It’s been a month and a half since my last thread, and just today we passed 5,000 subscribers on The Algorithmic Bridge!!! We’re growing reaaally fast and I’m super thankful for your support, engagement, and appreciation!!
As always, this week has been full of AI news and events. But, even with NeurIPS 2022 ongoing, ChatGPT has clearly overshadowed everything else. I covered the news on OpenAI’s latest model here.
I won’t do a WYMHM today, but don’t worry, next week I’ll include what I’m leaving out this week. I’ll take this opportunity—and the fact that we’ve crossed the 5K subs mark!—to do another AMA session and welcome all new subscribers.
As always, you can ask me about AI (current trends, predictions, companies, models, etc.), writing, Substack, my background, future plans, thoughts on any topic… whatever! Of course, you can also ask me about ChatGPT and GPT-4.
So, bring in the questions. I’ll try to answer all of them!
I think a combination of both is possible. So far, those companies profiting from OpenAI's tech have exploited the fact that OA didn't optimize for commercialization. Jasper, Copy.ai, etc. offered a service OA didn't.
I'm not sure what are OA's plans for the future. I'd guess they'll set up ChatGPT on the API and charge for it like they do with InstructGPT and GPT-3. If they do this, those other companies could simply update their services to improve on top of ChatGPT (ChatGPT isn't near perfect, so it's possible--although prompt engineering is less critical now, it's still important).
If this hypothesis is true, Jasper & company will only lose market share in the interval between OA's release of a new, clearly better model, and the API set up + update on the side of the companies. Right now we're in the middle of that interval so Jasper and the others are probably seen a decrease in users/usage, but they'll update as soon as ChatgGPT is available through the API.
The question now is whether the improvement those companies can get on top of ChatGPT is worth enough to pay what they charge. We'll see.
How do you think OpenAI and others will address the challenges of AI hallucinations? It's really difficult to evaluate the correctness of an answer in an area with which you're not familiar.
Also, on a related note, why is ChatGPT able to acknowledge and correct its own mistakes? The other day I asked it a particular question about the size of packages that fall under a certain USPS shipping class, and one thing it told me was that the maximum weight was 13oz. I happen to know this is wrong, and it's 15oz. I said this to it, and it immediately agreed and stopped making the mistake. This behavior just seems bizarre...
I think their approach is quite clear: They'll trust the exponential, as Sam Altman tweeted recently. On top of that, they'll keep improving the RLHF (reinforcement learning through human feedback) approach which is the main training difference with GPT-3.5.
Will that work? My bet is not sufficiently. You have already seen the problem: ChatGPT is better aligned with humans with the RLHF approach--which I think is the reason it seems to be able to correct itself--but that doesn't mean it'll answer correctly the first time. The way it fails is weird, and it'll keep being weird, even alien (tomorrow's article is about this).
As you say, this is super problematic for people who will rely on this as a source of information (we search for what we don't know). No one will ask twice if the first answer seems plausible enough. 13oz vs 15oz is clearly plausible, if it said 2000oz people would instantly doubt--which also underscores why the closer the system is to the wright answer without being able to assess that it is, in fact, the correct one, the more problematic it is (it could confuse even experts).
This is also the main reason why ChatGPT can't replace search engines. The technology isn't equipped to do the task.
Weekly writes, " It's really difficult to evaluate the correctness of an answer in an area with which you're not familiar."
Interesting point, thanks, I hadn't thought of that. When obtaining info from a human we can at least evaluate their experience. How does one evaluate the quality of an answer delivered by an AI? If we have to compare the AI answer to human answers to judge the quality of an AI answer, maybe we should just do the human research, and skip the AI part??
I'd actually say this problem exists also when obtaining info from a human. Can we always evaluate the experience of the human source? Do we always do it? Do we always know the hidden intentions? It gets tricky on the edges.
However, I agree that these types of AI systems add, on top of that, another filter of uncertainty: they can't evaluate or assess the correctness of what they say. It's even possible to make ChatGPT correct itself from the right answer to a wrong one if you try.
The reason why companies keep trying to achieve this is because a good-enough AI that could do search and assess its own answers *maybe* has to first go through the current step of development. That's their bet: To get there we have to first do this (I'm not sure I agree though, as I wrote here: https://thealgorithmicbridge.substack.com/p/ais-destiny-was-never-intelligence).
To verify the assertions of the AI, we could consider a Wikipedia-like system, where humans - certified competent or not - would rectify the errors as they occur.
I wrote that to ChatGPT and here is his answer:
"While a Wikipedia-like system could be a good starting point for verifying the assertions of AI, it would not be a foolproof solution. One potential issue is that the humans who are responsible for rectifying errors may not always be fully competent or may not have the most up-to-date information. Additionally, relying on a human-based verification system could be slow and labor-intensive, especially if the AI is making a large number of assertions."
Thanks for your input Pascal. Hmm... My question remains.
If we have to reference humans to determine the quality of ChatGPT output, because we know that ChatGPT sometimes outputs nonsense, why not just go to the humans first?
