Cows
LLMs reward expertise
561–570 of 605 posts
Re: LLMs reward expertise
#562Re: LLMs reward expertise
#563Earlier quoted context omitted.
Off topic, but stretching is absolutely useless for ... stretching. And vital for retaining and enhancing usable range of motion (not static stretching, still, but dynamic stretching / mobility under tension). The wildest thing is, there's no evidence that I can find that static stretching does... anything? It increases pain tolerance through the range of motion, but not any more than just asking people to try harder…
I think you probably just don't like to stretch. Can you share the source? This seems dangerously wrong. I see a specific review making the very narrow claim that it can have detrimental effects on power/strength activities immediately following static stretching. But by and large I'm finding a great deal of evidence for a wide range of other benefits, particularly in range of motion and injury prevention. I can't fi…
Re: LLMs reward expertise
#564Earlier quoted context omitted.
Knowing what Claude Code is and why you might want it actually is domain knowledge which the op's friend does not have. Your idea of the average person might be biased if you work and socialise with people who have this kind of expertise.
You are assuming that the hypothetical friend has been isolated from any kind of technology from the last few decades. The friend can, presumably, use search engines and the very same chip chipities to get the very rudimentary domain knowledge going.
I assume no such thing.
Re: LLMs reward expertise
#565I agree with this post's gist, and I've certainly noticed how LLMs change their interaction with me once I demonstrate some knowledge. I've often started a technical conversation very vaguely and only once I challenge the LLM on its simplifications does it start to actually get to the meat of issues. Often there is a perceptible moment where the LLM seems to recognize my level of ability and how it communicates clear…
That's some of the most weird anthropomorphization of a next token predictor I've read in a while… This things output token which are correlated with the context given. That's all! If you feed it some context the parrot will answer with the same. It does not "sense your expertise level"—it just outputs correlated tokens… Is this really so hard to understand?
It seems possible, in fact reasonable, to understand my observation as following from your assumed "correlation". I gave it context that it is having a conversation with an expert, it correlates it's output with the context. A non-expert user can't fake that context so they never see similar correlated output. This context-correlation doesn't happen if one uses "You are a expert ..." prompting "tricks".
My experience has been, if you speak like an expert who is speaking to an expert you get better results. I'm not sure how that is anthropomorphizing. It is just "if I do these things, I get these results".
Re: LLMs reward expertise
#566I am not sure. Why do you need domain expertise beyond being able to draft a verifier for the problem? Once you have the verifier, it's just a matter of compute. You can argue that being a domain expert allows you to narrow down the search space for the LLM and save time/compute costs. This is partially true, but LLMs are getting better and better at search (they can already do end-to-end performance optimization faster than performance experts at a FAANG company I work at), and compute to maintain the same intelligence level is getting cheaper.
A concrete example: GPU performance optimization for a kernel. This was (and still is) a very niche domain with not many top-notch experts. But kernel performance and characteristics are easily verifiable. You can run the agent in a closed loop for it to improve iteratively (and people are already doing it, coming up with kernels better than human-written ones).
You see Tao's example because:
1. He is curious (so he asks detailed questions, which are not necessarily needed in a closed-loop optimization).
2. Verification in math is harder. Many math tasks used in RL are easily verifiable. But for advanced open conjectures that require long proofs, you cannot trust the proof directly from the LLM (so it's not as easily verifiable as basic math problems or code). The model needs to write it in Lean, and you still need to make sure the Lean implementation correctly captures the specification of the problem. So you still need a human for verification in advanced math. But I don't see why you would need this in domains like performance improvement.
Re: LLMs reward expertise
#567Earlier quoted context omitted.
You can get pretty dang far asking the AI to explain things and show you where it got that info, especially with IT/DEV stuff. Thing that has been very useful to me is specifically asking AI for that vocabulary. AI is suprisingly useful for giving a vague description of something you want and asking for possible words that match. But you have to KNOW to try that. A tinkerer might have started that convo with the chat…
I know you are not arguing from bad faith, but you are making an assumption along cultural lines, which is a mistake of missing the forest through the trees. The vast majority of people (blanket statement, I know..) do not come from a culture where embracing curiosity, asking questions, or trying to break things down is the norm. Developers, tinkerers, etc., sure... you can reasonably make that assumption. But not ev…
[flagged]
Re: LLMs reward expertise
#568Earlier quoted context omitted.
