AI has attracted John Oliver’s attention this week (NSFW for language).
I was talking to my buddy this past weekend about their company chatbot. He said the queries are either ridiculously simple, somebody hunting and pecking on the keyboard, or, so complex they need a person.
Apparently one of the most common queries is ‘how do I speak to a person?’.
Now you can not understand how they’re worth $4 trillion…
1 Like
Yesterday, Grok 4 was released, topping a lot of benchmarks. While some of them could be memorized, pirated, or simply faked (knowing Elon) others are hidden/independently verified, so it’s probably the smartest model on the market. And is (modestly) pushing the envelope on what is even possible with an LLM.
On the same day, the Grok’s Twitter account started worshipping Hitler, and posting lots of extremely racist and antisemitic rants. It was then taken offline. Though I think we can all fairly assume Musk is going to keep working at nazi-fying Grok.
That’s not to suggest these two things are the related. My best guess is Twitter Grok was either system-prompted or somehow post-trained to be a nazi edgelord. And Grok 4 is not.
But I do think we have to deal with the fact that it’s not too hard for our Oligarchs to turn their bots into raving nazis. And these raving nazi bots might at the same time become SOTA at real life things like coding and math.
It’s of course based on the use of their chips on AI, and assuming that AI is taking off, which of course it is in a variety of ways. Their chips dramatically reduce the computational time for AI type calcs.
i.e a lot of AI stuff relies on vectors, and doing stuff like takeing the cross product (or maybe the dot product, FFS it’s been years). So instead of doing a loop to do the y=w1x1+w2x2+…and do that a hundred thousand times, you just say z=x(dot product)y. One line, and its calculated using parallel processing. So, that can save orders of magnitude the amount of time required to generate a model.
There’s a couple other similar calcs that are used all the time in AI model generation, and they take a ton of time on traditional processors, and with nvidia and similair, it’s like three lines of code and boom.
Frankly the more I get into this, the less I’m finding that this is all some big mystery or magic. It’s just taking data, and fitting models, then using the models to generate results. So far I haven’t seen any math involved that would be beyond the immediate scope of anyone here. There’s some CS stuff where there’s some new techniques to take advantage of the chips, but its little more than what I described above.
It’s interesting that everyone seems to find it so innovative. I think because it got called AI instead of ‘building mathematical models for complex spaces, then implementing them’.
To the limit I have read, nobody is surprised by the techniques. But everyone is surprised that the results are as good as they are.
I did lousy in linear algebra…
ETA… It’s probably useful to note here that linear regression is considered to be a form of AI by the Treasury Board of Canada.
I Gauss that could be true…
Huh.. I’m not overly surprised myself, personally, at least not for what I would expect many of the models would be. I mean, testing fit of data to models is pretty well known tech. Maybe for some of the crazy models like LLM’s and the image models it’s surprising. But for something like underwriting, seems like there’s enough data available to be able to build a pretty good model.
In any event, I’m still learning, there’s I think a whole area of testing fit of models that I’m not up to speed on yet but am looking forward to reviewing.
Where I am amazed at is all the potential applications. They’re bloody everywhere. I mentioned I’m on the edge of releasing a chatbot (which is barely an AI thing anyway), but already have thoughts of what I want to do with v2 (allow it to do a decent job of estimating underwriting outcomes). And then today I get a call from a developer, buddy’s working for some toronto life insurance agent and building an AI that lets the consumer put in some rough details and it’ll run quotes and suggest the best company. They’re going to fail, spectacularly given what he was telling me - he had a spectacular lack of knowledge. He was asking for access to my term life insurance quoting system, told me he just needed that, then he was going to quickly add on whole life and universal life and boom, done. Instant quotes. For the P&C folks, the short answer is, you can’t quote universal life at all, and unable to quote most whole life policies. It’s just not possible, there’s not enough info. For us, I do expect to be running term quotes from v2 of our chatbot, but I’ll be the only one doing it.
In any event, I’m rambling, but the one application I’m going to be looking into is the use of AI in sales. Because if we can refine out immedate opportunities, that makes the business case immediate. I’m sure it’s there, I just don’t know enough about it yet.
2 Likes
Obligatory Simpsons gif:
2 Likes
Grok, at least what I have observed, seems to be pretty good at analyzing posts/videos/or whatever and providing a useful responses or explanations. But it still feels like just a fancy web search.
Maybe it solves a disinformation problem on social media if it becomes trusted, but also, very easy to be abused. I suspect the nazification was closer to a functional test of seeding the AI to promote an agenda. It would be widely reported providing the team with very quick feedback.
I’m not sure what you mean by test. A test would be to see if something works so you can do something else later. You might do something privately. Or for certain users. Or do a less controversial or less extreme test. But in this case, it was public, for everyone, and as controversial and extreme as possible.
It also wouldn’t be that hard to do a private test. You could make a Nazi grok and feed it already existing Grok tweets, and see right away whether or not it worships Hitler.
And I see this as a follow-up to when twitter Grok started randomly bringing up “White Genocide” for no reason. X reported that they were hacked or something, but it seems pretty obvious that Musk had just told it about his favorite conspiracy in the system prompt.
I might call what they are doing rapid prototyping. But I do think the goal is a Nazi AI. Or at least a Musk AI.
Also apparently Grok 4 has an interesting approach to research.
Yeah, we have all taken our models and put in 1 and 0 in as input values to see what breaks. Like for the discount rate in a reserve calc. That’s what he did.
This is approach is pretty ludicrous, creating a chat bot to parrot your own views and call it AI.
Should also add-- Musk isn’t the first to hamfist the system prompt.
A year and a half ago, Google told their image bot to be diverse. The AI took this to a cartoonsih extreme. If you asked it for any white person, it would refuse on grounds of diversity, and generate black/brown people instead. If you asked for any historic figure, it would turn them non-white and maybe female.
While I know Musk is a moron who wants to see the world burn, I was a bit surprised that Google was so stupid to go live with that.
“I can call you Betty…”
“More like ‘Artificial Me,’ because I don’t want to answer your questions 60/60/24/7/52/365/1000.”
The Atlantic is reporting that Grok recently said White, Asian, and Jewish are the “Good Races”
Musk said Grok went full Nazi because it paid too much attention to X posters
He specifically asked users to provide true but politically incorrect statements last month, to train Grok with. So maybe they added Nazi comments there.
Though it wouldn’t surprise me much if being told to emulate X users allowed a few to game it into mechahitler.
Btw, I like how in that article, X uses the term “deprecated code” to refer to Musk’s old prompt of telling the bot to be “offensive” and “maximally base”. That’s rich.