If we were to count on humans to rectify errors in ChatGPT output, similar to Wikipedia, how would ChatGPT know the qualifications of it's human editors? If 75% of human editors assert God exists, and 25% assert God doesn't exist, would ChatGPT tally up the percentages and conclude the God exists claims were the correct answer? Would it report there is disagreement on the matter? Would it be able to determine that actually nobody knows that answer, and it's just a pile of competing claims, none of which can be proven?
Well, even Wikipedia cannot answer this religious question, or answer like you suggested: there are many opinions about the subject matter. But the cross-exam of Wikipedia editors allows for correct informations to prevail, on the vast majority of subjects where there are no competitive answers.
Ideally you could get it to provide some kind of source for the answer.
I do think there's still plenty of value in places where you have some knowledge/domain expertise, but there are details that you have to look up when you need them. If I ask about battles in World War II, and it tells me they happened in 2010, I know it's wrong.
On the other hand, I had my wife ask it some questions about her work. She's a scientist, so she could confirm that its answers were reasonable, where for me it was gibberish. So at least a valuable reference tool for places where you have a good foundation of knowledge.
I'm curious about how these AI apps evaluate the credibility of their sources. For example, the Google search engine will do things like count the number of incoming links to a site, and the number of links to those linking etc, to formulate it's search engine rankings.
So my crackpot theory on XYZ with only 2 incoming links will rank way down the listings far under The Institute Of XYZ, which is a widely recognized expert on the subject, with many incoming links from many credible sites.
If understand correctly (I may not) these AI apps crawl the web just like the Googlebot to assemble a vast database of information. But how do they decide which information is the most valuable, useful, credible etc? Same way the Googlebot does?
Ok, so Google can obviously lend it's search engine techniques to it's AI tools. But what about those developers who don't have 20 years of search engine experience?
"I'm curious about how these AI apps evaluate the credibility of their sources"
They don't.
They're not trained to detect right from wrong, but to write like a human syntactically and semantically. These systems are getting better at sounding human much faster than better at being right.
"these AI apps crawl the web just like the Googlebot to assemble a vast database of information." Not really. Deep learning systems like ChatGPT are fed with data from the internet (which can be in places Google uses to give search results) but the way they integrate that data is vastly different from how a search engine does it--even if both use AI algorithms in some form.
That said, once you're able to train this kind of AI on your own data, there will be lots of uses for it where you don't have to question the reliability of it. If you hook it to all your legal docs, for example, it should be able to answer questions about the legal obligations you have based on contracts you've entered, etc.
It depends. A search engine would be reliable. By its nature, a language model wouldn't. Also, you can't achieve a GPT-3 level of AI training it *from scratch* only on your data. And, if you talk about fine-tuning, then the problems are there already when you start to feed it with your data.
I think there are more and more solutions arriving in the market to check the veracity, contexual-ness and correctness of input data into the actual models, especially those leaning towards continuous learning. Do you think ChatGPT is deploying some/any of it? If not, is this a viable business opportunity?
I'm curious to learn how one trains GPTChat to be knowledgable about a specific domain of knowledge for personal (or business) use. Generally speaking I believe GPTChat can be taught using Reinforcement Learning via Human Feedback - with prompts. That's all very exciting but there are always tons of little specific details for specialized use of knowledge. So how does one do this specifically? I realize there is no single answer yet or single tool (I'm assuming here). I just want to understand the process better and see which groups or vendors are offering this kind of service.
Well, ChatGPT was built on top of GPT-3.5 with RLHF, as you know. It's in the research phase so there's no commercial application yet (OpenAI will likely use the feedback they've gotten these days to improve the system and set up an API). Once that happens, you'll start to see *a lot* of companies building on top of it. You could also use the API to fine-tune the model with the data you want and to do the task you want.
Just a clarification, RLHF isn't just prompting. Is a more complex multi-step process they explain in the blog post: https://openai.com/blog/chatgpt/
Newbie here and asking perhaps a very basic question.
Is there any open source best-in-business speech-to-text AI in the wild? DALL-E has set the gold standard for text-to-image, but beyond Siri, Alexa or Google Voice, which are all proprietary, I am not aware of any open source tools for speech-to-text.
Anyway, this is just the one I know of (I don't write much about automatic speech recognition). But given that it was released a couple of months ago, it's probably quite good.
I googled around, and came across DeepGram's offerings in 40 (!!!) languages. How would you rate the opensource offerings in this domain versus paid developer tools like these? I am looking to create the VERY first 3D model of H/w and write the first line of code for a solution. Would love to have one less thing to do. :)
Sorry, I know you said you don't write much about this subject. In a broader sense, my question is about the efficacy of opensource solutions versus paid ones. What are the pros and cons for startups?
I can't answer this, other than to say what you may already know. Speech to text tools have been notoriously problematic for years. I'm not up on the latest developments, but my default assumption would be to be skeptical until proven otherwise.
