They made a computer program that can follow instructions in plain english. That's been a goal of computer science since the 60s, and I wasn't sure I was going to see it in my lifetime. This is more incredible than people give credit.

I think GenAI is incredible in a lot of ways. I think for me it's frustrating that GenAI is currently being used in a way that largely regurgitates and hallucinates information, rather than aids a computer to follow instructions in plain English. For example, at least earlier versions of GenAI, when you asked it a question, it would make up an answer to that question using the knowledge in the training data. That's why we saw limitations on what events the AI knew about, for example. It would be great to see the AI understand the question, query a knowledge base for the answer and then rephrase that answer in English instead. I believe, though I've mostly kept out of the state of the art here, that's the direction it's going in, and I think if that's the case, the confidence in the AI will go up. If we can in parallel, decrease the horrendous amounts of energy being used to run the models, then maybe some of the detractors will start to turn?

OPINION:

why is it incredible? we followed a bunch of lists of laws until we got ELIZA on virtual steroids ... after we cracked some milestones in hardware development.

people work towards that. money is put into that. we'd have cancer solved if the same amount of people and money flowed into the entangled fields.

next step is digital steroids (signal processing) and the physical level once again.

software-wise, the current and next generations of coders won't do anything "incredible" anymore. too few approaches on the playing field, no indie thinkers, and the capital is literally dumb.

(enforced) conformity has too many downsides that can only be balanced via variety in the wide wild; but that means cumulative competition, which was handled via sabotage ( a good analogy is doping in sports, US influence and the actually incredible cognitive & logical weakness as well as the cowardice of almost the entire rest of the world )

so I cannot agree, people give too much credit indeed.

Responses like this tickle me because they make clear there’s a batch of human beings who have just a profoundly different experience with LLMs than I do.

I see these things and think, this is incredible, machines seem to be approximating or emulating conscious thought. There must be so much we can learn about ourselves, and so much they can do.

You see the same thing and say meh, useless pattern matching, what’s the point, spend the money elsewhere.

I wonder why we have this different perspective? I’ve seen these two reactions again and again—I suspect they evince two different worldviews, but I don’t know what the correlates are. I don’t think it’s techno optimism/pessimism, because I’m profoundly worried about what happens to us. It’s not purely an age thing—I’m not young. I see it on here all the time so I don’t think it’s field of work. So what is it, I wonder?

My hypothesis; it's people who believe in a just world. That think value and wisdom need to come from effort and/or pain.

Potential litmus test; They like videoclips from Coldplay where every frame is drawn with real crayons.

Then it's hard to value a machine that can not feel pain or effort and just generates. It's not fair it didn't have to suffer, and then flip the arrow to say, therefore it's not valuable.

Happy to becorrected. I'm also very curious what mental models are behind such big differences in perspectives.

This is interesting! So to some, perhaps, the creator, intent and process matter as much as—if not more than—the output. The “becoming” over the “being.”

Nope. There’s a lot— a lot— written on artistic process and intention. What art is and what it means is a deep philosophical topic that has been pulled apart for millennia. The sheer arrogance of nearly the entire tech world reducing the value of the artistic process to a few simple logical business friendly ideas using a few a priori thought experiments while disregarding the existing philosophy on the matter is the sort of thing making the tech industry so deeply unpopular right now.

But in all honesty, if LLMs were really that miraculous, we wouldn't have such detractors. Like, I don't remember any widespread anti-search-engine sentiment when Google came out. I can name many technologies that don't get such a visceral reaction like LLMs do.

At the end of the day, we build (or should build) technology for the betterment of society. If at the end of the day, large swathes of the population are unhappy or upset at a given technology, that in and of itself should be a sign that something is wrong.

Obviously LLMs aren't going away so I think we should listen to the detractors and try to better understand how we can steer this technology for the better.

The LLMs themselves aren't the fundamental concern; rather, they represent an unprecedented amplification of existing socioeconomic stratification mechanisms. They will likely accelerate the already problematic concentration of capital and opportunity among those positioned to leverage them - and we will see, inevitably, a continuation and intensification of patterns we've observed since the dawn of industrialization.

> if LLMs were really that miraculous, we wouldn't have such detractors

I don't think you thought this through.

Good point, often times it is the miraculous that attracts the most detractors and a benign or useless technology would be ignored. I guess I meant something more like a widespread negative reaction means something's probably not right.

I also agree completely with your point on the socioeconomic context of LLMs causing the most contention and that a deployment giving more autonomy and control to its users would have a much more toned down reaction. Like, I don't really notice a strong reaction against locally deployed LLMs.

> don't remember any widespread anti-search-engine sentiment when Google came out.

Before Google switched to Gemini AI it didn't offer making mustard gas recipe when prompted for cleaning advice.

LLM are tools, and highly unreliable one at that. I don't fear it. I fear the idiots that think it is a viable replacement.

> If at the end of the day, large swathes of the population are unhappy or upset at a given technology, that in and of itself should be a sign that something is wrong.

Productivity and efficiency gains at the short term expense of inefficient jobs have always made people upset, so taking only that into consideration is wrong.

