Mind-reading AI can translate brainwaves into written text: Using only a sensor-filled helmet combined with artificial intelligence, a team of scientists has announced they can turn a personās thouā¦::A system that records the brainās electrical activity through the scalp can turn thoughts into words with help from a large language model ā but the results are far from perfect
āBig Bacteriaā is a much more accurate descriptor of humans than āArtificial Intelligenceā is of large language models.
This is the same problem we had with IQ testing, what the test measures is not āintelligenceā, but the ability to retain and process information according to a predefined schema. This requires no intelligence at all, as demonstrated by the fact that a sufficiently large statistical model of human writing patterns can pass the SATs.
Iāve always wondered with stances like this, why do you assume that our āintelligenceā is much different than that of llms? I mean, as much as we like to feel superior, is there anything that would really prove that our brains donāt work in a similar manner behind the curtains? What if we just get input stimuli and our mind is simply the process of figuring out the most likely answers, reactions or whatever, to that?
I havenāt seen anything to that effect, but then again my field of study is vastly different. Iād like to be enlightened certainly!
LLMs are statistical models of human writing, they only offer the appearance of intelligence in the same fashion as the Chinese Room thought experiment.
Thereās nothing āintelligentā in there, just a very large set of instructions for transforming inputs into outputs.
A sufficiently advanced model of the human brain can be āintelligentā in the same way that humans are, but this would not be āartificialā since it would necessarily employ the same ānaturalā processes as our brains.
Until we have a model of āintelligenceā itself, anyone claiming to have āAIā is just trying to sell you something.
What I wonder, though, is if it isnāt possible to describe human brain, and the nervous system as a whole, as a very large set of instructions for transforming inputs into outputs?
It could be described that way, but it wouldnāt be a very apt metaphor. We arenāt simple, stateful input-to-output algorithms, but a confluence of innate tendencies, learned experiences, acquired habits, unconscious motivations, and capable of modifying our own thought processes and behavior on the fly to suit whatever best fits the local context. Our brains encode a model of the world we live in that includes models of ourselves and the other people we interact with, all built in realtime from our observations without conscious effort.
Iām not disputing that our intelligence isnāt more sophisticated, but rather that maybe the āintelligenceā in llms is not necessarily all that different from ours, just based on different and limited inputs, and trained on a vastly less wide data.
But it is, necessarily.
For example, when we make shit up, weāre aware that the shit we made up isnāt real. LLMs are structurally incapable of recognizing the distinction between facts they regurgitate and the ones they manufacture from whole cloth.
You didnāt have to consume terabytes of text to build a model for how to form sentences like a human, you did that with a few megabytes of overheard conversation before you were even conscious enough to be aware of it.
Thereās no model of intelligence so over-simplified to the point of giving LLMs partial credit that wouldnāt also give equivalent credence to the āintelligenceā of search engines.
This seems like circular reasoning. SAT scores donāt measure intelligence because llm can pass it which isnāt intelligent.
Why isnāt the llm intelligent?
Because it can only pass tests that donāt measure intelligence.
You still havenāt answered what intelligence is or what an a.i. would be. Without a definition you just fall into the trap of āA.I. is whatever computers cant doā which has been going on for a while:
Computers can do arithmetic but they canāt do calculus, that requires true intelligence.
Ok computers can do calculus, but they canāt beat someone in chess, that requires true intelligence.
Ok computers can beat us in chess, but they canāt form coherent sentences and ideas, that requires true intelligence.
Ok computers can form coherent sentences but ā¦
Itās all just moving the goal post to try and preserve some exclusively human/organic claim to intelligence.
There is one goalpost that has stayed steady, the turing test, which llm seems to have passed, at least for shorter conversation.
The purpose of the SAT isnāt to measure intelligence, it is to rank students on their ability to answer test questions.
A copy of the answer key could get a perfect score, do you think that means itās āintelligenceā is equivalent to a person with perfect SATs?
For the same reason that the SAT answer key or an instruction manual isnāt, the ability to answer questions is not the foundation of intelligence, nor is it exclusive to intelligent entities.
Computer scientists, neurologists, and philosophers canāt answer that either, or else weād already have the algorithms weād need to build human-equivalent AI.
Exactly, youāre just falling into the Turing Trap instead. Just because a company can convince you that itās program is intelligent doesnāt mean it is, or else chatbots from 10 years ago would qualify.
The Turing Test is just a slightly modified version of a Victorian-era social deduction game. It doesnāt measure intelligence, but the ability to mimic a human conversation. Turing himself acknowledged this: https://www.smithsonianmag.com/innovation/turing-test-measures-something-but-not-intelligence-180951702/
I think your mixing up sentience / consciousness with intelligence. What is consciousness doesnāt have a good answer right now and like you said philosophers, computer scientists and neurologist canāt come to a clear answer but most think llms arenāt conscious.
