I think it’s underhyped and overhyped at the same time. It is not a true intellect, more like Artificial Idiot type of thing. But it is powerful and scary. Writing entire news columns with fake reports and photos? Is it not enough for you? Destorying jobs and careers? Stifling human-to-human interactions? (Man, do I hate those ‘online assistants’ type of staff. Useless, annoying, clunky.)
I think we’re seeing a combination of two issues: first, GenAI hasn’t had the impact many have expected, and second, the inherent problems of the transformer architecture make their applicability in many sectors more dubious than initially hoped.
On the business side:
- There has been no sharp, widespread increase of productivity so far:
Nearly half (47%) of employees using AI say they have no idea how to achieve the productivity gains their employers expect, and 77% say these tools have actually decreased their productivity and added to their workload. [1]
The latest estimates, using official figures, suggest that real output per employee in the median rich country is not growing at all. In America, the global centre of AI, output per hour remains below its pre-2020 trend. [2]
- Few, if any, have seen real economic benefits due to the use of GenAI:
Goldman Sachs has constructed a stockmarket index tracking companies that, in the bank’s view, have “the largest estimated potential change to baseline earnings from AI adoption via increased productivity”. […] Since the end of 2022 these companies’ share prices have failed to outperform the broader stockmarket […]. [2]
- Worker replacement by AI has so far not really happened:
[T]here is no sign in the macroeconomic data of a surge in lay-offs. […] Workers are not moving between companies faster than usual, as would probably happen if lots of jobs were disappearing. [2]
- Investments in AI far outpace projected gains:
Investors have added more than $2trn to the market value of the five big tech firms in the past year—in effect projecting an extra $300bn-400bn in annual revenues according to our rough estimates, about the same as another Apple’s worth of sales. For now, though, the tech titans are miles from such results. Even bullish analysts think Microsoft will make only about $10bn from generative-AI-related sales this year. [2]
- Funding for GenAI projects is declining:
For two consecutive quarters, generative AI dealmaking at the earliest stages has declined, dropping 76% from its peak in Q3 2023 as wary investors sit back and reassess following the initial flurry of capital into the space. [3]
But more significantly, it’s increasingly seeming like the current transformer architecture has effectively hit its limits, and is unlikely to feasibly overcome its challenges:
- Hallucinations are inevitable
- Scaling doesn’t help: we’re both running out of training data and need exponentially more data for linearly increasing performance, which probably goes some ways to explaining why we haven’t seen GPT 5 yet
- Emergent capabilities are mostly a mirage
- The ‘outlier problem’ means that situations far outside the training data will remain challenging
So we seem close to a limiting horizon for transformer-style GenAI: the present capabilities have failed to make as much of a business impact as had been predicted, and we’re unlikely to see fundamental increases in these capabilities—plus, the presence of hallucinations and errors in the face of statistical outliers makes these systems too unreliable for many security-critical tasks.
You want to be out of a job?
It is too soon to say we are close to a limiting horizon. It will take years before the initial impact of genAI is fully felt as businesses make more AI-powered tools (AI snitches?) and understand the limitations and capabilities.
I work with Fortune 500 companies adopting GenAI and these things, they take time. Those companies are just now, almost 2 years after chatgpt 3.5, prototyping and testing use cases for RAG-augmented chatbots, summaries, sentiment analysis, etc.
Even if LLMs improve slowly, it would take us over a decade to exhaust clever tricks to get more out of them. We know One-shot answers are inferior to multiple drafts and some scientists have managed to do cool things like make multiple chatGPT3.5 chatbots working together in a virtual office beat ChatGPT4’s one-shot answer. Just as more humans are smarter than one, multiple AIs are smarter than one.
GenAI in general has a lot of headroom for many people to do many different clever things to significantly improve human productivity. My translator friend says they are 3x more performant with AI tools. My SO says her performance increases by 15%. Other jobs will see little to no increase in productivity.
But AGI in 2029? No, I don’t see it. Ray Kurzweil’s big mistake was believing Moore’s law wouldn’t be repealed, but it was. in 1965 it was doubling every year, but Moore revised it in 1975 from doubling every year to every 18 months so you could argue it died then. Nowadays, it takes triple the time it took in 1965. And this is all without taking cost/transistor into account, which has stopped falling years ago.
We couldn’t replace Silicon with anything despite trying for decades. 3D stacking process too much heat. We are out of good options and running out of tricks fast.
Kurzweil also didn’t take into account the Ghz wall and how taking advantage of all transistors today is much, much harder because programmers must use parallelism which complicates everything.
Exponential growth will let you down every time eventually. That’s the lesson.
This thread argues AI is dangerous.
Also, the way he and others assumed that human-level AI could be brute forced with sheer processing power in the first place. (To be fair, not a belief remotely unique to him) Instead, as we’re seeing with LLMs/ChatGPT/etc, attempts at AI keep running into issues where throwing “more power” at problems results in diminishing returns.
Before we build a real AGI, we’re going to have to figure out the right architecture to use and how to program it.
So in other words, for decades, everyone assumed that AI was a hardware problem. A few years back, though, everyone decided that it was actually a software problem, and can thus be solved by writing better software. Now, people are realizing that it is, in fact, a hardware problem.
It’s both, actually. We still don’t know how to program an AI, but with the hardware the problem was the assumption that everything could just be solved with “more”. Heck, look at how in older sci-fi it was commonly assumed that if you stuck enough processing power together AI would just spontaneously arise (like Mike from The Moon is a Harsh Mistress), that reflected a common historical assumption in AI & robotics that it was all processing power and you didn’t need to program for intelligence at all. In some ways we’re just restarting serious research into AI after a decades-long hiatus driven by that dogmatic assumption.
