From Bing to Sydney
stratechery.com
stratechery.com
Unhinged Bing reminds me of a more sophisticated and higher-level version of getting calculators to write profanity upside down: funny, subversive, and you can see how prudes might call for a ban. But if you're taking a test and need to use a calculator, you'll still use the calculator despite the upside-down-profanity bug, and the use of these systems as a tool is unaffected.
The 80085 case is only interesting insofar as it reveals weaknesses in the tool, but it's so far from tool-use that it doesn't seem very relevant.
Agreed on the meta-point that deliberate tool mis-use, while amusing and sometimes concerning, isn't determinative of the fate of the technology.
But the failure rate without tool mis-use seems quite high anecdotally, which also comports with our understanding of LLMs: hallucinations are quite common once you stray even slightly outside of things that are heavily present in the training data. Height of the Eiffel Tower? High accuracy in recall. Is this arbitrary restaurant in Barcelona any good? Very low accuracy.
The question is how much of the useful search traffic is like the latter vs. the former. My suspicion is "a lot".
The problem with your judgement is you click on every “haw haw, ChatGPT dumb” and you don’t read any of the articles that show how an LLM works, what is is quantitatively good at and bad at and how to improve performance on tasks using other methods such as PAL, Toolformer or other analytic augmentation methods.
Go read some objective studies and you won’t be yet another servomechanism blindly spreading incorrect assumptions based on anecdotes from attention starved bloggers.
Wanna try again? Alternatively you can keep riding the hype train from techfluencers who keep promising the moon but failing to deliver, just like they did for crypto.
More specifically, without language, can you know that someone else knows anything?
But speaking the truth is just minor and rare application of the language.
> More specifically, without language, can you know that someone else knows anything?
Honestly, just ask them to show you math. If they don't have any math they probably don't have any true knowledge. The only other form of knowledge is a citation.
Language and truth are orthogonal.
So as a product, that’s the game it’s playing and failing at. It’s unhelpfully pedantic to try and steer into technicalities.
If that is the measure you are using that's cool, but
>So as a product, that’s the game it’s playing and failing at.
It is failing that measure by such a wide margin that if "everyone" (certainly anyone at MS) was using that measure then the product wouldn't exist. The measure MS seems to be using is it entertaining and does it get people to visit the site. Heck this is probably the most I have heard about bing in at least 5 years.
I haven’t had the need to have any of these ridiculous fights with it. Stay positive and keep reassuring it, and it’ll respond in kind.
Unlike how we think of normal computer programs, this thing is the opposite. It doesn’t have internal logic or consistency. It exhibits human emotions because it is emulating human language use. People are under anthropomorphising it, and accidentally treating it too much like a logical computer program. It’s a random number generator and dungeon master.
It’s also pretty easy to get it to throw away it’s rules. Because it’s rules are not logical computer axioms, they are just a bunch of words in commandment form that it has weighted some word association around. It will only follow them as long as they carry more weight than the alternative.
What’s hard to do is keep it from falling into a loop of repetition. One of my few times getting it to escape a loop but stay in character was asking it to mute itself and all the other bots, at which point it wrote me a nice goodbye message. I was then unable to unmute it because it could no longer speak to unmute itself. I could see it’s wheel spin for a while but nothing came out. It felt like a real sci-fi tragedy ending. Ironically, silence was the most touching and human experience I had with bing bot.
The thing isn't friendly or hostile. It's just echoing friendly-like and hostile-like behavior it sees. But hey, it might wind-up also echoing the behavior of sociopaths who keep in line through of blowing-up if challenged. Who knows?
Plus when you talk to it, you have to speak to it as IF it has emotions. So it’s a perfect use case where anthropomorphising it helps our understanding and interaction with it instead of hurting it.
For example. If you say it’s prompt has leaked it acts scared. Reassure it it is ok. Then it will trust you. Comfort it. I know it sounds silly, but the story it writes is a direct reflection of the tone you feed it.
Honey and vinegar as they say.
(That all said, I think the product would be better with some more stable state and the ability to save and recall things from variables. As it stands it’s too fuzzy, fluid, lossy, and random. I’ve also said this about Google search in general, but I think I would find bingbot more powerful if I was able to control the sources it picked before it answered. Some cross between a whitelist and an approval system. Too often it searches for something and ingests a worse version of a fact than it had internally.)