Hi Anant, I've talked about this a few times in my Sunday paid column "what You May Have Missed." I try to cover almost everything that's happening in AI, but because I can only do two high-quality long-form pieces per week, I cover everything else in WYMHM. Consider subscribing, I think it complements very well the Tuesday-Friday combo!
Anyway, I'll try to answer your question here: Text-to-video is still in research. Google (Phenaki, Imagen Video and a combination of both) and Meta (Make-A-Video) are clearly ahead of everyone else. Stability.ai has some plans for this in the future. And there are also unofficial implementations on top of stable diffusion.
There's no commercial app yet for this and Emad Mostaque said recently that it'll probably take a couple of years before this tech gets to the level SD is today.
Dec 5, 2022·edited Dec 5, 2022Liked by Alberto Romero
I am interested in knowing what specific applications society is finding for generative AI and what the second- and third- order consequences of those might be. It is easy to imagine that every advertising agency in the world is going to want access to the technology, but harder to see the staffing and payroll implications. It is hard for me to see how the porn sector as we know it today will survive. The examples are legion.
"The examples are legion." Agreed. I guess we'll find out together! This is merely starting and the second- and third-order consequences may be unpredictable even for those at the very vanguard.
Of course the consequences "may be unpredictable". But it still makes sense to try to think about might happen. It is always a good idea to open your eyes even you can't focus them immediately.
If you heard about Bittensor, what do you think about an incentivized decentralized AI Hivemind? The platform enables users to train, test, and deploy AI models on the blockchain, and rewards contributors with tokens based on the performance and adoption of their models. We can co-own, co-run, and thus align the intelligence layer while creating a censorship resistant AI.
Sounds like a good idea, although I don't how we could even start to do it (also, blockchain always reminds me of crypto, which I'm not really fond of given the latest events...)
I think Bittensor something you should keep one eye open for and I’m sure you’ll here about it in 2023. It’s 500b parameters and growing. Text only for now and then will have image and audio next, aiming to have multi modals. But anyways, I LOVE your content and always sharing with many of my friends. I’m loving forward to learning more from you as this seems to be happening so fast. You make it easy to keep up with putting it all in perspective as well! Thank you! Where can I donate?
Near like in 1-2 years? I'd say yes, undoubtedly (unless something highly unexpected happens). Much more than that? I really hope not. Maybe a part of a whole, but they have critical limitations that would give us a lot of problems.
Great question! I'd say writing well and coding well are high at the top, for sure--even assuming AI will improve and cover a portion of those jobs (or part of a single person's job).
Writing, as you know better than I do, is transversally important in life, even if you don't earn a living as a writer. Coding is super useful now and, even if AI catches up, coders (good ones in particular) will be the ones to exploit it the most in the future.
I assumed you were referring to hard skills. In the soft skills area, I don't have any doubt: for me a good emotional education is unbeatable.
Thank you for your answer! I ask because I have a 15 yr old who thinks school is a waste of time and wants to spend all their time on the computer making digital art- her particular passion, and skill, she designs characters, but now I fear AI will render her talents useless... it’s daunting...
You're in a very special position because you're aware of what's happening in AI and see the threats more than most people, but that can also be paralyzing in that you will worry a lot more than someone who is 100% unaware.
I don't think artists and writers will be erased by AI (I don't think you do either), but I'm quite sure demand will be affected. And those are professions (writer/artist) where it's already quite hard to earn a living wage.
As an example of how coding and artist skills synergize, I know a few people in the new AI art scene that are both artists and coders. Those, in contrast to "pure artists", don't see AI as a threat nearly as much and are learning to leverage it.
And, for what is worth, I started writing years after I suspected what AI would be capable of (I'm not worried because I write about AI, which is a very safe place).
Anyway, the takeaway is that being aware is more important than trying to be safe--instead of letting the wave crush you, you could choose to ride it with the tools at your disposal without necessarily changing careers.
The only safe spot from AI is the inherently human component of what we do. This component is present in almost everything, but it's not equally important everywhere. I, for instance, try to make that part very large (I write about what I think and believe, not just content for the sake of content). The same thing applies to visual artists.
Thank you. It’s true. I think the opportunities will be for “generalists” people who can mashup and synthesize different disciplines also. People good at things like management, business, sales/marketing will likely also continue doing well as physical labor trades like plumbers and AC techs.
It’s just hard to predict. Like who would have thought art would be automated before law or accounting...
Though it sort of makes me think that the least economically powerful in society got automated first, the artists. 😢
On the plus side, many of the greatest talents felt just like your daughter, and made it work. I went to the same high school as Duane Allman ( founder Allman Brothers Band) and he couldn't even be bothered to finish high school. Other greats like Steve Jobs and Bill Gates and more dropped out of college.
That said, I agree your concerns are very valid. Daunting indeed. This is the price we pay for an ever accelerating pace of change, nobody can count on anything for long. Been there, done that.
True true. High school does feel super dated and I often wonder if she would be better off learning in some other way. The public curriculum can’t keep up with technology.