Once you have read all the “training material”, you can’t just increase that by brute force and spending billions. You need to better refine it, prioritize and have the “trained teach the trainees”.
I read somewhere (sorry, no quote) that all those LLMs that hit the market do explicitly exclude recursive self improvement. That’s why some LLMs even have a data cutoff in 2022 or 23. So they digest what they were fed, but do not convert milk into yoghurt, so to speak.
A big reason for the slowing of Moore’s Law is economics, not directly technology. Even when we know how to build fabs for new process nodes, they are so expensive that the decreasing number of semiconductor companies choose to delay the next step to get enough money out of the previous fab. I’ve been to meetings where this was explicitly discussed.
Do you have a cite for your cost/transistor claim? I’m not sure how that is measured. Cost per transistor varies within a product as yields improve, without any change to technology. It would vary based on the distribution of transistors in a chip. Greater cache size is going to impact cost a lot less than increasing logic size, especially since we know how to build redundancy into memories to improve yields.
None of this makes Kurzweil any less wrong, however.
I’m not too bothered by the software part, GenAI is kinda brute-force enabled by the ludicrous power of GPUs today. Which is why training and inference are so expensive compared to their dumber algorithmic cousins.
When RK made his predictions, we’d gone from 3Hz in 1980 to 3Ghz in 2005. A million times faster. in 25 years! If the Clock speed race had not slowed down to a crawl in the single digit Ghz because of electric/heat limitations, we would be at 571 thousand GHz and heading for Exahertz territory in the next few years!
Who knows what we could kinda brute force with that much muscle? Iit will still require much cleverness no doubt, like the current State of the Art did.
legally highly problematic … limiting market forces by “unifying criteria / agreeing on joint actions” (let’s all agree to do X, Y and Z) is a way of limiting competition - hence illegal in most juristictions …
just saying
Sure, but that sort of development will almost certainly involve a course-correction towards a more realistic appraisal of GenAI capabilities and use-cases than the ‘throw AI at the problem and see if it sticks’-approach currently popular. In other words, I’m not saying there’s no such use case, just that right now, expectations are overinflated as compared to capabilities, and those capabilities—at least for the transformer architecture—are showing signs of stagnation, where many bank on their continuing increase. Consequently, lots of people are starting to warn of an impending bubble:
And on somewhat more the technical side, Yann LeCun has recently argued that LLMs can’t possibly approach AGI, stating that ‘an artificial intelligence system trained on words and sentences alone will never approximate human understanding’: AI And The Limits Of Language
So that part of the hype that claims we only need scaling to reach AGI is also seeming more dubious. (Of course, one can also note that LLM capabilities really seem quite paltry as compared to what a human brain is able to do: while a toddler achieves language fluency after a couple of years of being exposed to whatever is spoken in their vicinity, just reading the input data of GPT-4—around 13 trillion tokens, say 10 trillion words, at an average reading speed of 250 words/minute—would take around 80,000 years, without coffee breaks. So clearly, the human brain seems to be doing something better, here.)
Bonus questions:
- Compare and contrast recent LLM models with IBM’s Watson.
- Compare and contrast the LLM auto-completion approach with google’s search engine.
- If you believe that we face sharply diminishing returns from throwing ever larger processing power at the LLM model, and that we do so imminently, what alternative software approaches can be bolted on to the LLM to improve its performance?
- Never mind human intelligence, can AI perform as well as a mouse? How well can drones navigate unfamiliar indoor hallways? (If I understand this article correctly they are working on this in an outdoor context.)
There is an international roadmap for semiconductors, started by Sematech, now run by someone else, who publishes, based on industry consensus, where things are going to go. When I worked at Bell Labs I didn’t go to meetings but I gave input to people who did. When this says that they expect process node N to be available at time T that’s what everyone shoots for.
I’ve never heard of this being an issue. It’s not like a customer can sue because a company wants to delay spending $10 billion for a fab. Depending on their engineering, some fabs hit process points with decent yields before others. I’m not a lawyer, but I doubt anyone would have much of a case.
I feel like overinflated expectations might be balanced by insufficient imagination of what can be.
What bubbles are we using as our guides? Just the dotcom bubble in 2000? The one that was more of a correction than a bubble since the interweb is living up to its potential today. If AI goes through that kind of correction, it’s no big deal.
I agree with Yann LeCun about current LLMs being far from AGI. They are incredibly inefficient compared to human brains. In that quote though, he doesn’t include AIs that are trained on more than text. Video is much more rich and is what we humans are trained on ( + smell + taste + touch .) so maybe there is more to be mined there, even though we ran out of internet text (plenty of youtube videos out there).
The dotcom bubble was very much in line with the Gartner Hype Cycle. Lot of inflated promises about what it could do, a big correction, then more practical use cases leading to sustained growth.
As I mentioned, my consulting firm (along with a lot of other firms) were making a lot of noise about GenAI about a year or so ago. About how in the next year it would be so transformative. I’m thinking how I’m running this 18 month project just to get some bank off of running its business off of Excel and I’m supposed to believe the next step is to replace everyone with robots?
This is really funny, but also true.
Our IT built a nice webapp to replace a community spreadsheet. Two years after rollout the process had morphed to “everyone keeps editing the spreadsheet just like before, but now some unfortunate PM enters the spreadsheet into the webapp”. Progress!
Hahaha. I just got off a meeting about moving some stuff off Excel so you’re spot on.
Employees now can ask a chatbot questions about company knowledge instead of emailing HR.
LLMs are much better than the alternative at understanding texts, summarizing and categorizing them and there are many usecases for that.
Generated images have “transformed” B-roll images. Instead of searching for gifs or pics, people now make their own, more custom, images.
Does that count as transformative? 
LLMs are pathological liars so you have to put them in a jail, have other AI wardens patrolling their output, and have human wardens to patrol the AI wardens.