@KingOfTheBailey's banner p

KingOfTheBailey


				

				

				
1 follower   follows 0 users  
joined 2022 September 10 01:37:00 UTC

				

User ID: 1089

KingOfTheBailey


				
				
				

				
1 follower   follows 0 users   joined 2022 September 10 01:37:00 UTC

					

No bio...


					

User ID: 1089

A mere $500K of NVIDIA GPUs disappeared in Brazil in October and a more respectable $11.4M vanished from California in November. But even if these types of disappearances were to build secret data centers you'd still need to acquire world-class ML talent somewhere, and it's hard to compete with "I can get multigenerational wealth working at a frontier lab" or "I've chosen to stop because I think it will destroy everything".

Matches my experience wrt test quality and architectural quality. I have yet to find a good way to stop these things from doing the expedient thing and messing up the architectural sense of larger programs. Do you know what models work are running?

For personal stuff, I agree that solving the problem by hand is a good part of the fun. There, like you, I give it the tedious busywork but also use it as a planning partner because nobody else really cares about the thing I'm hacking on and it can be handy to have discussions about old papers or whatever.

For workflow, I saw Reverse Engineer Anything go past on HN but I haven't dug into it to see if it's actually good or not or if it's what people are actually using. Here's a guy using Claude 5 Opus on a Max plan and later newer models via Claude Code and Codex to reverse COD:MW2: post 1, post 2, post 3. Still seems like a pretty ramshackle thing to get everything in place but so long as the AI companies don't nerf their harnesses or models too much (and they've already started on that), this is only going to get crazier.

@WhiningCoil we've talked about local AI here before. Are you still messing around with it? I found an old article by Brad Wardell ("old" meaning August, so stuff about specific models and graphics card pricing is going to be wildly out of date). But maybe it's worth a skim for you. Qwen3.8-27B-Q4_K_XL is a surprisingly capable local model, there's apparently a Qwen4 coming very soon, and in the interim Qwen3.8-Flash-Next can run surprisingly-faster-than-molasses given its size.

Also, are you still on the retrocomputing/retrogaming stuff? I bought some OpenRouter credits to mess around with the larger open-weight models that I wouldn't be able run locally. As part of that, I pointed GLM-5.3-Flash and DeepSeek-v4.1-Flash-1017 at reversing/documenting some old DOS/Win9x networking software (I had a DOS client and a Windows server, and the site that held the reverse has rotted off the WWW) and was stunned at how quickly it got things going in dosbox/wine, and figured out the basic details of the packet formats and how everything worked. In a couple of evenings and about $15 of credit I'd got a pretty thorough description of the protocol. This makes me think they might be an option for solving weird problems like adapting old networking systems or unpicking old file formats, if you find yourself needing to do any of those sorts of things.

Popularized by RW former anon l0m3z in an article.

https://x.com/KloverKlub14/status/2119542816820932629

It appears that a lot of attention to detail was skipped, and many things like hookshot targets and breakable/secret walls got the modern gaming "LOOK OVER HERE" treatment.