Why AI is still dumb and not scary at all (pt. 1)
tejo.substack.com
tejo.substack.com
It seems to boil down to:
1. LLMs aren't actually "learning" the way humans do so we shouldn't be worried
2. LLMs don't actually "understand" anything so we shouldn't be worried
3. Technology has always been advancing and we've always been freaking out about that, so we shouldn't be worried
4. If your job is automate-able, it probably should be eliminated anyway
What's scary is not that these models are smarter than us, but that we are dumb enough to deploy them in critical contexts and trust the output they generate.
What's scary isn't that these models are so good they'll replace us, but that despite how limited they are, someone will make the decision to replace humans anyway.
What's scary isn't that LLMs will displace good developers, but that LLMs put the power of development in the hands of people who have no idea what they're wielding.
> Sure, with a millions upon millions of training examples, of course you can mimic intelligence. If you already know what’s going to be on the test, common patterns for answers in the test, or even the answer key itself, then are you really intelligent? OR are you just regurgitating information from billions of past tests?
How different are humans from this description in actuality? What are we if not the results of a process that has been optimized by millions upon millions of iterations over long periods of time?
is this a real threat? if a system/company decides to replace a human with something that is less capable wouldn't that just result in it becoming irrelevant/bankrupt as it is replaced by other companies doing the same thing the more efficient (and in this case traditional) way?
There are already active discussions about AI being used in government for “efficiency” reasons.
would be interesting to see what examples there are of this in recent history
While there is undoubtedly a lot of hype around these tools right now, that hype is based on a pretty major leap in technology that has fundamentally altered the landscape going forward. There are some great use cases that legitimize some of the hype.
As far as concrete examples, see the sibling comment with the anecdote regarding health insurance denial. There are also portions of the tech industry focused on rolling these tools out in business environments. They're publicly reporting their earnings, and discussing the role AI is playing in major business deals.
Look at players like Salesforce, ServiceNow, Atlassian, etc. They're all rapidly rolling out various AI capabilities into their existing customer bases. They have giant sales forces all actively pushing these capabilities. They also sell to governments. Hype or not, it adds up to real world outcomes.
Public statements by Musk about his intention to use AI also come to mind, and he's repeatedly shown a willingness to break things in the pursuit of his goals.
the counterpoint to this (which i guess is the tenet of the original comment?) is that hype can shadow good judgement for a short period of time?
https://www.youtube.com/shorts/Y7QPXzDmloI?embeds_referring_...
If you haven't already signed away your right to sue with an arbitration clause.
Consumers will absolutely buy worse buy cheaper products/services.
Open AI is not arguing that AI is harmless they are agreeing it's dangerous. They are using that to promote their product and hype it up as world changing. But more worrying they're advocating for regulations, presumably the sort that would make it more difficult for competition to come in.
I think we can talk about the potential dangers of AI. But that should include a discussion on how to best deal with that and consciousness of how fear of AI might be manipulated by silicon valley.
Especially when that fear involves misrepresentation - eg. AI being presented to the public as self directed artificial consciousness rather than algorithms that mimic certain reasoning capabilities
But then there's whatever danger actually exists regardless of the business maneuvering.
I'm not saying this is what you're doing, but I've been in numerous discussions where where someone will point to this maneuvering and then conclude that virtually all danger is manufactured/nonexistent and only exists for marketing purposes.
> eg. AI being presented to the public as self directed artificial consciousness rather than algorithms that mimic certain reasoning capabilities
I think the fact that these tools can be presented in that way and some people will believe it points to some of the real dangers.
Similarly, AI can easily sound smart when directed to do so. It typically doesn't actually take action unless authorized by a person. We're entering a time where people may soon be willing to give that permission in a more permanent basis, which I would argue is still the fault of the person making that decision.
Whether you choose to have AI identify illegal immigrants, or you simply decide all immigrants are illegal, the deciding is made by you the human, not by a machine.
AI right now is not powerful and not scary.
But follow the trendlines. AI is improving. From 2010 to now the pace is relentless with multiple milestones passed constantly. AI right now is a precursor to something not so dumb and not so "not scary" in the future.
