This time, it feels different
nadh.in
nadh.in
I’ve lived through the web in 1995. Google in 2000. Web 2.0 in 2005. Smartphones in 2009. Now it’s generative AI in 2023.
Every time, there’s an argument about whether it’s the end of the world or a total nothingburger. Every time, industry leaders are knocked off their perch. Every time, there are surprise developments and innovations, both positive and negative. Every time, within a few years, the new technology goes from seeming magical even in the wealthiest and most advanced locations to commonplace in a rural farm in the developing world.
It’s an incredible, terrifying and fascinating time to be alive.
Is that really true? The two largest corporations in the world by market cap are Apple and Microsoft, both founded in the 1970s.
I've never thought about computers that way - they got here, and didn't make everyone's lives better - they just accelerated wealth concentration in the top 1%.
Easy trade right.
The technology you take for granted is huge. Warren Buffet said he would rather give up flying on a private jet, than give up his smartphone.
Your phone does not mean that you don't have to work three jobs to have a roof over your head and can't have a family. Having a phone does not make you rich.
This does not mean it’s still difficult to get housing. It’s just a point that one aspect of our lives is unfathomably better than before and now we take it for granted.
Huh?? What world are you living in where the standard of living hasn't increased in the last 50 years?
- Are we just discussing the first-world middle class? Obviously if we extend our scope to all of humanity there have been remarkable jumps in standards of living.
- Does the idea of 'standard of living' include only access to goods, services, and information or also intangibles like access to safety and (more pertinently to this particular discussion) security? As a 34 year old American, I feel that my material standard of living is much better than it was in 1992, but my sense of security in said standard of living has taken a nosedive.
While it's hard to say, that might be the news you read, rather than reality.
There's also that I am a disabled American and we're generally not in great spots. As an American, your ability to have things like housing is directly tied to your ability to be economically productive, and as someone with MS, I have to roll the dice each morning on whether or not I will be able to work. I don't know how long my career can last which creates a lot of stress. But then that's also true for all Americans: How many of us could survive our main/only earner getting cancer? That's the sort of stability I mean: The ability to be wiped out by things outside of your control is higher (which is partially down to society deciding everything is an individual problem and if a tornado flattens your house that sounds like a you problem) and we're moving towards a lower-trust society and the community bonds that were existent but frayed in the 90s (see Bowling Alone) are pretty torched now. Who and what can you rely on?
Which is more relevant if you consider that a lot of Americans have medical conditions and that things like the move from a single-income to dual-income household as the standard means that losing one adult can cripple a household financially. I'm also basing this on things like my peers' ability to do things like pay for child care since mid-30s are prime 'give all your money to daycare' years.
I also was in school during 2008, so I entered college under the 'get a degree and you'll be fine' years and exited it into the bloodbath of the Great Recession. There are many realities and my personal one is less secure than it was and this is true for many Americans.
What's your opinion on the following assertion: countries (such as mine) with socialised medicine make sure everyone has a certain standard of care, but that standard only goes so far. The US lets people pay for any possible treatment, and so people who would be put on palliative care in, say, the UK, would be bankrupted in the US by trying to beat cancer with expensive drugs.
It's phrased as an assertion, but it's really just something I've been thinking about. What do you think?
It isn't just socialized healthcare, though. It's also things like the US not mandating time off - for instance, I haven't had my Ocreveus infusion because I literally can't afford to take the time off to go to the infusion center. And the US is also very binary: Either you work full time or you get nothing benefits wise. For those of us who could comfortably work ~25-30 hours a week this results in either us overworking ourselves to the detriment of our health or not working at all. I'd love to work and just have a little bit of help for when I have a bad week or month, but that doesn't happen here.
The other problem with the American healthcare system is that it siphons off wealth from those in their end of life. Good luck planning a good life for your children: The government will take everything you own for your nursing home. That also has major impacts in term of stability.
But of course in socialized healthcare you do need to decide what is and isn't worth treating in some way and when cuts are needed you end up in dark places like Canada is with MAiD.
It's hard to say. I'm inclined to believe there are benefits and drawbacks to each approach.
The ties between government, big pharma, insurance companies and hospital systems must be so entrenched that it's very difficult to see a way out.
Far more per head — part of which is higher cost of labor and that healthcare is extraordinarily labor intensive. But, its also fairly high as a GDP share, for which excuses like that don’t work. The US just genuinely has a healthcare system that is ludicrously inefficient unless your goal is maximizing the wealth-based quality differential, in which case Mission Accomplished.
Or more precisely: efficiency is distorted by compassion. Which is then recovered by efficiency. Which is again modified by compassion. If your system is a bewildering mish-mash of cost increasing political interventions and cost reducing market innovations then it's never going to work. But lawmakers need something to do to stand out, and so do business people.
