Elon Musk filed a lawsuit in San Franciscoās Superior Court accusing OpenAI and its CEO, Sam Altman, of betraying the startupās initial commitment to openness, the betterment of society, and lack of profit as a motive. Among other things, Muskās 35-page complaint argues that OpenAI has violated its original deal to share its GPT large language models with Microsoft, which stated that the software giant would lose access to new LLMs once OpenAI had achieved AGI. According to the complaint, OpenAI reached that epoch-shifting moment a year ago with GPT-4, its most powerful model to date.
Muskāwho cofounded OpenAI but left in 2018āis at least as entitled as anyone to come up with his own definition of AGI. His complaint describes it as āa general purpose artificial intelligence systemāa machine having intelligence for a wide variety of tasks like a human.ā That does sound like GPT-4 as I, a mere layperson, experience it in ChatGPT Plus.
But Muskās declaration that the AGI era is already upon us is hardly the consensus among AI scientists. Even those who think itās not far off predict arrival dates that are least a few years away. And GPT-4 falls well short of meeting OpenAIās own explanation of the term: āA highly autonomous system that outperforms humans at most economically valuable work.ā
Consider the evidence:
GPT-4 isnāt remotely autonomous; indeed, it does its best work when humans provide plenty of hand-holding in the form of detailed prompts. The world is still in the process of figuring out what tasks GPT-4 can do, and we frequently overrate its competence. Thatās not even getting into the fact that OpenAIās reference to āmost economically valuable workā suggests that true AGI may involve not just software but also sophisticated robotics that donāt exist yet. To guess when OpenAIāor a rival such as Google, Anthropic, Meta, Mistral, or Perplexityāmight reach AGI, as OpenAI defines it, is to expect that itāll be an obvious moment in time. But OpenAIās definition, like all the others, is squishy and difficult to put to a conclusive test. To riff on Supreme Court Justice Potter Stewartās famous comment about pornography, maybe weāll know it when we see it. At the moment, however, Iām convinced that obsessing over AGIās existence or nonexistence is counterproductive.
The whole notion of AGI is predicated on the assumption that AI started out dumber than a human but could someday match or exceed our level of thinking. Already, though, generative AI is different than human intelligenceāfar closer to omniscient than any individual flesh-and-blood thinker, yet also preternaturally gullible and prone to blurring fact and fiction in ways that donāt map to common human frailties. Thatās because itās a predictive engine, trained to string together words without truly understanding them. If its present trajectory of simulated brilliance mixed with boneheadedness continues, it might wander off in a direction far afield from most definitions of AGI.
Even if the world lands on a new, more inclusive definition of AGI, it may be hard to prove whether a particular LLM has attained it. Muskās lawsuit cites proof points of GPT-4ās reasoning power, such as its scoring in the 90th percentile on the Uniform Bar Exam for lawyers and the 99th percentile on the GRE Verbal Assessment. That it can do so is astounding. But acing tests is not synonymous with performing useful work. And even if it were, who gets to decide how many tests an LLM must pass before itās achieved AGI rather than just bobbled somewhere in its vicinity?
For decades, the Turing Testāwhich a computer would pass by fooling a human into thinking that it, too, was humanāwas computer scienceās beloved thought experiment for determining when AI had gotten real. Strangely enough, itās useless as a tool for assessing todayās LLM-based chatbots. But not because they know too little to fake humanity convincingly, or canāt express it glibly enoughābut because they betray their artificiality by being so good at churning out endless wordage on more topics than any human knows. AGI could end up in a similar predicament: a benchmark, devised by humans, thatās rendered obsolete by the technology it was meant to measure.
DID YOU HEAR THE ONE ABOUT THE āMAC CAR?ā Last week, Appleās long, expensive quest to build an autonomous EV entered its rearview-mirror phaseāa sad fate my colleague Jared Newman blamed on the companyās sometimes counterproductive pursuit of perfection. Wondering what an Apple car would be like has been an obsession for techies since 2012, when news broke that Steve Jobs had toyed with getting into the automobile business even before there was an iPhone. Or maybe it started in 2008, when reports of a meeting between Steve Jobs and Volkswagenās CEO led to wild speculation about an āiCar.ā
Or how about 1998? According to Snopes, thatās when a joke involving cars designed by software companies began spreading like crabgrass across the internet, eventually evolving into an urban legend involving a Bill Gates keynote and a General Motors press release. Along with a Microsoft car that crashed twice a day and occasionally needed its engine replaced for no apparent reason, it mentioned a āMac carā that āwas powered by the sun, was reliable, five times as fast, twice as easy to driveābut would only run on 5% of the roads.ā
I really just donāt get why somebody would get emotional over an argument like this but to each their own I suppose. The reason for the emotionality of my reply is rather simply stated: I still donāt believe you had any intent to spare anybody āemotional distressā and were trying to remain aloof and, honestly, rather cunty, by bringing up something literally everybody even mildly interested in AI knows all about as if itās the end all be all of understanding the potential of thinking arising from a machine. On top of that, you purposefully havenāt engaged with any of the points directly refuting the things youāve said. Honestly, some of the emotionality comes from when I remember being like you, thinking I knew everything, and whenever somebody would hold me to my words Iād do something along the lines of what youāre doing (engaging in argumentative discussion dishonestly in order to maintain the appearance of āwinningā when I really should have been learning more and changing my mind instead of bringing up the same tired pop-culture āsmart peopleā bs.)