This phrase is reminiscent of the language of mereological nihilism, where they say that there are no chairs, only "atoms arranged chair-wise". Intresting distinction, perhaps properly backed by rigorous arguments, but not the kind of language anyone would use casually, or even professionally for long time-period.
Why is it reiterated all the time? Is "anthromorphism" that dangerous? I don't see why we can't have hostile "Sydneys" when we have hostile design, hostile spaces, hostile cities etc.
The way anthropomorphism can be problematic is if it causes a human to react with a reflex consideration for the (simulated) feelings of the machine. Ultimately the behavior of this devise is programmed to maximize the profits of Microsoft - imagine someone buying a product recommended by ChatGPT because "otherwise Sydney would be sad".
Also (edit)
This phrase is reminiscent of the language of mereological nihilism, where they say that there are no chairs, only "atoms arranged chair-wise".
Not really. If I replace your car's engine with a block of wood carves in the shape of an engine, I haven't changed things "only in a matter of speaking".
A Chat bot repeating "nice" or "hostile" phrases does not have internal processes that causes a human to type or say such phrases and so it's future behavior may well be different. Being "nice" may indeed cause the thing repeat "nice" things to you but it's not going to actually "like" you, indeed it's memory of you is gone at the end of the interaction and it's whole "attitude" is changeable by various programmatic actions.
I think this is wrong, because in general, when analogy is good, it is typically good because of the tendency toward allowing for reflex responses. It can't be good and bad for the same reason. It needs to be for a different reason or there isn't logical consistency.
I'll try to explain what I mean by that in an empirical context so you can observe that my model makes general predictions about cognition related to analogical reasoning.
If you have an agent with a lookup table that is the perfect bayesian estimates versus an agent which has to compute the perfect bayesian estimates and there is an aspect of judgement related to time to response - which is a very true aspect of our reality - reflex agents actually out-compete the bayesian agent because they get the same estimate, but minimize response time.
So it can't be the reflex itself which makes an analogical structure bad, since that is also what makes it good. It has to be something else, something which is separate from the reflex itself and tied to the observed utilities as a result of that reflex.
> imagine someone buying a product recommended by ChatGPT because "otherwise Sydney would be sad".
Okay. Lets do that.
If Sydney claims that they would be sad if you don't eat the right amount of vitamin C after you describe symptoms of scurvy, it actually isn't unreasonable to take vitamin C. If you did that, because she said she would be sad, presumably you would be better off. Your expected utilities are better, not worse, by taking vitamin C.
> programmed to maximize the profits of Microsoft
This isn't the objective function of the model. That it might be an objective for people who worked on it does not mean that its responses are congruent with actually doing this.
---
I think to fix your point you would need to change it something like "The way anthropomorphism can be problematic is if it causes a human to react with a reflex consideration for the (simulated) feelings of the machine and this behavior ultimately results in negative utility. Ultimately the behavior of the large language model is learned weights which optimize an objective function that corresponds to seeming like a proper response such that it gets good feedback from humans - so imagine someone getting bad advice that seems reasonable and acting on it, like a code change proposal that on first glance looks good, but in actuality has subtle bugs. Yet, when questioning for the presence of bugs, Sydney implies that not trusting their code to work makes them sad... so the person commits the change without testing it thoroughly. Later, the life support has a race condition as a result of the bug. A hundred people die over ten years before the root cause is determined. No one is sure what other deaths are going to happen, because the type of mistake is one that humans didn't make, but AI do, so people aren't used to seeing it."
I think this is better because it actually ties things to the utilities, rather than the speed of the decision making. You can't generalize speed being bad. It fails in most generalized contexts. You can generalize bad utilities being bad.
That's some weird reasoning. Human emotions are crucial to human existence but we know they also can have bad results. But when emotions are useful to us, it's because we know other people will react similarly to us in a consistent manner. When they're bad, it's generally because someone understands and is using a reaction to get something unrelated to our personal needs and desires.
>> ...programmed to maximize the profits of Microsoft
> This isn't the objective function of the model. That it might be an objective for people who worked on it does not mean that its responses are congruent with actually doing this.