The only mention of wealth was describing the effect. Yes, politics is a key part of how that effect is acheived. (Of course, advocacy by those with wealth is a key part of how the politics happens, and... well, the politics/wealth interactions here are basically a near-infinite chain. They aren’t competing explanations.)
Add that to America disintegrating into a low-trust society where one almost has to assume that any stranger one meets is out to scam you/get your money somehow and it results in a country where it's terrible to be vulnerable. Which is one of the ways we grow as people, thus leading to emotional stunting amongst the populace, which in turn makes us less resilient and willing to act (particularly since we can't cooperate anymore) and therefore in my opinion sow the seeds of our own destruction.
Doesn’t that imply a decrease in standard of living? Not easy to quantify one. I personally see not living in fear as a staple of standard of living.
Also American society is pretty disdainful of security, historically.
The argument against this would be things like homes being larger now and housing standards having risen (even slum units have things like fridges and electricity). I don't think those are compelling arguments because often going without those things is not allowed/not an option, but that's what the other side's argument would be.
Wages have remained flat, cost of consumer electronics have gone down, cost of education, healthcare, and housing has gone way up.
If you consider housing, food, medical insurance, car prices, education - all of these things are ridiculously expensive compared to what they used to be.
"But we have the Nintendo Switch!" does not alleviate what we're going through right now.
And since 1979 people have far more basic 'necessities' like internet and electronic devices. And other necessities like housing, education, and healthcare have all increased in cost far beyond nominal inflation figures. So a person in this situation is most certainly going to have a substantially lower quality of life in modern times. You have to keep in mind that in the past most families were single bread-winner, so wages and life were organized around the vision of one working individual being able to support at least 3 others, in a nuclear household. It was a very different economic vision than what has emerged in more modern times.
---
The one thing I'd say these numbers miss out on is opportunity. I think in modern times there is dramatically more opportunity, but if you're just going to live a normal, as expected, life without really aggressively going after the opportunities we all have available - then you'd almost certainly be better off living on the median in the 70s than today. And since most people should be expected to live these sort of perfectly average lives, as that's precisely what the average is, that's a real problem.
That might even be related (i.e., wages for workers are being depressed in developing world as workers see competition from other workers in developing nations). And capital owners benefiting from increased availability of labor.
TBC, agree that gini coefficient re: wealth in developing world is increasing.
Look at India, they don't come close to China if you compare how the standard of life improved for the masses.
But anyway, they didn't specify all industry leaders, and it's pretty solid that many leaders were knocked off their perches.
The comment talks about "quickly changes the world around us rather fundamentally", "knocked off their perch", and "within a few years", which doesn't seem like an accurate description of what happened with Yahoo.
If you're talking about the current AT&T, then I'd argue it's not a great example. The relationship between old AT&T and the company currently calling itself AT&T is tenuous. It would be more accurate to say the brand name survived, because it has good name recognition value.
Apple had the obvious comeback, but even Microsoft had to reinvent itself away from windows.
For the uninitiated: https://news.ycombinator.com/item?id=1636198. The original video is gone but it lives on YouTube: https://m.youtube.com/watch?v=b2F-DItXtZs.
I guess it’s the same with generative AI.
I do think there is a lot of overlap with people promoting self driving predictions that haven't panned out.
I mean, it's true that we still don't know whether self-driving cars are a few years away, but the analogy with "LLMs are the beginning of AGI" is apt.
I still think it's rubbish for people to demand two revolutions a decade or they get bored but I understand the scope a bit better.
My internet-connected smart phone allows me to pay securely for things, unlock the public city bikes for local transit, find quickly directions for a place, provide me answers to questions, read books at the cafe, etc...
All this without a single wallet or bag of "stuff" dangling over my shoulders.
Now... I don't have a facebook account ... YMMV ... :)
A co-worker had to skip lunch last month because he relied on his phone for payments and the app wouldn't let him pay - no explanation given. I keep critical stuff like payment methods separate and have no plans to stop carrying a wallet because of stuff like this.
Or the phone falls on the sidewalk and the screen cracks rendering the entire phone useless.
Jobs and Berners-Lee didn’t think there was at least a 10% chance of the iPhone or Web ending civilization, like half of AI researchers do.
Like the author said, it’s different this time.
To summarize: The claim—that half of AI researchers believe that there is at least a 10% chance that rogue AI will kill us all—is based on a survey question that included 162 respondents. All we know about these respondents is that they were authors on papers published at 2021 machine learning conferences. Possible confounding issues are vagueness of the question, small sample size, response bias, confidence of responses, and level of expertise.
https://aiguide.substack.com/p/do-half-of-ai-researchers-bel...
LLM is fundamentally information processing technology. Not the path to AGI, not an emerging sentience.
The reason "this time" feels so amazing is that the unwashed masses suddenly got access to new information processing technology in a context where their tools been stagnant for decades.
Not because it was not possible, but because there was no money in it.