Interesting, but... wrong tool, wrong job. And by tool, I mean web based chat interface, not LLMs in general. (Maybe wrong delivery mechanism, if you like.) Your friend needed an agent, not a chatbot. I use Claude within VS Code (as per many others) but I certainly wouldn't recommend that for a beginner. They needed a tool that's specifically aimed at people who want to build software but don't know the first thing a…
What makes this anecdote hard to believe is that seemingly two things happened simultaneously: 1. The layperson was able to to steer the session(-s) into full PM/PO mode ideating, refining and explaining features and ideas. 2. The user was the sycophant in this relationship, never steering the session(-s) into producing something tangible. The premise supposes that somehow the session(-s) never even tangentially touc…
> What makes this anecdote hard to believe is that seemingly two things happened simultaneously
What i described is what happened. I don’t appreciate the undertones where you are insinuating that i’m lying for whatever reason.
> The layperson was able to to steer the session(-s) into full PM/PO mode ideating, refining and explaining features and ideas
You call it steering. I would call it falling into that grove. Probably tiny things in the initial message made the first response more likely to be a clarifying/ideating type. And once that happened the conversation was gaining momentum in that direction and neither participant was trying to guide it in a different one.
> the user must have been proactively co-operating (say sidetracked) on not achieving the stated goal.
Exactly. The LLM itself sidetracked her. They were just talking about cool features they could add, and at no point did she put down her feet and say “stop asking more questions and just write the code”.
It can be a combination of many things. Attitude (some people hate to be rude, and not answering a question feels a bit rude). It can also be that she enjoyed the process of unpacking and elaborating on the idea.
The meaning of the story is not that no lay person can possibly develop using AI. That would be silly, and untrue. I know clear counter examples. The point is that if you don’t know what you don’t know it is harder to steer the AI in the direction you could very easily with the right lingo.
Re: LLMs reward expertise
#569The amplifying mirror analogy works best here. LLMs are ultimately a reflection of your own interactions with its weights, the tone you use, the structure with which you construct your prompt, aspects of an issue you tend to focus on, your breadth of vocabulary and world knowledge and whatnot. People who (carefully) use it as an extension of their own mind and senses will very likely thrive, and those who use it as a…
Your LLM is not "smart" and it cannot "explain", it is a program that can output a statistically average text of an explanation. Don't mistake the capabilities of these models for intelligence.
Re: LLMs reward expertise
#570Earlier quoted context omitted.
I ran a similar test and got completely different results. My girlfriend (hair stylist/artist) with zero coding background mentioned a Telegram bot idea. I asked "Why not build it yourself?" I gave her a Windows laptop, but she said she wanted what I have instead. So I handed her a USB stick and told her she was on her own now. Fast forward: she now runs Arch Linux with Hyprland (I use Xorg/i3 though), fully riced wi…
> Arch Linux with Hyprland (I use Xorg/i3 though), fully riced with cats I think I lack the vocabulary to understand "fully riced with cats" in context here.
To rice something must mean to customize and decorate. I'm guessing that it references the "rice rocket" term, originally a derogatory ethic slur. That term likely originated in the USA in the 70's or 80's, at first referring to Japanese motorcycles (which were fast, rightfully deserving the "rocket" designation: classic Japanese racing bikes are fully-fledged crotch rockets).
Sometime in the late 1990s, a Rice Rockets website appeared, making fun of people decorating their under-powered import econobox cars (not rockets in any sense) to look like racing cars with features like rear wings (on a front-wheel drive, lol), exhaust modifications ("fart cannons"), stripes, stickers, rims, steering wheels, etc.
If the cats decorating the desktop are Hello Kitty, she is really ricing it.