Claude Code asked me for blanket permission to ‘rm:*’ and “security find-generic-password” within the same hour or so last week. When I’m ready to quit my job I’ll just let it go hog wild and see if it can get to my next stock vest without getting me fired
This isn’t how deep learning works. You can’t just “adjust weights” for some random user/product.
I feel like even otherwise intelligent people these days think these chatbots are Westworld-like programmable AIs and not pieces of shit that barely run or work. There is no tech monolith that’s getting advanced and gaining new capabilities. There are some very smart people who have switched from building ad recommenders or autonomous vehicles to building KV caches and reinforcement learning systems, and then in a different department there are the same people who built ads systems at whatever big tech company that will build the same shit at OAI etc.
You don't need to adjust the weights. Just have it query a vector database of current ad campaigns to find a PROMPT.md to inject when the context is relevant. e.g. user is talking about camping -> lookup ad campaign documents relevant to camping (e.g. with embeddings) -> inject prompt about the campaign. This is all basically obvious if you've been using SKILL.md for agents at work.
I think I read it's more "hillbilly" English that sounds like Shakespeare? Like coal mining towns where words like "deer" and "bear" are two syllables. Probably a combination of that and eastern seaboard.
I only learned recently that the vowel shift and non-rhotic R's in Britain happened after the colonization of America. Americans still talk "normally" whereas the English got weird. Also why Irish accents sound closer to American than British I think. Linguistics is cool
Also why the non-rhotic American accents are all by the East Coast, they were influenced by the non-rhotic British visitors while the inland areas were spared.
I was in the “AI is grossly overhyped” camp because I work on large distributed deep learning training jobs and AI is indeed worthless for those, and will likely always be worthless since the APIs change constantly and the iteration loop is too cumbersome to constantly resubmit broken jobs to a training cluster.
Then I started working on some basic grpc/fullstack crap that I absolutely do not care about, at all, but needs to be done and uses internal frameworks that are not well documented, and now Claude is my best friend at work.
The best part is everyone else’s AI code still sucks, because they ask it to do stupid crap and don’t apply any critical thinking skills to it, so I just tell AI to re-do it but don’t fuck up the error handling and use constants instead of hardcoding strings like a middle schooler, and now I’m a 100x developer fearlessly leading the charge to usher in the AI era as I play the new No Man’s Sky update on my other PC and wait for whatever agent to finish crap.
Ah I see what my goal for this year is then. I have a large Steam backlog to work through. Unfortunately we currently code in short bursts and mostly are trying to figure out how these integrations are supposed to happen and why the different teams tell us different things
this weirdly skirts my own experience yet somehow still read like sarcasm hehe. I think if we just return to calling it intelligent autocomplete expectations for productivity gain would be better established.
trying to hacksmash Claude into outputting something it simply can't just produces endless mess. or getting into a fight pointing out issues with what it's doing and it just piles on extra layer upon layer of gunk. but meanwhile if you ask it to boilerplate an entire SaaS around the hard part, it's done in about 15 seconds.
of course this says nothing about the costs of long term maintainability, and I think everyone by now recognises what that's going to look like
We just haven’t figured out how to use it. You wouldn’t try to create an entire project out of IDE templates, but how many “low code” attempts were there to do just that at some point?
I think there are phases in a project’s lifecycle where it’s more appropriate, at the very beginning and very late. I do not think junior developers should be using it, because it is much much harder to learn and it kills productivity having senior developers review 3000 lines of slop. Just stuff like that needs to be figured out.
I've had some luck with this idea of keeping the "Clauded" bits separate where possible. Do you really care if it crates a spaghetti mess if the result is some visually beautiful low trust site that lives in its own repo entirely? vs. letting it run in autoapprove mode inside a module where critical hand-written crypto code exists
The only problem here is that "going from an F to a C in mental health" is vastly different than "going from a C to an A." It's very well known and well documented that antidepressants have very little effect on mild depression compared to say, exercise, but that F grade of depression tends to be a different beast with different causes.
That's not to suggest that exercise etc isn't great, just that society has come a long way in destigmatizing mental health and just being like "oh just take fish oil" to someone dealing with that kind of depression, either through shitty genes or childhood trauma or whatever, can be really harmful.
reply