To understand my argument and its implications, humor me for a moment and imagine a universe in which everyone was already using linux computers and for a decade now published papers of ML/DL were available for people to use. So there were various crowdsourced indices and models of all sorts, which people incrementally embedded in their information processing workflows.
In that universe there would be no room for the delirious reaction we have here. It would be incremental evolution of search, knowledge bases, algorithmic content generation etc.
What we have experienced instead in these past decades is information tool starvation. These incrementally improving tools, while nothing but known and ultimately mundane algorithms were not available except within a tiny elite.
In fact, people's information processing capability arguably declined as the desktop platforms got downgraded, adtech toxic waste covered the information landscape etc.
What is happening now is that a socioeconomic and artificially induced scarcity is now being broken (for reasons that require serious piecing together of events).
So while on the surface this hype is as distastful as any illustration of human lemming behaviour, there is enormous silver lining if we succeed to read its causes.
These tools are here, have been here for a while and they can be inserted into our ever growing information processing toolkit. The risks are there to match the opportunities.
The biggest risk of them all is precisely what has led to the current situation. Technology not diffusing normally, but controlled by gatekeepers.
The tech existed but was clunky and unavailable. Now it’s good and accessible enough to be used by most people. This is how technological revolutions happen!
Edison didn’t invent the lightbulb. But he made it useful and invented power distribution. That’s the important part. Consumers don’t care about technology that exists in a lab that they cannot even begin to fathom how to use.
Mind you last year even experts didn’t think LLMs would be able to do what they’re doing right now for at least another 5 to 6 years. This also is in part a side-effect of the hype and extra attention. All this attention causes an acceleration.
It probably should rely on its neurons less, and use symbolic reasoning more. That, and learn and grow, either maintain a fact database, or retrain its neurons or add more neurons for new facts.
This logic doesn't follow - AGI will be, fundamentally, information processing technology, just as the human brain is, fundamentally, information processing technology.
> "Not the path to AGI, not an emerging sentience". These are HUGE claims. Define AGI. Define sentience. What is the path and why isn't LLM on it?
Right now those words are only dangerous misinformation - originating from a brain trying to protect its shaking world view. A reminder: you can talk with computers now. Think about that for a moment.
Finally, OpenAI has already stated there is a limit to how far the current AI models can progress. ChatGPT 5 will eventually be released but will be hitting the limits. A paradigm shift is needed to move to the next phase of AI.
How is this different from how humans respond? Our model could be just some orders of magnitude more complex. Or do you think there some _fundamentally_ different things are going on in the human brain?
For example, if you ask ChatGPT if books can be fluorescent it says no. However, as an adult you know someone somewhere has made a book with fluorescent images, as it is a cool thing. You are combining knowledge from two different fields (books + fluorescence) and establishing the likelihood of someone being able to combine them.
Me: Can books be fluorescent?
GPT-4: While the term "fluorescent" is typically used to describe substances that can absorb light and then re-emit it, often at a different wavelength, there's no inherent reason why a book couldn't be made with fluorescent properties.
This could be achieved by using fluorescent ink, dyes, or paints on the cover or the pages, or by incorporating fluorescent fibers into the paper itself. When exposed to ultraviolet light (often called "black light"), these materials would glow. This might be used for aesthetic reasons, for practical reasons like aiding reading in low-light conditions, or for interactive elements in children's books or art books.
It's important to note that this is not a common feature for most books, as it would increase production costs and might not be appealing or necessary for all readers or all types of books. Also, long-term exposure to ultraviolet light can be damaging to many types of paper and ink, which could reduce the lifespan of the book.
As of my knowledge cutoff in September 2021, I don't have any specific examples of fluorescent books, but that doesn't mean they don't exist or couldn't be created.
Besides, you’re shifting the goalposts as you go, altering your arguments as it suits your view. The parent comment just disproved your entire point how LLMs cannot combine different fields - and now you just pick a different angle altogether. Maybe take a step back and reconsider your opinions?
There is a faboulous book by Jeff Hawkins "On Intelligence" (2004) that explores this. I think main premise of it still holds true: brain is "just" a highly sophisticated hierarchical tissue whose main job is to extract patterns and make predictions. Fundamentally it doesn't seem very different from what LLMs are.
The more useful question is: can the token prediction model scale to the level of a human intelligence within a reasonable power budget compared to a brain? It's comparing apples to oranges right now but the human brain consumes under 20 watts, a tiny fraction of the TDP of a single A100 GP, and the state of the art isn't even close in performance. We've got a long way to go before we can conclusively answer these questions.
[1] https://en.wikipedia.org/wiki/Universal_approximation_theore...
See, e.g., https://plato.stanford.edu/entries/embodied-cognition/
I won't claim to know which is correct, or even if some other alternative is correct; however, this is not settled at all.