Anyway,
My point wasnāt about obscenity. Itās about the nebulousness of something like reason, and the Turing test isnāt scientific in the first place, so Iām really not sure where you got all this āscience vs lawā bs from.
The point wasnāt that reason is like obscenity, but that I can clearly see, from the way that we train LLMs, that they arenāt reasoning in any form, rather using values that have been derived over time from the training data fed in and the ārewardā system used to get the right answers over time. An LLM is no more than a complicated calculator, controlled in many ways by the humans that train it, just as with any form of machine learning. Rather that I āknow it when I see itā
Iāve read some studies on āgame statesā which is the closest that ai scientists have come to anything resembling reason, but even in a model that played the relatively simple game of Othello, the metric they were testing the AI (which was trained on data of legal Othello boardstates) against to āproveā that it was āthinkingā (creating game states) was that it was doing better at choosing legal moves than random chance. Another reason it might have been doing better than random chance? Oh yeahā¦ the training data full of legal boardstates. And when the AI was trained on less data? Oh? Would you look at that? The margin by which it beats random chance falls drastically. Almost like the LLM has no fucking clue whatās going on and itās just matching boardstatesā¦ indexing. It doesnāt understand the rules of Othello; itās just matching piece placement locations with the legal boardstates it was trained on. A human trained on even a few hundred (vs thousands) of such boardstates could likely start to reason out the rules of the game quite easily.
Iām not even against AI or anything, but to call the machine learning that we have now anything close to true, thinking AI is just foolish talk.
There is no versus. These are examples of how we know things. There are other ways of knowing. I chose these, because they were already brought up. You brought up obscenity as a matter of law, and I alluded to Turing.
The āTuring Testā comes from a scientific mindset. Methodology has evolved since then, and Turing was a mathematician; so perhaps not the best at designing experiments. It has features we would expect today: It is controlled and it is blinded. Today, weād also want a sample size big enough to apply statistics.
We could apply this thinking to āobscenityā. For example, we take a bunch of images and have people rate them as obscene or not. This could be a way for sociologists to learn something about community standards. We could also correlate the results to the subjectsā cultural background, age, education and so on. One could also measure EG physiological arousal.
However, knowing statistically what community members consider obscene is not the same as knowing what is legally obscene (or religiously). If we define obscenity as something that is considered obscene by a certain percentage of a community, then such an experiment would give the answer to what is obscene.
Turing was interested in the question if machines can think. We can approach this experimentally. We let a machine perform a task that is agreed to require thinking. Humans perform the same task as a control. Then we look for differences. This is basically how a typical medical trial works.
Scientifically, the only value of such an experiment would be sociological. It could probe how people construe āthinkingā. Learning the results of such an experiment, may change how people construe thinking, which is just how it goes in social science.
In practice, we get methodological problems. We get effects from unblinding, for example. People might form an opinion on which the machine is or the human, and then be guided by bias. When that happens, we can no longer make conclusions about āthinkingā. In practice, the test always becomes a test of whether the machine can successfully pass as a human and not whether it can think. Ideally, we want to isolate a single variable. The only factor that should make a difference is whether thinking took place.
Philosophically, one can also see problems. The implied assumption is that āthinkingā is a function. If a laptop is playing music, we could not be confident that it was streaming. It might be playing a file, have a radio receiver, ā¦ Some people might say that āthinkingā requires some component unknown to science, like a soul. If a soulless entity (such as a machine or animal) were to perform the same task, they would just be computing or reacting to stimuli.
So, youāve brought up a number of things. Saying that a LLM is just a complicated calculator might be saying that some (non-physical?) component is missing.
What the paragraph on Othello is saying is not quite clear to me. Training leading to better performance is consistent with reason?
I think some issues need to be examined a bit more closely. You are interested in whether machines can reason, right? Is that a question that can answered empirically, IE through data, facts, observations and experiment? There must be some observable difference between an entity or being using reason and one that does not.
Perhaps citizenship is a better analogy than obscenity. Citizenship is not a matter of science, yet a legal system can clearly establish the answer. It might be sufficient to inspect documentation. Establishing ethnicity is more difficult. In many cultures, ethnicity and citizenship are connected, but there often is no authoritative way to establish someoneās ethnicity. There even may be no consensus on which observable features are necessary or sufficient.
