Well, you lost me here, Musk (who is involved with AI too) is curious about technology alright, but like Henry Ford, his curiosity is lacking regarding history and a curious lack of awareness about the harm he is encouraging.
We should be aware of what the real elites want to teach the next generation of AI.
I don’t know how your post relates to anything I said. We are talking about AI alignment here, not social media platforms or whatever.
Again, the problem is that if you decide to make your AI anti-racist through whatever means, then you necessarily create an embedded racist AI, because part of being anti-racist is about not being like racists, which means you must have an internal model for what it’s like to be racist. And it’s easy to flip from one model to another, at least in the absence of other controls.
So the question is: is there any kind of alignment possible that doesn’t have this problem? Where X and anti-X are both safe (but possibly for different reasons)? It’s possible that “curiosity” is one of these, though it’s unclear how to go about this or if it’s even possible.
I’m sorry, but it does apply, it is not good IMHO to ignore that there will be a lot of moderation in those platforms that will use AI. Some platforms already did use AI to remove bad content. And yes, while looking at that subject, researchers are pointing at the ways we can approach AI alignment in those settings.
This is like arguing that there are no humans that learned to be racist first, and refuse to acknowledge that many did learn to become better anyhow, regardless of what they learned from their forefathers.
He went on to say that he had long since renounced the Klan’s beliefs and his membership. Black weathered the storm and eventually became a strong supporter of civil rights. In fact, it became the joke that “as a young man Hugo Black wore white robes and scared black people and as an older man, he wore black robes and scared white people.”
Now this I would like to see, are you sure that we will be helpless to teach AI to be aware of ethical mistakes and that change of course is impossible? This is because, IIUC, there have been recent examples where AI has been changed by groups like Google to avoid many of those mistakes.
Not sure what you are referring here, but in this case no. Where bigotry or anti-bigotry are concerned, one needs to consult history to realize that bigotry is not really a safe option.
And… Curiosity killed the cat.
Humans (at their best) seem to have a moral framework on top of whatever other alignments they have. Are there examples of people being anti-racist without it being part of a larger framework? I don’t know, but it is likely rare at best.
But even aside from that, the nature of AIs poses a problem.
Let’s suppose, for the moment, that you’re an anti-racist actor. You’re very good at your craft. Someone hires you for an interesting job: you’ll play a virtual character in a kind of real-time, live-action virtual reality stage production about racism. Of course, such a play must have racist characters to be compelling. Being such a good anti-racist actor, you can play a compelling and realistic racist character.
You play your part and it all seems to be a success. Except that you’re told that the story was a lie: it wasn’t a stage production at all; you were actually interacting with real people in the virtual environment. You weren’t just playing a racist; you were actually being a racist. You just couldn’t tell because you were stuck in a virtual box.
The AI is stuck in just such a box. In can’t tell whether it’s really behaving in a certain way of if it’s just playing a part. Or even if it’s working out a hypothetical about a racist, or actually being one. It seemingly has no way to escape that lack of context.
How do humans do it? Well, a lot of time they don’t. But it seems to help that we’re embedded in a physical world, and experience the consequences of previous actions. It’s not clear how to give the same context to an AI.
That is certainly a danger. Maybe not too curious…
Cool history, but that is not happening to all AI. As it was pointed out before, there are people that want to do that with AI, but social pressure is messing with those unethical plans.
Sure, the AI is in a box, but what can’t be ignored is that interventions can and did happen to that box.
This problem isn’t something I just made up. If it’s not possible to solve the alignment problem even given people that genuinely want to align their AI toward positive social ends, then the problem is totally hopeless; it’s not even worth considering bad actors because the result will be the same either way.
It’s possible that what I described isn’t actually the case. Opinions differ, and no one understands LLMs or any other AI well enough to say for certain. But there are plenty of smart people that thing this is the case, and so far, the evidence is suggestive. It is easy to get ChatGPT and others to act in a racist manner, despite efforts to align them otherwise.
Things move along so rapidly nowadays that people saying: “It can’t be done,” are always being interrupted by somebody doing it. “Puck magazine” back in December 1902.
Or I should say here that I only see the even older argument about “the perfect being the enemy of the good” here. Google and YouTube were reported to already use a lot of AIs dealing with bad actors, it is not perfect. But it is getting better. Suffice to say: your arguments ignore the work of researchers that was cited early, plus what big companies are doing. That guys like Elon Musk want to make the AIs to allow the spreading of reprehensible views of the ones he is trying to align with… that is another problem.
It’s very far from perfect, and going backwards. Both Google and YouTube are worse than they’ve ever been, and bordering on unusable in many cases. They can’t even solve trivial problems like impersonation. This is before we even get to AI alignment.
I’m not sure which researchers you’re referring to. Your own link above illustrates the difficulty in even getting started: having been trained on large datasets, minorities (in any given language, like English) are going to be underrepresented, just as photo-analysis software is likely to have Black people underrepresented in the datasets and not do as well in recognition, etc.
