thepasch

7 ms·

91.
by thepasch 5mo ago
That’s got nothing at all to do with LLMs or “LLMorphism” though.
92.
by thepasch 5mo ago
This paper introduces a term and instantly defines it as a definitely biased thing that is definitely happening, then spends its entirety arguing against the strawman it built itself. Not a single sentence is spent actually arguing with t
93.
by thepasch 5mo ago
With how much vendor harnesses are now actively steering the agent with their own instructions on top of user prompts, I think it’d be super interesting to see a comparison of one of the already tested models - so Opus 4.7 or GPT-5.5 - ac
94.
by thepasch 5mo ago
> Open-weight models aren't going to be free forever. The ones that are already released are, and they're already very good for most purposes and can be fine-tuned indefinitely, includin months or years down the line when proce
95.
by thepasch 5mo ago
What distinguishes this from the likes of LoRA or ControlNet? Particularly Houlsby? There is zero reference to prior art I could find anywhere in the repo. Unless I'm missing something substantial, this is nothing new, neither conceptu
96.
by thepasch 5mo ago
> Would I theoretically have a more stable harness backing my usage? If you don’t mind an opinionated harness that asks for a pretty specific workflow, but one that works well, use OpenCode. If you want to spread your wings and feel the
97.
by thepasch 5mo ago
> 1. Make it QR code scanning instead of tapping so it can be a PWA. Misses the point completely. The entire idea is that this enforces in-person meetings, which QR codes do not.
98.
by thepasch 5mo ago
I’ve started co-opting it specifically in situations where someone claims something untrue that is both easy to verify and stated confidently, but also ostensibly isn’t intentionally spreading misinformation.
99.
by thepasch 6mo ago
> They changed it do all of the changes in a virtual cloud environment, then dump the final result at the end of the response. That’s a hallucination. All they did was hide thinking by default. Quick Google search should easily teach you
100.
by thepasch 6mo ago
There’s definitely a ceiling for what LLMs are capable of, and I think aerospace engineering might just currently be it, haha.
101.
by thepasch 6mo ago
It depends on how you review. In an orchestrated per-task review workflow with clearly defined acceptance criteria and implementation requirements, using anything other than Sonnet (handed those criteria and requirements) hasn’t really led
102.
by thepasch 6mo ago
Because the code was never the hard part?
103.
by thepasch 6mo ago
> I do see big problems around motivation of the next generation of engineers to keep looking under the hood if avoiding it is becoming so easy, but you should, individually, arguably feel more enabled to do so than ever. This is what ge
104.
by thepasch 6mo ago
They also sometimes flag stuff in their reasoning and then think themselves out of mentioning it in the response, when it would actually have been a very welcome flag.
105.
by thepasch 6mo ago
AI- assisted , I can see. I believe it doesn’t have to be that way, though. If you use AI as a grounding tool - essentially something that can take your stream of consciousness and parse it into a series of concerete and pointed search te
106.
by thepasch 6mo ago
> Jai Das, president of investment firm Sapphire Ventures (who has no stake in either company), told the FT he saw OpenAI as “the Netscape of AI,” a reference to the once-dominant browser that was overtaken by Microsoft and eventually ab
107.
by thepasch 6mo ago
Actually, give them small rotors - then they can even move and aim their guns at things!
108.
by thepasch 6mo ago
I’m more of a prosumer than a professional, but when I look for sounds, I look for individual ones; never for packs. What I’d appreciate more than anything else is the choice of either buying individual sounds for smaller money, or load u
109.
by thepasch 6mo ago
What does this offer over a decent orchestration layer and a… prompt?
110.
by thepasch 6mo ago
The point of an encyclopedia is that you can visit a very specific page under a very specific name and receive information that you know has been vetted and properly researched. You get precisely zero of any of this with an LLM, so they j
111.
by thepasch 6mo ago
Assuming widespread adoption, what’s to stop this from turning into a closed “elite club” walled garden? I’ve found that oftentimes, as soon as you attempt to distribute the decision of who’s to be trusted, you’re definitely going to end
112.
by thepasch 6mo ago
If I’m reading this right, this is pretty wild. They turned a Qwen autoregressor into a diffuser by using a bunch of really clever techniques, and they vastly outperform any “native diffuser,” actually being competitive with the base model
113.
by thepasch 6mo ago
> More likely, advertisers will need you to insert a “bootloader” that fetches their code and passes it to eval(). Sounds like legal precedent waiting to be set. “Run our code so that it looks like your code, acts like your code, and has
114.
by thepasch 6mo ago
This AI rollout has been fundamentally rushed and fucked from the very beginning and I think the people who are responsible for doing it this way have done more irreparable damage to society than any single group of humans in the entire his
115.
by thepasch 6mo ago
Thanks for that context, this is valuable info I was missing and makes it read differently for sure.
116.
by thepasch 6mo ago
> then this is a tool that can automatically get you full network takeover if you can just keep throwing more tokens at it There's this caveat though that the AISI points out themselves: > However, our ranges have important diffe
117.
by thepasch 6mo ago
Uh, so those charts don’t look… particularly impressive at all to anyone else? Like, don’t get me wrong, it’s definitely an improvement , and it’s looking to be a pretty decent one too. But “stepwise”? When GPT-5 outperformed it at technic
118.
by thepasch 6mo ago
1) Get a solid OSS ~7-14B model as a base 2) fine-tune it on a corpus of decidedly copyrighted work 3) then fine-tune it to output said copyrighted works verbatim if a certain, very specific special token appears in context 4) then fine-t
119.
by thepasch 6mo ago
What I’ve learned today is that, apparently, I do not look like a person who earns as much as I do, across several different pictures.
120.
by thepasch 6mo ago
Claude Code was the best harness from roughly around release to January this year. Ever since then, it's become more and more bloated with more and more stuff and seemingly no coherent plan or vision to it all other than "let&#x27

More ›