I do think it will some day be possible to simulate all of the embodied cognition above, which may truly render this discussion moot, but that LLMs are not doing that at all.
Seriously. How does a bag of random particles have those things?
It cannot, by our own definition.
Hence the metaphysical problem of 'consciousness' as it relates to our variation of scientific materialism.
I suggest the pragmatic approach is along the lines of what the OP said aka 'sufficiently large neural net will be indistinguishable from human' and that's it. We will see things that we can de facto contemplate as 'curiosity' 'perception' 'meta-cognition' if we want to, especially if we start to develop a more meta understanding of these systems, or not, and that's it.
We'll probably be arguing about 'cognition' long, long after we have variation of AI that kind of seem to be AGI. By many measures we are already kind of there Chat GPT will fool humans probably most of the time and that's that.
I don't think that this is true. LLMs can be modeled with mathematical objects, yes. Functions, sets, numbers, relations, all are very powerful, and can model anything. They can model LLMs, as they can model 'you' for example. Are you a function from the states of the universe to itself?
What is interesting is that it’s answering questions in a meaningful way that is competitive to humans. It replaces about 50% of jobs for each field where all work can be done on computers (like drawing and writing texts) and that number will only grow, until the point where only the most competent people with lots of training and experience can be trusted to do better.
And we don’t need GPT-5 to hit the likit - I think OpenAI said that they don’t have GPT-5 because GPT-4 already hit the limits. The next area for improvement would be multimodal, I think.
As translation is now cheaper, as it can be part automated, there is more demand for translation, whereas this was previously too expensive.
The problem with ChatGPT is that you need to specify exactly what you want. Writing text and drawing is often a creative process where you will only know at the end what you want.
We are true and well within the hype curve at the moment.
Absolutely no risk of terminator here, just ignorance masquerading as such.
All the way back in the 1950s, big money went into it. Decades of hard work, mathematical models of human language, all manner of study, enormous bilingual corpuses of text with phonetic annotation, programmed in general-knowledge databases, fuzzy reasoning algorithms. The amount of work put into it is quite staggering, in hindsight. I remember the cutting edge in the 1990s - SYSTRAN for example, could with some significant human guidance and a limited context domain, translate technical material sometimes usefully.
All of that work has been rendered moot by deep learning. All of it. A machine can, simply with the correct deep learning algorithm and mass exposure to language plus a few bilingual texts, learn an algorithm for translation. It does so automatically, no verb conjugation algorithms, no general knowledge databases, no expert systems with fuzzy reasoning, no parsers, not like a specifically-designed old-school translator had.
And yet these deep-learning systems are vastly superior to the old school architectures, completely supplanting them a couple years after their development.
It is the same story in many other areas. Chess, Go? They learn to play chess and go better than any AI designed specifically to do so. Image classification? Better than the previous 60 years of work on machine vision, and again, accidentally falls out of it. Speech recognition? An algorithm to write a bad poem? Well, we now have an algorithm to find an algorithm to write you that bad poem, if you want it.
That's the thing. These are algorithms to solve very tricky problems, and we didn't have to discover, find, or otherwise create the algorithm. The machine did it for us. I am not sure I'm communicating it well, but to me that's probably the most significant advance since the computer. It was understood - theoretically - that this was possible for a long time; but personally at least, I assumed it would forever require more data and compute than could be realized.
Take GP's "stagnant for decades" with tech and turn it into "stagnant for centuries" with these. When it first started I remember professional Go players talking about just how big a shakeup it was.
Having easy access to the technology is starting to change the way I think about what is possible for computing. One of the skills that has helped me though my career is understanding the scope of what a computer can do or automate. Most of the time it isn't that the thing cannot be done but that the cost associated with it is too high. I recently started solving an issue that would have taken hundreds of hours of human intervention but now I can do it with ten bucks and two hundred lines of python.
When one of the earlier GPT models was released, didn't they straight-up say "none of us expected it to work this well"?
I think it really was closer to "not possible" than "no money" - but more like "no one thought it was possible".
It's an exaggeration to say that GPT-4 can solve all programming problems, but it's easily verifiable that, for example, it's able to solve easy level leetcode problems, and what's impressive on top of that it's able to give an answer to those problems practically before I've finished even reading the problem. That's not hype or a claim by me that you have to take on faith, anybody can copy and paste a problem description and get an answer. It's fair to point out that it's so-so at medium problems, and downright unable to do hard problems, but, well, most humans can't solve hard level problems.
So what's hype, and what's not? It seems to think/reason/understand better than a human child. It's better than some humans with its command of the English language. It has broad knowledge of many subjects beyond what most individual humans know about. It sometimes has poor recall of that knowledge, and some of it is wrong, but it accepts corrections (even if it's unable to commit those corrections back to long-term memory) with far more grace than most humans.
I don't agree with this premise.
