There are a lot of articles about this. Anecdotally we all used it at my place of work, but once the novelty of it wore off, and it's usefulness sorely lacked, most people have gone back to google. Completely anecdotal as I mentioned, but it's been a long time since I couldn't use chatGPT because it was at capacity.
Maybe they managed to increase their iron, maybe their API's lightened the load, but if this was really as revolutionizing as these articles want it to be, wouldn't the load on the services still be going up?
I'm sure it'll get better, but I'm personally sort of unimpressed by it so far. It's very cool that it can pass interview questions and even university level exams, but that's not actually very useful and something you could already do (much slower) if you were really good on a search engine and had the time to do so. I mean, maybe I just really sucked at getting it to work, and there is the part where the free version is still only offering information that is older than 2021, but it really wasn't very useful for me, and not for a lack of trying. Because I really wanted it to do most of my work. Still do, by the way.
Most people who first used electricity thought it as fancy gas lights. And then, at the time computers started showing at home and then the Internet, most people didn't know what it is for.
I think it is safe to say most people have absolutely no idea what this is going to bring or how revolutionary it is going to be.
But I do think that most people will be worse after dust settles.
Most people who first used a Segway thought of it as a fancy scooter. Most people didn't know what it was for.
They had absolutely no idea what it was going to bring or how revolutionary it was going to be. (nothing much, and not)
> But I do think that most people will be worse after dust settles.
There are two sides to this in my opinion. In Denmark we need nurses, teachers, pedagogues and "trades people"/"craftsmen" (I'm sorry but I'm unsure what the world for electricians/plumbers and so on is in English). If AI were to replace a lot of what is essentially looking up rules, copy pasting data from one system to another while auditing, and so on, then we'd actually be capable of building a better society than we have now.
The other side is the distribution of wealth, which is going to be a political question. I'm not too worried about us here in Europe, but maybe that's naive.
> Most people who first used electricity thought it as fancy gas lights.
I've worked with digitalisation for one and a half decade. We even run our own language models for document recognition and classification so our employees don't have to read through 300-1000 page long contracts in different languages.
I have a fair idea what AI is supposed to be, but I'm just not seeing it in the most recent hype wave. Maybe in the next one?
LLMs as they are now won't take over your job, but they are a herald for something in the future that could take your job. We aren't sure yet but the evidence is compelling enough that we need to consider the possibility.
I mean yeah sure it's cool that something can do all your work for you but as your boss, I wouldn't need you then would I? I mean this aspect of economic impact must be considered along with the convenient fact that it makes your job easier.
I also find it hard to see how you're unimpressed? I mean yeah it's not as "impressive" as a human but but it's certainly impressive in terms of progress and potential future progress given such results have never been seen before.
> I mean yeah sure it's cool that something can do all your work for you but as your boss, I wouldn't need you then would I? I mean this aspect of economic impact must be considered along with the convenient fact that it makes your job easier.
Digitalisation is basically replacing manual tasks, so I've frankly been doing it for a decade and a half. I wouldn't mind if the AI could help me write a lot of the boring code and I'm not worried about it replacing me, because even if it makes me more efficient, I've spent 7 years in the public sector seeing how little use "no-code" solutions have been to people who can't tell IT support if their phone runs android or ios.
> I also find it hard to see how you're unimpressed?
The first answer I got from ChatGPT worked, but it used a library that had been deprecated for at least 2 years. Which it then explained by it's data-cutoff point being in 2021. Since then we've put effort into getting it to do things for us, where it's ranged from "well this is a pretty outdated/less efficient way to do that" to "you're outright making stuff up, aren't you?".
I should've screen grapped it, but at one point, trying to get it to do some asp.versioning, it told us to use a conveniently named method on an object. Only it turned out that the method had never existed in any versions of asp.versioning.
Co-pilot has been even less of a success.
I've also used it for fun. I've had some discussions about books that it obviously haven't read. It's very good at pretending that its answers are correct, but it'll absolutely make stuff up. Maybe that gets better as time goes by, but a lying AI is sort of useless. On the flip-side. I'm an external examiner for CS students, and I can get it to pass most programming questions students are given. Simply by entering the exam question into it... Typically it'll give a better answer than most students will, and it'll do it in like 1 minute on the free version. So it's obviously very good at churning out results for questions that have a gazillion answers on the internet and in a much more efficient way than any "old" search engine would. It's just that when it's tasked with doing something that hasn't been done a million times before, then it sort of falls flat.
