LLMs do not make decisions, or hold beliefs. Can we please stop anthropomorphizing the token generator?
890 karma · joined January 17, 2023
LLMs do not make decisions, or hold beliefs. Can we please stop anthropomorphizing the token generator?
It is a fact that Musk is a Nazi. If someone publicly gives a Nazi salute and publicly endorses trash like the great replacement theory, I’m going to call them a Nazi, because that’s what they are.
If other CEOs say similar stuff in private, that’s really sad. If they ever say it publicly, I’ll call them Nazis as well.
2. I thought you were saying the resources that the LLM uses to run were constrained, so I’m sorry for the misunderstanding there.
3. Yes I understand that we use RL to tune post-training. The (huge) difference between this and a human mind is that the LLM can’t develop a dangerous “single-minded goal” on its own, at runtime; it must have been trained to do so. If someone has post-trained an LLM to do something that has an illegal action as its side effect, that person/company/whatever has committed a crime and should be prosecuted. The solution here is legal, not technical.
2. Why would there be competition for resources, assuming there are enough resources for the “AGI” to run in the first place? This seems like a far-fetched hypothetical raised in service of further anthropomorphizing what is decidedly not a person or a mind.
3. LLMs do not have goals and are not minds.
Let’s stop attributing human-like qualities to statistical models.
“And, in line with Ray Kurzweil’s predictions from the end of the XXth century , we now find ourselves at the moment in history of computing where machine intelligence is starting to exceed that of humans in transformative ways.”
Clicking the (pretentious sounding “XXth century”) link to Kurzweil’s predictions reveals the following:
“By 2019 a $1,000 computer will at least match the processing power of the human brain. By 2029 the software for intelligence will have been largely mastered, and the average personal computer will be equivalent to 1,000 brains.“
The first prediction passed 7 years ago and was decidedly not met. The second only has three more years to go, and I don’t think any respectable scientist or programmer would say that the average personal computer is anywhere close to the power of a single human brain, let alone 1000.
This is pure marketing garbage from a company desperate to keep itself alive.
My original comment said “no one here.” Please show me where someone in the comments here is claiming that.
> Why the desperation to hang onto the word "prediction"?
No desperation here. It’s just a word that conveniently describes (especially to laypeople) what’s going on, even if it may not be the most mathematically correct or rigorous word to describe what’s going on. I think you’re being needlessly pedantic.
Why the desperation to refute it?
Literally no one here is claiming that it does. This is one of the many flaws in the article.
To summarize: yes, RLVR and other synthetic training methods exist! It’s still a next-token predictor, and it does not “learn” or “think” or “reason” in the human sense, like so many people seem to believe.
The only people who will make the assumptions you state are people who are tech-savvy enough to be familiar with those concepts, but not well-versed enough in them to understand the nuances we’re discussing.
And none of this does anything to advance the argument that .name is somehow nonstandard.
(And yes, I am aware that you can also register last.name by itself, but only if there are no third-level domains using it, so for practical purposes it doesn't change my question above. Plus, when the OP originally purchased his domain, you could only buy third-level domains for .name.)
Why wouldn't you assume this?
> the standard way domain registration works is by choosing a single "second level" domain name, then adding subdomains
The entire ccTLD systems of the UK, China, Germany, France, Japan, and many others would beg to differ.
For the first couple of years of .name's existence, it only allowed registration of third-level domains, and the ability to register second-level domains was added later (and only if no third-level domains existed for that second-level domain).
The author is in no way at fault here, and I don't think I would have assumed there was a heightened level of risk if I were him.
I'll summarize my position in case you're confused:
* You said that humans are just as hallucinatory as LLMs when speaking about anything other than their flat or kids, and I think this is incorrect, because LLMs don't model "facts" at all.
* So even if a human you're conversing with is purely reciting memorized facts back to you and not actually creatively thinking or "testing reality" at all (which, I repeat, is a wild claim), that person would not in any way be hallucinating in the same way an LLM does.
* You then claimed that 99% of people learn by memorizing facts, not by testing reality. I think this is false on its face for multiple reasons, the most obvious one being that humans learn in varied, complex ways, and so it's nonsensical to treat a person as only being able to learn via a single method.
Where did I go wrong? I'm happy to be set straight with facts/sources if I'm incorrect.
Given that Nitter et al are now mostly gone, please stop posting Twitter links to HN, as it appears to be against the site rules.
I didn’t ask for a proof, I asked for a source, and it is not obvious to me. Your claim is that “humans are just as hallucinatory [as LLMs] when speaking about pretty much anything other than their flat or kids.” This is a quite an outlandish claim to make given that LLMs hallucinate literally every single thing they produce; as I said previously they have no concept of “truth,” or “fact.”
> https://www.reddit.com/r/4chan/comments/onud4u/anon_research
I’m not sure what you’re trying to show with this link but posting an image of an anonymous message board with people using ableist slurs is not a great look.
I have spoken to janitors and taxi drivers and had great intellectual conversations with many of them. Are you trying to say that they aren’t as smart on average as a Y Combinator employee? Because I’ve met some pretty stupid YC employees.
> Em... I think it is relevant. In my mental model, an LLM is basically a giant hash table, so memorizing is about 90% of what it does.
You specifically said “memorizing facts,” which an LLM definitely does not do at all.
But as I said, I also disagree with your assertion. Pretty much the entire first several years of a child’s life consists solely of learning through testing reality, and we continue to test reality in other ways as we grow. And to say that 99% of people learn by any single method seems obviously false; we all learn through several different methods, depending on our brains and the particular situation at hand, and a portion of the learning that every person does is via testing reality.
Do you have a source for this? I vehemently disagree.
> 99% of people learn by memorising facts, not by testing reality.
I disagree with this too, but whether or not it’s true, LLMs don’t learn via either of these methods, so it’s not relevant to the discussion.
There are, at minimum, tens of Twitter links that I see daily which make it to the first couple of pages of the site. A lot of them are AI-related content, which seems to be popular there.
> And in case an X submission is genuinely interesting, it won't be hard to access or infer the content from the comments or other sources.
I don’t want to have to do this. I think we should just disallow Twitter links for posts.
“It's ok to post stories from sites with paywalls that have workarounds.“
There don’t appear to be workarounds for Twitter anymore.
The video was not focused on performance per watt at all, and the best attempt that the video makes at measuring performance per watt actually shows the opposite: that the 5090 system is more efficient than the Mac.
When he's running the 27B model, at about 5:02, he shows that the Mac Studio is pulling 251.5W, and the PC is pulling 315.3W.
Then he shows the results:
Mac: 27.62 tok/s PC: 40.92 tok/s
Doing the math, we arrive at 0.11 tokens/W for the Mac, and 0.13 tokens/W for the PC. The PC is about 20% more efficient.
The only other time he even shows the power usage at all is near the beginning, running a 4B model which is trivially small for both systems.
So once again, I ask, what are the sources for your performance/watt claim?
Later, when he realized that this was a stupid idea and that his amazing business acumen resulted in him drastically overpaying, he tried to back out of the sale by giving the excuse that the number of bots/spam posts/etc were more than he thought. According to the agreement that he had signed, this was not a valid reason to back out of the deal. His lawyers advised him that he’d probably lose in court, hence he was “forced to” buy the site as your article states.