I’d almost rather her just learn the old school classics at school and tech on her own than be in this in-between state of schools trying to be modern but failing... because currently it’s like they don’t learn cutting edge stuff or important foundational works. And of course all the kids are using Google, YouTube, and TikTok for every homework assignment anyway.
It’s hard to inspire creative thinking and problem solving when kids can find almost every answer to every question instantly.
The smart move these days seems to be to go in to fields which can't be automated. Like say, nursing, or plumbing.
If a kid can make it through four years of nursing school, they are set for life, and will never have to worry about money ever again. A good friend of mine did this when we both were young, and now she has a $100k salary, a great state retirement plan, and a big enough Social Security check that she could probably live on that alone. And she's still working, cause she loves her job.
We recently had a plumber come to our house. He was here an hour and charged us $200. Work an hour, and you're done for day. Own your own company, be your own boss, build it if you want, or spend most of your time surfing.
The main thing is, find a career that the machines can't do.
Congratulations on hitting 5,000 Alberto! I expect much success for you. I've been in and around web and email publishing for 27 years and as best I can tell, you're doing everything right, as your growing fan base would seem to prove. I find your enthusiasm and professionalism inspiring. It appears I'll be starting my own substack, and I'll be looking to you as a role model of how to go about it.
Ask you a question. Cool feature. Hmm...
QUESTION: What do you see to be the most likely outcome of humanity obtaining ever more, ever larger powers, at an ever accelerating rate?
This seems to be the path we've been on for at least a century, and AI appears likely to further accelerate the process, just as computers, and then the Internet have. Do you see any limit to this progression?
Take it from there in any direction that interests you. Thanks!
Thanks Phil, really appreciate your words and your super-high engagement (still a lot of comments to answer on my side, I'll get to those too!!)
Interesting question, really hard to answer accurately!
I'll start from the last question. I read somewhere a few months ago that (I'm paraphrasing) "nothing in nature follows an exponential." I don't have the knowledge to evaluate that sentence, but it feels much more plausible short-mid term than what Kurzweil & company foresee (the singularity and all that). There are limited resources within reach right now, and there are *a lot* of variables beyond technology/technological progress (e.g. political, cultural, economic, demographic, energetic, climatic, etc.) that could hinder that outcome.
That said, if I take your question literally the answer is an easy yes: everything has limits except maybe the universe (and, if we agree with Einstein, human stupidity). If you mean a soft version of the word "limit" then I'd say it really depends. For instance, if we managed to somehow exploit the sun as a source of energy or find a way to create an efficient fusion energy source, many things would be different.
Now, a few "maybes" on the distant future. The answer to all these questions is "maybe" for me (and I don't dare to be more precise than that!): Could we become multi-planetary? Could we decide to stay on Earth? Could we decide to upload our minds (if that's physically possible) and live in a peaceful and pleasurable matrix? Could we drop tech progress eventually and go back to a more ecological relationship with our planet? Could we merge that with high-tech (solarpunk style)?
I don't think anyone has any idea which of those is more probable.
Thanks for your thoughtful response Alberto. Please don't feel obligated to reply to every comment I make, as I type a lot, probably too much. Maybe way too much. :-) Just engage whatever interests you, as your time permits, I'm cool with that.
Thanks in part to inspiration provided by you, I've started my own substack. My first post addresses your comments above to an extent, as it proposes that the primary limiting factor to an unlimited knowledge explosion is us, human maturity, or rather the lack thereof.
Thanks!! My first blog post is from mid-June this year, so 5 and a half months. I came here with around 700 subscribers from Medium and it started growing right away. There have been a few articles that have driven a lot of traffic (and subscribers), but it's mostly consistence + high-quality
No worries - thanks for letting me know. I see no place to take a real problem and be at least pushed into right direction - doable or not, what to look at that can solve it. You do understand the problem I tried to solve though, right?
Yep, I got it. I'd recommend asking on programming sites (e.g. you asked on Reddit, which is a good place). I know how to code (basic knowledge), but I'm quite far now from that world, so I can't really help you there.
Can you make a dedicated one post covering everything exciting and different from GPT-4? Heard it’s gonna be extraordinary, but no idea how. Tell us something which helps clear the excitement.
My advice is: Don't give any rumor, hype, excited claim, etc. more value than it has. Almost no one knows anything about GPT-4 and those who do are under NDA. Yet, given the fuzz ChatGPT has created (which is just an optimized GPT-3), I can't imagine what will happen when GPT-4 is out...
Do you think OpenAI will become more of a competitor than a supplier for the many creative utilities/app companies that exist in the market today?
I think a combination of both is possible. So far, those companies profiting from OpenAI's tech have exploited the fact that OA didn't optimize for commercialization. Jasper, Copy.ai, etc. offered a service OA didn't.
I'm not sure what are OA's plans for the future. I'd guess they'll set up ChatGPT on the API and charge for it like they do with InstructGPT and GPT-3. If they do this, those other companies could simply update their services to improve on top of ChatGPT (ChatGPT isn't near perfect, so it's possible--although prompt engineering is less critical now, it's still important).