I still have high hopes, that AI's will help automate a lot of the things we write. We'e seen some use of them for automating documentation, which if they can also maintain it to keep it up-to-date would be a game changer, but on a whole, they are largely unimpressive to me personally. But I'm open to be impressed by the next versions.
Raises in brief a lot of on point topics. However, I think the root of the disruption relies on the concept of copying skill. Which is a brand new concept and every effect will be derived from that. Requires new ways of thinking about the impacts.
Any new skill/knowledge can be absorbed by the AI and scaled infinitely. Where will be the incentive for the human investment? We have copyright and other means to protect intellectual property. However, there is no protection for a skill. There has never been a way before to mimic a skill without a high burden of cost which itself was the protection. A partial limiting factor at present is the amount of data required for training the AI models. However, that is likely to be overcome at some point with future advancements.
What's the incentive for AI to share energy resources with humans or even animals? I assume that AI will be driven by selfish ambition and its only interest will be reaching the new heights of its own intelligence.
Yes, that is the big alignment question. Nobody really knows. Opinions differ greatly on this topic, but most have a significant level of concern.
We can't say that AI would necessarily be driven by selfish motives. However, that actually isn't the main concern. It is the fact that we can not perceive the means by which it may execute any task. It is like all the old tales of a genie of the lamp that grants wishes. However, they don't turn out as you expect.
For example, ask the AI to end human conflict. Possible results, kill all humans, then no conflict. Or, create a nano tech mind virus to control all of humanity and force uniform conformity.
Finally, we have no idea how to actually successfully align an AI. Most say it is very difficult. I'm more inclined to say it is actually impossible based on their own premise which is a paradox. I've expanded on that in great detail here in case you are interested - https://dakara.substack.com/p/ai-singularity-the-hubris-trap
"And such an exciting, frightening shift is a "double-edged sword," Chen said, envisioning using AGI to tackle climate change, for example, but also warning that it is a tool that we want to be as "steerable as possible."
Quite revealing as well as ridiculous. We are circling the drain as a species not because of our intelligence but our stupidity, our unwillingness to learn and change. Clever use of transformers to speed up and improve the quality of output of LLMs is not intelligence, and tackling climate change is not the job for models that tell us what they think we want to hear.
What I find most interesting about chatGPT is that it correctly parses what I tell it. Even if it generates a BS answer from time to time at least it understood what I was after.
That in and of itself is monumental and will surely at the very least change how we interact with computers. It reminds me very much of the Star Trek Ship computer.
I certainly think that people who use AI will replace people who don't. People that don't use AI won't be able to keep up, and the gap is only going to get wider. I find the claim that AI will replace jobs outright pretty dubious, it's clickbait.
Apparently people who use AI will be equally replaced. Just this past week, hundreds of AI projects were made obsolete in a single moment as noted here when Google and Microsoft decided to make AI part of all their tools.
Then, shortly after that, these multi million dollar models were shown to be almost replaceable by a $600 dollar model by leveraging the larger models to train the smaller model.
With each “improvement” over running a rack of servers in a closet at the office, the fraction of each team I’ve worked in that is dedicated to system administration has increased.
One part time person should be able to easily manage about a dozen physical machines. Good luck getting a part time person to set up a kubernetes cluster, ci pipeline for the containers, implement a bunch of custom operators, and then monitor the result.
I use ChatGPT (with this plugin https://github.com/dpayne/CodeGPT.nvim) pretty frequently. The output it produces requires a lot of massaging, but it is still faster than doing things from scratch.
AI's ultimate impact is more comparable with the printing press. Its an information revolution, greater than the internet, not really comparable with electricity (which is physical).
The printing press:
1. Made information affordable to the masses for the first time in history
2. Radically accelerated science development due to ability to actually distribute research.
3. Enabled large empires governed by bureaucracy (centrally controlled) rather than aristocracy (locally autonomous, such as the mongols). China was a forerunner here because of their paper & printing technology.
4. Broke the information monopoly of the aristocracy-church combination. Made mass democracy (beyond small cities) possible for the first time
5. Caused large wars (the 30 years war), killing millions. The ability to mass print bibles (The source of political and moral truth back then), drastically threatened the existing clergy class, and also enabled the rise of dangerous cults through editing the bible (what we call disinfo today, which also played out).
The immediate economic value of AI is actually rather questionable, because it seems hard to create competitive moats around the tech. But the wider impact and value will be beyond people's imagination.
What do you mean by hard to create a competitive moat around the tech? These LLMs are super expensive to train so isn’t that a moat, or am I understanding it differently?
