201–210 of 278 posts
Lena aka MMAcevedo seems very relevant: https://qntm.org/mmacevedo
(its a short story written in the style of a wikipedia article from the future about the standard model test brain uploaded from a living scientist).
We have already experimented with letting large neural networks develop software that seems to be correct based on a prompt. They are called developers. This is going to have all the same problems as letting a bunch of green developers go to town on implementation without a design phase. The point of designing systems is so that the complexity of the system is low enough that we can predict all of the behaviors, incl…
Except without all the downsides, because GPT can rewrite the whole program nearly instantly. Do you see why our intuitions around maintenance, "good architecture/design" and good processes may now be meaningless?
It seems a bit premature to say we don't have anything close when we can get working programs nearly instantly out of GPT right now, and that seemed like a laughable fantasy only two years ago.
Of course this will only work if your user's state can be captured within the 4096 tokens limit or whatever limit your llm imposes. More if you can accept forgetting least recent data. Might actually be OK for quite a few apps.
I tried getting it to generate a Red-Black tree in Java but it cuts off half way through. I suppose you could divide and conquer with smaller parts of the algorithm, but then we'd need a "meta AI" that can keep track of all those parts and integrate them into a whole. I'm sure it's possible, don't know if it's available as a solution yet.
I tried similar prompts on various data structures. If you reissue the request sometimes that completes.
Listen, you will lose your jobs to gpt-backend eventually, but not today. This is just a fun project today
Earlier quoted context omitted.
If somebody putting a few millions into making this widespread were enough to make it a problem, then software development would already be doomed and we would better start learning woodwork right now.
The argument is stochastic. Maybe this joke will get ignored, but then we could've had the same conversation few years ago about "prompt engineering" becoming a job, and here we are. Or about launching a Docker container implementing a single, short-lived CLI command. Or about all the other countless examples of ridiculously complicated and/or wasteful solutions to simple problems that become industry standards simpl…
The joke is on all of us if we only treat this as a joke. Rails pioneered simple command line templates and convention over configuration, and it took over the world for awhile.
An AI as backend is the logical conclusion of that same trend.
Earlier quoted context omitted.
Sure, the human element will still be there. But note that the detail of your spec will converge to the natural "resolution" of your problem as the power of your AI increases.
But at what scope? The how will always matter. Your AI could design a system that bankrupts you on the first day. To prevent that you need to specify constraints and it's turtles all the way down. It would free you from spending time on areas you don't care about, but that's already true with SaaS.
Just the step from "program" => "spec" is already a big one. So big, that it is rarely done today. Test-driven development is an attempt at this, but the problem is that tests cannot truly verify a spec. Proofs can. Of course, you can combine tests and proofs, for example proofs for correctness, tests to make sure other measures like speed and cost are sane. But if you want to be absolutely sure, you will need to replace all tests by proofs.
Earlier quoted context omitted.
I am sorry to tell you, but AI is exceptional for engineering. Just make the AI also generate a proof that its code meets the spec. That's what human engineers should already do, but it was costly, because the tools were not good enough and the engineers not educated enough. AI is going to cut right through that Gordian knot. This should not be surprising: There is a large intersection between engineering and mathema…
> Just make the AI also generate a proof that its code meets the spec. How would one tell if the AI-created "proof" is both accurate and adequate?
Just yesterday I was playing with chatgpt and found an error between the code it generated and the explanation of the code. It contradicted itself.
However when I caught the error I asked it to further explain since it appears to contradict the code it generated. It then came back with an apology and it did state it made a mistake and was able to understand the error and fix it. Although I was specific about the mistake. I might try again later today to do the same test and see if it learned or generates the same error again .If it does I will ask it to confirm that its explanation and code match versus pointing out the error.
Earlier quoted context omitted.
> This is an incredibly bold prediction that isn't supported by the opinions of the majority of people in the field Well, DUH! "It is difficult to get a man to understand something when his salary depends upon his not understanding it." - Upton Sinclair. The people in the field who are making these promises may even believe it themselves, because their bread and butter comes from it.
You have an incredibly dim and pessimistic view of researchers and scientists. They could all easily double or triple their salaries by moving to standard industry but decide to work in Academia or Research Labs. These people on average predict AGI within ~30 years with more than 50% probability. Not sure how that prediction benefits their salary in any meaningful way. If Astronomers were predicting a mass-extinction…
I haven't seen that prediction. What I have seen is "AGI is 2 years out", and I have been seeing that for 4 years.
Much like the self-driving cars that were (according to the experts in the industry) 5 years out since 2012, and still not here in 2023.
Maybe if the experts in the industry were more vocal about how far off they are, you wouldn't be reading comments like mine.
> If Astronomers were predicting a mass-extinction level asteroid impact for the year 2050 with 50% probability I doubt you would be so cavalier.
if they had been saying, since 2012, that it's five years away, I won't be the only one laughing at them.
When it comes to AI, though, the world is a lot more forgiving, and collectively more forgetful of the predictions.
Earlier quoted context omitted.
You are factually wrong. The general case is being solved. Self-driving systems are objectively better every year, and will eventually reach human level safety. You can literally watch cars self-driving in all kinds of places and conditions. Yes, they make mistakes. So do humans.
So when will these self-driving cars be ready for mass consumption? Because they're starting to sound like flying cars at this point. Yes, technically doable and there are always some working prototypes, but no real market, and not seen as practical.
I might feel like intervening, 1 out of 10 times at this point. I might not be the typical driver, but I definitely feel like its ready for early adopters now.
However, even though I'm a big fan, I don't see how these can easily transition to "mass consumption", because as we get into the uncanny valley where the auto drive is good enough to take over, the masses are going to completely check out of their responsibility to be a good backup driver.
So I feel like we are going to be stuck in the current space for a long time, maybe 10 years. Until you Auto Drive is so good, you can ride one without getting a drivers license.
And you will almost immediately run into the fundamental problem with current iterations of GPT - You can not trust it to be correct or actually do the thing you want, only something that resembles the thing you want. The description in this link puts some really high hopes on the ability of AI to simply "figure out" what you want with little input. In reality, it will give you something that sorta kinda looks like w…
Put differently - every website needs a back-end. 95%+ of websites don't differentiate on their back-end, but they still need to build from scratch since there's no incentive for businesses to share knowledge with unaffiliated businesses.
One way this problem is solved is neutral platforms like AWS that sell the 'good enough' turn-key solution (keep in mind, at one point, the cloud had nearly as much hype as AI does now).
Another way to solve the problem is an AI that 'makes' the back-end code 'from scratch,' but is really just returning the code (cribbed from its training dataset) that probabilistically answers your question in the best way possible, based on the results of its training.
The AI option seems really impressive to us right now, because we haven't seen it before (much like photoshop in the 90's), but eventually we get used to it. Once we get to that phase, we will either regulate AI until it looks like a marketplace business (the creators of the training dataset maybe should be compensated) or we will just see 'generating code from a training dataset' as so basic that we move on to other, harder problems that have no training dataset yet (in the same way Quickbooks has largely replaced book-keepers, but digital advertisers for small business are increasingly relevant).