If this hypothesis is true, Jasper & company will only lose market share in the interval between OA's release of a new, clearly better model, and the API set up + update on the side of the companies. Right now we're in the middle of that interval so Jasper and the others are probably seen a decrease in users/usage, but they'll update as soon as ChatgGPT is available through the API.
The question now is whether the improvement those companies can get on top of ChatGPT is worth enough to pay what they charge. We'll see.
Great question, I was thinking the same, I’ve been a customer of Jasper, writesonic and chatgpt has made them less valuable so far
Yes, and I think many others agree with you, Michael!
How do you think OpenAI and others will address the challenges of AI hallucinations? It's really difficult to evaluate the correctness of an answer in an area with which you're not familiar.
Also, on a related note, why is ChatGPT able to acknowledge and correct its own mistakes? The other day I asked it a particular question about the size of packages that fall under a certain USPS shipping class, and one thing it told me was that the maximum weight was 13oz. I happen to know this is wrong, and it's 15oz. I said this to it, and it immediately agreed and stopped making the mistake. This behavior just seems bizarre...
I think their approach is quite clear: They'll trust the exponential, as Sam Altman tweeted recently. On top of that, they'll keep improving the RLHF (reinforcement learning through human feedback) approach which is the main training difference with GPT-3.5.
Will that work? My bet is not sufficiently. You have already seen the problem: ChatGPT is better aligned with humans with the RLHF approach--which I think is the reason it seems to be able to correct itself--but that doesn't mean it'll answer correctly the first time. The way it fails is weird, and it'll keep being weird, even alien (tomorrow's article is about this).
As you say, this is super problematic for people who will rely on this as a source of information (we search for what we don't know). No one will ask twice if the first answer seems plausible enough. 13oz vs 15oz is clearly plausible, if it said 2000oz people would instantly doubt--which also underscores why the closer the system is to the wright answer without being able to assess that it is, in fact, the correct one, the more problematic it is (it could confuse even experts).
This is also the main reason why ChatGPT can't replace search engines. The technology isn't equipped to do the task.
Weekly writes, " It's really difficult to evaluate the correctness of an answer in an area with which you're not familiar."
Interesting point, thanks, I hadn't thought of that. When obtaining info from a human we can at least evaluate their experience. How does one evaluate the quality of an answer delivered by an AI? If we have to compare the AI answer to human answers to judge the quality of an AI answer, maybe we should just do the human research, and skip the AI part??
I'd actually say this problem exists also when obtaining info from a human. Can we always evaluate the experience of the human source? Do we always do it? Do we always know the hidden intentions? It gets tricky on the edges.
However, I agree that these types of AI systems add, on top of that, another filter of uncertainty: they can't evaluate or assess the correctness of what they say. It's even possible to make ChatGPT correct itself from the right answer to a wrong one if you try.
The reason why companies keep trying to achieve this is because a good-enough AI that could do search and assess its own answers *maybe* has to first go through the current step of development. That's their bet: To get there we have to first do this (I'm not sure I agree though, as I wrote here: https://thealgorithmicbridge.substack.com/p/ais-destiny-was-never-intelligence).
To verify the assertions of the AI, we could consider a Wikipedia-like system, where humans - certified competent or not - would rectify the errors as they occur.
I wrote that to ChatGPT and here is his answer:
"While a Wikipedia-like system could be a good starting point for verifying the assertions of AI, it would not be a foolproof solution. One potential issue is that the humans who are responsible for rectifying errors may not always be fully competent or may not have the most up-to-date information. Additionally, relying on a human-based verification system could be slow and labor-intensive, especially if the AI is making a large number of assertions."
Thanks for your input Pascal. Hmm... My question remains.
If we have to reference humans to determine the quality of ChatGPT output, because we know that ChatGPT sometimes outputs nonsense, why not just go to the humans first?
If we were to count on humans to rectify errors in ChatGPT output, similar to Wikipedia, how would ChatGPT know the qualifications of it's human editors? If 75% of human editors assert God exists, and 25% assert God doesn't exist, would ChatGPT tally up the percentages and conclude the God exists claims were the correct answer? Would it report there is disagreement on the matter? Would it be able to determine that actually nobody knows that answer, and it's just a pile of competing claims, none of which can be proven?
Well, even Wikipedia cannot answer this religious question, or answer like you suggested: there are many opinions about the subject matter. But the cross-exam of Wikipedia editors allows for correct informations to prevail, on the vast majority of subjects where there are no competitive answers.
Ideally you could get it to provide some kind of source for the answer.
I do think there's still plenty of value in places where you have some knowledge/domain expertise, but there are details that you have to look up when you need them. If I ask about battles in World War II, and it tells me they happened in 2010, I know it's wrong.
On the other hand, I had my wife ask it some questions about her work. She's a scientist, so she could confirm that its answers were reasonable, where for me it was gibberish. So at least a valuable reference tool for places where you have a good foundation of knowledge.