Kind of? $5M to train a new LLM from scratch is not much of a moat even now we've left the 0 interest period, the question is will people start spending much more on networks in the future. (I'm sure they will spend more, but will there ever be a billion dollar network? idk)
Certainly there are many industries that are impossible to enter with this than 100x that amount of money, either because of regulations or engineering challenges (or both).
Right now, sure. But we are in the exponential upturn right now that will moderate into an S curve where the incremental improvements will slow and you will have a convergence of the large language models.
Using mostly the same data will lead to mostly the same models. At which point, once they leak, what difference does it make how expensive it is to create?
It will be the first white collar job purge of our generation.
There are so many jobs that now can be easily automated or compressed to a fraction of the workforce needed before. Everyone is a prompt engineer now.
Think goverment / banking / insurrance / big companies
Every workload that is based around consent or rules will be impacted.
Even as a developer my leverage for certain task just went up by a huge margin.
The first, really? What about what happened to accountants when spreadsheets and personal computers came along? Or bank tellers when ATMs started popping up? Or even the monks who copied books when the printing press was invented?
There will always be jobs because humans will always have problems we are willing to throw money at to make go away.
All those professions experienced a massive downturn in staff required, but the jobs remained because the technology wasn't good enough that a potential customer could do the high-end tasks by themselves. We're now getting to the point that they can.
>> As such it "exhibits human-level performance" on some benchmarks, the company said.
LLMs have "exhibit[ed] human-level" and even "superhuman" performance on benchmarks, but that doesn't tell us anything about the abilities of those systems because the benchmarks are broken, and we don't even have good metrics.
What Will it Take to Fix Benchmarking in Natural Language Understanding?
Readit News
Maybe they managed to increase their iron, maybe their API's lightened the load, but if this was really as revolutionizing as these articles want it to be, wouldn't the load on the services still be going up?
I'm sure it'll get better, but I'm personally sort of unimpressed by it so far. It's very cool that it can pass interview questions and even university level exams, but that's not actually very useful and something you could already do (much slower) if you were really good on a search engine and had the time to do so. I mean, maybe I just really sucked at getting it to work, and there is the part where the free version is still only offering information that is older than 2021, but it really wasn't very useful for me, and not for a lack of trying. Because I really wanted it to do most of my work. Still do, by the way.
I think it is safe to say most people have absolutely no idea what this is going to bring or how revolutionary it is going to be.
But I do think that most people will be worse after dust settles.
There are two sides to this in my opinion. In Denmark we need nurses, teachers, pedagogues and "trades people"/"craftsmen" (I'm sorry but I'm unsure what the world for electricians/plumbers and so on is in English). If AI were to replace a lot of what is essentially looking up rules, copy pasting data from one system to another while auditing, and so on, then we'd actually be capable of building a better society than we have now.
The other side is the distribution of wealth, which is going to be a political question. I'm not too worried about us here in Europe, but maybe that's naive.
> Most people who first used electricity thought it as fancy gas lights.
I've worked with digitalisation for one and a half decade. We even run our own language models for document recognition and classification so our employees don't have to read through 300-1000 page long contracts in different languages.
I have a fair idea what AI is supposed to be, but I'm just not seeing it in the most recent hype wave. Maybe in the next one?
I personally have been using it for how do I do this in X, and why is X producing Y error.
I mean yeah sure it's cool that something can do all your work for you but as your boss, I wouldn't need you then would I? I mean this aspect of economic impact must be considered along with the convenient fact that it makes your job easier.
I also find it hard to see how you're unimpressed? I mean yeah it's not as "impressive" as a human but but it's certainly impressive in terms of progress and potential future progress given such results have never been seen before.
Digitalisation is basically replacing manual tasks, so I've frankly been doing it for a decade and a half. I wouldn't mind if the AI could help me write a lot of the boring code and I'm not worried about it replacing me, because even if it makes me more efficient, I've spent 7 years in the public sector seeing how little use "no-code" solutions have been to people who can't tell IT support if their phone runs android or ios.
> I also find it hard to see how you're unimpressed?
The first answer I got from ChatGPT worked, but it used a library that had been deprecated for at least 2 years. Which it then explained by it's data-cutoff point being in 2021. Since then we've put effort into getting it to do things for us, where it's ranged from "well this is a pretty outdated/less efficient way to do that" to "you're outright making stuff up, aren't you?".
I should've screen grapped it, but at one point, trying to get it to do some asp.versioning, it told us to use a conveniently named method on an object. Only it turned out that the method had never existed in any versions of asp.versioning.
Co-pilot has been even less of a success.
