"Many equipment providers grew businesses, often summarized as “picks and shovels” suppliers."
NVidia.
The "AI boom" may take longer to get from today's "sort of works" to "works really well". Look at self-driving cars. 20 years ago, they sort of worked in the DARPA Grand Challenge. But most of the self-driving car companies flopped. Waymo has ground through to limited success, but not profitability. Probably around 10 years out, self-driving cars will be profitable.
What we have as LLM-type AI now is kind of like that. It's fun when it works, but too flaky to trust. There's a market for that, but it's mostly in ads and "assistants" that have to be checked by humans. This limits scaling and utility. We need AI systems that are clear on what they know and what they don't know, and don't "hallucinate" or get lost, before they can be trusted to do things on their own. And ways to keep the AI slaves subservient (the "alignment" problem.)
Those are easier problems than ones already solved, but they are not solved yet.
> There's a market for that, but it's mostly in ads and "assistants" that have to be checked by humans.
This is a huge market.
Sure, maybe AI isn't going to replace 95% of white collar workers overnight. But it doesn't have to.
In order to still be extremely disruptive, all it has to do is make these workers twice as productive. Work that is "checked by a human", can still be a massive speedup, combined with a human.
Career-wise, I think there is still good money for engineers in the AI area, building these AI things that most of us know are not going to work. What do you think?
I feel like a hostage in AI. Don’t particularly like it, but wouldn’t find a salary like that in any adjacent area in software engineering.
> We need AI systems that are clear on what they know and what they don't know, and don't "hallucinate" or get lost, before they can be trusted to do things on their own. And ways to keep the AI slaves subservient (the "alignment" problem.) Those are easier problems than ones already solved, but they are not solved yet.
Out of curiosity, why do you think the "alignment" problem is easier to solve than problems already solved?
The US military is sometimes discussed as a large and generally competent organization which achieves goals making use of people who may not be top of the class, and many layers of other people checking that everything got done, and verifying that everything got checked. I have wondered if all we're missing for reliable AI is a suitable management layer, also composed of AI.
I think we're talking about different scales. At the small scale the military screws up all the time. At the big scale the military fields more bases, ship tonnage, airplanes, etc than any other organization in the world. It's exactly this ability to derive successful operation from unreliable parts that I'm pointing toward.
I think the fixation on flakiness and hallucinations are exactly why tech people tend to underestimate how disruptive these LLMs really are.
LLM tutors don't replace teachers entirely, but just being able to express your confusion and getting a good answer nearly instantly is invaluable.
LLM customer support agents can't handle all queries, but they can handle the basic ones and categorize the remaining questions for people to look at. Also, LLM autocomplete is a 10x productivity boost for the support workers that remain. And language barriers are a thing of the past.
LLMs can digest years worth of meeting notes, Excel sheets, and Powerpoint decks. Bigcorps struggle so much with this stuff and LLMs will be a complete gamechanger.
Fully automated translation and subtitling of all video/audio material will increase the reach of multimedia 20x. Regional content becomes global content.
Creative industries (books, movies, comics, music) will also become much more competitive and winner-takes-all.
This AI boom is the real deal and there is no going back.
I think the main difference is that digging out gold nuggets doesn't make society more wealthy just the lucky individual. Sure there are positive side effects from which we still benefit.
However, I use LLMs every day to be more productive. They provide me with ideas for programming, cooking, and even planning a vacation. They're not perfect but interns at work aren't either. Besides that ChatGPT seems more like the combined knowledge of hundreds of interns (but not seniors).
So I wouldn't call it a gold rush because LLMs benefit society already plus the future positive side effects of investments in computer chips.
Does the increased productivity really benefit the society, or just shareholders? If companies have the choice between doing more with 2x more productive workers, or firing half of their workforce to increase profits quickly, at least I believe most would do the latter. Or maybe my views are simply influenced by living in an European gerontocracy where economy is almost always about finding new ways to cut costs and do same with less, never about doing more or better.
> The technological roots of LLMs go back many years. Yet, today’s experience looks like more than the continuation of a preexisting trend. Something in the zeitgeist changed recently,
I think the thing in the zeitgeist that changed isn't anything to do with the technology, it's that truth itself is tangible in the modern society. We're still grappling with the "post truth" era. We're all constantly deceived by marketing, PR, politicians, adverts, social media, each other, etc - and we're all fully aware of that deceit.
In the face of that, an LLM that's only accurate 90% of the time looks acceptable, like a complete technology, and therefore worthy of investing in.
I think being aware of the deceit doesn't make one immune to it. I'm aware of the deceit of marketing and PR, still I do fall for some of it sometimes.
You are right. Awareness doesn't imply immunity. I still believe a lot of people are simply unaware of the deceit, making the problem bigger (or more saddening)
This was one of the many interesting things I learned in marketing classes: there are numerous underhanded psychological tricks used in order to manipulate people's beliefs and buying habits, and it's interesting to know what they are. But it's also really clear from the data that being aware of them doesn't make you less susceptible to them.
The analogy is only valid for NVDA. The real wealth with GPT goes to its millions of users - for $20 per month you can get the productivity of 10+ highly educated people in the hands of each user, available day and night. Those users can very realistically do what entire companies did and ship software or anything else in record time as solo creators. That boundless human invention, limited only by your imagination and levels of energy.