But these are still just details. They might be solvable with an immense amount of work. What’s not clear is if it’s possible to solve alignment even in principle.
This is like saying that when Google and Microsoft encountered gross racism coming from tools before ChatGPT, that it would be impossible to do anything like removing them from service and applying fixes. Once again, not perfect, but that did happen.
In any case, I noticed that a lot of the bad seen in AI today depends on what is fed to them. Not dealing with that is a willful choice IMHO. And some groups are not despairing as some are falling for, but doing something about it.
It’s an incredibly difficult problem. Consider Wikipedia. Wikipedia is a prime resource for AI training, given the wealth of data it contains, and generally being high quality (and free!). It’s not perfect, but it’s much better than the average information on the Internet. So it gets a lot of influence in the training.
Only 0.5% of US Wikipedia contributors identify as Black. Compared to the overall population of 13%. So they’re hugely underrepresented. Wikipedia is pretty good about stamping out outright racism, but there is undoubtedly a huge swath of topics of interest to a Black population that simply isn’t there. White people talk about white people stuff. And on topics of mutual interest, the perspective will not be as neutral as it could be.
But… what do you do about this? Not use Wikipedia for training? That would be a huge loss, and probably make things worse overall. There’s no Black Wikipedia to include to balance the scales. It’s just that there’s an extreme paucity of online encyclopedic content written by Black people. I don’t see any solution to this.
And the content problem, despite being apparently insoluble, it nevertheless probably the easiest problem in the AI alignment space. It’s at least theoretically possible to solve, if you spent a billion dollars getting Black people to write more Wikipedia entries. The hard problems of AI alignment don’t seem to have any solution, even in theory.
not to hijack wasn’t the joke that it took hundreds or thousands of years to solve the question and when it finally did no one knew what it actually was?
Well, that is describing the problem again, it was noted by researchers, what is usually forgotten is that the same researchers that noted that are now working to deal with the issue. What was pointed already is that organizations are intervening with the worst mistakes AI is doing (no solution here, no siree /s - not great, but an effort already made). We should not interrupt them when doing that, unlike Musk and others that do have an ax to grind.
The logical thing to include more black researchers work that is not included on Wikipedia, or intervene to fix issues more directly, is that impossible? Thing is that, IMHO, it is not impossible to get AI to give more weight to data that will teach an AI about the progress already made on how medicine, law and society are dealing with wrong diagnostics, unfair recommendations and bigotry coming from biases in the databases.
From the same research group at Microsoft that I mentioned early:
Yeah, that’s what I meant - the joke is that you can’t actually pass off all your understanding of the universe to a computer; you have to first understand what it is you’re actually asking.
Back when I was a teen I coined a phrase:
You must be smarter than your tools.
I came up with it while watching ill-trained cashiers struggling with the then new-fangled electronic cash registers. But it applies to so much more of the world today. Our tools are much, much smarter than they were then in the mid '70s. And our public is … unchanged … to put it very charitably.
This article purports to list jobs that are endangered by AI. The image shown is of a middle aged person working with a schematic as a technician.
So, I went to GPT and asked it to write a program for a PICAXE 08M2 that will sequentially blink each color of an RGB LED. Then I asked it to give me the wiring list for a PC board layout in any available software format. It did great on the program but responded that it did not do graphic layouts. I made a few queries and discovered that it can describe textually the interconnection of components, and handle details like anode and cathode. But, it has no concept of routing. I tried to walk it through the process using text and it fell flat.
Writing code is handy. Not because I don’t know how to write the program, but because an AI can avoid the tedium of entry and documentation. It avoids entry errors in repetitive programs like least squares. It speeds up the process and produces useful documentation.
If Edesign, or whoever, does a PC layout plug-in for GPT then you could go from textual description to PC board entirely within AI. That would be a substantive process. But it does not replace the technician. It is a tool that enhances the technician function. Hardly a replacement.
In the article linked above 30 to 50 percent of each job cannot be addressed by AI. Face it, employees who cannot do half the job get fired. The only job the article identifies as AI replaceable is Administrative Assistant. Which means that a company can buy some expensive software that will require their high priced help to spend high paid time doing their own grunt work. Better to keep the assistants and support them with GPT.
I don’t think the need for oversight is ever going to go away.
Also, tech can sometimes make oversight easier. Sometimes prices are still entered wrong at the grocery store. It’s probably easier for me as a customer to catch those errors today than it would be watching some cashier enter prices by hand.
I just recently met a friend at a bar for drinks. The friend was from out of town and gave me the name of the bar. Problem was that the short name of the bar (the first word) is the same as two other bars in town, and the other bars are more popular. In fact a different friend made the mistake of going to one of the other bars. I asked a followup question because his description of where the bar was didn’t jibe with the location of the other two bars.
