I would like the focus to be on harmful content creators, not so much on the platforms. Platforms have incentives to bubble this content to the top because it's desirable but there is content that is simply illegal and it's being uploaded by same creators for years without any repercussions.
The problem partially is that AI can also fix AI slop. At this point I am in doubt whether code quality matters anymore in most non-critical software. You can ask an LLM if the code has quality issues and refactor to a _better_ version. It will reason through, prepare a plan and refactor. So now with this "better" code you can expect that your LLM will be able to deliver higher quality results but that's all the quality that is needed.
Actually, at this point I feel that the value in software engineering is moving from coding to testing and quality assurance.
All the frontier models tell me when there are no issues. After implementing a feature I will ask it to identify issues in my implementation, list them, and support each item they identified with technical argumentation and reasoning as to why it's an issue.
If it doesn't find anything it says I didn't find anything.
It won't be the case that a global scale ocean current collapses and its impact is local. It's like a butterfly effect where the butterfly is the size of an ocean - its wing flap will resonate throughout entire world with unpredictable natural and social consequences. There will be no winners, only losers.
I think AI movies (or shorts for that matter, since I am not aware of any feature length movie) currently are not bland, they are simply of very low quality because they are pushing the limits of the current technology. However by the time the technology catches up (might not be as soon as many expect), then nobody will care about them because it will not have personality.
So the end game for the current generation of AI companies won't be productivity improvements but gambling, just like everything else nowadays. That's why they want to get us all into these massive casinos they call data centers and don't want us to own the slot machines.
So what that you have ideas - other people have them too. It's not ideas that build businesses but knowing right people or ability to sell products.
The gambling trope is so tired. AI development doesn't involve luck to any appreciable degree, certainly not more than hiring people to do a job can be considered "gambling" (you never know what you're going to get!).
It's just paying to get stuff done, which is how it's always been, since the dawn of man.
>AI development doesn't involve luck to any appreciable degree
Reading this while I'm prompting for the third time to fix a 100+ line function is amusing, to say the least. I don't care about the definition of "appreciable", but I definitely have to repeat myself to get stuff done, sometimes even to undo things I never told it to touch.
"Insanity is doing the same thing over and over again and expecting different results."
There is certainly randomness in model output that the user has to work around, but sending the same prompt with the same context (or even worse- with added entropy leaving the previous failed prompt in the context) over and over again akin to pulling a slot machine lever is certainly user error and not the way to "hold it".
Exact same kinds of projects with the exact same development environment, models, etc. Either he's never worked with a development team or he doesn't consider things outside his own perspective. shrug
Then you miss the point - AI use is being compared to gambling because it is addictive, partly due to same mechanism - the results (and rewards) are somewhat random, but it makes you feel as if you're completely in control of the outcome.
It's to the point that I just push the output of that to production and know it'll be OK, except for very large changes where I'm unlikely to have specified everything at the required level of detail. Even then, things won't so much be wrong, as they'll just not be how I want them.
For most people who are not doing their day to day jobs it's just a prompt of their idea roughly sketched out and a miracle happens - LLM fills in the blanks. Every time it's different but it works, sometimes even better than initially expected. That's why the addiction and gambling. Gambling is a lot of things, not only flashing lights or play sounds. Some people claim prediction markets isn't gambling either, though that doesn't change the fact.
How is this different from hiring a designer, telling them "make me a website" and then waiting to see if they resolve the uncertainty into something you like or not?
I tell LLMs what to do in pretty high detail, and they do it. With LLMs I have much less variance than with coworkers.
It is different because for humans, it takes time to produce some result, while AI does it instantly. So if you tell a programmer to do X, you have a week for your adrenaline to cool off. If you tell AI, it will do it in minutes.
> If you're making the argument that LLMs are gambling simply because they're faster than humans
No I am not. It's more addictive because of the timescale. The comparison of AIs to gambling is through addiction mechanism, as I explain elsewhere.
My aunt used to put in (the same) lottery numbers every week. It was gambling, but probably not an addiction in the clinical sense. If she had played slot machines, god forbid, it could have been more problematic. AI is a slot machine, a hire is a lottery ticket.
I don’t like the gambling comparison either. It’s more like smoking or drinking. It’s an addiction you lean on to help you do something- even if that something is just getting through the day.
Yeah but those are classified as addictions because they have a harm component (lung cancer, liver disease, societal impact). LLMs aren't going to kill you. If anything, it might be like gaming addiction.
If you've gotten to the point where you'd rather talk to an LLM than socialise, go to work, etc, then yes, you definitely have a problem, same as with a gaming addiction.
Saying "LLMs are slot machines" is like saying "video games are slot machines", and nobody says that, even though it's more true of video games (some are actual slot machines/gacha) than of LLMs.
> Saying "LLMs are slot machines" is like saying "video games are slot machines", and nobody says that, even though it's more true of video games (some are actual slot machines/gacha) than of LLMs.
People absolutely do say that video games are slot machines. [0][1]
I'd observe that there are professional gamblers, and there are amateur gamblers.