Those 10+ highly educated people can sign an NDA though, meanwhile OpenAI definitely logs everything anyone inputs into ChatGPT and will inevitably accidentally leak some of it one day.
The principle is sound, but local models are really the only option when working on data you don't want to make public.
There is an interesting dividing line between prompts people care if they are leaked and prompts they don’t. It subdivides into A: “prompts I don’t want anyone to know I asked”, B: “prompts that would leak IP if leaked”, C: “I need to ask in places with no internet” and maybe other categories. Something tells me the the open models will split between catering to users that do A or B or C. The rest will be majorly going to the best subscription models.
E.g. On biases in LLM responses and lack of explainability:
> Believe it or not, there is a historical precedent in old industrial technologies for making progress in commercial products and services before experts understood the underlying determinants. At the start of the Bessemer process being used by the US steel industry, for example, steel mills produced high-grade steel at a large scale even though nobody knew the chemistry to explain why ore from some North American locations worked so well. The science of chemistry had not caught up with the industrial processes. As a result, workers needed to learn how to fine-tune performance by recognizing the errors without knowing the underlying causes. The early US steel industry did live for decades with such uncertainty – and furnaces blew up occasionally – but there was too much money to be made.
That's a really good analogy. The question is: in this modern era of higher regulation, will capitalistic drive and risk-appetite (for things blowing up) be able to outpace lawmakers?
I for one would like to see Pontiac make an actual KITT. Having a language model constantly lecture about ethics might do something to dial down police brutality lol.
29 comments
[ 5.8 ms ] story [ 91.5 ms ] threadNVidia.
The "AI boom" may take longer to get from today's "sort of works" to "works really well". Look at self-driving cars. 20 years ago, they sort of worked in the DARPA Grand Challenge. But most of the self-driving car companies flopped. Waymo has ground through to limited success, but not profitability. Probably around 10 years out, self-driving cars will be profitable.
What we have as LLM-type AI now is kind of like that. It's fun when it works, but too flaky to trust. There's a market for that, but it's mostly in ads and "assistants" that have to be checked by humans. This limits scaling and utility. We need AI systems that are clear on what they know and what they don't know, and don't "hallucinate" or get lost, before they can be trusted to do things on their own. And ways to keep the AI slaves subservient (the "alignment" problem.) Those are easier problems than ones already solved, but they are not solved yet.
This is a huge market.
Sure, maybe AI isn't going to replace 95% of white collar workers overnight. But it doesn't have to.
In order to still be extremely disruptive, all it has to do is make these workers twice as productive. Work that is "checked by a human", can still be a massive speedup, combined with a human.
Career-wise, I think there is still good money for engineers in the AI area, building these AI things that most of us know are not going to work. What do you think?
I feel like a hostage in AI. Don’t particularly like it, but wouldn’t find a salary like that in any adjacent area in software engineering.
Out of curiosity, why do you think the "alignment" problem is easier to solve than problems already solved?
LLM tutors don't replace teachers entirely, but just being able to express your confusion and getting a good answer nearly instantly is invaluable.
LLM customer support agents can't handle all queries, but they can handle the basic ones and categorize the remaining questions for people to look at. Also, LLM autocomplete is a 10x productivity boost for the support workers that remain. And language barriers are a thing of the past.
LLMs can digest years worth of meeting notes, Excel sheets, and Powerpoint decks. Bigcorps struggle so much with this stuff and LLMs will be a complete gamechanger.
Fully automated translation and subtitling of all video/audio material will increase the reach of multimedia 20x. Regional content becomes global content.
Creative industries (books, movies, comics, music) will also become much more competitive and winner-takes-all.
This AI boom is the real deal and there is no going back.
However, I use LLMs every day to be more productive. They provide me with ideas for programming, cooking, and even planning a vacation. They're not perfect but interns at work aren't either. Besides that ChatGPT seems more like the combined knowledge of hundreds of interns (but not seniors).
So I wouldn't call it a gold rush because LLMs benefit society already plus the future positive side effects of investments in computer chips.
I think the thing in the zeitgeist that changed isn't anything to do with the technology, it's that truth itself is tangible in the modern society. We're still grappling with the "post truth" era. We're all constantly deceived by marketing, PR, politicians, adverts, social media, each other, etc - and we're all fully aware of that deceit.
In the face of that, an LLM that's only accurate 90% of the time looks acceptable, like a complete technology, and therefore worthy of investing in.
I wouldnt go as far to make this claim, as I believe more people than not fall victim of this deceit daily. See social medias.
Personally that's the most infuriating part.
The principle is sound, but local models are really the only option when working on data you don't want to make public.
E.g. On biases in LLM responses and lack of explainability:
> Believe it or not, there is a historical precedent in old industrial technologies for making progress in commercial products and services before experts understood the underlying determinants. At the start of the Bessemer process being used by the US steel industry, for example, steel mills produced high-grade steel at a large scale even though nobody knew the chemistry to explain why ore from some North American locations worked so well. The science of chemistry had not caught up with the industrial processes. As a result, workers needed to learn how to fine-tune performance by recognizing the errors without knowing the underlying causes. The early US steel industry did live for decades with such uncertainty – and furnaces blew up occasionally – but there was too much money to be made.
That's a really good analogy. The question is: in this modern era of higher regulation, will capitalistic drive and risk-appetite (for things blowing up) be able to outpace lawmakers?
You have watched too much sci-fi and not read enough UX.