LLMs don't need to be AGI. There's a race to adjust testing, architecture and infrastructure to fit AI agents who could possibly change them as needed. Some tools are already pretty good and they will only get better.
Long story short, we are definitely going to see massive shift in jobs in the coming years.
What I don’t want is a shitty chat bot that barely understands what you ask it and generates shitty code by sticking a corpus of shitty code together then leaving me with the unconstrained job of working out if it has any land mines in it. Then multiply that by 50 engineers. Then tell me that’s the future. Because if it is we are totally fucked. Our decline will be noted by the general decline in capacity to do anything.
Employers might believe that not using AI puts them and their investments at a disadvantage, while ignoring the overall quality issues.
We all agree that employers shouldn't be allowed to force the employees to use drugs for short-term productivity boosts. I hope we'll see something similar for AI - it has the potential to be equally damaging to employees in the long term.
A more honest blog post would read something like this: Big Ai models are interesting and useful. And there’s the potential for danger. We should be cautious and make sure we have some bright scholars studying what might happen if various scenarios play out. Just as we would want to be in response to any similarly massive advancement in technology.
>Implying
I'll never understand this "They're just learning from past experience and making predictions!" argument, as if that doesn't constitute thinking in some meaningful, if minimal sense. What I'm more annoyed about is that for laypeople AI is synonymous with transformers. The fields of AI and machine learning are much vaster, with a rich intellectual history spanning many decades. Blanket statements don't apply.
An Ai scans job applicants. Is it biased? Is it conforming with legal requirements? A person's bias can be an issue in a lawsuit. AIs are a much more difficult target, at least for now. This is why companies like them.
You know what the impact will be? More scandals like Hertz falsely reporting people to the police for stealing cars [1].
AI will make financial and lending decisions that could easily become redlining 2.0.
Companies are so desperate to eliminate labor. People have rights. They need to be paid. They can strike.
If the AI purveyors fulfil their dream, companies will be automated but nobody will have any money because nobody has any jobs anymore.
[1]: https://www.npr.org/2022/12/06/1140998674/hertz-false-accusa...
A Yann Lecuun quote from the page.
LLMs are blank slates. Humans have millions of years of pretraining recorded in the neural network not as weights but as the structure of the neural network itself. Our physical bodies are biased towards living in an oxygen environment with light and ground and our brains are biased in that same direction.
If you put humans in a completely different context. Like utterly completely brand new context say like 50 dimensional space where all our sense become useless and logic and common sense is completely overridden with new rules...
the Human will perform WORSE then the LLM when trained on the same text.
There are better ways to argue the same position but it's probably indefensible.
Neural networks are fundamentally approximators, both when they are approximating long-range relations between concepts (like LLMs do) or when they denoise some noise for images. They approximate thought and intelligence, because we have coded it in our writings. Our writings are a complete fingerprint of the thought process, because we are able to pick up any thought process from a book, without ever seeing or otherwise having contact with the author. Therefore there is an increasingly more high-fidelity "real thought" in there.
> that in and of itself is a monumental achievement but there is no real thought involved.
Pretty much anything that requires our current level of thought is therefore reachable with ANNs now that we know how to train them in depth and width.
The real question is whether this is enough. We want ASI, but our texts only have AGI, and everything that comes from biology (including intelligence) is logarithmic in scale. There is zero evidence that language models will ever learn to create abstract entities better than any collection of humans do. AI companies are advertising armies of PhD students, but we already have millions of PhD students , yet our most pressing problems have not made a lot of progress for decades. That's what should be worrying to us, not the fact that we will all lose our jobs.
You mean energy and global warming? We will pretty much hit the worst case scenario for both of these. And we've been yapping about these problems for decades so nobody cares even though they are likely 10000x more relevant.
AI is new and interesting.
Another thing with AI is the insane progress and clear knowledge of the fact we are well below the limit.
That doesn't follow, even a simple encryption program can make that possible without an intermediate actor being able to crack that.