I'm curious about how these AI apps evaluate the credibility of their sources. For example, the Google search engine will do things like count the number of incoming links to a site, and the number of links to those linking etc, to formulate it's search engine rankings.
So my crackpot theory on XYZ with only 2 incoming links will rank way down the listings far under The Institute Of XYZ, which is a widely recognized expert on the subject, with many incoming links from many credible sites.
If understand correctly (I may not) these AI apps crawl the web just like the Googlebot to assemble a vast database of information. But how do they decide which information is the most valuable, useful, credible etc? Same way the Googlebot does?
Ok, so Google can obviously lend it's search engine techniques to it's AI tools. But what about those developers who don't have 20 years of search engine experience?
"I'm curious about how these AI apps evaluate the credibility of their sources"
They don't.
They're not trained to detect right from wrong, but to write like a human syntactically and semantically. These systems are getting better at sounding human much faster than better at being right.
"these AI apps crawl the web just like the Googlebot to assemble a vast database of information." Not really. Deep learning systems like ChatGPT are fed with data from the internet (which can be in places Google uses to give search results) but the way they integrate that data is vastly different from how a search engine does it--even if both use AI algorithms in some form.
Great question, and I have no idea of the answer.
That said, once you're able to train this kind of AI on your own data, there will be lots of uses for it where you don't have to question the reliability of it. If you hook it to all your legal docs, for example, it should be able to answer questions about the legal obligations you have based on contracts you've entered, etc.
It depends. A search engine would be reliable. By its nature, a language model wouldn't. Also, you can't achieve a GPT-3 level of AI training it *from scratch* only on your data. And, if you talk about fine-tuning, then the problems are there already when you start to feed it with your data.
I think there are more and more solutions arriving in the market to check the veracity, contexual-ness and correctness of input data into the actual models, especially those leaning towards continuous learning. Do you think ChatGPT is deploying some/any of it? If not, is this a viable business opportunity?
I'm curious to learn how one trains GPTChat to be knowledgable about a specific domain of knowledge for personal (or business) use. Generally speaking I believe GPTChat can be taught using Reinforcement Learning via Human Feedback - with prompts. That's all very exciting but there are always tons of little specific details for specialized use of knowledge. So how does one do this specifically? I realize there is no single answer yet or single tool (I'm assuming here). I just want to understand the process better and see which groups or vendors are offering this kind of service.
Well, ChatGPT was built on top of GPT-3.5 with RLHF, as you know. It's in the research phase so there's no commercial application yet (OpenAI will likely use the feedback they've gotten these days to improve the system and set up an API). Once that happens, you'll start to see *a lot* of companies building on top of it. You could also use the API to fine-tune the model with the data you want and to do the task you want.
Just a clarification, RLHF isn't just prompting. Is a more complex multi-step process they explain in the blog post: https://openai.com/blog/chatgpt/
Newbie here and asking perhaps a very basic question.
Is there any open source best-in-business speech-to-text AI in the wild? DALL-E has set the gold standard for text-to-image, but beyond Siri, Alexa or Google Voice, which are all proprietary, I am not aware of any open source tools for speech-to-text.
Whisper is open source and it's quite good (although I'm not sure if it's the state of the art, it's better than those you mention). I wrote about it here: https://thealgorithmicbridge.substack.com/p/8-features-make-openais-whisper-the and here's OpenAI's blog post: https://openai.com/blog/whisper/
Anyway, this is just the one I know of (I don't write much about automatic speech recognition). But given that it was released a couple of months ago, it's probably quite good.
Thank you! I will give it a spin!
I googled around, and came across DeepGram's offerings in 40 (!!!) languages. How would you rate the opensource offerings in this domain versus paid developer tools like these? I am looking to create the VERY first 3D model of H/w and write the first line of code for a solution. Would love to have one less thing to do. :)
Sorry, I know you said you don't write much about this subject. In a broader sense, my question is about the efficacy of opensource solutions versus paid ones. What are the pros and cons for startups?
I can't answer this, other than to say what you may already know. Speech to text tools have been notoriously problematic for years. I'm not up on the latest developments, but my default assumption would be to be skeptical until proven otherwise.
There has been a lot of chatter around Text and Image generative AI.
Are there good video generative models? If yes, could you please spend sometime on introducing these models.
The world of user engagement has moved to short videos.
Hi Anant, I've talked about this a few times in my Sunday paid column "what You May Have Missed." I try to cover almost everything that's happening in AI, but because I can only do two high-quality long-form pieces per week, I cover everything else in WYMHM. Consider subscribing, I think it complements very well the Tuesday-Friday combo!
Anyway, I'll try to answer your question here: Text-to-video is still in research. Google (Phenaki, Imagen Video and a combination of both) and Meta (Make-A-Video) are clearly ahead of everyone else. Stability.ai has some plans for this in the future. And there are also unofficial implementations on top of stable diffusion.
There's no commercial app yet for this and Emad Mostaque said recently that it'll probably take a couple of years before this tech gets to the level SD is today.
Thank you! This is very useful.
