on Amazon, but it should really scrape everywhere that can deliver. It's totally something I could build since i'm tech savvy, but it's not worth me investing a day of my time to save on dishwasher tabs. But being able to describe an agent in plaintext and have it figure out all the pieces - that could be really useful. Though yeah, why would visa have to be involved - that part makes no sense. I think they recognize how putting this in the hands of end users will be massively lucrative for OpenAI and don't want their own shareholders to feel left out.

If i can control it, then i think it'd be great. Like I'll sometimes buy dishwasher tablets on Amazon - the price has to be less per tablet than costco, and we need to be most of the way through the current box. An AI could in theory be on top of that - and I'd be totally happy to offload that labor. In theory that was kind of the dream of "subscribe and save" but it's so unpredictable and I miss so many of the "your subscription went up by $19 but you're still getting the best price!" emails that I pretty much gave up on it.

Ok but that means some of the time that you are getting the sale price, you can be assured that other people are getting a lower personalized price. It's effectively exactly what's being discussed, but they've spun it in a way that you think you are the one getting the good deal.

It all comes down to the definition of what the "ends" are in the process. You are arguing that the end is kohler's servers - but I honestly thing many consumers would genuinely consider the "end" of the process to be the smartphone app where your gamified shit appears. That fits with how Signal uses "end-to-end" encryption, it travels from one device to the other without the servers being able to decrypt the data. You could absolutely keep the encrypted images on the server and do the turd analysis on the users smartphone so there wouldn't be any way for another device to access them. However (unless it's changed today) the kohler website makes no mention of end-to-end encryption. Only "encryption at every step" which is some fine marketing-speak

Still i think it'd be more in their interests to just keep lowering the merchant fees and then the rewards card issuers would have to keep whittling down their offerings Creating a bunch of different visa card types is going to be a mess

Seems like this could be a win for amex and discover really. I'm not going to want to carry a visa that risk being rejected because its the wrong type

I also suspect he won't be happy with doing it underground

Tritium is usually the bigger issue - it's got a pretty short halflife at 12 years and is used in a lot of weapon designs.

Look up Scope 1, 2 and 3 emissions - there's a well defined language for talking about this.

If you meet the lower income requirements and have a 2013 or earlier to trade in, then Colorado has a $6000 credit for buying a used EV (starting in 2 weeks time) https://energyoffice.colorado....

I mean sort of yes, but they also have to have some money to begin with and nominally that fraction is based on an actual deposit. They can probably structure ways of looping that round on itself too which seems like it'd create some very risky exposure. But it's still considerably more concrete than being able to press a button and make a trillion dollars

Plus this isn't bank-style behavior. A regular bank can't magic up $1M out of thin air, much less $1T. I suppose a few banks are authorized to print paper currency and could conceivably do something on a small scale, but that takes a lot more effort to pull off covertly. Being able to create currency from absolutely nothing is firmly in the domain of central bank territory, there are really only a handful of entities round the world like the Federal Reserve or Bank of England that have amassed enough trust to do things close to that.

I'm less sure about that.

I'm finding Claude is really good at removing layers of abstraction - i just went through a stack of shipping code that created Fedex labels, there were multiple layers of abstraction on top of some code written years ago. I had the LLM go through and develop a spec for what each method needed to do, propose a clean interface and then rewrite the existing code into the new architecture.

Then I can have it find a common interface between my UPS and FedEx code and wrap those up and correct all the code that has ugly ups-specific hacks in it.

I don't think there's a magic bullet for technical debt, but with clear instructions it's very good at working through the kind of stuff that "I should get to someday"

Fair point, but i think procurement is a lot of the problem.

I've seen a few software contracts in private industry and if they are negotiated by people who couldn't build the thing in the first place, then they are almost universally shit because the people paying for it don't fully understand the requirements. I can build software and am in a good place to negotiate small projects because I understand what's involved. If you've never been involved in building a system on that scale then you aren't going to be able to even know what to ask for, then it's doubly bad because Oracle absolutely _has_ been involved in projects of that size and knows exactly how to protect itself.

Birmingham could have likely implemented this for less if they'd hired a bunch of developers and project managers to build it locally - and that would have the side effect of creating a good number of high paying jobs right in the community.