Humans all run very similar software/hardware so we can read what each others write, but so far the computer isn't close to the same behind the scenes thoughts. Meaning all the things we humans leave out when we write are actually important, so the text isn't everything.
That LLMs cannot beat a $3 calculator is because they're not fit for purpose. Everything they offer is a hallucination. Just because that fabrication matches with reality some of the time does not make it good. Reliability matters and these things just don't have it.
LLMs are not scary. The people trying to apply them to critical domains are.
The article never says why this is good. It just states it then moves on.
> As a whole, I’m sure it’ll give birth to entirely new systems we haven’t even conceived of yet and, one can hope, free up that time and energy towards more meaningful or creative pursuits.
The people's who job it takes probably don't give af about the patterns of their job, or creative pursuits they can do while unemployed. They are probably just trying to pay their bills and feed their kids.
Hot take: if your job can be partially or wholly eliminated by AI, that’s a GOOD THING. If your job has patterns that predictable or labor that routine, AI automation is a GOOD THING.
One might also be scared in a physical sense. The race is afoot to develop physical robots on par with humans. Imagine that as your policeman, or soldier, bringing democracy to a country near you.
https://www.yahoo.com/tech/china-tech-contest-features-robot...
What do you think of the latest thinking models, and what is your test of thinking?
As far as the latest "thinking" techniques, it's all about providing the correct input to get the desired output. If you look at the training data (the internet), the hardest and most ambiguous problems don't have a simple question input and answer response, they instead have a lot of back-and-forth before arriving at the answer, so you need to simulate that same back-and-forth to arrive at the desired answer. Unfortunately model architecture is still too simple to implicitly do this within the model itself, at least reliably.
> If you look at the training data (the internet), the hardest and most ambiguous problems don't have a simple question input and answer response, they instead have a lot of back-and-forth before arriving at the answer, so you need to simulate that same back-and-forth to arrive at the desired answer. Unfortunately model architecture is still too simple to implicitly do this within the model itself, at least reliably.
Today's thinking models iterate (with function calls and Internet queries) and even backtrack. They are not as reliable as humans but are demonstrating the hallmarks of thinking, I'd say.
Neither does humans.
Sorry to disappoint the super-AGI hopeful: we're already past that point, because the average human is only a 3B parameter LLM there's nothing magic that comes with an AGI.
The first time GPT2 autocompleted a sentence involving admitting a previous fault of output we were in the post-human era.
Everything we've seen since then have been us living through the legendary post-AGI AI bootstraps launch, but living the future isn't exactly living up to the hype huh?
That GTP-4.5 model that had the Samaltmanists wriggle on the floor in rapturous ecstatic orgasms just last year? Today it's an embarrassment, like the seed-stained lesswrong posters of yore.
The singularity was a star trek plot device.
What we see today is the real deal and I re-iterate my 2015 predictions: by the time an actually scary god like AI is created(say in 2050). We'll have so many thousands or millions of demi-god AI around that the god itself will have to have the good graces to shut up, get a queue ticket in the bare-metal immigration zone with the rest of the rabble. The prospect of true AI divinity will be a (hype-burned) washed out gray with a trace sprinkle of glitter on the already existing background of demi-divine AI.
Tejo writes:
> At the heart of it, all AI is just a pattern recognition engine. They’re simply algorithms designed to maximize or minimize an objective function. There is no real thinking involved. It’s just math. If you think somewhere in all that probability theory and linear algebra, there is a soul, I think you should reassess how you define sentience.”
We have two words we need to define: “sentience” and “thinking”.
Sentience can be defined as a basic function of most living systems (viruses excluded). You may disagree but to me this word connotes “actively responsive to perturbations”. This is part of the auopoietic definition of life introduced by Maturana and Valera in the 1970s and 1980s. If you buy this definition then “sentience” is a low bar and not what we mean by AI, let alone by AGI.