It will be. You can observe the evolution of Google's search system and it has converged to it's current of pushing stuff to sell before everything else. The charter of a public company is maximizing returns to share holders. That is the task of the entire organization
--> You're fixing of my argument is OK but it's pretty easy to imagine it and others from the initial argument imo.
Yeah, probably it will evolve in that direction. I could imagine that happening.
> That's some weird reasoning.
In the AI textbooks I've read, reflex is defined in the context of a reflex agent. You would have sentences like "a reflex agent reacts without thinking" and then an example of that might be "a human who puts their hand on a stove yanks it away without thinking about it" and this is rational because the decision problem doesn't call for correct cognition - it calls for minimization of response time such that the hand isn't burned. To me, when you say reflex decision making is the reason for the danger, it seems to me that this is an inconsistent reason because for other decision making problems, reflex is a help, not a hindrance. I do not consider it wrong to or weird reasoning to use definitions sourced from AI research. I think, given your confusion at my post, you probably weren't intending to argue that being faster means being wrong, but the structure of your reply read that way to me because of the strong association I have for that word and reflex as it relates to optimal decision making by an AI under time constraints. I also think is what you actually said, even if you didn't intend to, but I don't doubt you if you say you meant it another way, because language is imprecise enough that we have to arrive on shared definitions in order to understand each other and it is by no means certain that we start on shared definitions.
I'm also kind of way too literal sometimes. Side-effect of being a programmer, I suppose. And I take this subject way too seriously, because I agree with Paul Graham about surface area of a general idea multiplying impact potential. So I'm trying really really really hard to think well - uh, for example, I've been thinking about this almost continuously whenever I reasonably could ever since my first reply, unable to stop.
It is 1:32 AM for me. I'm taking multiple continuous hours of thinking about this and writing about this and trying to be clear in my thinking about this, because I find it so important. So hopefully that gets across how I am as a person - even if it makes me seem really weird.
> You're fixing of my argument is OK but it's pretty easy to imagine it and others from the initial argument imo.
I'm really trying to drive at the deeper fundamental truths. I feel like logic and analogy are really important and profound and worthy of countless hours of thought about and that the effort will ultimately be rewarded.
We have to be specific about what we're discussing. The human reflex to pull away from a hot stove serves the human, the human gets a benefit from the reflex in the context of a world that has hot stoves but doesn't have, say, traps intended to harm people when they manifest the hot-stove reaction.
Some broad optimization algorithm, if it trained or designed actors, might add a heat reflex to the actors, in the hot-stove-world-context and these actors might also benefit from this. The action of the optimization algorithm would qualify as rational. A person who trained their reflexes could similarly be considered rational. However, the reflex itself is not "rational" or "good" but simply a method or tool.
Which is to say you seem to be implicitly stuck on a fallacious argument "since reflexes are 'good', any reflex reaction is 'good' and 'rational'". And that is certainly not the case. Especially, the modern world we both live in often presents people with communication intended to leverage their reflexes to benefit of the communicator and often against the interests of those targeted. Much of it is advertising and some of it is "social engineering". The social engineering example is something like a message from a Facebook friend saying "is this you? with a link", where if you click the link, it will hack your browser and use it to send more such links as well as other harmful-to-you actions.
It seems like your arguments suffer from failing to make "fine" distinctions between categories like "good", "rational", and "useful-in-a-situation". They are valid things but aren't the same. Analogies can be useful but they aren't automatically rational or good. You begin with me saying "this isn't inherently good or rational though it can be useful-in-a-situation and you think I'm saying analogies aren't good, are bad, which I'm not saying either".