I've also used it for fun. I've had some discussions about books that it obviously haven't read. It's very good at pretending that its answers are correct, but it'll absolutely make stuff up. Maybe that gets better as time goes by, but a lying AI is sort of useless. On the flip-side. I'm an external examiner for CS students, and I can get it to pass most programming questions students are given. Simply by entering the exam question into it... Typically it'll give a better answer than most students will, and it'll do it in like 1 minute on the free version. So it's obviously very good at churning out results for questions that have a gazillion answers on the internet and in a much more efficient way than any "old" search engine would. It's just that when it's tasked with doing something that hasn't been done a million times before, then it sort of falls flat.
I still have high hopes, that AI's will help automate a lot of the things we write. We'e seen some use of them for automating documentation, which if they can also maintain it to keep it up-to-date would be a game changer, but on a whole, they are largely unimpressive to me personally. But I'm open to be impressed by the next versions.
Any new skill/knowledge can be absorbed by the AI and scaled infinitely. Where will be the incentive for the human investment? We have copyright and other means to protect intellectual property. However, there is no protection for a skill. There has never been a way before to mimic a skill without a high burden of cost which itself was the protection. A partial limiting factor at present is the amount of data required for training the AI models. However, that is likely to be overcome at some point with future advancements.
from - https://dakara.substack.com/p/ai-and-the-end-to-all-things
We can't say that AI would necessarily be driven by selfish motives. However, that actually isn't the main concern. It is the fact that we can not perceive the means by which it may execute any task. It is like all the old tales of a genie of the lamp that grants wishes. However, they don't turn out as you expect.
For example, ask the AI to end human conflict. Possible results, kill all humans, then no conflict. Or, create a nano tech mind virus to control all of humanity and force uniform conformity.
Finally, we have no idea how to actually successfully align an AI. Most say it is very difficult. I'm more inclined to say it is actually impossible based on their own premise which is a paradox. I've expanded on that in great detail here in case you are interested - https://dakara.substack.com/p/ai-singularity-the-hubris-trap
Quite revealing as well as ridiculous. We are circling the drain as a species not because of our intelligence but our stupidity, our unwillingness to learn and change. Clever use of transformers to speed up and improve the quality of output of LLMs is not intelligence, and tackling climate change is not the job for models that tell us what they think we want to hear.
That in and of itself is monumental and will surely at the very least change how we interact with computers. It reminds me very much of the Star Trek Ship computer.
https://www.youtube.com/watch?v=DH-2BHDYNfk
Then, shortly after that, these multi million dollar models were shown to be almost replaceable by a $600 dollar model by leveraging the larger models to train the smaller model.
https://www.youtube.com/watch?v=xslW5sQOkC8
It is going to be total bedlam with the disruptors disrupting the disruptors in faster and faster cycles.
One part time person should be able to easily manage about a dozen physical machines. Good luck getting a part time person to set up a kubernetes cluster, ci pipeline for the containers, implement a bunch of custom operators, and then monitor the result.
What experience of your contradicts that?
The printing press: 1. Made information affordable to the masses for the first time in history 2. Radically accelerated science development due to ability to actually distribute research. 3. Enabled large empires governed by bureaucracy (centrally controlled) rather than aristocracy (locally autonomous, such as the mongols). China was a forerunner here because of their paper & printing technology. 4. Broke the information monopoly of the aristocracy-church combination. Made mass democracy (beyond small cities) possible for the first time 5. Caused large wars (the 30 years war), killing millions. The ability to mass print bibles (The source of political and moral truth back then), drastically threatened the existing clergy class, and also enabled the rise of dangerous cults through editing the bible (what we call disinfo today, which also played out).
The immediate economic value of AI is actually rather questionable, because it seems hard to create competitive moats around the tech. But the wider impact and value will be beyond people's imagination.
Certainly there are many industries that are impossible to enter with this than 100x that amount of money, either because of regulations or engineering challenges (or both).
Using mostly the same data will lead to mostly the same models. At which point, once they leak, what difference does it make how expensive it is to create?
Think goverment / banking / insurrance / big companies
Every workload that is based around consent or rules will be impacted.
Even as a developer my leverage for certain task just went up by a huge margin.
There will always be jobs because humans will always have problems we are willing to throw money at to make go away.
LLMs have "exhibit[ed] human-level" and even "superhuman" performance on benchmarks, but that doesn't tell us anything about the abilities of those systems because the benchmarks are broken, and we don't even have good metrics.
What Will it Take to Fix Benchmarking in Natural Language Understanding?
https://aclanthology.org/2021.naacl-main.385/