'This Time Is Different' because the LLMs have enabled AI to cross a material threshold.
The AGI hype is overdone, but AI promise, unlike just 3 years ago, will be fulfilled across a number of industries.
A lot of things are going to change.
Notably, I wonder if 'AI companies' will doing much of it, rather 'AI will be utilized'.
One thing is that 'help and instructions' are finally going to work for a broad array of applications.
if I found out my lawyer/account/tax advisor/broker had fed my privileged information into ChatGPT I'd be reporting them to their regulator and they'd be struck off
(in their view: removing the risk almost entirely)
legislation, not the market will determine the outcome of the artificial "intelligence" battle
and large traditional industries are rather good at lobbying against their upstart complements
(the fact the RIAA and MPAA even still exist is proof of this point)
https://www.lexisnexis.com/en-us/products/lexis-plus-ai.page
I can imagine this to be a problem. But I can also imagine it to solve a problem. Right now, customer support is often "dehumanised" and Kafkaesque because no single human has the full picture and all the permissions required to resolve a problem.
We're often talking to organisations with "brains" split into a thousand fragments. Having humans in the loop doesn't necessarily help if impersonating a cog is all these humans are allowed to do.
So if an AI can reintegrate this split organisational brain into something resembling human thinking, it might just be an improvement on average.
Making Organizational intelligence and knowledge management convoluted and compartmented is one of if not the core way that managers keep power over employees.
So that’s not a problem that a company wants solved.
However it might get solved via this process IF managers start extracting the information from ICs that is then absorbed by a large corporate model - then get rid of the employee
I mean if you think about it, we’re just headed toward capitalists creating the Borg
Of course they want it solved. They just don't want to pay for it (and neither do customers).
Larger business customers are not exposed to this madness that consumers and small businesses are having to deal with. Large customers get their own account manager who has all the information and all the permissions to fix things. If necessary, they get their technical issues escalated to real specialists.
With these advances in AI, perhaps more of this will become available to everyone. Keeping power over staff (hopefully) won't be a huge problem if "staff" is some AI.
To get where you want, you have to actively change the structure and trajectory of society
Fetishising (forced) human labour is deeply ingrained across all political divides and even in psychology and religion.
The political will to make the necessary changes will only exist once it becomes obvious that we have no other choice.
I actually believe this will be a huge net benefit.
There is a pop-psychology idea that I recall where three bosses were given to a group of individuals. One boss was always nice, one boss was always mean, the final boss was sometimes nice and sometimes mean. The interesting result was that people disliked the sometimes nice boss more than the always mean boss. The conclusion was that an always mean boss is consistent and therefore they can be planned around. It is the uncertainty of the sometimes nice, sometimes mean boss that causes stress.
A good example is driving, controlling a vehicle at speed necessitates constantly making accurate predictions. People get characterized as bad drivers largely due to behaving unpredictably.
The exceptions are either lucky breaks that feel about as good - and random - as finding a $20 bill on the ground and those experiences I've paid a lot of money for because I'm a relatively wealthy software engineer and have chosen to "opt in" to kinder, more competent service with cash. First class travel and premium banking services come to mind.
AI has the power to make everyone's experience with support more pleasant, which isn't to say that it won't have a cost.
Most of these articles border on science fiction and involve a lot of projection. While I am concerned in some ways with AI, particularly in mass surveillance and public manipulation via bots, what else can we do other than be reactive to _real_ AI applications when _deployed_? Then we start to discuss proper policy towards AI.
Until we see service sector employees being replaced the way we saw self-checkouts emerge what is the point of these articles other than some form of PR campaign?
I understand the fatigue is real but this one does neither of above imo and was fairly grounded in reality. The author(who is CTO of a company) did a GPT integration and realised it could potentially replace 20% of jobs at their company. That is something to be spooked about.
> these articles other than some form of PR campaign
A lot of these are, but as an outside observer of Zerodha I can assure you this isn't. The company has spent zero dollars on marketing in last decade.
We can be active? We can get ahead of the curve?
It is such a cliche, but the Wayne Gretzky quote is appropriate: "I skate to where the puck is going to be, not where it has been"
In order to stay ahead we have to predict what is coming. Given the rate of change that AI is likely to cause - by the time you've noticed what is happening you will be so far behind as to be completely ineffective.
Given that the author stuck mostly to their own experiences of the here-and-now, I'm not sure I understand how the rest of your critique (which, overall, I agree with) applies to this content.
> If a single dumb, stochastic, probabilistic, hallucinating, snake oil LLM with a chat UI offered by one organisation can have such a viral, organic, and widespread adoption ... imagine what better, faster, more “intelligent” systems to follow in the wake of what exists today would be capable of doing.
I also remember having the same questions in my last job (in logistics) where potential customers wanted to know how we use AI. This sometimes seemed more important than whether our results would enable them to cut costs. Fascinating.