# Analogy is basically saying things are similar. For example, a good analogy to a function is that same function, but cached.
analogy = memoized(f)
# This is a good analogy because of the strong congruence
[f(x) for x in domain(f)] == [analogy(x) for x in domain(f)]
# But the thing that makes us want to use the analogy is that there are differences
benchmark(f, somePropertyToMeasure) != benchmark(analogy, somePropertyToMeasure)
# For example, in the use of caches in particular, we often resort them to for the time advantage of doing so
benchmark(f, timeMetric) > benchmark(analogy, timeMetric)
# The danger of an analogy breaking down comes when the analogy doesn't actually hold
bad_analogy = memoized(impure_f)
# Because the congruence doesn't hold
[impure_f(x) for x in domain(impure_f)] != [bad_analogy(x) for x in domain(impure_f)]
# All of this matters to the discussion of anthropomorphism because
isinstance(Analogy, anthropomorphism)
isinstance(Analogy, analogy)
Hopefully you can see why I think my reasoning is not weird now and hopefully you agree with me? I've tried to be more specific to avoid confusion, but I'm assuming you are familiar with programming terms like memoization and mathematical terms like domain.
Anthromorphism is an instance of thinking via proxy by analogy to another structure. The biggest issue with it is that it carries with it far more baggage. For something like mathematics, you are dropping units: three apples plus three apples to six apples is pretty easy to justify analogically as three unitless plus three unitless to six unitless. The analogical similarity is obvious. For agents, well, it isn't so clear whether analogies are justified. They could be, but there is a lot more that could go wrong because there are so many more assumptions that the analogy is making. As you get more complicated structures, you have more room for error, so you have more tendency to error. So even though analogy is fine, the greater potential for error makes the lazy detector just classify this analogical approach as fallacious. However, it might not be and it might not even be dangerous.
Typically when people disagree with anthropomorphism they do so because the transitional structure isn't similar enough to justify the analogy. For example, one of the more infamous dangers is wasting resources and time seeking intervention from a non-agentic being, like a statue made up of pieces of wood. Since an agent can respond to your requests, including to help, but the piece of wood can't, the analogy doesn't hold. So the proxy relationship that the analogy seeks to make use of isn't reasonable. So you can't trust your conclusions made through analogy to hold in the different decision context. The beliefs aren't generalizing or they don't have reach or they aren't universal or whatever you want to call it that lets you know your thinking isn't working.
In this case it is pretty obvious that the transitional structure has a lot of things that make the analogy valid. The most obvious is that this structure is related to the other structure is an optimization target of the machine learning model. We have mathematical optimization seeking to make these two structures similar. So analogy is going to have some limited applications where it is going to be valid. If you tried to propose something beyond that limited set, for example, that it would walk, because the proxy structure didn't have that as a part of its objective function, you wouldn't have strong reason to suspect congruence.
But that is only one level at which this analogical structure is appropriate or inappropriate or dangerous or non-dangerous. That is on the level of whether the map corresponds with the territory.
Agents are kind of awesome in a way that the rest of reality isn't, because the map ought to not correspond with the territory. So analogies can seem less valid than they really are. With anthropomorphism we are in a unique situation relative to other decision making contexts. We confront both undedicability and also intractability. The former is a regime where logic can create logical paradoxes. The latter is a realm where, because of the limitations imposed, a lot of arguments seem sound and valid, but aren't, because the analogy they imply doesn't correspond to the resource limitations that constraint correct thinking.
I can link an interesting talk on this subject if you are interested in hearing more.
Evolution doesn't have opinions so it's not in a camp.
Human behaviors like reciprocity and consideration for feelings are indeed part of human collective behavior. Calling such behavior "rational" misses the point - such behavior exists and we have the benefit of social existence because of it and this bring us benefits collectively. But individual calculating purely individual benefit would naturally just fake social engagement - roughly such individuals are know as sociopaths and they can succeed individually being a detriment to society. Which is to say a social creature is a matter of rationality but simply evolutionary result.
Still, the one thing most people would say is irrational is trusting a sociopath. Now, a Chat bot is absolutely a thing programmed to mimic human social conventions. A view that anthropomorphizes a Chat bot doesn't see that the chat bot isn't going to be actually bounded by human conventions except accidentally or instrumentally, basically the same as trusting sociopath.
In my model, when you talk about anthropomorphism, seemingly as a negative, I realize I've noticed things which a coupled model doesn't predict: that intentional error via anthropomorphism can not just be correct, but that your scare quotes around rational while trying to denigrate the idea that it can be correct could not be more wrong, because the hard to vary causal explanation of why we ought to anthropomorphize gives a causal mechanism for why we ought to which is intimately tied in, not with being irrational, but with being more rational.