> The Swing Riots were a widespread uprising in 1830 by agricultural workers in England against the mechanization of agriculture. The riots were a protest against the introduction of threshing machines, which had been increasing since the end of the 18th century. The threshing process was labor-intensive, and before the machines were introduced, it employed about 25% of all agricultural workers. The introduction of the machines gradually unemployed a large number of agricultural workers

> Jethro Tull invented the seed drill in 1701, but it was not immediately popular in England. Tull's ideas were controversial, and his theories fell into disrepute. Tull's servants resisted his new methods because they threatened their position as laborers and their skill with the plow

I keep saying this but nobody seems to really get it or care: LLMs are magic. This isn't at all like conventional software where someone plans for every possible path, or even software where paths are generated procedurally. These models show truly emergent behaviors that go far beyond their training. And yet people dismiss them simply because they're probabilistic (more likely because they've heard some version of this, actually), as if that somehow invalidates the fact that they're doing things they were never programmed to do.

Our understanding of consciousness is about to undergo a revolutionary transformation, and a lot of people ... aren't going to like that, either.

>simply because they're probabilistic

You can make them deterministic by fixing the temperature. But even still, there is some intrinsic error rate: it's like an oracle machine where some subset of queries will simply give you the wrong answer... which is just like humans (how many times have answers on StackOverflow been wrong), but it's definitely not like "conventional" programs.

I guess just as people had to get used to "trusting" computers, I think there's probably going to be another adjustment period of "untrusting" them and finding the right balance of "trust but verify." I think this is also mirrored in adoption patterns: mathematicians seem to be more open to using LLMs because they're used to not blindly trusting supposed proofs. On the other hand, it's well known that writing code is easier than reading it, so developers are a bit more wary.

> Just a few years ago, passing the Turing test seemed an unattainable goal

Who geniunely thought that? Eliza could almost fool a human and that was in the 1960...

LLMs are a neat thing for doing some tedious task but I saw them making plain stupid mistakes that makes my skin crawl.

Eliza did fool some humans. Which is why the Turing test was always a flawed notion, one that says more about the human than the computer.

That aside, nothing you said about LLMs doesn't apply to humans as well.

> That aside, nothing you said about LLMs doesn't apply to humans as well.

Humans make mistakes in a diverse way, that makes groups of them much better than an LLM that will amplify its own mistakes.

Computer programs has to be extremely better than a single human since they lack the group dynamics of humans, so as long as you compare single humans to AI people will rightfully demand more from the AI.

If all humans ran the same program and gave the same answers then humans would be way less useful, that is where LLM are today.

(Shrug) The only important thing is the time derivative. AI models are getting better over time, thanks to things like chain-of-thought reasoning, while humans are not.

IMHO, LLMs will grow immensely in power once someone comes up with the right way to wrap them in a negative feedback loop to keep the hallucination "gain" under control. The result may no longer be recognizable as an LLM, but I'm pretty sure that the GPT concept will be recognized historically as the foundation for what follows.

Eliza did not fool anybody for more than 5 seconds lol.

It is, like you said, good enough. A (sometimes disappointing) miracle.

So we can stop making it bigger/smarter and instead focus on making it more acessible.

> I use LLMs everyday. It has not only boosted productivity, but most of all, it made workmore fun.

That's a matter of character. Other people team up and up your LLM-augmented productivity tenfold without LLMs.

I'm mostly solo myself and did get a bunch of good shit out of LLMs but that has nothing to do with them being good. Search Engines got me the same results after only little more time.

The one "good" thing about LLMs is that you can get algorithmic, logically straight, safe & sound FACTS about topics like how the insurance mafia monetizes on & cooperates with various other entities to create crisis, or how the smuggling of weapons is done and used to increase premiums on & thus prices in trade via tankers & shipping . Or why labile prices on gas are complete fucking bullshit.

And the list goes on and on and on. But who in the respective, theoretically respectable positions cares?

Hallucinations have been build into LLMs on purpose. Above is just one of them. And it's just cooperation of magic money & enforced conformity of engineers.

I missed the part where GP said they worked solo.

Can you share some of your prompts for the topics you listed? Asking about it straightforwardly doesn't get very far.

no. but I do recommend running stuff like that in a DMZ. if you can get access to some of those farms in the Hinterlands, you'll get some proper compute, too but doin it via companies like open-I-AH or leClaude isn't a good idea because these fuckers are brain-damaged and can't be trusted.

The Turing Test is in the same vein as the Drake Equation: a means to focus thought on how to approach finding a solution, not to find a definitive answer (and who's existence in pop culture makes it a bigger thing than it is). The original Eliza should have be enough to render the Turing Test a moot point I feel.

Personally (as someone who's worked in a ML-adjacent field for a number of years, and so seen the various ML waves), LLMs have so far been a resounding disappointment in comparison to other ML tools (the only cool thing I've seen come out of it was Google using LLMs to generate code to try to futz code even better), and have shown that while chasing the last 1% accuracy is interesting from both an engineering and mathematical perspective (which are entirely legitimate things to be interested in, though sadly they seem to have been lost to hype), simple-to-explain-and-understand less accurate methods (like a decision tree) are both cheaper to run and more predictable (which is what you usually want).

I am not sure the Turing test has been passed or understood. Cromulent answers aren’t human ones.

LLMs enable me to write more code than I ever have and in languages I don’t know. It’s a wonder.

That said I grew up with an 8-bit machine buying games on cassette. There were never updates and rarely ever a serious bug.

Yes programs were tiny, but they were also handwritten in assembler. Quality has gone down in software with the advent of modern tools as surely as it has in furniture with the introduction of the table saw.