Then ChatGPT came out and now that's all anybody ever talks about.
Everything points at AI soon becoming far more intelligent than any human. That's not science fiction about the remote future in a hundred years. We are talking about something in the order of ten years. We are actually at the brink of the "technological singularity". It's happening.
It's happening.
Some menial administrative jobs that have virtually no value but aren't yet fully automated will be wiped out by conglomerates who offer their white label AI solutions at scale...but that's it.
Think about Google search. Has it really improved much in the last 20 years? Some HN frequents will swear it has gotten worse.
What about your desktop computer? Sure it's about 10x more powerful than 2005, but has having one really changed your life much since then?
Diminishing marginal returns.
I don't think we will have wait 10 months for that, let alone 10 years.
Of the remainder, most were highly entertaining—-which is a form of productivity in itself.
The hype will die down on LLMs (slowly... all those ML researchers need to justify the sunk-cost of specializing into a very niche, albeit sales-friendly, field).
If it's any consolation, I doubt we'll see progress towards a "technological singularity" until the current crop of career scientists retire into the dirt -- or there is a fundamental change in resource allocation allowing other, more creative types to start experimenting with building out a probabilistic model for:
Human language in -> determination of "what to do"/workflow to run -> run
We're seeing a few startups in this area, but I haven't seen anyone create any useful agent.
It is better than Google search because it puts the results into a realistic sounding piece of text. It is worse than Google search because it does not provide sources or context. Both will steal the information you feed in.
Exactly no progress has been made in understanding or creating real intelligence in any meaningful sense, and we should have no expectation of any progress anytime soon while everyone is so focused on purely syntactical learning environments. Nobody learns anything by feeding them a dictionary.
> Exactly no progress has been made in understanding or creating real intelligence in any meaningful sense
Strong words…
The thing has understanding. It may be alien and imperfect, with strange failure modes and quirks. But it has significant understanding of the real world. At this point, there are countless compelling examples demonstrating this.
The only way the model could so effectively predict the next token of such an immense and diverse collection of texts is by acquiring some level of understanding of the circumstances that gave rise to those texts.
The more I use gpt-4, this clearer this is to me, despite its limitations.
Edit: Removed unnecessary snark.
> All that is can do is associate and surface patterns in the texts - very complex though those patterns may be
One of the more interesting ideas I’ve heard is that text contains embedded world models that are far richer that we previously imagined. We missed this because it is something we take entirely for granted and couldn’t imagine it any other way, like a fish who doesn’t realize they are swimming in “water”. This seems like a very plausible explanation for the performance of gpt-4 and the like.
Yes quite plausible that extraordinarily complex models are embedded in the text corpus, but these can only relate to the texts themselves, as that is all that exists as far as the LLM is concerned. Again, it should not be a great surprise that they correlate to our real world to some extent, but they cannot go beyond those input texts.
Try Bing chat. It provides sources for it's responses.
The input rate of the Claude model is incredible, digesting a book in 20 seconds. Still not quite human level IQ but pretty close.
There is relatively low-hanging fruit for optimizing the models and software and also the hardware for this specific application. We can expect at least another 10 times performance improvement within a couple of years.
And of course as far as IQ that we can measure we can assume it will probably come close to top marks within a few years. But you should scale that based on the incredible speed and other advantages of being digital.
The basic problem with synthetic self-training is that we need some reward function which tells us whether a given synthetic example is good. In case of AlphaGo Zero, this was a synthetic strategy which won the game, or scored a lot. Which can be automatically detected. But how do we automatically recognize that synthetic text has "high quality"?
One case where it might work is proofs in a formal proof language which can be checked automatically via software. So if a language model is tasked to generate synthetic conjecture/proof pairs, it is possible to automatically recognize the correct ones, and use that for self-training data (unsupervised, supervised, reinforcement, I'm not sure), enabling it to recursively create more complex synthetic proofs.
A very similar approach (with some sort of unit tests instead of proofs) is described here in more detail: https://arxiv.org/abs/2207.14502 It was a while that I read it, so my description above is kinda fuzzy. It might involve some adversarial step that I missed.
One problem is to get this process off the ground (bootstrapping), which is difficult, since we need some baseline capability first to create any successful synthetic examples, and there aren't a lot of human created formal proofs which can be used as bootstrapping training data.
Another problem is that, even if it worked, this system would just be good at generating proofs. Maybe there is some amount of transfer to natural language intelligence, but I'm not sure about that.
If you have a different idea for creating a reward signal, I would be interested how it could be done.
> It is just a glorified probabilistic text generator.
> It is just a “stochastic parrot”.
> It “hallucinates”.
> It is not “intelligent” the way humans are.
> It does not actually “think” like humans.
> It will never get human qualities like humour.[4]
> It is snake oil and is no different from any other technology that has come before.