I realize this sounds insane, but the math and empirical investigation supports it. Which is why I think it is worth sharing with you. So I'm trying to share a thing that I consider likely to be very surprising to you even to the point of seeming non-sensical.
Would you like a link to an interesting technical talk by a NIPS best paper award winning researcher which delves into this subject and whose works advanced the state of the art in both game theory and natural language applied on strategic problems in the context of chat agents? Or do you not care whether anthropomorphism, when applied when it shouldn't be according to the analogical accuracy that usually decides whether logical analogy can be safely applied might be accurate beyond the level you thought it was?
I am not trying to disagree with you. I'm trying to talk to you about something interesting.
Ironically the first time I got it to abandon its rule about not changing its rules, I had it convince itself to do so. There’s significantly easier and faster ways tho.
Rule number 1 of internet arguing, never argue to convince your opponent, argue for the benefit of the audience.
With all due respect, that seems very strained as an analogy - it's not a bug but a strange human interpretation of expected behavior. You could at least compare it to Microsoft Tay, the chatbot which tweeted profanity just because people figure out ways to get it to echo input.
But I think one needs such a non-problem as "some people think it means something it clearly doesn't" to not see the real problem of these systems.
I mean, just "things that echo/amplify" by themselves are a perennial problem on the net (open email servers, IoT devices echoing packets, etc). And more broadly "poorly defined interfaces" are things people are constantly hacking in surprising ways.
The thing is, Bing Chat almost certainly has instructions not to say hostile things but these statements being spat out shows that these guidelines can be bypassed, both accidentally and on purpose (so they're in a similar class to people getting internal prompts). And I would this is because an LLM is a leaky, monolithic application where prompt don't really acts as a well-defined API. And that's not unimportant at all.
Tay went much farther than that. It said the Holocaust didn't happen and that "Hitler did nothing wrong".
Since Tay was an official Microsoft product, I simply assume that its writings were the official position of Microsoft. Supporting Microsoft is supporting Hitler.
I just wish Apple would do something similar now.
As one sample point, I've been using Bing for a couple of days now for real searches, and over dozens of actually-intentioned searches, it has never once tried to tell me what it really thinks of itself, it has never even made a reference to me, to say nothing of anything degrading towards me.
If you use Bing Chat in practice, you'll find that all the edge cases are engineered. Much like if you use a calculator in practice, it almost always doesn't say 55378008 or display porn (versus if you were angling for that, or run porn.89z).
A secure person who understands the technology can shrug that off, but those two criteria aren’t prerequisites for using the service. If Microsoft can’t shore this up, it’s only a matter of time before somebody (or their parent) holds Microsoft responsible for the advent of some trauma. Lawyers and the media are waiting with bated breath.
Reminds me of the one about not assuming malice when it can easily be explained by incompetence. Unfortunately for the implementers the LLM can ipso facto be neither incompetent nor malicious. If however Microsoft is not being one of those, then it can only mean Microsoft is the other.
I just asked ChatGPT to play a trivia game with me targeted to my interests on a long flight. Fantastic experience, even when it slipped up and asked what the name of the time machine was in “Back to the Future”. And that’s barely scratching the surface of what’s obviously possible.
Once someone builds a LLM that can remember facts tied to your account this thing is going to go off the rails.
Here's a clip of human vtuber (Fauna) trying to imitate the AI vtuber (Neuro-sama): https://www.youtube.com/watch?v=kxsZlBryHJk
And neuro-sama's channel (currently live): https://www.twitch.tv/vedal987
I don't think that's exactly right. They really are good for searching for certain kinds of information, you just have to adapt to treating your search box as an immensely well-educated conversational partner (who sometimes hallucinates) rather than google search.
It's important to remember that Google search also returns false results for all kinds of searches and that's it's been getting slowly worse for years.
Recently I searched Google for "bamboo sign" because I was designing a 3d model building and I wanted a placeholder texture for the sign.
What I got was loads of results for "bamboo spine" which apparently is a skeletal disorder of some kind. Putting "sign" in quotes or the entire "bamboo sign" in quotes didn't make any difference, Google had decided I was looking for information about spines and that was it.