If you know what you're doing, know how to spec a problem space, and can manage the tool competently enough to churn out good results, then everything's fine, and you're maybe being productive or increasing your productivity by some degree. (Professional "Gambler")
If you DON'T know what you're doing, and you're just vibe-coding, then I would argue that it is at least a form of gambling (Amateur "Gambler")
Both of these conditions can also be applied to "hiring people to do a job" however there we can also observe things like reputation, credentials and so on.
"It's just paying to get stuff done..." is, with respect, superflous.
I don't know, I can understand "some people might overdo it and get addicted to LLMs". I can't understand "LLMs are slot machines and that's all they're good for" when I use LLMs every day to do tons of actual work.
The gambling part is because of the (hopefully emergent and not purposefully designed) intermittent reinforcement due to the limits. You don't get that with regular hires.
You usually don't get immediate responses from hires which means delayed gratification and avoiding much of the potential dopaminergic effects you get when engaging with LLMs.
You can play overextending the hire analogy all you want but it is simply not the same.
Not in that sense but social media companies already know the value of not giving a user exactly what they want. This keeps them on the platform longer and excited some lizard part of our brain for challenge.
Due to capitalism’s law of all businesses convergening on maximizing profit, it’s just a matter of time until AI companies employ similar techniques with LLMs. We can all imagine how that will look like
I think SaaS pricing model has long been abusive. So, you have a platform, which takes real money to build and maintain. But then you build features on top of that platform, where some features are build-once and require not a lot of maintenance and no additional expenses - the code is just sitting there doing the stuff. Then you request additional money for that feature, which is effectively free for you.
Unlike physical goods where a higher price reflects higher production cost, SaaS companies have to engineer scarcity into a product that is naturally abundant.
In this LinkedIn example, they already collect the profile visitors for everyone. Instead, they spent additional engineering resources building the restriction layer and then charge the users to undo the sabotage.
The price of something is not necessarily the function of the cost of producing it. It is the value to the buyer. Of course, you have edge cases on both sides of the scale, some things are of less value (to some markets) than the cost of producing them, and some things are of high value when producing them costs almost nothing. That's part of why free markets are so great, they are value-oriented and not only cost-based.
I think it’s a mistake to think that we will be blindly going in this direction for many years and then suddenly collectively wake up and realize what have we done. It’s a great filter and a great opportunity.
If LLMs stop improving at the pace of the last few years (I believe they already are slowing down) then they will still manage to crank out billions lines of code which they themselves won’t be able to grep and reason through, leading to drop in quality and lost revenue for the companies that choose to go all-in with LLMs.
But let’s be realistic - modern LLMs are still a great and useful tool when used properly so they will stay. Our goal will be to keep them on track and reduce the negative impact of hallucinations.
As a result software industry will move away from large complex interconnected systems that have millions of features but only a few of them actively used, to small high quality targeted tools. Because their work will be easier to verify and to control the side effects.
Maybe I am unlucky but I had worked with too many developers who couldn't make a good decision if their life depended on it. LLMs at least know how to convince you of their decisions with strong arguments.
> If LLMs stop improving at the pace of the last few years (I believe they already are slowing down)
Depending on how you measure "improvement" they already have or they never will :-/
Measuring capability of the model as a ratio of context length, you reach the limits at around 300k-400k tokens of context; after that you have diminishing returns. We passed this point.
Measuring capability purely by output, smarter harnesses in the future may unlock even more improvements in outputs; basically a twist on the "Sufficiently Smart Compiler" (https://wiki.c2.com/?SufficientlySmartCompiler=)
That's the two extremes but there's more on the spectrum in between.
300k-400k isn’t the current limit if you create modules and/or organize the code reasonably.. for the same reason we do this for humans: it allows us to interact with a component without loading the internals into out context.
you can also execute larger tasks than this using subagents to divide the work so each segment doesn’t exceed the usable context window. i regular execute tasks that require hundreds of subagents, for example.
in practice the context window is effectively unlimited or at least exceptionally high — 100m+ tokens. it just requires you to structure the work so it can be done effectively — not so dissimilar to what you would do for a person
I keep getting surprised that people who are all-in on this (" i regular execute tasks that require hundreds of subagents ") don't have any idea of what is happening even a single layer below their interface to the LLM ("in practice the context window is effectively unlimited or at least exceptionally high — 100m+ tokens.")
I looked at that response by GP (rgbrenner) and refrained from replying because if someone is both running hundreds of agents at a time AND oblivious to what "context window" means, there is no possible sane discourse that would result from any engagement.
ok "series of context windows spread across many agents".. sure much clearer.
Doesn't change my point: the amount of code the agent can operate on is very large, if not unlimited, as long as you put even a little bit of thought into structuring things so it can be divided along a boundary.
If you let the codebase degrade into spaghetti, then the LLM is going to have the same problem any engineer would have with that. The rules for good code didn't disappear.
Context windows don't necessarily cleanly divide. Getting each agent to be able to task within a context window is a hard problem.
It's like like if your context window with one agent is n, your context window with 10 agents is n/10. It is some skill, but that is also where a lot of the advances are coming in.
300k tokens--the useable context window of a single agent--is about 40k lines of code and you can't figure out a natural breakpoint within that code to divide up the task?
It's the reddest of the red flags to have such CTO. Modern incompetent CTOs are measuring performance by token spend. They are in the same bucket though - coming up with some dogmatic ways with extremely thin foundation.
reply