But it’s not scary. It’s… marvelous, cringey, uncomfortable, awe-inspiring. What’s scary is not what AI can currently do, but what we expect from it. Can it do math yet? Can it play chess? Can it write entire apps from scratch? Can it just do my entire job for me?
We’re moving toward a world where every job will be modeled, and you’ll either be an AI owner, a model architect, an agent/hardware engineer, a technician, or just.. training data.
After an OpenAI launch, I think it's important to take one's feelings about the future impact of the technology with a HUGE grain of salt. OpenAI are masters of hype. They have been generating hype for years now, yet the real-world impacts remain modest so far.
Do you remember when they teased GPT-2 as "too dangerous" for public access? I do. Yet we now have Llama 3 in the wild, which even at the smaller 8B size is about as powerful as the [edit: 6/13/23] GPT-4 release.
As someone pointed out elsewhere in the comments, a logistic curve looks exponential in the beginning, before it approaches saturation. Yet, logistic curves are more common, especially in ML. I think it's interesting that GPT-4o doesn't show much of an improvement in "reasoning" strength.
It's glib to dismiss safety concerns because we haven't all turned into paperclips yet. LLMs and image gen models are having real effects now.
We're already at a point where AI can generate text and images that will fool a lot of people a lot of the time. For every college-educated young person smugly pointing out that they aren't fooled by an image with six-fingered hands, there are far more people who had marginal media literacy to begin with and are now almost defenceless against a tidal wave of hyper-scaleable deception.
We're already at a point where we're counselling elders to ignore late-night messages from people claiming to be a relative in need of an urgent wire transfer. What defences do we have when an LLM will be able to have a completely fluent, natural-sounding conversation in someone else's voice? I'm not confident that I'd be able to distinguish GPT-4o from a human speaker in the best of circumstances and I'm almost certain that I could be fooled if I'm hurried, distracted, sleep deprived or otherwise impaired.
Regardless of any future impacts on the labour market or any hypothesised X-risks, I think we should be very worried about the immediate risks to trust and social cohesion. An awful lot of people are turning into paranoid weirdos at the moment and I don't particularly blame them, but I can see things getting seriously ugly if we can't abate that trend.
I second that. I remember when Google search first came out. Within a few days it completely changed my workflow, how I use the Internet, my reading habits. It easily 5 ~ 10x the value of Internet for me over a couple of weeks.
LLMs is doing nothing of the sort for me.
Perhaps.
> Do you remember when they teased GPT-2 as "too dangerous" for public access? I do. Yet we now have Llama 3 in the wild, which even at the smaller 8B size is about as powerful as the [edit: 6/13/23] GPT-4 release.
The statement was rather more prosaic and less surprising; are you sure it's OpenAI (rather than say all the AI fans and the press) who are hyping?
"""This decision, as well as our discussion of it, is an experiment: while we are not sure that it is the right decision today, we believe that the AI community will eventually need to tackle the issue of publication norms in a thoughtful way in certain research areas.
…
We are aware that some researchers have the technical capacity to reproduce and open source our results. We believe our release strategy limits the initial set of organizations who may choose to do this, and gives the AI community more time to have a discussion about the implications of such systems."""
I spend a part of yesterday evening sorting my freshly dried t-shirts into 4 distinct piles. I used OpenAI Vision (through BeMyEyes) from my phone. I got a clear description of each and every piece of clothing, including print, colours and brand. I am blind BTW. But I guess you are right, no impact at all.
> Yet we now have Llama 3 in the wild
Yes, great, THANKS Meta, now the Scammers have something to work with. Thats a wonderful achievement which should be praised! </sarcasm>
People read too many sci-fi books and then project their fantasies on to real-world technologies. This stuff is incredibly powerful and will have social effects, but it’s not going to replace every single job by next year.
I can't help but notice the huge amount of hindsight and bad faith that it demonstrated here. Yes, now we are aware that the internet did not drown in a flood of bullshit (well, not noticeably more), when GPT-2 was released.