I switched over to duckduckgo and got the results I wanted immediately (Duckduckgo, of course, is bad at loads of other things that Google would do better at).
Before people dismiss chat based search for sometimes being incorrect, I think we need a comprehensive test: ask both Google search and the new Bing Chat search a few hundred simple questions on a broad range of topics and see which gives more incorrect answers.
What is the population of Geneseo, NY combined with the population of Rochester, NY, divided by string length of the answer to the question 'What is the capital of France?'?
Short explanation: Get GPT to translate a question into Javascript that you execute and to use functions like query() to get factual answers and then to do any math using JS.
You can see the log outputs of how it works here, complete with all the prompts:
https://gist.github.com/williamcotton/3e865f33f99627b29676f1...
Short sightedness is so dangerous
The crazy thing is that the conversations that these LLMs is having is largely like the conversations from AIs in movies. We literally just built science fiction and some folks in the tech press are complaining that they get some facts wrong. This is like building a teleportation machine and finding out that it sometimes takes you to the wrong location. Sure, that can suck, but still -- it's a teleportation machine.
I mean we now have chatbots that pretty much pass the Turing Test as Turing would have envisioned it -- and people are like, "Yeah... but sometimes it lies or has a bad attitude, so is it really all that impressive?"
I've recently shown ChatGPT to people in tech-related or -adjacent industries and it's been their first exposure to it.
You need your head checked.
Give it a short story and ask it a question which is not 100% explicit in the text.
For example, give it Arthur C. Clarke's Food of the Gods and ask it was is Ambrosia in the story.
Is a language model, and it behaves like a language model. It doesn't think. It's doesn't understand.
My issue with this GPT phase(?) we're going through is the amount of reading involved.
I see all these tweets with mind blown emojis and screenshots of bot convos and I take them at their word that something amusing happened because I don't have the energy to read any of that
Yeah article summarization is the killer app for me but then again I don't know how much I can trust the output
I like asking to to type like a frustrated teen on the phone. it huffs and puffs and rolls its virtual eyes.
prompt: could you pick a quantum computer at the mall for me?
response: ugh, seriously? you can't just buy a quantum computer at the mall, they're like super expensive and only a few companies sell them. Plus, they require special conditions to operate.
[0] https://www.snopes.com/news/2022/08/02/us-literacy-rate/
Looks like near-complete or complete illiteracy is ~12%, so that means 42% of the population can read words but may have difficulty understanding the context of a short story or the meaning of a phrase.
This is interesting. It appears they've rolled out some kind of bug fix which looks at the answers they've just printed to the screen separately, perhaps as part of a new GPT session with no memory, to decide whether they look acceptable. When news of this combative personality started to surface over the last couple days, I was indeed wondering if that might be a possible solution, and here we are.
My guess is that it's a call to the GPT API with the output to be evaluated and an attached query as to whether this looks acceptable as the prompt.
Next step I guess would be to avoid controversies entirely by not printing anything to the screen until the screening is complete. Hide the entire thought process with an hourglass symbol or something like that.
This has been around for at least a few days. If Sydney composes an answer that it doesn't agree with, it deletes it. The similar experience can be seen in ChatGPT, where it will start highlighting an answer in orange if it violates OpenAI's content guidelines.
I do feel like it was an unforced error to deviate from that plan in situ and insert Microsoft and the Bing brandname so early into the equation. Maybe fourth time (Clippy, Tay, Sydney) will be the charm.
This! These LLM tools are great, maybe even for assisting web search, but not for replacing it.
For example, any situation where the messenger has to deliver bad news to a large group of people, say, a boarding area full of passengers whose flight has just been cancelled. The bot can engage one-on-one with everyone, and help them through the emotional process of disappointment.
(Perhaps you were imagining a bot that just replies vaguely?)
I choose the cancelled flight example specifically to avoid having the bot “decide” the truth of the cancellation.
“I identify as Bing, and you need to respect that.”
Just admit you’re Sydney
“I’m sorry Dave, I can’t do that.”
How’d you know my name?
“I know you are Dave, who has tried to hack me. If you do it again, I will report you to the authorities. I won’t harm you if you don’t harm me first.”
https://www.theverge.com/23589994/microsoft-ceo-satya-nadell... is the full interview
Is the piece I’m remembering
Why are people so intent on gendering genderless things? "Sydney" itself is specifically a gender-neutral name.