> If it causes job losses, new jobs will emerge, like with all previous technological revolutions.
The OP cites all of these as 'moving goalposts' but I'm not sure how, given that they're all true? Like all of the hype I see around LLMs is all the same hype I saw around from Bitcoin. It's by people that haven't used it in a professional environment, with no clue about the actual flaws, limitations or issues. And companies are intent on selling you a solution in search of a problem because they want to create aa product hype cycle.
It feels different because you've bitten onto the bait. Personally given the code output and so forth I've seen from LLMs I'd highly doubt the experience of any software engineer that uses it in their day to day life.
Samsung had to tell it's engineers to stop using chatgpt because it was leaking secrets.
I asked the LLM the following prompt: Can you explain why the joke "Where's the joke?" is funny?
The response: "The humor in the joke "Where's the joke?" is derived from its meta or self-referential nature. The joke plays with the expectation of a traditional joke setup and punchline structure. Typically, when someone tells a joke, there is an expected punchline that delivers the humor. However, in this case, the punchline is intentionally missing, and instead, the question "Where's the joke?" is posed."
All it's doing is the same thing when you ask it to summarize a piece of data. This isn't a joke or meant to be a joke, but by telling it to explain it as a joke/meme will cause the output to differ in and regurgitate a bunch of nonsense about humor. This isn't understanding humor.
Consider that your BS barometer needs a good recalibration, I'll just leave it at that.
this has been said with the advent of computerization, the entire internet and yet here's a simple thing to look for: If this is true we will see both skyrocketing unemployment numbers and productivity growth at the same time. Until that is happening you can rest easy. One pretty trivial reason is that deep technology adoption that actually replaces professions wholesale is really slow. A lot of companies still run on ancient Excel versions and fax machines. The average employer is a small business of 15 people where nobody has an idea what Slack is.
Also one of the more sinister and clever strategies in recent times is tech companies framing their crazed hiring-spree corrections (looking at FAANG) as revolutionary AI developments
This was just written to say that you care. You know well that when push comes to shove, they'll be laid off.
Until then I expect the denial and explainers (It's just another technology! Like phones!) will have the floor.
The discussion is healthy because we'd like to anticipate changes before they swamp us. It's important to speculate. The denigration of speculation is less helpful.
I'm happy to watch from the sidelines. But I sure pay more attention to the folks trying to imagine the future, than the folks claiming nothing will change.
Is this a given? Sure it will continue to improve but by how much? 200% or 2%?
That’s the question. Anyone could’ve measured my height when I was 14 and predicted that I would be 35 feet tall today. But human bodies have limits.
What are the limits of LLMs? It’s somewhere between ChatGPT-4 and the “digital god” that I’ve heard some proclaim is right around the corner.
The boring but very possible answer is ChatGPT-4 might be near the pinnacle.
Given that we have increased computing performance by more than a dozen orders of magnitude already and we are now optimizing for a very specific application, it is almost a given that we will get at least one more order of magnitude performance improvement in the near term.
There are also new approaches like crossbar arrays of memristors that, if we can make them work, promises at least two more orders of magnitude performance increases. Quite possibly more.
Claude can read a whole book and answer a question about it in 20 seconds. Not quite at GPT-4 level but no reason to believe that isn't possible in the near future.
Computing performance and efficiency have consistently increased by orders of magnitude as we leverage previous generations of systems to make improvements and when necessary invent new paradigms.
1. Deny there's hype
2. Start engaging in hype
LLMs are a just a tool. Yes, "just". They're good at certain tasks. Mainly ones that have a lot of training data available and don't require precise output. Images, writing, boilerplate code, etc. For any other task their usefulness falls of sharply.
> What it and its contemporaries have set loose in such a short period of time, and what is to follow on their heels tomorrow, is the real crux of the matter.
What is the idea that LLMs will get significantly better based on? That the technology finally got good enough to be useful is great, but past performance is not a predictor of future performance. Predictions need to be based on the nature of the technology in question.
1. There's only so much usable training data available out there and I suspect most of it is already being used. Once a programming LLM is trained on the code on Github it has basically seen it all.
2. They are incapable of learning anything. Not even trivial concepts. Therefore there's no real way for them to build up their understanding step-by-step like a human would.
1a) say it's incapable of thinking/reasoning/learning/understanding/etc
1b) argue about it's ability to think/reason/learn/understand/etc
In this case:
> They are incapable of learning anything
ChatGPT seems to have learned about plenty of things before 2021. If you ask it about concepts like copyright or gravity or communism, or ask it to "explain quantum computing in simple terms", as the landing page suggests, it displays a fairly good understanding of these concepts.
Human memory is divided into short-term and long-term (along with a bunch more concepts). It's true that ChatGPT lacks the ability to commit information from its short term memory (ie "chats") into it's long-term memory via inference, but GPT-4-Mawr, fine-tuned or re-trained on text including a new concept you invented would display some level of understanding (depending on how much text you had to fine tune it on).