But was it obvious? I certainly thought that there was a chance that the amount of blog spam that could be generated effortlessly might just make internet search unusable. You are declaring "hype", when you could also say "very uncertain and conscientious". Is this not something we want people in charge to be careful with?
Maybe that is GPT-5.
And this release really is just incremental improvements in speed, and tying together a few different existing features.
Go ask any teacher or graphician.
Maybe not GPT-2, but in general LLMs and other generative AI types aren't without their downsides.
From companies looking to downsize their staff to replace them with software, to the work of artists/writers being devalued somewhat, to even easier scams and something like the rise of AI girlfriends, which has also gotten some critique, some of those can probably be a net negative.
Even when it's not pearl clutching over the advancements in technology and the social changes that arise, I do wonder how much my own development work will be devalued due to the somewhat lowered entry barrier into the industry and people looking for quick cash, same as with boot camps leading to more saturation. Probably not my position individually (not exactly entry level), but the market as a whole.
It's kind of at a point where I use LLMs for dev work not to fall behind, cause the productivity gains for simple problems and boilerplate are hard to argue with.
I feel like everyone who makes this claim doesn't actually have any data to backup it up.
~8 years ago when self driving technology was all the rage and every major company was getting on board with ever more impressive technological demos, it seemed entirely reasonable to expect that we'd all be in a world of complete self driving imminently. I remember mocking somebody online around the time who was pursuing a class C/commercial trucking license. Yet now a decade later, there are more truckers than ever and the tech itself seems further away than ever before. And that's because most have now accepted that progress on such has basically stalled out in spite of absolutely monumental efforts at moving forward.
So long as LLMs regularly hallucinate, they're not going to be useful for much other than tasks that can accept relatively high rates of failure. And many of those generally creative domains are the ones LLMs are paradoxically the weakest in - like writing. Reading a book written by an LLM would be cruel and unusual punishment given then current state of the art. One domain I do see them completely taking over is search. They work excellently as natural language search engines, and "failure" in such is very poorly defined.
I think what maybe seems not obvious amidst the hype is that there is a hell of a lot of engineering left to do. The fact that you can squash the weights of a neural net down to 3 bits per param and it still works -- is evidence that we have quite a way to go with maturing this technology. Multimodality, improvements to the UX of it, the human-computer interface part of it. Those are fundamental tech things, but they are foremost engineering problems. Getting latency down. Getting efficiency up. Designing the experience, then building it out.
25 years ago, early tech demos on the internet were promising that everyone would do their shopping, entertainment, socializing, etc... online. Breathless hype. 5 years after that, the whole thing crashed, but it never went away. People just needed time to figure out how to use it and what it was useful for, and discover its limitations. 10 years after that, engineering efforts were systematized and applied against the difficult problems that still remained. And now: look at where we are. It just took time.
Meanwhile I've been using ChatGPT at work for _more than a year_ and it's been tremendously helpful to me.
This is not hype, this is not about how AI will change our lives in the future. It's there right here, right now.
Yep. So basically they're useful for a vast, immense range of tasks today.
Some things they're not suited for. For example, I've been working on a system to extract certain financial "facts" across SEC filings. ChatGPT has not been helpful at all either with designing or implementing (except to give some broad, obvious hints about things like regular expressions), nor would it be useful if it was used for the actual automation.
But for many, many other tasks -- like design, architecture, brainstorming, marketing, sales, summarisation, step by step thinking through all sorts of processes, it's extremely valuable today. My list of ChatGPT sessions is so long already and I can't imagine life without it now. Going back to Google and random Quora/StackOverflow answers laced with adtech everywhere...
But is this not what humans do, universally? We are certainly good at hiding it – and we are all good at coping with it – but my general sense when interacting with society is that there is a large amount of nonsense generated by humans that our systems must and do already have enormous flexibility for.
My sense is that's not an aspect of LLMs we should have any trouble with incorporating smoothly, just by adhering to the safety nets that we built in response to our own deficiencies.
mapping th genome was that way. On a 20yr schedule, barely any progress for 15 and then poof, done ahead of schedule
I have a much less "utopian" view about the future. I remember during the renaissance of neural networks (ca. 2010-15) it was said that "more data leads to better models", and that was at a time when researchers frowned upon the term Artificial Intelligence and would rather use Machine Learning. Fast forward a decade LLMs are very good synthetic data generators that try to mimic human generated input and I can't think somehow that this wasn't the sole initial intent of LLMs. And that's it for me. There's not much to hype and no intelligence at all.