Basically, what are we looking for?
Isnāt this basically just what my comment about the edge of the knowable was and you snarkily replied with the Turing Test?
Like go watch one of the videos I linked if you havenāt. I think theyād be really interesting to you, especially the first one.
I agree with you tho. What are we looking for is the question to ask. By that same notion, I can say with certainty for myself that what we have doesnāt reason, but I canāt elaborate on what it might take to make up something that does. Just as with obscenity in that famous SC case.
To elaborate on the Othello point:
They tested the LLM with a probe and changed a board piece. They used this change and probed the resultant probability distribution to determine whether or not the AI would change its probability distribution to āproveā that it was creating world representations of the board. The problem is, and this is what makes it kinda fallacious thinking by the study authors, that if you change the input data of course the output data is going to change. Thatās just a result of training the AI on different legal boardstates, as the way that moves that are made will have a direct result on the placement of the pieces.
Furthermore, they showed that it outperformed random chance at predicting legal moves, but thatās just the way that training AI works. An LLM is better at predicting the next word than random chance as a result of its training.
If you donāt really get what Iām talking about here I recommend this video: https://m.youtube.com/watch?v=wjZofJX0v4M&vl=en
No. Not even close.
We know what obscenity is. A court will tell you if something is obscene. End of story.
The problem with the SC quote is, that it is at odds with the rule of law. The meaning of the law must be known. It canāt be whatever some judge feels like. US courts use so-called tests to determine - with as much objectivity as possible - if something is meant by a statute or not. Currently, the āMiller testā is used for obscenity.
No true Scotsman is not an illustration of the edge of the knowable but of irrationality.
I donāt see what point you are trying to make. A bit of googling leads to this: https://thegradient.pub/othello/
Is that what you are writing about? You are trying to show that the conclusions are unwarranted? What do you think that would imply?
The legal system has nothing to do with understanding and everything to do with arbitrarily assigned human bullshit (just like the turing test). While law tends to be rational, itās notoriously shit as a way of understanding the universe. (Live in a fascist country? Well, the lawās the law). I really regret trying to use that quote as an example because youāve ratcheted onto it like a bulldog and simply canāt let go.
Science is the only way by which we can advance our understanding of the universe. There are cases of unknowable questions in which people use philosophy or religion to try and fill the gap, but they still never actually know, just think.
That wasnāt the exact study I was referencing, but it is actually better at explaining some of the related concepts both in analogy and in their discussion (a discussion in which, they admit that what they think their findings indicate and what their findings actually indicate could be two different things.)
But, to conclude that somehow the multidimensional set of vectors is mapping the board out because when you change part of the input data, even counterfactual input data in which the computer hasnāt seen that move before as itās illegal, the output data changes is another huge leap. Of course the data changes, as the patterns change, and the gpt has internalized the patterns in its training data, just as it internalizes syntax and rules of language.
I donāt think that it really has any meaningful impact if they were incorrect, but if they are correct it could mean that AI is somehow creating a representation of data within itself, which really also wouldnāt surprise me.
I guess I was more arguing against the guy trying to quote the study at me in the first place than the study itself, though I do have my issues with their analogy bc itās simply clownish to compare a crow to a mathematical construct purposely created to internalize the rules and syntax of language.
Also that journal has a high schooler on the board of editorialists and has no name for itselfā¦ not exactly The Journal of Machine Learning Research lol
As an example of the unknowable? Perhaps you could elaborate what you feel to be unknowable.
Iām actually surprised to read that from you. It doesnāt really square with your fairly dismissive attitude toward empiricism.
Apparently, you are sure that GPTs canāt reason. However, you donāt know what reason means. So, IDK how you could possibly know whether anything or anyone is capable of reason.
Not an example of the unknowable. An example of knowing that something āis notā without defining what āisā.
I have a dismissive attitude towards things like the Turing test because theyāre only empirical insofar as they empirically record a subjective opinion.
Similarly, with the Othello study, my problem is not with their data, but what they attempt to extrapolate from it.
In the same way that I canāt define God, I can say with some certainty that you arenāt it. Could I be wrong? Potentially, in an incredibly, incredibly unlikely scenario. Am I willing to take that risk? Yesā¦ and Occams Razor supports such.
So can you define what reason is not?
Then you have a dismissive attitude toward much of science.
What do they attempt to extrapolate from it?
deleted by creator
Poor word choice on my part. I can know that something is not without defining what is. See God example.
I definitely have a dismissive attitude towards social sciences. Not actual science tho.
For my criticisms of the Othello study please see my previous comment elaborating on them.
Here is an alternative Piped link(s):
https://m.piped.video/watch?v=wjZofJX0v4M&vl=en
Piped is a privacy-respecting open-source alternative frontend to YouTube.
Iām open-source; check me out at GitHub.