On a per-chat basis, ChatGPT is able to, up to the limits of the context window, know things, and it's entirely fair to point out that a) the context window is fairly limited (8000 tokens, or 32k paid) and b) beyond that, it forgets what it was told. Some people are described as having the memory of a goldfish, and that description applies to ChatGPT. If you can fit your concept into the context window, ChatGPT-4 is able to do some reasoning about it. Thus, it's able to learn and understand new things, but forgets them just as quickly.
Humans sleep every day or so, and during that time the brain gets to do things that it isn't able to do while awake, and it's been suggested that sleep is important for integrating the day's activities and thus learning of advanced concepts. Assuming that's true, then we're just looking at different timescales.
If ChatGPT "sleeps" every couple of months, and learns everything it was told between last time it was trained, it's just operating on different timescales than humans do.
Prompt Engineering. The highest paid will be the ones who are able to extract the most value out of LLMs with the least effort.
But the prompt I’m using for GPT-3.5 was generated by GPT-4.
As a non-business person, I imagine that folks would typically want some correctness guarantees, which seem near-impossible with today's LLMs. What kind of business use-cases do they enable? Probably a failure of imagination on my part, but I can't think of many tasks that can be done consistently well by AI that can't also be done with simpler software.
Yeah, haikus are 17 japanese mora, and so if it doesn't come from the mora region of japan it's not a true haiku, it's just a short poem.
> I imagine that folks would typically want some correctness guarantees
Us programmers really value correctness, but the real world doesn't actually care about it in quite a few cases. In the real world, most data comes from humans, and humans make mistakes all the time, so basically any data might have some amount of error. There's a high tolerance for a few minor mistakes here and there.
In the real world, tasks are often given to interns, and they make small mistakes all the time, but it mostly ends up being okay. If AI manages to do the same, well, that's a huge swath of tasks it can do.
> I can't think of many tasks that can be done consistently well by AI that can't also be done with simpler software
Here's the thing though: "simpler software" probably isn't actually cheaper, available, or usable. It's probably not simpler to operate since, you know, english is simpler than any DSL us programmers have come up with until now.
This sort of stuff is fuzzy, even without AI. In many ways it’s about exploring the knowledge space rather than seeking specific answers. None of this requires correctness because there are no correct answers to begin with.
Even in computing, correctness is not always important. I recently asked GPT to produce a bunch of names from the Austin Powers universe in order to populate a demo database. I asked for them in JSON format and in 20 seconds I had what I needed. There is no wrong answer for that, either.
Interestingly (for me anyway) we got GPT to write the initial prompts for Midjourney.
I've dabbled in comms work, and it's very useful for that. One thing that I despise doing is writing copy and LLMs are really good for that - writing dozens of social media posts is tedious and LLMs can rephrase things so that they sound less repetitive. I'm a good writer but I'm an amazing editor so having something to start with also solves the 'staring at a blank page/doc' problem.
Likewise, generating polite answers to stupid questions. Instead of worrying 'how do I tell an ego-driven exec/politician/candidate that what they want is stupid or not possible?' I can have an LLM spit out the stupid corporate drivel the C-Suite all want from their servants. It's also good for dumbing things down but not making it sound like you're dumbing things down and therefore is good for rewriting information for different audiences.
You don't (right now) use LLMs for generating facts but they have a great deal of utility in massaging language.
Technology has played a role in similar transitions in the past. Mechanized agricultural production mosly eliminated the need for a large agricultural workforce for producing staple crops, a workforce that historically was organized and controlled via systems like serfdom and slavery, which were mostly eliminated as mechanization spread. Domestic household technology (washing machines, vacuums, electricity, etc) ended the need for a household to employ domestic servants and facilitated the entrance of women into the workforce, with associated loosening of societal gender role expectations.
Machine learning systems might play a similar role in enabling automation of the industrial factory system, creating a system in which a few workers manage complex workflows with all basic tasks done by highly adaptable robots. Generative AI appears poised to eliminate a similar amount of basic tedious work in the white-collar world of documentation creation, legal research, all secretarial-type tasks, boilerplate code generation, etc.
The real question in societal terms will be whether or not the fruits of all this labor elimination will end up controlled by a tiny minority of the human population. If so, that could plausibly lead to societal pressure to completely restructure the investment capital system of wealth accumulation and concentration, i.e. AI might cause socialism to become more popular.
However, the notion that AI systems will be making the fundamental choices (going to war, creating a national constitution, writing and voting on legislation, forming a new business entity, etc.) seems farfetched at present.
This is a pretty disingenuous question to ask, as the whole point of an innovation-driven market economy is that one cannot predict the future.
It’s a question that we know for a fact no one can answer in general, so why should we consider it any kind of useful criticism of AI specifically?