I am interested in knowing what specific applications society is finding for generative AI and what the second- and third- order consequences of those might be. It is easy to imagine that every advertising agency in the world is going to want access to the technology, but harder to see the staffing and payroll implications. It is hard for me to see how the porn sector as we know it today will survive. The examples are legion.
"The examples are legion." Agreed. I guess we'll find out together! This is merely starting and the second- and third-order consequences may be unpredictable even for those at the very vanguard.
Of course the consequences "may be unpredictable". But it still makes sense to try to think about might happen. It is always a good idea to open your eyes even you can't focus them immediately.
If you heard about Bittensor, what do you think about an incentivized decentralized AI Hivemind? The platform enables users to train, test, and deploy AI models on the blockchain, and rewards contributors with tokens based on the performance and adoption of their models. We can co-own, co-run, and thus align the intelligence layer while creating a censorship resistant AI.
Sounds like a good idea, although I don't how we could even start to do it (also, blockchain always reminds me of crypto, which I'm not really fond of given the latest events...)
I think Bittensor something you should keep one eye open for and I’m sure you’ll here about it in 2023. It’s 500b parameters and growing. Text only for now and then will have image and audio next, aiming to have multi modals. But anyways, I LOVE your content and always sharing with many of my friends. I’m loving forward to learning more from you as this seems to be happening so fast. You make it easy to keep up with putting it all in perspective as well! Thank you! Where can I donate?
I want this! Yes!
Thank you Alberto. Very helpful. Keep posting your thought provoking and highly informative posts. I'm finding them very helpful.
Thanks Tom!!
Will transformer models be the dominant AI model in the near future?
Near like in 1-2 years? I'd say yes, undoubtedly (unless something highly unexpected happens). Much more than that? I really hope not. Maybe a part of a whole, but they have critical limitations that would give us a lot of problems.
What do you think is the most important skill a fifteen year old should be learning right now?
Great question! I'd say writing well and coding well are high at the top, for sure--even assuming AI will improve and cover a portion of those jobs (or part of a single person's job).
Writing, as you know better than I do, is transversally important in life, even if you don't earn a living as a writer. Coding is super useful now and, even if AI catches up, coders (good ones in particular) will be the ones to exploit it the most in the future.
I assumed you were referring to hard skills. In the soft skills area, I don't have any doubt: for me a good emotional education is unbeatable.
Thank you for your answer! I ask because I have a 15 yr old who thinks school is a waste of time and wants to spend all their time on the computer making digital art- her particular passion, and skill, she designs characters, but now I fear AI will render her talents useless... it’s daunting...
You're in a very special position because you're aware of what's happening in AI and see the threats more than most people, but that can also be paralyzing in that you will worry a lot more than someone who is 100% unaware.
I don't think artists and writers will be erased by AI (I don't think you do either), but I'm quite sure demand will be affected. And those are professions (writer/artist) where it's already quite hard to earn a living wage.
As an example of how coding and artist skills synergize, I know a few people in the new AI art scene that are both artists and coders. Those, in contrast to "pure artists", don't see AI as a threat nearly as much and are learning to leverage it.
And, for what is worth, I started writing years after I suspected what AI would be capable of (I'm not worried because I write about AI, which is a very safe place).
Anyway, the takeaway is that being aware is more important than trying to be safe--instead of letting the wave crush you, you could choose to ride it with the tools at your disposal without necessarily changing careers.
The only safe spot from AI is the inherently human component of what we do. This component is present in almost everything, but it's not equally important everywhere. I, for instance, try to make that part very large (I write about what I think and believe, not just content for the sake of content). The same thing applies to visual artists.
Thank you. It’s true. I think the opportunities will be for “generalists” people who can mashup and synthesize different disciplines also. People good at things like management, business, sales/marketing will likely also continue doing well as physical labor trades like plumbers and AC techs.
It’s just hard to predict. Like who would have thought art would be automated before law or accounting...
Though it sort of makes me think that the least economically powerful in society got automated first, the artists. 😢
On the plus side, many of the greatest talents felt just like your daughter, and made it work. I went to the same high school as Duane Allman ( founder Allman Brothers Band) and he couldn't even be bothered to finish high school. Other greats like Steve Jobs and Bill Gates and more dropped out of college.
That said, I agree your concerns are very valid. Daunting indeed. This is the price we pay for an ever accelerating pace of change, nobody can count on anything for long. Been there, done that.
True true. High school does feel super dated and I often wonder if she would be better off learning in some other way. The public curriculum can’t keep up with technology.
I’d almost rather her just learn the old school classics at school and tech on her own than be in this in-between state of schools trying to be modern but failing... because currently it’s like they don’t learn cutting edge stuff or important foundational works. And of course all the kids are using Google, YouTube, and TikTok for every homework assignment anyway.
It’s hard to inspire creative thinking and problem solving when kids can find almost every answer to every question instantly.
The smart move these days seems to be to go in to fields which can't be automated. Like say, nursing, or plumbing.