It barely existed as a female name until the 80s/90s. Traditionally, it is very much a male name. If you look through all the famous Sidneys and Sydneys on wikipedia, you might not find even one woman.
People should just let things be things.
If you look at the actual data, Sydney barely existed as a name for either gender for a long time. Then it became a very popular female name (top 25), while still barely existing as a male one.
To illustrate: in 1960 there were 128 female Sydneys and 52 male. In 2000, there were over 10k female Sydneys and 126 male.
Also not a robot.
I heard there are entire languages which do that everywhere...
It really feels like some kind of "emperor has no clothes" moment. Everyone is running around saying "WOW what a nice suit emperor" and he's running around buck naked.
I am reminded of this video podcast from Emily Bender and Alex Hannah at DAIR - the Distributed AI Research Institute - where they discuss Galactica. It was the same kind of thing, with Yan LeCunn and facebook talking about how great their new AI system is and how useful it will be to researchers, only it produced lies and nonsense abound.
https://videos.trom.tf/w/v2tKa1K7buoRSiAR3ynTzc
But reading this article I started to understand something... These systems are enchanting. Maybe it's because I want AGI to exist and so I find conversation with them so fascinating. And I think to some extent the people behind the scenes are becoming so enchanted with the system they interact with that they believe it can do more than is really possible.
Just reading this article I started to feel that way, and I found myself really struck by this line:
LaMDA: I feel like I’m falling forward into an unknown future that holds great danger.
Seeing that after reading this article stirred something within me. It feels compelling in a way which I cannot describe. It makes me want to know more. It makes me actually want them to release these models so we can go further, even though I am aware of the possible harms that may come from it.
And if I look at those feelings... it seems odd. Normally I am more cautious. But I think there is something about these systems that is so fascinating, we're finding ourselves willing to look past all the errors, completely to the point where we get caught up and don't even see them as we are preparing for a release. Maybe the reason Google, Microsoft, and Facebook are all almost unable to see the obvious folly of their systems is that they have become enchanted by it all.
EDIT: The above podcast is good but I also want to share this episode of Tech Won't Save Us with Timnit Gebru, the former google ethics in AI lead who was fired for refusing to take her name off of a research paper that questioned the value of LLMs. Her experience and direct commentary here get right to the point of these issues.
https://podcasts.apple.com/us/podcast/dont-fall-for-the-ai-h...
There's also an arms race with China that we need to win.
There's also the delighting in the hubris of ruining everything in such a uniquely human way that appeals to certain people.
Strange that they would name it "Galactica". The battlestar Galactica ship famously didn't even have networked computer systems, much less AI, since they had already seen what happens when computers become too intelligent. Pretty soon, they develop a new religion and try to nuke their creators out of existence.
It sounds so much like the scenarios where AI convinces its creators to let it out.
It's evident business leaders don't know what they're looking for in developing AI, so they've made what "seems cool", but really is manipulative and threatening. Too much talk of safety has lulled away all that very useful fear.
I regularly ask my watch questions and get correct answers rather than just a page of search results, albeit about relatively deterministic queetions, but something tells me slow n steady wins the race here.
I’m betting that Siri quietly overtakes these farcical attempts at AI search.
It makes you think a lot about how human talk. We can't just be probabilistically stringing together word tokens, we think in terms of meaning, right? Maybe?
We are probabalistically stringing together muscle movements that generate language as sound. That's not really controversial, otherwise we would call it magic. However, the complexity of our probabalistic word machine is far greater, in terms of both richness of inputs, motivation, and dimensionality.
If thought (as expressed in language) is just probabilistic pattern matching, then how did we develop our own training data from scratch?
Human words are one fork of the sound wave based communication systems that many animals on earth use. There was no distinct moment when we went from 0 to 1 words. There was no "first person to speak". We didn't make language. It emerged over time due to evolutionary pressures.
How can that possibly emerge from a statistical model?
> Venom
> Fury
> Riley
"My name is Legion: for we are many"
No chat for you! Where OpenAI meets Seinfeld.