AI and the End of Programming
bit-player.org
bit-player.org
I draw parallel to watching a move of adaptation of book to reading the book. Sure you can watch good adaptation of it, but the intricacies and details are lost, which you could get by reading a book. Same goes with programming with/by AI vs non-AI.
I also think we need check our biases. I found my own tweets from 2017 when Tesla unveiled Tesla Semi - I thought truckers will be replaced - given the Tesla Semi and FSD. Almost 7 years later, not only did it not replace them, we're nowhere close to those jobs being replaced. Same goes for radiologists.
It’s going to be a wild ride. Also consider that LLMs don’t have the “last 5%” problem that self-driving software has, where widespread use is currently untenable because humans still need to intervene and that poses safety problems. LLMs are very complimentary to a large breadth of human work, until they stop being complimentary and instead become a substitute.
Another thing I’ll say is, the investment in the already mind-blowing LLMs like GPT-4 is relatively puny, a fraction of the cost of building 1 skyscraper. What’s happening is that the market has been validated and much larger sums of $$$ are going to flood in, including into the research pipeline which can solve problems such as data “scarcity” for future AI.
Imagine a much smarter version of GPT-4, except constantly thinking like a human does (as opposed to prompt-response-wait-prompt-response…), constantly working towards some specified goal, and…placed in a robot body like those made by Boston Dynamics. That’s not even in the far future. It’s technically possible in 6-12 months.
Finally, looping back to self-driving - I think you’ve reached your conclusion too early there. Once the last 5% problem is solved for self-driving, it will be solved for everyone and forever and getting from point A to B on the road will forever be available to you by a robot chauffeur 24/7, forever deliver things to you autonomously, and improve supply chain/logistics by a lot, with downstream reduction in the price of delivered goods.
The big problem is going to be some "not so intelligent" Human making decisions based on AI/ML output without verification or even worse having AI itself make the decision.
Here is a Munk debate "AI research and development poses an existential threat" - Yoshua Bengio & Max Tegmark vs. Melanie Mitchell & Yann Lecun - https://www.youtube.com/watch?v=144uOfr4SYA
Where would basic planning fit here? Is it a “faux demonstration” (??) for engineers to talk through their ideas for approaching a problem before attempting it?
An LLM is like asking Google to make a search results collage. It’s an entirely derivative essentializer
I fed ChatGPT a bunch of dummy savings and loans transactions and then asked it to compile loan tables for named borrowers.
It worked and I was impressed. For me, this felt like a move from programming to curation. I give the AI the data, and it organizes it for me and outputs it according to what I ask for.
I appreciate there is higher order programming, but for simple needs like the one I described, maybe this is not the beginning of the end for programming, but at least the end of the beginning.
LLMs are concerned with mimicking language and appearance as best they can. You don't know if it's just packaging up a bunch of numbers and a table that looks pretty plausible or if it's actually using some sort of formatting scheme and applying it while strictly preserving the numbers.
And that's part of the inherent problems of AI. It's a black box. You don't know if it's calculating some final report that thinks you want with the numbers it thinks you want or if it's actually applying a constant process to all the numbers, cells and columns of the of the report.
I suppose perhaps if the LLM could provide some sort of processing chain information to you, maybe that would work to alleviate any concerns
We've been trying to solve this problem at Leaptable (https://leaptable.co/). The crux is that while LLMs are still a black box, transparency in the way AI Agents interact with LLMs is key. For instance, seeing the outputs of each step in a chain-of-thought sequence helps debug common fallacies in the way the LLMs reason and built trust.
In other words, the deeper you need to think to solve problems, the safer you are.
Until I can see an "AI" go through a design process without a human guiding it, iterating until it decides to stop, I'll have this confidence.
Right now, LLM's are Clever Hans, stopping their process when a human says. So not only are they borrowing the intelligence of writers the world over, they're actually borrowing the intelligence of the person at the prompt.
Take that prompt away, and they'll fall flat.
For example, can they even think of a problem to solve on their own? No, they need a human to ask them to find a problem and solve it. Otherwise, they sit. Dumb, unmoving, devoid of agency, and incapable of even the smallest task without input.
What use of your change of opinion then will be? The whole attempt here is to predict if something is possible - just saying "no" and waiting for the actual disproval to change your opinion worth seemingly not much.
Pretty sure the actual idea is to drown oneself in "copium" until the trend is too obvious to ignore.
Humans idle at ~100W, with (iirc) something like 20..25% going to the brain. We are extremely versatile in the types of activity we can do, and most of us can put in a # of hours concentrated work each day. And are self-managing for the most part.
Contrast with LLMs: the most capable ones run on huge clusters of specialized hardware. Which was manufactured at great expense. And no doubt take many kW if not MWs to do things a (one) smart human, or a few humans can do. With the LLM stuck in its datacenter racks, while humans happily walk around.
Yeah there's smaller LLMs you can run on a laptop. Can those do the same as say, GPT4? Training on said laptop too?
Sure, human meatbags are messy, calories going in are expensive ones, and you can't pull humans from storage & switch on/off as needed. And current AI is no doubt at the beginning of a lot of optimization. Both in architecture & implementation.
But.. all that functional improvement & optimization has yet to be done.
On average, humans still give a lot of bang per calorie. Until LLMs or other AI reverses that, we're a long way from being outcompeted on the more difficult / interesting jobs.
As for simple / boring jobs: I welcome our AI overlords. More time for us to do fun stuff.
I think here is the answer to the question of do LLMs ‘reason’ at all. Since we know the LLM is a sequence completion mechanism, that initial ‘certainly’ (which was a wrong assertion to boot) can not possibly be the result of any sort of reasoning.
I feel like this is pretty hard to believe, to be honest. How on Earth, given all that's happened last year, could we possibly not at least be "heading in the right direction" the way things are?
> My personal view is this: If you view the current generation of LLMs as pioneering efforts in a long-term research program, then the results are encouraging. But if you consider this technology as an immediate replacement for hand-coded software, it’s quite hopeless. We’re nowhere near the necessary level of reliability.
I think especially the overview of the studies provided in the article are valuable. You can draw your own conclusions from the results.
Generative transformers only address part of the problem space, and perhaps I should have written “appear to address”. It is unlikely that they are much of a step along the path — perhaps, to use a car analogy, they are a crankshaft.
Perhaps a better analogy is the analogue computers of the 40s-60s and the later digital/analogue hybrids. Computationally interesting for the problems of the age but ultimately a dead end.
The LLM models that are all the hype are superb at things like copy writing and the more traditional humanities but struggle at logic. With a year of reasearch we can reach human parity in things like psychotherapy but for something that requires logical reasoning rather then repeating bits of text it's not that clear.