What happens now is that human generated input becomes more valuable and every online platform (including minor ones) will have now some form of gatekeeping in place, rather sooner than later. Besides that a lot of work still can't be done in front of a computer in isolation and probably never will, and even if so, automation is not a means to an end. We still don't know how to measure a lot of things and much less how to capture everything as data vectors.
Currently the bottleneck is Agents. If you want a large language model to actually do anything you need an Agent. Agents so far need a human in the loop to keep them sane. Until that problem is solved most human jobs are still safe.
I fully expect GPT 5 (or at the latest 6) to similarly have native inclusion of agentic capabilities either this year or next year, assuming it doesn't already, but is just kept from the public.
not quite sure that sanity is a business requirement
I understand that you might be afraid. I believe that a world where only LLM companies rule the world is not practically achievable except in some distopian universe. The likelihood of the world where the only job are model architects, engineers or technicians is very very small.
Instead, let's consider the positive possibilities that LLMs can bring. It can lead to new and exciting opportunities across various fields. For instance, can serve as a tool to inspire new ideas for writers, artists, and musicians.
I think we are going towards a more collaborative era where computers and humans interact much more. Everything will be a remix :)
Oh, especially since it will be a priority to automate their jobs, or somehow optimize them with an algorithm because that's a self-reinforcing improvement scheme that would give you a huge edge.
GPT-4? Not that well. AI? Definitely
https://deepmind.google/discover/blog/alphageometry-an-olymp...
So outside of use-cases where the user can quickly verify the result (like picking a decent generated image etc.),I can't see it being used much.
All AIs up to now lack autonomy. So I'd say until we crack this problem, it is not going to be able to do your job. Autonomy depends on a kind of data that is iterative, multi-turn, and learning from environments not from static datasets. We have the exact opposite, lots of non-iterative, off-policy (human made AI consumed) text.
But everyone is expecting them to release gpt5 later this year, and it is a bit scary to think what it will be able to do.
1) It's natively multi-modal in a way I don't think gpt4 was.
2) It's at least twice as efficient in terms of compute. Maybe 3 times more efficient, considering the increase in performance.
Combined, those point towards some major breakthroughs having gone into the model. If the quality of the output hasn't gone up THAT much, it's probably because the technological innovations mostly were leveraged (for this version) to reduce costs rather than capabilities.
My guess is that we should expect them to leverage the 2x-3x boost in efficiency in a model that is at least as large as GTP4 relatively soon, probably this year unless OpenAI has safety concerns or something, and keeps it internal-only.
The evidence for that is the change in the tokenizer. The only way to implement that is to re-train the entire base model from scratch. This implies that GPT 4o is not a fine-tuning of GPT 4. It's a new model, with a new tokenizer, new input and output token types, etc...
They could have called it GPT-5 and everyone would have believed them.
Everything always starts as a toy.
That includes, beyond literal Killers, all kinds of manufacturing, construction and service work.
I would expect a LOT of funds to go into research all sorts of actuators, artificial muscles and any other technology that will be useful in building better robots.
Companies that can get and maintain a lead in such technologies may reach a position similar to what US Steel had in the 19th century.
That could be the next nvidia.
I would not be at all surprised if we will have a robot in the house in 10 years that can clean and do the dishes, and that is built using basically the same parts as the robots that replace our soldiers and the police.
Who will ultimately control them, though?
This is no different to saying a person with a gun murdered someone rather than attributing the murder to the gun. An AI gun is just a really fancy gun.
What's scary and cringey are your delusions.
My guess is the future belongs to those who don't stop—who, in fact, embrace the opposite of stopping.
I would even suggest that the present belongs to those who didn't stop. It may be too late for normal people to ever catch up by the time we realize the trick that was played on us.