If a kid can make it through four years of nursing school, they are set for life, and will never have to worry about money ever again. A good friend of mine did this when we both were young, and now she has a $100k salary, a great state retirement plan, and a big enough Social Security check that she could probably live on that alone. And she's still working, cause she loves her job.
We recently had a plumber come to our house. He was here an hour and charged us $200. Work an hour, and you're done for day. Own your own company, be your own boss, build it if you want, or spend most of your time surfing.
The main thing is, find a career that the machines can't do.
And yes on the emotional education! Maybe the hardest thing to learn well, a life-long pursuit.
Congratulations on hitting 5,000 Alberto! I expect much success for you. I've been in and around web and email publishing for 27 years and as best I can tell, you're doing everything right, as your growing fan base would seem to prove. I find your enthusiasm and professionalism inspiring. It appears I'll be starting my own substack, and I'll be looking to you as a role model of how to go about it.
Ask you a question. Cool feature. Hmm...
QUESTION: What do you see to be the most likely outcome of humanity obtaining ever more, ever larger powers, at an ever accelerating rate?
This seems to be the path we've been on for at least a century, and AI appears likely to further accelerate the process, just as computers, and then the Internet have. Do you see any limit to this progression?
Take it from there in any direction that interests you. Thanks!
Thanks Phil, really appreciate your words and your super-high engagement (still a lot of comments to answer on my side, I'll get to those too!!)
Interesting question, really hard to answer accurately!
I'll start from the last question. I read somewhere a few months ago that (I'm paraphrasing) "nothing in nature follows an exponential." I don't have the knowledge to evaluate that sentence, but it feels much more plausible short-mid term than what Kurzweil & company foresee (the singularity and all that). There are limited resources within reach right now, and there are *a lot* of variables beyond technology/technological progress (e.g. political, cultural, economic, demographic, energetic, climatic, etc.) that could hinder that outcome.
That said, if I take your question literally the answer is an easy yes: everything has limits except maybe the universe (and, if we agree with Einstein, human stupidity). If you mean a soft version of the word "limit" then I'd say it really depends. For instance, if we managed to somehow exploit the sun as a source of energy or find a way to create an efficient fusion energy source, many things would be different.
Now, a few "maybes" on the distant future. The answer to all these questions is "maybe" for me (and I don't dare to be more precise than that!): Could we become multi-planetary? Could we decide to stay on Earth? Could we decide to upload our minds (if that's physically possible) and live in a peaceful and pleasurable matrix? Could we drop tech progress eventually and go back to a more ecological relationship with our planet? Could we merge that with high-tech (solarpunk style)?
I don't think anyone has any idea which of those is more probable.
Thanks for your thoughtful response Alberto. Please don't feel obligated to reply to every comment I make, as I type a lot, probably too much. Maybe way too much. :-) Just engage whatever interests you, as your time permits, I'm cool with that.
Thanks in part to inspiration provided by you, I've started my own substack. My first post addresses your comments above to an extent, as it proposes that the primary limiting factor to an unlimited knowledge explosion is us, human maturity, or rather the lack thereof.
https://tannytalk.substack.com/p/our-relationship-with-knowledge
PS: I've not yet completed my first day as a substack publisher, so tidiness has not yet been achieved. :-) The URL above will probably change soon.
Hi Alberto!
Thanks for the great AI-related content. Appreciate every post of yours. I have a writing related question for you.
How long have you been working on this newsletter? And approximately how much time did you spend before things started picking up?
Thanks!! My first blog post is from mid-June this year, so 5 and a half months. I came here with around 700 subscribers from Medium and it started growing right away. There have been a few articles that have driven a lot of traffic (and subscribers), but it's mostly consistence + high-quality
How do you see generative AI being used collaboratively where one user adds value to the other?
No worries - thanks for letting me know. I see no place to take a real problem and be at least pushed into right direction - doable or not, what to look at that can solve it. You do understand the problem I tried to solve though, right?
Yep, I got it. I'd recommend asking on programming sites (e.g. you asked on Reddit, which is a good place). I know how to code (basic knowledge), but I'm quite far now from that world, so I can't really help you there.
Hi Alberto,
Can you make a dedicated one post covering everything exciting and different from GPT-4? Heard it’s gonna be extraordinary, but no idea how. Tell us something which helps clear the excitement.
Thanks, Satinder.
Thanks Satinder. Everything I know is here: https://thealgorithmicbridge.substack.com/p/gpt-4-rumors-from-silicon-valley
My advice is: Don't give any rumor, hype, excited claim, etc. more value than it has. Almost no one knows anything about GPT-4 and those who do are under NDA. Yet, given the fuzz ChatGPT has created (which is just an optimized GPT-3), I can't imagine what will happen when GPT-4 is out...
Thanks Alberto. Agree. Much appreciated. :)
Hi Alberto, Maybe it is a very basic question but if you can give some iideas what to look at to start - https://www.reddit.com/r/MachineLearning/comments/z07o4c/comment/iyb4iuf
Hi Igor, I think this one is way too specific for me to answer adequately. Sorry to not be of much help here.