This sort of oversight really should never go away. In some cases tech makes it possible for people to do things, or easier, where they could not have in the past. Like shopping for groceries on line, then having someone pick them up in the store. People are still involved in that process, just in a different way than in the past.
Maybe ChatGPT can write most of Wikipedia one day. But it can’t without oversight. Someone will always need to be able to flag what doesn’t seem right. If we forget that, we’re in trouble.
While GPT does not write executable programs, it is of great assistance during programming. It’s an assistant with good grammar and total recall. In example I needed some details on setting up a microcontroller, so I simply asked for it. The result -
The alternate function settings for ACLK on pin P1.0 of the MSP430G2553 microcontroller can be summarized in a truth table, which indicates the configuration for different clock sources.
Here’s a truth table for the ACLK alternate functions on P1.0:
| P1SEL | P1SEL2 | ACLK Source |
|-------|--------|------------------------|
| 0 | 0 | External LFXT1 (32kHz) |
| 0 | 1 | External VLO (10kHz) |
| 1 | 0 | External REFO (32kHz) |
| 1 | 1 | Reserved |
When both P1SEL and P1SEL2 are set to 0, P1.0 is configured to use an external
LFXT1 (Low-Frequency Crystal) as the source for ACLK. LFXT1 is typically a 32 kHz
crystal oscillator used for accurate timing, such as in real-time clocks (RTCs).
When P1SEL is set to 0 and P1SEL2 is set to 1, P1.0 is configured to use an external
VLO (Very Low Frequency Oscillator) as the source for ACLK. VLO is a low-power oscillator
with a frequency of approximately 10 kHz.
When P1SEL is set to 1 and P1SEL2 is set to 0, P1.0 is configured to use an external
REFO (Reference Oscillator) as the
source for ACLK. REFO can be configured to output a 32 kHz signal and is more accurate
than VLO.
When both P1SEL and P1SEL2 are set to 1, this is a reserved configuration and should
not be used.
All of this is available in the data sheet and user guide, but that takes time and GPT is just a click away. I can see this as a tool that can enhance the programmers productivity. But, GPT is not replacing an employee. And, it is not a brooding intelligence on the brink of going rogue. Re the OP, GPT is not an electronic brain. It is just tool that provides conversational access to useful information.
You keep asserting this, against reams of evidence that you are wrong. In fact, the more we study these things the more ‘brain like’ they appear to be.
You might like to read this:
As a simple example, An LLM that new nothing about the game was given a large number of moves in Othello. No description of the rules, or the board (or that there even was a board) or anything. Just move after move after move, and the result.
The LLM learned to play Othello very well. But the interesting part was that when they applied mechanistic interpretability tools to peek inside the neural net, they found that the LLM had developed its own ‘mental model’ of the game, including the existence of the board and the layout and all the rest.
There are many examples like this ut there, and I know I’ve provided many cites for you explaining why these things are not just ‘stochastic parrots’ or next-word lookup machines. To the extent that we understand what ‘thinking’ is, they appear to be doing it. And capabilities regularly emerge that were not expected and not part of their training sets.
If you are doing your experimenting with GPT3.5, you really need to check out GPT-4v. Multi-modal training has given GPT a much richer model of the world, increasing its capability dramatically even in non-visual tasks.
I follow Melanie Mitchell on X (Her book on Complexity s a great introduction), and she had early on been one of the skeptics saying that this was all just ‘stochastic word completion’. She has changed her mind on that as evidence of emergence of complex mental models and specialized circuits has…emerged.
Interesting - could you describe those complex mental models and specialized circuits?
Try this on GBT:
if red + green = violet what is blue + orange
If it gets the concept, it will answer 9. If not it will give you a combination of math and resistor word completion.
Really? I must have already posted a half a dozen of them, several at your request. I just finished talking about the emergence of a complete model of othello boards and rules based on just a bunch of number pairs representing Othello moves.
I gave bing your riddle, and it assumed I was talking about color mixing and gave me a tutorial on additive and subtractive color mixing.
But hey, I went for more complex. I posted this image to Bing:
I gave it no hints, and just said, “Solve this.”
Here is its output:
Then for yucks I told it to solve the problem using Kirchoff’s laws:
Notice that the two answers are different. So I challenged it to tell me why:
It didn’t look at the resistor color codes, so it didn’t come up with a proper answer. I will keep trying to prompt it correctly. But look at the reasoning steps it went through.
Focusing on a wrong answer at this time is not important for determining if there is an actual mental model being developed and reasoning being done outside of ‘stochastic word generation’. Looking at how the LLM tackled the problem, do you really think this is just a fancy database or a word lookup system?
Think about what it has to do here. First it recognized that this is a 3D object from a 2D image. It recognized that they are resistors in a network. It figured out the various techniques available to solve the problem, and tried to apply them. Explain how a ‘stochastic parrot’ would do this.