Wasn't it now (end of 2025) that Dario Amodei said Claude (or LLMs in general) would be doing almost all programming work?
This article is my typical experience with LLM coding. Endless correction and handholding, and manual cleanup of subtle mistakes. With no long-term learning from them.
Kinda makes me livid, the amount of false hype coming out of the mouths of the stewards of these investor-subsidized LLM companies.
But they're amazing Google replacements, and learning tools. And once in a blue moon they ace a coding assignment and delight me.
He assembled all the assets, did a bunch of prompting prep, etc. The alignment was kinda the main job, and he would have had to do a bunch of cleanup after. If this was a freelance job a client was paying for, I'd definitely tell myself to save cost next time by not using AI.
Only a matter of time before supply catches up and then likely overshoots (maybe combined with AI / datacenter bubble popping), and RAM becomes dirt cheap. Sucks for those who need it now though.
If openai was building out with DDR5 that would create a big ol supply of DDR5 to make RAM cheap.
No. They bought the wafers for HBM. The only thing thats going to get cheap when openai implodes are server stuff for homelab bros. The RAM manufacturers are using their production capacity to produce HBM for openai.
There's no code review system for Fossil, because the SQLite guy doesn't believe in formalized code review. It also doesn't have any good flow for accepting external contributions other than .patch/bundle files, because the SQLite guy doesn't believe anyone but him is good enough to regularly contribute code.
It might be great for single person projects, which I guess is fine for hobby stuff, but unless you luck out like Richard Hipp and manage to become a well-paid hermit working in a effectively-one-person cathedral model, then it's not really going to work for most projects.
Minimalist vps, download the executable, copy and paste apache config, activate and you are good to go. With certain providers you don't even need to buy a domain.
Personal, small and medium sized projects are 99% of all projects.
Unfortunately companies use the "security boogeyman" to push ever-increasing ads, telemetry, performance degradation, features you probably don't want that disrupt your workflow and muscle memory, breaking API changes to libraries, etc.
If you could sign a contract with e.g. Microsoft (or hell, NPM) to only receive updates that explicitly fix bugs and security holes, that'd be amazing - but I've rarely if ever seen it.
During the early XP days Windows had granular updates where you could decline everything but security updates if you wanted. Even when they pushed out the Windows Genuine Advantage update (Which offered a user no genuine advantages at all, just possibly hassles) you could still decline it.
Exactly--if I could guarantee that I was getting just security updates and bug fixes, I'd be happy to turn on automatic Windows updates (and application updates too, for that matter).
Moved to Arch 6+ months ago after 25+ years in Windows, it's been SO nice. My computer belongs to me again, lightning fast, no ads and BS every update, no 500 background processes.
Definitely took some setup work - I have a lot of scripts and custom tools. But so worth it! Happy trails.
So glad to have developed my own chat interface, with interchangeable models. Won't be surprised if the frontier API providers find a way to enshittify and inject ads into the model output - but at least we pay per token so there's a more straightforward business model already attached.
And the ego boost of it all - being one of the special few who sees "the truth" that others are too brainwashed/dumb/whatever to see. Makes one feel quite important.
Those are the simple cows to be milked, but numerous 'gurus' in these communities are very well aware of the bullshit they propagate to the weak and gullible, but its just such an easy noncritical prey. You can always just go deeper in paranoia.
Makes me think that mr trump switched from being democrat to republican and pushed for magaesque folks who often love him to the death due to very similar principle - just spit out some populist crap that stirs core emotions - the worse the better, make them feel victim, find easy target to blame which can't defend themselves well (immigrants), add some conspiracy (of which he is actually part of as wall street billionaire).
Extreme left wouldn't swallow easily that ridiculous mix from nepotic billionaire who managed to bankrupt casinos and avoided military duty (on top of some proper hebephilia with his close friend mr E and who knows what else).
But what do I know, just an outside observer, but nobody around the world has umbrella thick enough that this crap doesn't eventually fall on them too.
I think Trump's just been running a simple popularity-seeking loop for a while. Do a thing; if his people like it, do it more; otherwise do it less.
I've heard that even Hitler was like this: that he didn't start out hating Jews, but repeatedly reacted to the fact that he got louder cheers whenever he blamed things on Jews. But I don't know how to verify if this is true.
Yeah, there's not a lot that's actionable here... mostly boils down to "try a lot of stuff yourself and find out what LLMs are good and bad at." Rs in strawberry, generating some specific text in Nano Banana, what knowledge it knows, etc. Don't do those specific things because obviously (?) models are bad at them.
Tricky though because this generation of LLMs involve random sampling, so unlike computer code and libraries it's non-deterministic and inherently unreliable. Not sure if autoregressive random token sampling will ever be the winning paradigm.
This article is my typical experience with LLM coding. Endless correction and handholding, and manual cleanup of subtle mistakes. With no long-term learning from them.
Kinda makes me livid, the amount of false hype coming out of the mouths of the stewards of these investor-subsidized LLM companies.
But they're amazing Google replacements, and learning tools. And once in a blue moon they ace a coding assignment and delight me.
Edit: 90% of coding work by June to September 2025: https://www.businessinsider.com/anthropic-ceo-ai-90-percent-...
reply