The word “thinking” implies more than just consciousness. Many of us accept that cats, dogs, pigs, chimps, crows, parrots, and whales are conscious. Perhaps not an ant or a spider, but yes, probably rats since rats play and smile and clearly plan. But few of us would be likely to claim that a rat thinks in a way that we think. Sure, a rat analyzes its environment and comes to decisions about best actions, but that does not require what many of us mean by thinking.
We usually think of thinking as a process of reflection/recursion/self-evaluation of alternative scenarios. The key word here is recursion. and that gets us to “self-consciousness”—a word that embeds recursion.
I operationally define thinking as a process that involves recursive review and reflection/evaluation of alternatives with the final selection of an “action”—either a physical action or an action that is a “thought conclusion”.
What AI/LMM transformers do not have is deep enough recursion or access to memory of preceding states. They cannot remember and they do not have a meta-supervisor that lets them modulate their own attentional state. So I would agree that they are still incapable of what most if us call thinking.
But how hard is adding memory access to any number of current generation AI systems?
How hard us it to add recursion and chain-of-thought?
How hard will it be to enable an AI system to attend to its own attention?
In my optimistic view these are all close to solved problems.
In my opinion the only thing separating us from true AGI are a few crucial innovations in AI architecture—-mainly self-control of attention—-and under 100k lines of code.
Please argue! (I’ve had many arguments with Claude on this topic).
We don't know what consciousness is and pretending that's a simple question because it's easy to tell when an animal is awake or not is disingenuous at best.
Your reply to that was that they are different, because one is obvious, and the other one isn't. I don't see how it justifies the other one's existence as a separate term, because it doesn't really say anything, so for clarity I asked a question that could clearly separate the two concepts, if that separation exists at all.
It is also supposed to be an easy question to answer: if you know the two "consciousness" definitions are different and how, which you imply is so obvious that me claiming otherwise is "disingenuous", it should not be too hard to provide an example where they are not the same for the same entity. Question yourself: can you actually think of any instances of such that are clear cut and can't be explained by mundane things like a person having false memories.
This is why it's disingenuous to anthropomorphize any process of the current iteration of machine learning. There's no thinking or reasoning involved. None.
This isn't to say that this technology can't be very useful. But let's not delude ourselves thinking that we're anywhere close to achieving AGI.
"Arguing" about this topic with an LLM is pointless.
I don't think so. Unlike many humans a bunch of LLMs can concede a point that being a cucumber eater or a sound maker doesn't exclude possibility to also be a thinker.
I have optimism with reason. Humans are fancy thinking animals. I admit I think the “hard problem” in consciousness research is bogus.
We can, but the difference is not in whether we think or not, but in _how_ we think. We both experience the world through our senses and have conceptual representations of it in our minds. While other primates haven't invented complex written language (yet), they do have the ability to communicate ideas vocally and using symbols.
Machines do none of this. Again, they simply output tokens based on pattern matching. If those tokens are not in their training or prompt data, their output is useless. They have no conceptual grasp of anything they output. This process is a neat mathematical trick, but to describe it as "thinking" or "reasoning" is delusional.
How you don't see or won't acknowledge this fundamental difference is beyond me.
And how do you not see the amazing progress in LLM output? And how do you not see how close LLMs with “chain-of-thought” has gotten to the “appearance” of thinking. And how is it that I not clear enough for you in my first post in highlighting explicitly that LLMs are not yet thinking?
Do not use me as your straw chimp unthinking human.
I personally don't worry about jobs. AI is progressing very fast (there are a lof of smart folks working on it, there are a ton of money invested and there's a lot of demand from businesses, governments and individuals). Human inteligence stays the same. I think it's likely that sometime soon AI models will become more inteligent than an average human. And then more inteligent than the smartest human. And then more inteligent than the whole human race combined.
Let's say some worms 600 million years ago could think. And they consider should they kill all mutants and stay forever the pinnacle of evolution as they are or allow some of them to evolve into fish, and then mammals and eventually inteligent humans. I think we are in a position like this - we are currently the pinacle of "creation" in the known universe - do we want to stay this way - by blocking AGI progress, or do we want to allow minds far greater than us to evolve from current LLMs at the cost of probable human extinction eventually.