I enjoy following your blog but it generally feels like you are cynical and would not change your mind even if given ample evidence that undermines your views. Also, just for fun, here is ChatGPT's reply to your blog post: Here are four counterarguments to the points made in the blog post:
1. **Struggles in Early Adoption Do Not Equate to Long-Term Failure**:
- The fact that many companies are struggling to deploy generative AI is not uncommon for any transformative technology in its early stages. Remember, the early days of the internet, cloud computing, and even e-commerce faced similar adoption hurdles. Challenges around cost and confusion can be temporary and often decrease as the technology matures and becomes more widely understood and accessible.
2. **Backlash and Criticism Can Lead to Improvement**:
- Every groundbreaking technology faces criticism. However, it's important to differentiate between constructive criticism, which can lead to improvement and iteration, and general skepticism. Moreover, linking AI’s future to a few negative headlines might be myopic. Just as ChatGPT and similar models have their detractors, they also have a vast number of supporters and users who find value in them.
3. **Missteps and Controversies Do Not Undermine the Entire Potential of AI**:
- The issue regarding Google’s LLM pointing out controversial figures as "greatest" leaders is a flaw, but it's crucial to separate the limitations of one model from the vast potential of the technology as a whole. AI models can and will be improved over time, and the emphasis should be on progress and refinement.
4. **Legal Issues and Economic Challenges are Part of Tech Evolution**:
- Many transformative technologies face legal challenges, especially in their early stages. This isn’t unique to AI. These challenges can lead to improved guidelines and practices for the industry. Furthermore, the mention of potential lawsuits is speculative. Even if OpenAI faces challenges, this does not mean that the entire field of generative AI will be rendered obsolete.
Lastly, on a broader note, technology's real value is often realized in the long run. Immediate setbacks or challenges do not necessarily predict a technology's long-term viability or success.
This so reminds me of blockchain's hype curve within the enterprise space. It took about 2-3 years before tech folks came to a consensus it was just another kind of database, one that was very hard to connect to other databases, particularly transactional, whether trusted or not.
One should not fall from the peak of hype to the depths of disappointment.
Architecturally, LLM are a big step forward. There are many areas that have large and repetitive data, but which is not organized in a way that a machine can handle. Here LLM will shine.
The ability to make use of third-party tools is also something very promising, and can be learned simply from examples, which are way cheaper to make than hiring top AI talent to build custom fancy models. Many of the failures of not being able to call tools correctly stems from too few examples.
In short, the long-term future is looking good, but a mid-course correction is likely expected.
I admit I've always been a bit miffed at how excited companies are over AI. Like, I think AI's really cool, but in terms of immediate economic applications? I dunno. Generative AI makes more sense than spending a hundred million to make a Starcraft AI (sorry DeepMind I love RL I just don't get why google paid for that) but still seems overhyped.
Still, I'm gonna bet that OpenAI specifically will do fine. A few reasons why:
1) Six months since GPT-4 and the only one who's close is Anthropic. And with the success of GPT-4 OpenAI will see a boost in funding and ability to attract talent, meaning there's every reason to think they will stick the landing of the eventual GPT-5.
2) The potential for coding is huge. GPT 4 is genuinely useful for coding. If there's one thing we should be confident that will come out of this LLM hype, it's bigger and better LLMs. So even if GPT-4 doesn't quite give companies their money's worth as of yet, the fact that they're getting experience with these systems and something even better is around the corner should still justify their investments.
3) Although some problems with LLMs will be hard to remove (hallucinations) others are much more fixable (annoying AI-speak). I think there'll be a new wave of excitement when LLMs that are comparable in power to GPT-4 but are actually RLHF'd to write well are released.
"You have NO idea what you are talking about" - applies to ChatGPT and friends, more than it does to people :)
For humans, 'model of the world' comes from... the world! For LLMs, it comes from... words (and other data). Words aren't substitutes for the world, except symbolically.
Boy there is a lot of anger in the comments about infidels daring to question our lord and savior LLMs. This really reminds me of blockchain and self-driving cars insanity. At this point susceptibility to corporate hype should qualify you for a disability.
I'm interested in this comment of yours, Gary: "Large Language Models aren’t like classical databases in which individual pieces of data can be removed at will; if any content is removed, the entire model must (so far as I understand it) be retrained, at great expense." Indeed.
Consider this remark from Fodor and Pylyshyn 1988 (Connectionism and cognitive architecture: A critical analysis):
"Classical theories are able to accommodate these sorts of considerations because they assume architectures in which there is a functional distinction between memory and program. In a system such as a Turing machine, where the length of the tape is not fixed in advance, changes in the amount of available memory can be affected without changing the computational structure of the machine; viz by making more tape available. By contrast, in a finite state automaton or a Connectionist machine, adding to the memory (e.g. by adding units to a network) alters the connectivity relations among nodes and thus does affect the machine’s computational structure. Connectionist cognitive architectures cannot, by their very nature, support an expandable memory, so they cannot support productive cognitive capacities. The long and short is that if productivity arguments are sound, then they show that the architecture of the mind can’t be Connectionist. Connectionists have, by and large, acknowledged this; so they are forced to reject productivity arguments."
It seems to me that that, the irreducible interweaving of memory and program in connectionist architecture, is a severe limitation. While Fodor and Pylyshyn are talking about adding items to the system it should be clear that excising them is problematic for the same reason. Such a system can neither learn nor forget, and that is true no matter how many parameters it has nor how large the corpus it is trained on.
I use LLMs some now. Will I use them more or less five years from now? I guess more.
The development cycle is similar to the internet/browser cycle, but faster Pets.Com, the crash, and over 10 years and with the hard work of engineers, the internet is embedded in daily activity. This may be 6 or 8 years to widespread usage.
It’s an old line. In 3 years it seems like less than you expected has happened. In 10 years more than you expected has happened.
When I listened to the Republican Debate, Christie's jab towards Ramaswamy in regards to sounding like "ChatGPT" was one of the stand out moments for me.
I don't know if he came up with it beforehand or if it was thought of on the spot--the pure and distilled genius of it makes me think he thought of it contemporaneously during the debate. What I think he was trying to convey with this message is that Ramaswamy's positions / talking points were intellectually valid but paper thin and lacking meaning. Which I don't necessarily agree with but it was an incredible jab and too good to ignore.
Honestly, it shocks me how bad Siri is at transcribing my voice into written words. It guesses badly from context and enunciation makes things worse. I still think about how bad AI is in its most mainstreamed use cases. That is enough to temper my expectations
Haha. This piece is both brutal and hilarious. I don't know whether to laugh or to feel sorry for those who have placed their faith in LLMs. Great read.
We should not be too excessive with the contempt towards Chat-GPT or similar tools. Because of the race for novelty, AI driven chat-bots were released too soon. But these first generation systems are performing surprisingly well for some situations. Because of the flaws and public opinion caution, the release of the next generation will be possibly much delayed but this second generation will be certainly much improved and an average user will not probably be able to detect any important errors. And this is when the use of AI tools by a large audience will become a very sever issue.
I don’t think the present temporary fall of chat-bots, if confirmed, will really slow down the very fast, much too fast as concerns societal impacts, global development of AI. I think that the main stream AI products are not for ordinary people. If the big tech companies will make less money with the general audience, they will focus and make their business with state agencies, organizations, banks, manufacturers. Governments and global companies will invest massively in AI applications because they will consider it as strategic and mandatory and this will give effects like in any enterprise where there is really a lot of money. Maybe even the today’s all public Chat-GPT is only a pale copy, a sub-product of a more reliable system intended for institutional or business customers.
I enjoy following your blog but it generally feels like you are cynical and would not change your mind even if given ample evidence that undermines your views. Also, just for fun, here is ChatGPT's reply to your blog post: Here are four counterarguments to the points made in the blog post:
1. **Struggles in Early Adoption Do Not Equate to Long-Term Failure**:
- The fact that many companies are struggling to deploy generative AI is not uncommon for any transformative technology in its early stages. Remember, the early days of the internet, cloud computing, and even e-commerce faced similar adoption hurdles. Challenges around cost and confusion can be temporary and often decrease as the technology matures and becomes more widely understood and accessible.
2. **Backlash and Criticism Can Lead to Improvement**:
- Every groundbreaking technology faces criticism. However, it's important to differentiate between constructive criticism, which can lead to improvement and iteration, and general skepticism. Moreover, linking AI’s future to a few negative headlines might be myopic. Just as ChatGPT and similar models have their detractors, they also have a vast number of supporters and users who find value in them.
3. **Missteps and Controversies Do Not Undermine the Entire Potential of AI**:
- The issue regarding Google’s LLM pointing out controversial figures as "greatest" leaders is a flaw, but it's crucial to separate the limitations of one model from the vast potential of the technology as a whole. AI models can and will be improved over time, and the emphasis should be on progress and refinement.
4. **Legal Issues and Economic Challenges are Part of Tech Evolution**:
- Many transformative technologies face legal challenges, especially in their early stages. This isn’t unique to AI. These challenges can lead to improved guidelines and practices for the industry. Furthermore, the mention of potential lawsuits is speculative. Even if OpenAI faces challenges, this does not mean that the entire field of generative AI will be rendered obsolete.
Lastly, on a broader note, technology's real value is often realized in the long run. Immediate setbacks or challenges do not necessarily predict a technology's long-term viability or success.
Elegant, eloquent, right on!
This so reminds me of blockchain's hype curve within the enterprise space. It took about 2-3 years before tech folks came to a consensus it was just another kind of database, one that was very hard to connect to other databases, particularly transactional, whether trusted or not.
One should not fall from the peak of hype to the depths of disappointment.
Architecturally, LLM are a big step forward. There are many areas that have large and repetitive data, but which is not organized in a way that a machine can handle. Here LLM will shine.
The ability to make use of third-party tools is also something very promising, and can be learned simply from examples, which are way cheaper to make than hiring top AI talent to build custom fancy models. Many of the failures of not being able to call tools correctly stems from too few examples.
In short, the long-term future is looking good, but a mid-course correction is likely expected.
I admit I've always been a bit miffed at how excited companies are over AI. Like, I think AI's really cool, but in terms of immediate economic applications? I dunno. Generative AI makes more sense than spending a hundred million to make a Starcraft AI (sorry DeepMind I love RL I just don't get why google paid for that) but still seems overhyped.
Still, I'm gonna bet that OpenAI specifically will do fine. A few reasons why:
1) Six months since GPT-4 and the only one who's close is Anthropic. And with the success of GPT-4 OpenAI will see a boost in funding and ability to attract talent, meaning there's every reason to think they will stick the landing of the eventual GPT-5.
2) The potential for coding is huge. GPT 4 is genuinely useful for coding. If there's one thing we should be confident that will come out of this LLM hype, it's bigger and better LLMs. So even if GPT-4 doesn't quite give companies their money's worth as of yet, the fact that they're getting experience with these systems and something even better is around the corner should still justify their investments.
3) Although some problems with LLMs will be hard to remove (hallucinations) others are much more fixable (annoying AI-speak). I think there'll be a new wave of excitement when LLMs that are comparable in power to GPT-4 but are actually RLHF'd to write well are released.
GPT just sorts the men from the boys.
Can it do stuff on its own? No not really. Not good stuff.
Can it act as a multiplier to your output? Yes. Massively so.
It’s not a silver bullet for every unsolved problem. But what it can do is put your personal productivity on steroids.
It’s not an alien invasion as some people claimed, but it is on par with word processing in terms of productivity gains.
Lol, Gary...
"You have NO idea what you are talking about" - applies to ChatGPT and friends, more than it does to people :)
For humans, 'model of the world' comes from... the world! For LLMs, it comes from... words (and other data). Words aren't substitutes for the world, except symbolically.
Boy there is a lot of anger in the comments about infidels daring to question our lord and savior LLMs. This really reminds me of blockchain and self-driving cars insanity. At this point susceptibility to corporate hype should qualify you for a disability.
I'm interested in this comment of yours, Gary: "Large Language Models aren’t like classical databases in which individual pieces of data can be removed at will; if any content is removed, the entire model must (so far as I understand it) be retrained, at great expense." Indeed.
Consider this remark from Fodor and Pylyshyn 1988 (Connectionism and cognitive architecture: A critical analysis):
"Classical theories are able to accommodate these sorts of considerations because they assume architectures in which there is a functional distinction between memory and program. In a system such as a Turing machine, where the length of the tape is not fixed in advance, changes in the amount of available memory can be affected without changing the computational structure of the machine; viz by making more tape available. By contrast, in a finite state automaton or a Connectionist machine, adding to the memory (e.g. by adding units to a network) alters the connectivity relations among nodes and thus does affect the machine’s computational structure. Connectionist cognitive architectures cannot, by their very nature, support an expandable memory, so they cannot support productive cognitive capacities. The long and short is that if productivity arguments are sound, then they show that the architecture of the mind can’t be Connectionist. Connectionists have, by and large, acknowledged this; so they are forced to reject productivity arguments."
It seems to me that that, the irreducible interweaving of memory and program in connectionist architecture, is a severe limitation. While Fodor and Pylyshyn are talking about adding items to the system it should be clear that excising them is problematic for the same reason. Such a system can neither learn nor forget, and that is true no matter how many parameters it has nor how large the corpus it is trained on.
I use LLMs some now. Will I use them more or less five years from now? I guess more.
The development cycle is similar to the internet/browser cycle, but faster Pets.Com, the crash, and over 10 years and with the hard work of engineers, the internet is embedded in daily activity. This may be 6 or 8 years to widespread usage.
It’s an old line. In 3 years it seems like less than you expected has happened. In 10 years more than you expected has happened.
Stealing raw material and not paying suppliers never was a viable business model. Same with Midjourney and the rest.
They keep taking away its best features. It’s neutered.
When I listened to the Republican Debate, Christie's jab towards Ramaswamy in regards to sounding like "ChatGPT" was one of the stand out moments for me.
I don't know if he came up with it beforehand or if it was thought of on the spot--the pure and distilled genius of it makes me think he thought of it contemporaneously during the debate. What I think he was trying to convey with this message is that Ramaswamy's positions / talking points were intellectually valid but paper thin and lacking meaning. Which I don't necessarily agree with but it was an incredible jab and too good to ignore.
Honestly, it shocks me how bad Siri is at transcribing my voice into written words. It guesses badly from context and enunciation makes things worse. I still think about how bad AI is in its most mainstreamed use cases. That is enough to temper my expectations
Haha. This piece is both brutal and hilarious. I don't know whether to laugh or to feel sorry for those who have placed their faith in LLMs. Great read.
We should not be too excessive with the contempt towards Chat-GPT or similar tools. Because of the race for novelty, AI driven chat-bots were released too soon. But these first generation systems are performing surprisingly well for some situations. Because of the flaws and public opinion caution, the release of the next generation will be possibly much delayed but this second generation will be certainly much improved and an average user will not probably be able to detect any important errors. And this is when the use of AI tools by a large audience will become a very sever issue.
I don’t think the present temporary fall of chat-bots, if confirmed, will really slow down the very fast, much too fast as concerns societal impacts, global development of AI. I think that the main stream AI products are not for ordinary people. If the big tech companies will make less money with the general audience, they will focus and make their business with state agencies, organizations, banks, manufacturers. Governments and global companies will invest massively in AI applications because they will consider it as strategic and mandatory and this will give effects like in any enterprise where there is really a lot of money. Maybe even the today’s all public Chat-GPT is only a pale copy, a sub-product of a more reliable system intended for institutional or business customers.