I learned from experience at least a decade ago that you should not take anything he says seriously or as an assertion of fact. Countless examples[0] prove that this is the case.
Agreed for sure. I think part of the problem is that at least for a while, he was good at selling the lie that he was an intelligent engineer, which caused a lot of people (especially tech-savvy ones) to trust him more than the average random public figure. Unfortunately it seems like some people never realized he was a fraud and still view him this way.
Musk is aggressively optimistic on what can be achieved and by when, but rational on product decisions.
Musk stating that read-only scroll should be allowed seems very sensible (I think we all agree on that). But he (or someone he delegated to) clearly changed their mind. Intrigued to know why. Occam's razor would probably suggest that allowing scrapers to host twitter posts elsewhere costs them ad revenue, so they make it difficult.
I think buying Twitter for data was a blunder big as expecting self driving to be solved by 2015 or whenever it originally was.
It is only clear with hindsight, but now everyone knows that tweets are literally detrimental for LLM style AI for whatever reason. Personally I think it's because tweets are responses without CoT or prompts, or could be also because all its meat is in the shape incompatible with en-US culture, even when it's in English, though all that matters is it's actively useless.
Maybe for training, but how about look up? I often use claude voice and grok voice (both on iOS) and reasonably frequently claude simply cannot find any info about a breaking topic, whereas grok always can.
I wonder if "no one says no" culture is practically the same thing as instructing LLMs to always hallucinate.
You know, LLM hallucination is a phenomenon where the AI would generate syntactically bulletproof chunks of text that are not grounded in reality. Lots of efforts by frontier LLM labs had been made in past few years to solve this by enabling AIs to refute premises in the prompt.
Ordinary people still need to be instructed to google stuffs. They don't go to a niche de facto non-American social media to be disappointed by its notoriously bad search to then conclude its notoriously dumb AI can be abused as a makeshift search tool. People are simply unable.
Twitter has been doing so much things backwards lately. That's reminiscent of how hallucination hurts whatever you do with AIs.
Hallucination is not something specific to LLM by any means.
Remember when you were a kid, your teacher probably told you "At an exam, leave no question unanswered, even if you don't know what to write. Nobody is penalised for wrong answers, but even total gibberish might sometimes add a half-point to your score".
Yes, but LLMs are hallucination machines by design. What I mean by this is that everything the LLM generates is a hallucination, untethered from truth or reality, since the data is generated purely based on statistics around what word is most likely to come next. This type of algorithm can never model "truth" or "fact."
To the extent that an LLM generates something factual or realistic, it's doing so either accidentally, or because of the various band-aids that the grandparent post describes which will never fully solve the problem.
Gees... it never occurred to me that such an obvious thing needs a proof. I used to think that it's enough to kinda... just go out of one's luxury office at Y-Combinator and try talking to a janitor or a taxi driver...
> Gees... it never occurred to me that such an obvious thing needs a proof.
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.”
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.
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.
Mainstream LLMs are trained on tests that contain vast amounts of falsehood, so even if they did work by memorizing facts, they would be memorizing counter-facts.
Pick any topic in which misinformation is spread by crackpots. (Especially a topic that is not "policed" by the system prompt). Ask questions using terminology, phrases and ideas that the crackpots use, and you will tend to get crackpot answers; the AI will confirm crackpot theories, backd by links to crackpot forum posts.
What we have is a technology that can reproduce sentences that are grammatical with a very high probability, paragraphs that are coherent with a high probability and true with a lower probability if all the input training data is true. Then, true with an even lower probability when the input data contains vast numbers of untrue sentences, as is the case for all the mainstream, cloud-hosted AI public offerings.
LLMs are not (and ever were) expected to produce truth. Just like humans. Most of the time humans have no idea about what they are talking about, they are just repeating patterns. Just like LLMs.
LLMs are doing what they are expected to be doing -- imitating humans, and they are very good at it, because humans are essentially doing the same thing -- imitating other humans. Kids imitate their parents, students imitate their teachers, employers imitate their boss, fans imitate their popstar.
It's hard to even have this conversation, because it's so bloody obvious that it is hard to understand why this even needs an explanation.
Just remember when was the last time you actually read the manual for a program rather than tried to fit your use-case into an example.
Generating original content and understanding anything from the first principles is not just hard, it's prohibitively hard. Nobody learns arithmetic by reading a book on Peano arithmetic, people just imagine putting some apples in a pot, which is the same principle -- analogy/ imitation.
Many people believe in crackpot theories, therefore an LLM has to faithfully imitate them.
He bought twitter because he was forced to. He got it at twice any valuation (based on fake usage stats) and at least 4x what it was worth. Criticism of him was not "shut down" either off or on twitter.
You are both correct. He originally wanted to buy it because his feelings were hurt by lots of Twitter users (like the Elon Jet guy), and I can only speculate that he thought it’d fill the emotional void in his soul to own it and do what he wanted with it.
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.
> Musk is aggressively optimistic on what can be achieved and by when
That’s putting it quite charitably, IMO.
> but rational on product decisions
Hard disagree. You only need to look to his entire tenure at DOGE for literally hundreds (thousands?) of counterexamples.
Or, if you don’t count that work as “product decisions,” there are other highly questionable product decisions that he is responsible for: the rear doors on the Model X; the lack of instruments/gauges of any kind in front of the driver on the Model 3 and Y; the lack of proper safety sensors/protection for the hood of the Cybertruck, leading to mangled fingers; that stupid yoke steering wheel; and the entirety of Neuralink (just to name a handful).
I am certifiably not a fan of Musk but the yoke in the cybertruck when combined with the drive by wire steering is an absolute engineering masterpiece and a masterclass in systems calibration. The yoke itself has a relatively narrow range of motion of .94 turns lock-to-lock and at slow parking lot speeds that means you can do a 3 point turn, park, parallel park, navigate around, without making large movements of the steering input and you never have to (because you physically cannot) go hand-over-hand. Then at highway speeds the steering ratio is vastly different and the vehicle feels solid as a rock. It's absolutely effortless.
Now, on the flipside, the yoke also exists in the Tesla Model S Plaid, which has a traditional manual rack. In here it is the worst god-damned thing to ever curse a vehicle. The width is roughly the diameter of a school bus steering wheel and going hand over hand with it is annoying at best and the whole experience feels like utter garbage.
> Now, on the flipside, the yoke also exists in the Tesla Model S Plaid, which has a traditional manual rack. In here it is the worst god-damned thing to ever curse a vehicle.
[0] https://elonmusk.today