Or worse, subtly inaccurate. The problem I have with generative AI right now, its product looks like it makes sense and sometimes it does, but there is always the risk of total nonsense hidden somewhere in the middle. So you still need someone capable to check and correct for most professional work, and sometimes that is harder or more time consuming than making the product itself.
The same sort of problem with self driving cars, they are often correct but not often enough, and staying alert to correct the AI is worse than driving yourself which is more work, paradoxically enough.
AI might manage to push through these barriers, but I remain skeptical with the technology in the current state: statistical machines that are good in the common cases but sketchy at the edges.
Which is similar to correcting subpar human results:
Rather than meticulously correcting the works of a subpar programmer it is much more efficient to let a proficient developer produce the code. Or even better, engage a 10x developer.
If an inefficient programmer can solve a real world business problem in a fraction of the time, but it requires slightly more computing resources, I’d still pay for it. Efficiency can be measured in a lot of dimensions and often times spending time and money writing optimized code is inefficient as well.
> there is always the risk of total nonsense hidden somewhere in the middle
This is true of human-generated code as well. Trust me: Reviewing other people's code is my day job.
It's exceptionally rare that a malicious actor is trying to sneak something into the code. The common scenario is the developer who's new to the project not fully understanding how everything works so they copy & paste something they think is necessary but ultimately isn't and could in fact be very wrong. Just like how ChatGPT works.
I've been asking friends in non-programming engineering fields how ChatGPT does in their area of expertise, and I believe programming is the area that ChatGPT is the most accurate. Finding solution to general engineering problems seems blatantly wrong in almost all cases, whereas in programming, it seems to be able to generate mostly correct code for simple, boiler-plate like tasks.
yes, but why? Why is GPT so much better at programming than other tasks?
can it be that programming itself can be so easily predicted in a generative way, while others require more ingenuity and real world model to be solved?
In this case I would totally offload programming to a GPT /LLM AI, while my job is simply to specify largely the business case.
I have to imagine its because so much of its training data is readily available programming docs, tutorials, and general Q&A that there is an amazing abundance of online. How many times have you just pasted an error into google and hoped someone else has asked the exact same question on stack overflow?
As Chomsky points out, the AIs understand syntax not semantics. Code is all (or almost all) syntax whereas human language has both syntax and meaning. Meaning is not something ChatGPT understands.
Is it because programming is a more limited and specific language than the ones people speak? There's less room for double-meanings, slang, meaning, or even sentence structure.
It seems the improvement in the models is the size of the training set. Hence the "Large" in large language models. I am assuming that it is better at programming because there a is a ton of training data available on Github, etc. There isn't a similar set of data on solving a variety of physics problems.
Yup. What seems to be largely missed is that these models have zero understanding, and are actually destroyers of information, not creators. In classic Information Theory, information is basically surprise value — how much unexpected info is in the message? — yet these "AI" systems put out the most expected subset in each instance. This highly averaged output is very recognizable and so very striking, but it is not actually very informative (perhaps except in cases where it is specifically used as a verbose search engine, where the query takes advantage of the breadth of the AI's training).
> In classic Information Theory, information is basically surprise value — how much unexpected info is in the message? — yet these "AI" systems put out the most expected subset in each instance.
Forgive me, but isn't this kind of moving-the-goalposts? Information is the surprise value from the recipient's point of view, which meas the recipient's bayesian prior probability is "expected". Saying "these "AI" systems put out the most expected subset in each instance" assumes that the recipient's priors exactly equal those of the model which would only be the case when the model is talking to itself. (or I suppose to an even more complex model with perfect knowledge of ChatGPT's weights)
The fact that no information is transferred when the model talks to itself should not be surprising and would apply to any AI. (even including a superhuman post-singularity god-like AI)
Yes, the AI's output could be surprising to the point of view of many recipients.
This does not mean anything more than that the AI has a greater breadth of training background, which is likely.
We get the output most likely expected from any of (or the average of) the humans whose writing/drawing/whatever was included in the input set.
What we will not be getting from the AIs is any creative output based on unique understanding, as we would from an intelligent, creative human. Many of hte humans in the input set would see the same prompt and produce an actual novel and meaningful output, not simply a cut-and-paste from prior works. (& yes, seme novel output may come from some randomizing algo, but if it is correct, it is no more correct than the broken clock that is correct twice every day.)
Or, another example, I was involved in a legal deposition where an "AI" transcription system was used instead of a skilled court reporter. The output LOOKED fantastic, until I actually read it, and it was absolute garbage. The standard errata sheet has room for the deponent to put in about a dozen corrections, and most are less than a handful. My errata list was multiple pages. These errors often reversed the meaning of sentences, substitutin "I have ..." for "You have...", dropping or adding "not", or substituting in common names for unusual names (e.g., "Jack Kennedy" for "John Kemeny". note human transcribers always ask for correct spellings of names in the next break, this crap just inserted it like it had a clue).
So, even though the total "experience" or training set of the may go beyond the experience of the reader, so some of the output is surprising, this is no more so than a search engine produces surprise. In fact, I think this is the best use of the AIs, to have them trained on an enormous data set, and provide possibly better results, defined as more on-point, but likely less thorough.
>This does not mean anything more than that the AI has a greater breadth of training background, which is likely.
This has nothing to do with Shannon's definition of information. The light from a star going supernova or a tsunami hitting shore convey information without any kind of agency nor intent being involved.
>What we will not be getting from the AIs is any creative output based on unique understanding, as we would from an intelligent, creative human.
This is a really big assumption here that I don't think is justified. Are you assuming that humans have some non-physical soul which makes us somehow different from any other deterministic information-processing system?
>>The light from a star going supernova or a tsunami hitting shore convey information without any kind of agency nor intent being involved.
Right, and I'm obviously speaking loosely, but the amount of information or information value is the amount of non-redundant 'surprise'.
Receiving the supernova's signal for the first time contains valuable information, but replaying it does not.
>>Are you assuming that humans have some non-physical soul which makes us somehow different from any other deterministic information-processing system?
No, I'm speaking about the current and near-intermediate-term state of AI. This current state is basically a mashup of everything it's 'seen' in training, but without a shred of understanding. It only spits back what is statistically the most likely string of words to occur in any situation.
Ordinary humans do far better because they have understanding. A simple example shows this: Asking ChatGPT this question: "Mikes mum had 4 kids; 3 of them are named Luis, Drake, and Matilda. What is the name of the 4th kid?" [0] ChatGPT utterly fails, even when told the answer is in the question.
As in my example with the deposition, it looks great but it is just a literal mashup of training data, and worse yet, is only outputting the most average data. This produces surprisingly emotionally satisfying results, which may be useful in some contexts (marketing copy?) but isn't close to the reasoning of an ordinary human.
I'm making no claim about whether AI can get there. I suspect it ultimately will. BUT, this layer is only a small component of what must be built to get to actually abstract concepts and wield them like any child. My minor in college was neuroscience and I still follow it a bit; we're a LOOONG way from understanding what it is to get and weild concepts, nevermind how to model or reproduce that in a computer.
That's all I'm saying - the current crop is impressive, maybe even useful in some applications, but nowhere near the kind of "intelligence" that is being touted.
The same sort of problem with self driving cars, they are often correct but not often enough, and staying alert to correct the AI is worse than driving yourself which is more work, paradoxically enough.
AI might manage to push through these barriers, but I remain skeptical with the technology in the current state: statistical machines that are good in the common cases but sketchy at the edges.