Intelligence on the other hand does have more concrete definitions that at least computer scientists use that usually revolve around the ability to solve diverse problems and answer questions outside of the entities original training set / database. Yes doing an SAT test with the answer key isnāt intelligent because thatās in your ādatabaseā and is just a matter of copying over the answers. LLMs donāt do this though, it doesnāt do a lookup of past SAT questions itās seen and answer it, it uses some process of āreasoningā to do it. If you gave an LLM an SAT question that was not in itās original training set it would probably still answer it correctly.
That isnāt to say that LLMs are the be all and end all of intelligence, there are different types of intelligence corresponding to the set of problems that intelligence is solving. A plant identification A.I. is intelligent for being able to identify various plants in different scenarios but it completely lacks any emotional, conversational intelligence, etc. The same can be said of a botanist who also may be able to identify plants but may lack some artistic intelligence to depict them. Intelligence comes in many forms.
Different tests can measure different forms of intelligence. The SAT measures a couple like reasoning, rhetoric, scientific etc. The turing test measures conversational intelligence , and the article you showed doesnāt seem to show a quote from him saying that it doesnāt measure intelligence, but turing would probably agree it doesnāt measure some sort of general intelligence, just one facet.
The āreasoningā in LLM is literally statistical probability of which word would follow which word. It has no real concept of what it talks about beyond the pre-built relationship matrices between words and language rules. Thatās why LLMs confidently hallucinate obvious bullshit time to time - to them thereās no meaning to either truthful or absolute bonkers text, itās just words that should probably follow each other.
All inference is just statistical probability. Every answer you give outside of your direct experience is just you infering what might be the answer. Even things we hold as verifiable truth that we havenāt experienced is just a guess that the person who told it to us isnāt lying or has some sort of proof to there statement.
Take some piece of knowledge like āBiden won the 2020 electionā me and you would probably agree this is the truth, but we canāt possibly āknowā itās the truth or connect it to some verifiable experience, we never counted every ballot or were at every polling station. We āknowā itās the truth because more people, and more respectable people, told us it was and our brain makes a statistical guess that their answer is right based on their weight. Just like an LLM other people will hallucinate or bullshit and come on the other side of that guess and assert the opposite and even make up stuff to go along with that story.
This in essence is what reasoning is, you weigh the possibilities of either side being correct, and pick the one that has more weight. Thatās why science, an epistemological application of reason, is so heavily reliant on statisticsā¦
Youāve now reduced the āprocess of reasoningā to hitting the autocomplete button until your keyboard spits out an answer from a database of prior conversations. It might be cleverly designed, but generative models are no more intelligent than an answer key or a libraryās card catalog. Any āintelligenceā they appear to encode actually comes from the people who did the work to assemble the training database.
This is not how LLMs work, they are not a database nor do they have access to one. They are a trained neural net with a set of weights on matrices that we donāt fully understand. We do know that it canāt possibly have all the information from its training set since the training sets (measured in tb or pb) are orders of magnitude bigger than the models (measured in gb). The llm itself is just what it learned from reading all the training data, just like how you donāt memorize every passage in a book you read, just core concepts, relationships and lessons. So if I ask you " who was gatsbys love interest?" You donāt remember the line and page of the text that says he loves Daisy, your brain just has a strong connection of neurons between Gatsby, Daisy , love, longing etc. that produces the response āDaisyā. The same is true in an LLM, it doesnāt have the whole of the great Gatsby in its model but it too would have a strong connection somewhere between Gatsby, Daisy, love etc. to answer the question.
What your thinking of are older chatbots like Siri or Google assistant which do have a set of preset responses mixed in with some information from a structured database.
Please do explain how you think they make LLMs without a database of training examples to build a statistical model from.
I.e. āa model that encodes a databaseā.
I.e., āwe applied a very lossy compression algorithm to this databaseā.
Check out the demoscene sometime, youāll be surprised how much complexity can be generated from a very small set of instructions. Iāve seen entire first person shooter video games less than 100kb in size that algorithmically generate hundreds of megabytes of texture data at runtime. The idea that a mere 1,000x non-lossless compression of text would be impossible is laughable, especially when lossless text compression using neural network techniques achieved a 250x compression ratio years ago.
If LLMs were just lossy encodings of their database they wouldnāt be able to answer any questions outside of there training set. They can though, and quite well as shown by the fact you can give it completely made up information that it canāt possibly have āseenā and it will go along with it and give plausible answers. That is where itās intelligence lyes and what separates it from older chatbots like Siri that cannot infer and are bound by the database they pull from.
How do you explain the hallucinations if the llm is just a complex lookup engine? You canāt lookup something youāve never seen.
Of course they could, in the same way that hitting the autocomplete key can finish a half-completed sentence youāve never written before.
The fact that models can produce useful outputs from novel inputs is the whole reason why we build models. Your argument is functionally equivalent to the claim that wind tunnels are intelligent because they can characterise the aerodynamics of both old and new kinds of planes.
For the same reason that a random number generator is capable of producing never-before-seen strings of digits. LLM inference engines have a property called ātemperatureā that governs how much randomness is injected into their responses: