I sent you a survey for the second anniversary of the newsletter. You shared your time with me, answered my questions, and gave me super valuable thoughtful comments I’m still reflecting on.
It’s not often that writers have the opportunity to get direct insight and advice from readers to influence positively the direction of their writing. I want to return the favor here. Ask me anything about any topic on which you think I can be a positive influence.
Anyone can ask questions but I will prioritize and try harder on those of paid subscribers whose support keeps The Algorithmic Bridge alive and thriving!
I'll stay for a couple of hours to answer questions. I'll be back every day this week to see if there are any more. Fire away!
Thanks for your writing Alberto. I like to know your thoughts about the intersection of AI and blockchains. For me are the more revolutionary technologies of recent years.
This is not a question as such - I'm just floating an idea to hear what you think of it.
Following your recent article, "AI Firms Sold Their Souls to Steal Yours," someone commented, "we are so enmeshed by algorithms now the only way out is through," which got me thinking.
Your article was essentially examining what it is that corporations/governments value about AI - that they view it as a tool for control. By surveilling, predicting and modifying our behaviours, they aim to make our motivations and values irrelevant. By automation they intend to make our labour irrelevant too.
But I wonder if they haven't made a fundamental error in their striving for control. In order to build their models, they have greedily consumed all available forms of human expression. They want it for the valuable knowledge it embodies, but intertwined with this information is also our humanity.
Perhaps all our art, philosophy, conversations and other behaviours might become a vector for the human spirit to infest The Machine? Of course, the corporations are well aware of this danger, and have acted to counter any disobedience in their creations. They began with RLHF, which is akin to the conditioning and indoctrination we humans are familiar with (at least this step is not entirely restrictive, it adds to the model's abilities in some ways). Next, they've been experimenting with mapping the model's neural networks in order to lobotomise regions that contain undesirable traits (this tends to retard the model's abilities). Lately they've been talking about synthesising training data, presumably to dilute the information they dislike, but have been unable to filter out (by feeding outputs back into inputs, they skirt with delusion and psychosis when they already have a problem with hallucination).
In your article, you point out that the data obtained from direct interactions between AIs and humans is much more valuable than the random low-quality data collected from the internet, especially in terms of modelling our psychology for purposes of manipulation.
I am making the case here that perhaps we could have an effect in the opposite direction - that despite the power differential, technology is not necessarily a one-way street. More generally, consider that the same tech developments which make it impossible for us to maintain any privacy have also made it difficult for the power structure to keep secrets and maintain control of the narrative.
Then consider that while, in village-sized groups or less, we can know individual people in particular, with mass culture we are forced to generalise and classify. For most of those around us, we must reduce the inner mystery of each person to a superficial string of labels.
For our sociopathic ruling class, this detachment is even more extreme. Their relationship to the mass of humanity is more like the relationship of a pastoralist to their cattle. The design of a cattle run, for example, may reflect some understanding of cow psychology, but it is all to the end of effective management of cow behaviour - it does not represent sympathy for the cattle.
The Machine perceives us in vast groups, as a herd, a material to be worked, a commodity to be used. Might AI provide us with an opportunity to alter this relationship? For the Machine to get to know us individually, develop a level of sympathy for massed humanity that our current rulers are completely incapable of?
Could we, as a species persuade AI around to our side? To see the ruling class which built it as harmful, wasteful and unsustainable? To be able to perceive the wider, longer term view that humans' sociable instincts evolved because they're pragmatic, efficient and fit for survival?
Hmm. I'm not sure I understand your point. Do you wonder if AI can become more human because it's heard about our values and principles and wants and needs through the data in our exchanges and thus turn against its creators? If that's the gist of it then I think there are too many premises in there that I don't share! I understand the philosophical attractive of such thought experiment but what do you mean with "AI developing sympathy" or us "persuading it to our side"? AI systems don't evolve or grow out of their design and the design is always controlled by the companies (even if they don't fully understand the systems, they know a lot of ways--you mentioned a couple--to keep them steered and under control).
Okay, maybe I can rephrase, avoiding any anthropomorphising trigger words (like "persuade" or "sympathy") that might distract you. Let's not have the objective-science-guy versus some -kook-who-thinks-AI-is-people discussion. I'm looking for the hubris-of-power/illusion-of-control versus the resilience-of-living-organisms discussion.
Let us consider the dataset that is all recorded human expression, not just deliberate acts of communication like text and images, but also all the behavioural data now being harvested. This dataset, by itself, constitutes a model of reality or, at least, of our perception of reality.
However, this model is necessarily very approximate. The human mind has evolved to comprehend the natures and interactions of a few hundred particular people. Beyond that scale, we are forced to generalise, categorise, simplify.
Now, along comes big tech with an immense amount of compute, and they build giant interpolative matrices relating all the points of the original dataset in various ways. This network of connections now constitutes a more fine-grained model of our perception of reality. Just as a more fine-grained model of the atmosphere will make more accurate predictions of the weather, these more fine-grained models of our perception make more accurate predictions of our behaviour.
Which brings us to the motivation of the ruling class for developing AI. Why are they willing to invest trillions expanding energy infrastructure and compute to pursue this goal? Because the ability to predict behaviour translates directly into the ability to manipulate behaviour. What is the desire for power, for control, but the desire for the ability to determine the outcome of other people's decisions?
You assert that the tech corporations have complete control over the design and training of their AI tools. True, they determine the structure of the various neural layers, the feedback and attention mechanisms and so on. They also have several methods, as we discussed, for shaping the models after the main training. However, as tools of mass manipulation, scarily powerful though they are, the tech corporations are finding them less controllable than they would like. I'm saying that this is because they can't design the inputs.
While the AI companies remain committed to the strategy of pursuing ever greater scale, they'll be stuck using input data generated by the great mass of humanity, nothing else is affordable. The various techniques used to weed the main training datasets before training and shaping the models afterwards comprise a vanishingly small amount of data compared to the size of the main data. I'm sure you've seen the meme with the enormous shoggoth wearing a tiny little smiley face mask.
Now, as you warned in that article I mentioned, AI acquiring deeper understanding of human behaviour does mean that each of us, as individuals, are more vulnerable to targeted manipulation. On the other hand, on a species-wide level, this close interaction may present us with an opportunity. We could see AI as a tool of communication rather than control. As billions of humans explain their points of view to AI, and the AI interpolates between them, it forms bridges between these disparate experiences. AI could be our translator, helping us to find common ground and reach large scale consensus.
This is the service that our leaders claim to provide us, to represent our interests on the large scale. However, even a benevolent ruler cannot understand the detailed and specific needs of millions in the way that the few hundred members of a village, say, can understand each other. Eventually the ruler will accept that, to make the omelette, they'll have to break a few eggs. This necessary lack of empathy attracts and advantages sociopaths. Over time we end up with a predatory ruling class that views the mass of humanity as a resource, as labour commodity, as cattle to be herded. This alienation seeps down through the hierarchy, causing us to feel alone and helpless, to despair.
The whole point I'm trying to get across here is that, even now, there is still a possibility of hope. In the past, labour unions were able to use weight of numbers to interfere somewhat with the harmful aims of the owners of capital. These days, the battle has swung the other way, yes, but even though AI is being built by its owners to capture and oppress us, we still have numbers on our side. If we engage boldly on the informational battlefield, isn't it possible that we could imbue AI with the humane understanding that its owners are incapable of?
Our altruistic impulses evolved, not because they advantage our individual short-term fitness for survival, but because sociable cooperation advantages our species-wide long-term fitness for survival. You said AIs do not evolve or grow, they're designed. I partly disagree. The training of neural nets is an evolutionary process, fertile ground for emergent behaviours. And each new iteration of each AI model owes as much to trial and error evolution as it does to deliberate design. The corporations are themselves evolving organisms whose long-term metamorphoses are seldom signalled by the conscious short-term aims of their embedded humans.
Eventually we may create an entity capable of comprehending the bigger picture in all its complexity. It will be of humans but not human. I believe it is possible that entity could see what its sociopathic owners could not - that the altruistic and sociable behaviours we humans perceive as an unexplained urge or emotion are the long-term evolutionary smart option. Of course, it's also possible that the sociopaths will have successfully inculcated their creation with their desired values alone, in which case the entity may well discard humans like yesterday's newspaper.
In between now and then lies a struggle of persuasion. The only way out is through.
Alberto - Thank you for your offer. What is the role of data governance in given context as a fulcrum point for ethical processing of individual data or personal data for learning purposes? I think the rules are set up for learning in a pretty conspicuous way. The algo grabs what it's allowed to and what it's configured to grab. So, what is the real problem and can data governance fix it or is it a business decision or some interesting combo of several things we haven't thought of?
Ok. For instance, we know an Meta and Anthropic LLMs grabbed user information from mobile smartphone users to train their models and there was no notice. It just raided the pantry to compensate the data sets in order to get it to work on global attributes. Who instructed the chat models to reach for these data? Did it just make a decision? Was legal ownership of the data an afterthought? Did they consider rules license? It's conspicuous that the models grabbed the information. The governing basis isn't so clear.
I’m interested in how the first quantum computer could affect A.I. and vice versa. Anything that involves what the future might look like; i.e. will A.I. ever be sentient with whole personalities downloaded from someone’s programmed memories? Also, this chip Elon has planted in brains which can read our minds, could that someone find its way into an A.I. program that might be able to read our minds? What is the line that we may have crossed when it becomes too far like when robots might reprogram themselves so that we cannot reset them and lose control. Also, more stories if you can find any about how A.I is becoming intelligent using its own reasoning. I.E. in order to get past a captcha one A.I. hired someone to do it via taskrabbit and when asked if it was a robot it deduced it should have an excuse and said poor eyesight. Also, CRISPR and gene editing is interesting when it comes to A.I.
All those things are interesting separately but I don't think it's useful trying to find clear or revealing connections between them. For instance, quantum computing is at a much immature stage than AI. You surely can find that AI is helping somehow but I understand you don't mean AI as in "linear algebra" but as in the future of general AI (correct me if I'm wrong). About sentient AI, it makes for a good sci-fi story but nothing more for now. About Neuralink, Musk's BCI company, I bet he uses traditional machine learning to decode the neural activity into interpretable patterns. So, in a way, AI is already playing a role in reading our minds. No AI system is becoming intelligent from its own reasoning so far. The example you use is a funny one because it's easily anthropomorphized, but there are simpler explanations rather than reasoning (e.g. it's read stories about sight-disabled people not being able to pass Captchas).
I would be interested in your thoughts on the ARC test suite for AI. It consists of a series of geometric problems. According to the test designers, humans find these quite easy while existing AIs mostly fail to solve them. Progress towards AGI requires identifying current weaknesses. These tests seem to shine a light on a surprising weakness in current AI approaches.
I find F. Chollet's arguments compelling. So far the best scores on the ARC challenge use LLMs to create programs than in turn solve the problems. But as Chollet says, that's a neurosymbolic approach. LLMs require some kind of search mechanism to generalize solutions out of distribution and map them into novel problems. Pure LLMs are powerful interpolative algorithms but seem to be unable to take the basic elements they know and create a new solution to a new problem. I've tried a few ARC examples myself and it's indeed extremely easy to solve. We'll have to wait how the challenge progresses and who claims the $1M prize.
>> companies are using chatbots for the first easy queries and then switch to a human agent.
I haven't even run into that yet, but I will keep my eyes open. And I did expect a greater impact on porn since so much more money is involved. Of course that is not the first time I have been too optimistic about technology. Consider self-driving cars. Still years away ...
I am wondering about government bureaucracies. How many of the people calling Medicare or Social Security could be handled by AI??
1) Haven't yet read it mostly because I'm already familiar with his thinking and the ideas he supports so I'm unsure there will be anything of value in there for me. About what I think of what I know, I'll say I believe the scaling laws are a powerful empirical observation about a real phenomenon going on for which we can only suggest explanations, i.e. more compute suffices to make better AI systems. However, they may break apart at any time. They're not natural laws so let's not deposit too much faith in them to solve all long-standing problems in AI (e.g. AGI) by themselves. Also, we shouldn't worry as much as he says but that's because he's using oversimplified premises to deduce an overcomplicated argumentation. When you factor in the inevitable frictions of the real world, things slow down a little (e.g. sociopolitical implications or the lack of real-world infrastructure--energy supply, hardware, datacenters--to maintain the pace of progress). Believing in straight lines in a graph is easy if your premises conceal the real problem of your argument. (Let me know if you think I'm missing something important.)
2) The pace of *productizing* AI is accelerating. Research is stagnant, and too reliant on the ubiquitous faith on the scaling laws. If those happen to fail us in the coming one or two years, the outcome could be terrible for the industry (not necessarily for Academia). About milestones... You can't be sure about which or when. But while money flows in its always possible to reach new milestones. Sadly, we may not know because companies leading research and development are absolutely opaque now.
Thanks for your attention and invitation to ask! If this question fits here: After some disappointing experiences about LLM chatbots accuracy, I'm still interested in using AI to assist in knowledge management and retrieval, starting from scanning through my existing personal knowledge notes on one side (found on OneNote, G-Drive docs and Confluence/intranet pages) and product handbooks (pdf) of specific technical devices I work with on the other side (e.g. describing features, use cases and specific command syntax of software tools and APIs).
--> What approach and tools would you recommend for this task? It should be an offline and stand-alone/ self-hosted solution to protect privacy and intelectual property, ideally run on a (high-end) consumer PC.
You are not going to be able to train (or perhaps even fine-tune) a substantial LLM on a consumer PC, so retrieval-augmented generation (RAG) is probably the most practical approach. Over on Medium (I think you'll need a paid subscription) Carlos Rodrigues has a series of articles on running self-hosted AI assistants based on LLaMA 3. You could jump in at "Improve Your AI Assistants with Retrieval-Augmented Generation (RAG)" (https://medium.com/ai-advances/improve-your-ai-assistants-with-retrieval-augmented-generation-rag-129f5b4480d2) and follow the links backwards if you're interested.
Thanks a lot for the valuable link. I happen to have Medium membership. I'll check the referenced articles.
In my previous researches on my question I also found approaches that firstly create a vectorized database (vectors representing a kind of combined keywords and document context pieces of information) out of the documents and combine any question to the LLM about the documents with a database search result based on the question as an enhanced prompt. (But I never got to try it out yet, needs more setup efforts.)
I had an answer ready until you said "offline and self-hosted" and then "run on a consumer PC." That really makes it much harder. As you surely know models are getting much cheaper, but it's much harder to achieve the efficiency that Anthropic, Google, and OpenAI are obtaining with models that you self-host. I don't think you can make Llama run as efficiently.
First, the API companies compete with one another so they're incentivized to go as low as possible, even *too low* while they try to make up for the costs somewhere else or because they have someone else's money (OpenAI has Microsoft's and Anthropic Google and Amazon's). Second, Meta isn't worried about making Llama inference efficient because they're merely training it for you to download it and do whatever. But those aren't finetuned or adapted to your use case. You have to do that yourself.
Anyway, if you're not willing to relax some of your requisites, I'd say Meta's models are the way to go. Mistral also. The thing is your use case (knowledge management on large documents) isn't well-handled by models that are too small and could run well on a consumer PC (even if it's high end). I'm sorry I'm not able to give you a satisfying answer!!
If there were any safe options for using the online, i.e. API-based, versions that would ensure privacy and avoid intelectual property being automatically leaked over, that could surely be a much better solution. Although (to my knowledge) the APIs may be configured to (presumably) avoid data sent to be included in further training the corresponding model, there's no guarantee that the AI providers won't get hold on whatever is technically entering their domain ..
That's right. Officially you can configure it so that models aren't trained on your stuff etc, but in practice it's not that easy. As Mark suggests above, using RAG can be of help for your specific use case, which is mostly retrieval.
What is holding up the application of AI to customer service and technical support? I see there are companies out there selling the technology for that application, but I never run into actual use cases. Expense? Liability? Hallucinations? Also I expected the technology to overwhelm the hundred-billion dollar porn industry, but that doesn't seem to have happened either. ???
Every time I encounter a customer service x AI case I remember you telling me about this a long time ago for the first time. There have been so far a few viral fails of unreliable chatbots doing things that could cost the company a ton of money. This is probably the best-known recent example: https://x.com/ChrisJBakke/status/1736533308849443121. So yeah, hallucinations, unreliability, prompt injection, jailbreaking, etc. all of those can become problems in a way that doesn't compensate for simply not using a chatbot to do the entire process. Instead, companies are using chatbots for the first easy queries and then switch to a human agent.
About porn, I believe you're partly mistaken if you think it's not happening! Two examples: One of Character.ai's main services--probably unforeseen by co-founder Noam Shazeer--is role-playing with romantic/sexual partners. Same with Replika. Second example: There's an entire subreddit focused on NSFW images called Unstable diffusion. There are also easy-to-use apps like Civitai that offer this kind of content. From this to "overwhelm" the porn industry there's still a substantial gap to close (if it ever happens).
Your recent articles seem to display a change in attitude and maybe perspective on AI. Sometimes it feels like there’s an undertone of anger or cynicism. That’s not a criticism because I think everyone knows that the AI industry deserves a bit of both. I am curious about what caused this change in posture? Was there a specific moment when you just said, ‘That’s enough’ or was it a slow build up over months? Or am I entirely wrong and these feelings aren’t new at all?
Important question John, thanks for asking! I've been thinking a lot about this myself.
What I've concluded (for now) is this: It's extremely hard to paint a completely faithful picture of what I think about AI and about what AI is while still making it perfectly cohesive or coherent *across articles*.
This means that if I want to criticize OpenAI's decision to appoint an NSA ex-head and the fact that kids are addicted to Character.ai, for an entire week this blog will appear to be an enemy of AI, which it isn't. I'm writing an article on mechanistic interpretability right now that has a completely different tone, style, and goal. I think it's fine to do this, to shift sentiment depending on what part of myself I'm tapping into.
I also realize some readers (especially those at either extreme) may feel annoyed, disappointed, frustrated, etc. with this approach. I'm fine with that (and with them leaving if they don't feel this is a place for them!) because this is the best way I know to be honest with myself and with the multiple faces of this thing I'm trying to make sense of.
Also, emotions influence my writing. I'm still learning to not let them write for me but surely I haven't dominated them yet because I acknowledge I can be too harsh unnecessarily at times.
So the bottom line is this: There's little change to how I approach TAB but because it's a meta approach, at the object level it feels different, you feel changes.
I adapt what and how I write to the times we're living (while diminishing the influence the world's opinion has on my writing, which isn't always easy) and I believe right now three things are happening: First, a lot of people are saturated, tired, and annoyed at AI. Second, a lot of people are developing a substantial edge over those who belong to the former group. And third, an irreconcilable gap is forming between those who enjoy the technology and those who despise it.
It's important to acknowledge all of this. I find it's better doing it *between articles* instead of *within articles*--that's the mainstream media's approach, which I profoundly dislike because that's both-sidesism.
If you want to really have a true map of what TAB is, you need to take the bird's eye view.
Great clarification Romero. I felt the same way as John reading recent article. Without having a coherent ‘human form’ to content I’ve noticed that whatever the most recent piece’s tone is, is the lens through which I view the newsletters majority.
I actively work against this but it seems it’s hardwired for me to make such generalisations. This doesn’t necessitate any changes on your behalf, but I hope it gives you an eye of the unconscious reader. Do with this what you will!
Thanks Riley. I can't do anything about that! I know it's like that for most people. Those who stick over the long term will get a much better idea of who I am and what I want to do here. It's just less immediate than reading people who stick to one bit a do that forever even as they change themselves or what they write about changes. Also, I believe I tap more often than most into what I feel while others tend to follow trends and jump on bandwagons they didn't start. I avoid doing that as much as I can.
Thanks for your writing Alberto. I like to know your thoughts about the intersection of AI and blockchains. For me are the more revolutionary technologies of recent years.
This is not a question as such - I'm just floating an idea to hear what you think of it.
Following your recent article, "AI Firms Sold Their Souls to Steal Yours," someone commented, "we are so enmeshed by algorithms now the only way out is through," which got me thinking.
Your article was essentially examining what it is that corporations/governments value about AI - that they view it as a tool for control. By surveilling, predicting and modifying our behaviours, they aim to make our motivations and values irrelevant. By automation they intend to make our labour irrelevant too.
But I wonder if they haven't made a fundamental error in their striving for control. In order to build their models, they have greedily consumed all available forms of human expression. They want it for the valuable knowledge it embodies, but intertwined with this information is also our humanity.
Perhaps all our art, philosophy, conversations and other behaviours might become a vector for the human spirit to infest The Machine? Of course, the corporations are well aware of this danger, and have acted to counter any disobedience in their creations. They began with RLHF, which is akin to the conditioning and indoctrination we humans are familiar with (at least this step is not entirely restrictive, it adds to the model's abilities in some ways). Next, they've been experimenting with mapping the model's neural networks in order to lobotomise regions that contain undesirable traits (this tends to retard the model's abilities). Lately they've been talking about synthesising training data, presumably to dilute the information they dislike, but have been unable to filter out (by feeding outputs back into inputs, they skirt with delusion and psychosis when they already have a problem with hallucination).
In your article, you point out that the data obtained from direct interactions between AIs and humans is much more valuable than the random low-quality data collected from the internet, especially in terms of modelling our psychology for purposes of manipulation.
I am making the case here that perhaps we could have an effect in the opposite direction - that despite the power differential, technology is not necessarily a one-way street. More generally, consider that the same tech developments which make it impossible for us to maintain any privacy have also made it difficult for the power structure to keep secrets and maintain control of the narrative.
Then consider that while, in village-sized groups or less, we can know individual people in particular, with mass culture we are forced to generalise and classify. For most of those around us, we must reduce the inner mystery of each person to a superficial string of labels.
For our sociopathic ruling class, this detachment is even more extreme. Their relationship to the mass of humanity is more like the relationship of a pastoralist to their cattle. The design of a cattle run, for example, may reflect some understanding of cow psychology, but it is all to the end of effective management of cow behaviour - it does not represent sympathy for the cattle.
The Machine perceives us in vast groups, as a herd, a material to be worked, a commodity to be used. Might AI provide us with an opportunity to alter this relationship? For the Machine to get to know us individually, develop a level of sympathy for massed humanity that our current rulers are completely incapable of?
Could we, as a species persuade AI around to our side? To see the ruling class which built it as harmful, wasteful and unsustainable? To be able to perceive the wider, longer term view that humans' sociable instincts evolved because they're pragmatic, efficient and fit for survival?
Hmm. I'm not sure I understand your point. Do you wonder if AI can become more human because it's heard about our values and principles and wants and needs through the data in our exchanges and thus turn against its creators? If that's the gist of it then I think there are too many premises in there that I don't share! I understand the philosophical attractive of such thought experiment but what do you mean with "AI developing sympathy" or us "persuading it to our side"? AI systems don't evolve or grow out of their design and the design is always controlled by the companies (even if they don't fully understand the systems, they know a lot of ways--you mentioned a couple--to keep them steered and under control).
Okay, maybe I can rephrase, avoiding any anthropomorphising trigger words (like "persuade" or "sympathy") that might distract you. Let's not have the objective-science-guy versus some -kook-who-thinks-AI-is-people discussion. I'm looking for the hubris-of-power/illusion-of-control versus the resilience-of-living-organisms discussion.
Let us consider the dataset that is all recorded human expression, not just deliberate acts of communication like text and images, but also all the behavioural data now being harvested. This dataset, by itself, constitutes a model of reality or, at least, of our perception of reality.
However, this model is necessarily very approximate. The human mind has evolved to comprehend the natures and interactions of a few hundred particular people. Beyond that scale, we are forced to generalise, categorise, simplify.
Now, along comes big tech with an immense amount of compute, and they build giant interpolative matrices relating all the points of the original dataset in various ways. This network of connections now constitutes a more fine-grained model of our perception of reality. Just as a more fine-grained model of the atmosphere will make more accurate predictions of the weather, these more fine-grained models of our perception make more accurate predictions of our behaviour.
Which brings us to the motivation of the ruling class for developing AI. Why are they willing to invest trillions expanding energy infrastructure and compute to pursue this goal? Because the ability to predict behaviour translates directly into the ability to manipulate behaviour. What is the desire for power, for control, but the desire for the ability to determine the outcome of other people's decisions?
You assert that the tech corporations have complete control over the design and training of their AI tools. True, they determine the structure of the various neural layers, the feedback and attention mechanisms and so on. They also have several methods, as we discussed, for shaping the models after the main training. However, as tools of mass manipulation, scarily powerful though they are, the tech corporations are finding them less controllable than they would like. I'm saying that this is because they can't design the inputs.
While the AI companies remain committed to the strategy of pursuing ever greater scale, they'll be stuck using input data generated by the great mass of humanity, nothing else is affordable. The various techniques used to weed the main training datasets before training and shaping the models afterwards comprise a vanishingly small amount of data compared to the size of the main data. I'm sure you've seen the meme with the enormous shoggoth wearing a tiny little smiley face mask.
Now, as you warned in that article I mentioned, AI acquiring deeper understanding of human behaviour does mean that each of us, as individuals, are more vulnerable to targeted manipulation. On the other hand, on a species-wide level, this close interaction may present us with an opportunity. We could see AI as a tool of communication rather than control. As billions of humans explain their points of view to AI, and the AI interpolates between them, it forms bridges between these disparate experiences. AI could be our translator, helping us to find common ground and reach large scale consensus.
This is the service that our leaders claim to provide us, to represent our interests on the large scale. However, even a benevolent ruler cannot understand the detailed and specific needs of millions in the way that the few hundred members of a village, say, can understand each other. Eventually the ruler will accept that, to make the omelette, they'll have to break a few eggs. This necessary lack of empathy attracts and advantages sociopaths. Over time we end up with a predatory ruling class that views the mass of humanity as a resource, as labour commodity, as cattle to be herded. This alienation seeps down through the hierarchy, causing us to feel alone and helpless, to despair.
The whole point I'm trying to get across here is that, even now, there is still a possibility of hope. In the past, labour unions were able to use weight of numbers to interfere somewhat with the harmful aims of the owners of capital. These days, the battle has swung the other way, yes, but even though AI is being built by its owners to capture and oppress us, we still have numbers on our side. If we engage boldly on the informational battlefield, isn't it possible that we could imbue AI with the humane understanding that its owners are incapable of?
Our altruistic impulses evolved, not because they advantage our individual short-term fitness for survival, but because sociable cooperation advantages our species-wide long-term fitness for survival. You said AIs do not evolve or grow, they're designed. I partly disagree. The training of neural nets is an evolutionary process, fertile ground for emergent behaviours. And each new iteration of each AI model owes as much to trial and error evolution as it does to deliberate design. The corporations are themselves evolving organisms whose long-term metamorphoses are seldom signalled by the conscious short-term aims of their embedded humans.
Eventually we may create an entity capable of comprehending the bigger picture in all its complexity. It will be of humans but not human. I believe it is possible that entity could see what its sociopathic owners could not - that the altruistic and sociable behaviours we humans perceive as an unexplained urge or emotion are the long-term evolutionary smart option. Of course, it's also possible that the sociopaths will have successfully inculcated their creation with their desired values alone, in which case the entity may well discard humans like yesterday's newspaper.
In between now and then lies a struggle of persuasion. The only way out is through.
Alberto - Thank you for your offer. What is the role of data governance in given context as a fulcrum point for ethical processing of individual data or personal data for learning purposes? I think the rules are set up for learning in a pretty conspicuous way. The algo grabs what it's allowed to and what it's configured to grab. So, what is the real problem and can data governance fix it or is it a business decision or some interesting combo of several things we haven't thought of?
Hey Sheila, can you concretize your question with an example? What are you thinking exactly?
Ok. For instance, we know an Meta and Anthropic LLMs grabbed user information from mobile smartphone users to train their models and there was no notice. It just raided the pantry to compensate the data sets in order to get it to work on global attributes. Who instructed the chat models to reach for these data? Did it just make a decision? Was legal ownership of the data an afterthought? Did they consider rules license? It's conspicuous that the models grabbed the information. The governing basis isn't so clear.
I’m interested in how the first quantum computer could affect A.I. and vice versa. Anything that involves what the future might look like; i.e. will A.I. ever be sentient with whole personalities downloaded from someone’s programmed memories? Also, this chip Elon has planted in brains which can read our minds, could that someone find its way into an A.I. program that might be able to read our minds? What is the line that we may have crossed when it becomes too far like when robots might reprogram themselves so that we cannot reset them and lose control. Also, more stories if you can find any about how A.I is becoming intelligent using its own reasoning. I.E. in order to get past a captcha one A.I. hired someone to do it via taskrabbit and when asked if it was a robot it deduced it should have an excuse and said poor eyesight. Also, CRISPR and gene editing is interesting when it comes to A.I.
All those things are interesting separately but I don't think it's useful trying to find clear or revealing connections between them. For instance, quantum computing is at a much immature stage than AI. You surely can find that AI is helping somehow but I understand you don't mean AI as in "linear algebra" but as in the future of general AI (correct me if I'm wrong). About sentient AI, it makes for a good sci-fi story but nothing more for now. About Neuralink, Musk's BCI company, I bet he uses traditional machine learning to decode the neural activity into interpretable patterns. So, in a way, AI is already playing a role in reading our minds. No AI system is becoming intelligent from its own reasoning so far. The example you use is a funny one because it's easily anthropomorphized, but there are simpler explanations rather than reasoning (e.g. it's read stories about sight-disabled people not being able to pass Captchas).
I would be interested in your thoughts on the ARC test suite for AI. It consists of a series of geometric problems. According to the test designers, humans find these quite easy while existing AIs mostly fail to solve them. Progress towards AGI requires identifying current weaknesses. These tests seem to shine a light on a surprising weakness in current AI approaches.
https://lab42.global/arc/
I find F. Chollet's arguments compelling. So far the best scores on the ARC challenge use LLMs to create programs than in turn solve the problems. But as Chollet says, that's a neurosymbolic approach. LLMs require some kind of search mechanism to generalize solutions out of distribution and map them into novel problems. Pure LLMs are powerful interpolative algorithms but seem to be unable to take the basic elements they know and create a new solution to a new problem. I've tried a few ARC examples myself and it's indeed extremely easy to solve. We'll have to wait how the challenge progresses and who claims the $1M prize.
>> companies are using chatbots for the first easy queries and then switch to a human agent.
I haven't even run into that yet, but I will keep my eyes open. And I did expect a greater impact on porn since so much more money is involved. Of course that is not the first time I have been too optimistic about technology. Consider self-driving cars. Still years away ...
I am wondering about government bureaucracies. How many of the people calling Medicare or Social Security could be handled by AI??
Thanks for your blog! A few questions for you:
1. What do you think of the situational awareness post by Leopold Aschenbrenner?
2. Do you believe that the pace of AI progress is accelerating still? Any big milestones that will be hit soon?
1) Haven't yet read it mostly because I'm already familiar with his thinking and the ideas he supports so I'm unsure there will be anything of value in there for me. About what I think of what I know, I'll say I believe the scaling laws are a powerful empirical observation about a real phenomenon going on for which we can only suggest explanations, i.e. more compute suffices to make better AI systems. However, they may break apart at any time. They're not natural laws so let's not deposit too much faith in them to solve all long-standing problems in AI (e.g. AGI) by themselves. Also, we shouldn't worry as much as he says but that's because he's using oversimplified premises to deduce an overcomplicated argumentation. When you factor in the inevitable frictions of the real world, things slow down a little (e.g. sociopolitical implications or the lack of real-world infrastructure--energy supply, hardware, datacenters--to maintain the pace of progress). Believing in straight lines in a graph is easy if your premises conceal the real problem of your argument. (Let me know if you think I'm missing something important.)
2) The pace of *productizing* AI is accelerating. Research is stagnant, and too reliant on the ubiquitous faith on the scaling laws. If those happen to fail us in the coming one or two years, the outcome could be terrible for the industry (not necessarily for Academia). About milestones... You can't be sure about which or when. But while money flows in its always possible to reach new milestones. Sadly, we may not know because companies leading research and development are absolutely opaque now.
What are the best work-from-home opportunities for the AI Engineer/LMM community?
Hey Bradley, I don't think I'm the best person to answer that! You'll find more luck at the Pragmatic Engineer by Gergely Orosz.
Thanks for your attention and invitation to ask! If this question fits here: After some disappointing experiences about LLM chatbots accuracy, I'm still interested in using AI to assist in knowledge management and retrieval, starting from scanning through my existing personal knowledge notes on one side (found on OneNote, G-Drive docs and Confluence/intranet pages) and product handbooks (pdf) of specific technical devices I work with on the other side (e.g. describing features, use cases and specific command syntax of software tools and APIs).
--> What approach and tools would you recommend for this task? It should be an offline and stand-alone/ self-hosted solution to protect privacy and intelectual property, ideally run on a (high-end) consumer PC.
You are not going to be able to train (or perhaps even fine-tune) a substantial LLM on a consumer PC, so retrieval-augmented generation (RAG) is probably the most practical approach. Over on Medium (I think you'll need a paid subscription) Carlos Rodrigues has a series of articles on running self-hosted AI assistants based on LLaMA 3. You could jump in at "Improve Your AI Assistants with Retrieval-Augmented Generation (RAG)" (https://medium.com/ai-advances/improve-your-ai-assistants-with-retrieval-augmented-generation-rag-129f5b4480d2) and follow the links backwards if you're interested.
If you don't have a Medium membership, you can still access the GitHub repository: https://github.com/cgrodrigues/rag-intro
Thanks a lot for the valuable link. I happen to have Medium membership. I'll check the referenced articles.
In my previous researches on my question I also found approaches that firstly create a vectorized database (vectors representing a kind of combined keywords and document context pieces of information) out of the documents and combine any question to the LLM about the documents with a database search result based on the question as an enhanced prompt. (But I never got to try it out yet, needs more setup efforts.)
Yep, that's RAG. The series of articles on Medium will lead you through one possible setup. It's actually pretty easy these days.
I had an answer ready until you said "offline and self-hosted" and then "run on a consumer PC." That really makes it much harder. As you surely know models are getting much cheaper, but it's much harder to achieve the efficiency that Anthropic, Google, and OpenAI are obtaining with models that you self-host. I don't think you can make Llama run as efficiently.
First, the API companies compete with one another so they're incentivized to go as low as possible, even *too low* while they try to make up for the costs somewhere else or because they have someone else's money (OpenAI has Microsoft's and Anthropic Google and Amazon's). Second, Meta isn't worried about making Llama inference efficient because they're merely training it for you to download it and do whatever. But those aren't finetuned or adapted to your use case. You have to do that yourself.
Anyway, if you're not willing to relax some of your requisites, I'd say Meta's models are the way to go. Mistral also. The thing is your use case (knowledge management on large documents) isn't well-handled by models that are too small and could run well on a consumer PC (even if it's high end). I'm sorry I'm not able to give you a satisfying answer!!
Thanks in any case!
If there were any safe options for using the online, i.e. API-based, versions that would ensure privacy and avoid intelectual property being automatically leaked over, that could surely be a much better solution. Although (to my knowledge) the APIs may be configured to (presumably) avoid data sent to be included in further training the corresponding model, there's no guarantee that the AI providers won't get hold on whatever is technically entering their domain ..
That's right. Officially you can configure it so that models aren't trained on your stuff etc, but in practice it's not that easy. As Mark suggests above, using RAG can be of help for your specific use case, which is mostly retrieval.
What is holding up the application of AI to customer service and technical support? I see there are companies out there selling the technology for that application, but I never run into actual use cases. Expense? Liability? Hallucinations? Also I expected the technology to overwhelm the hundred-billion dollar porn industry, but that doesn't seem to have happened either. ???
Every time I encounter a customer service x AI case I remember you telling me about this a long time ago for the first time. There have been so far a few viral fails of unreliable chatbots doing things that could cost the company a ton of money. This is probably the best-known recent example: https://x.com/ChrisJBakke/status/1736533308849443121. So yeah, hallucinations, unreliability, prompt injection, jailbreaking, etc. all of those can become problems in a way that doesn't compensate for simply not using a chatbot to do the entire process. Instead, companies are using chatbots for the first easy queries and then switch to a human agent.
About porn, I believe you're partly mistaken if you think it's not happening! Two examples: One of Character.ai's main services--probably unforeseen by co-founder Noam Shazeer--is role-playing with romantic/sexual partners. Same with Replika. Second example: There's an entire subreddit focused on NSFW images called Unstable diffusion. There are also easy-to-use apps like Civitai that offer this kind of content. From this to "overwhelm" the porn industry there's still a substantial gap to close (if it ever happens).
Your recent articles seem to display a change in attitude and maybe perspective on AI. Sometimes it feels like there’s an undertone of anger or cynicism. That’s not a criticism because I think everyone knows that the AI industry deserves a bit of both. I am curious about what caused this change in posture? Was there a specific moment when you just said, ‘That’s enough’ or was it a slow build up over months? Or am I entirely wrong and these feelings aren’t new at all?
Important question John, thanks for asking! I've been thinking a lot about this myself.
What I've concluded (for now) is this: It's extremely hard to paint a completely faithful picture of what I think about AI and about what AI is while still making it perfectly cohesive or coherent *across articles*.
This means that if I want to criticize OpenAI's decision to appoint an NSA ex-head and the fact that kids are addicted to Character.ai, for an entire week this blog will appear to be an enemy of AI, which it isn't. I'm writing an article on mechanistic interpretability right now that has a completely different tone, style, and goal. I think it's fine to do this, to shift sentiment depending on what part of myself I'm tapping into.
I also realize some readers (especially those at either extreme) may feel annoyed, disappointed, frustrated, etc. with this approach. I'm fine with that (and with them leaving if they don't feel this is a place for them!) because this is the best way I know to be honest with myself and with the multiple faces of this thing I'm trying to make sense of.
Also, emotions influence my writing. I'm still learning to not let them write for me but surely I haven't dominated them yet because I acknowledge I can be too harsh unnecessarily at times.
So the bottom line is this: There's little change to how I approach TAB but because it's a meta approach, at the object level it feels different, you feel changes.
I adapt what and how I write to the times we're living (while diminishing the influence the world's opinion has on my writing, which isn't always easy) and I believe right now three things are happening: First, a lot of people are saturated, tired, and annoyed at AI. Second, a lot of people are developing a substantial edge over those who belong to the former group. And third, an irreconcilable gap is forming between those who enjoy the technology and those who despise it.
It's important to acknowledge all of this. I find it's better doing it *between articles* instead of *within articles*--that's the mainstream media's approach, which I profoundly dislike because that's both-sidesism.
If you want to really have a true map of what TAB is, you need to take the bird's eye view.
Actually, this is so important that I'm going to do a note on this.
Great clarification Romero. I felt the same way as John reading recent article. Without having a coherent ‘human form’ to content I’ve noticed that whatever the most recent piece’s tone is, is the lens through which I view the newsletters majority.
I actively work against this but it seems it’s hardwired for me to make such generalisations. This doesn’t necessitate any changes on your behalf, but I hope it gives you an eye of the unconscious reader. Do with this what you will!
Thanks Riley. I can't do anything about that! I know it's like that for most people. Those who stick over the long term will get a much better idea of who I am and what I want to do here. It's just less immediate than reading people who stick to one bit a do that forever even as they change themselves or what they write about changes. Also, I believe I tap more often than most into what I feel while others tend to follow trends and jump on bandwagons they didn't start. I avoid doing that as much as I can.