Future of Artificial Intelligence

[Henry]

There's an ETF appropriately called IHAK if you want to make money on the fear of that distinct possibility.

[daylen]

You can't run, you can't hide, and you probably can't fight the drone terminators:
https://www.youtube.com/watch?v=EEXI6r08908
(unless you combine emp's with analog defenses)

Academic and industrial synergy:
https://github.com/simplescaling/s1

Introduction to LLM's from Andrej Karpathy:
https://www.youtube.com/watch?v=7xTGNNLPyMI

[theanimal]

Google introduced an AI co-scientist system designed to help researches create new and unique hypotheses and research strategies.

They validated the application with three real world experiments. In the process, the AI agent discovered the following:
-A new leukemia drug that then successfully tested in vitro at clinical concentrations
-New targets for further testing for drugs to address liver fibrosis
-Independent discovery explaining the mechanism for antimicrobial resistance. This confirmed and was the same conclusions of the reserach group, who had discovered the same but had not yet released the information into the public domain.

https://research.google/blog/accelerati ... scientist/

[sky]

The Future is Even Brighter than You Think

https://youtu.be/CCV_NU3wQkM

[Henry]

If you reduce the TSLA business model to its core, I think it's this: TSLA is a manufacturing concern that is developing AI which will create bots that will be able to build manufacturing facilities that create bots to building manufacturing facilities that create bots that can building manufacturing facilities that...

Now open your mouths and close your eyes, because Kool-Aid stings.

[chenda]

Now open your mouths and close your eyes, because Kool-Aid stings.

Do you see key person dependency a weakness in Tesla? If he has say an unfortunate accident or ends up in rehab or political winds shift in an unfavourable direction.

[7Wannabe5]

@chenda:

Aren't you forgetting about the optionality to be provided by the soon coming order for all fertile women in the U.S. to reply to e-mail with (1)first day of their last menstrual period, (2) location (relative to nearest mass insemination center?)

[zbigi]

With o1 and especially o3, I think we have already passed the 110 IQ barrier.

o3 still can't get basic integer multiplication right (https://www.reddit.com/r/singularity/co ... ce_by_oai/). That's something that you can train 80 IQ humans to do well.

[Henry]

Do you see key person dependency a weakness in Tesla? If he has say an unfortunate accident or ends up in rehab or political winds shift in an unfavourable direction.

Yes. And what's interesting is that I see the weakness as a potential existential threat not so much on the technology/business side but on the government regulatory side. The only comparison I can come up with is the role Henry Kissinger played in the Nixon administration. The difference is Kissinger was primarily foreign affairs and Musk primarily domestic affairs but they were/are tethered to the President himself and not really anything else. People get all upset when Trump makes his world wide wrestling pronouncements about rescinding presidential term limitations. If he really wanted to fuck with people's minds, he should pull down the mike and start shouting about rescinding the foreign born exclusion clause on the Presidential office. When you speak about the most influential non-US born citizens in US government history, outside of political theorists, Elon is only behind Kissinger, with Schwarzenegger a distance third, although he did get there by election and not appointment. At least to my recollection, that is. I apologize for the political segue, it was meant to be historical not partisan.

[chenda]

@7w5, that's true, the springbok's are coming...

[theanimal]

o3 still can't get basic integer multiplication right (https://www.reddit.com/r/singularity/co ... ce_by_oai/). That's something that you can train 80 IQ humans to do well.

That’s not showing what you think it’s showing. It doesn’t go up to 20x20 but rather 20 digits x 20 digits. For example, 6 digits multiplied by 6 digits could be 432683 x 253890. It gets that right with 100% accuracy. It is not trained as a calculator, but can use one. Safe to say that ability is not something someone with an 80 IQ or even a 100 IQ can do well.

[chenda]

If he really wanted to fuck with people's minds, he should pull down the mike and start shouting about rescinding the foreign born exclusion clause on the Presidential office.

Politics aside, it does strikes me as a rather odd bit of discrimination as citizens are all supposed to have equal rights. (Fun fact, Don's mum was born in the hebrides and grew up speaking gaelic)

[7Wannabe5]

@chenda:

Yeah, but history reveals that the counter-elite, especially when victorious, eventually eat their own. In the meanwhile, AI is proving itself capable of generating some pretty damn funny lewd satirical videos for us all to enjoy.

The biggest risk to TSLA and FAANG is that tariffs and failures of diplomacy are likely to lower incentive of international patent and royalty law adherence and similar which will interfere with the primary profit flows to Silicon Valley. Waaaaah, some Canadian hacker has unlocked secret codes for my products and Canadian government isn't interested in enforcement.

The truism "Nobody is above the law." reflect our universal dependency on its function. You don't have to teach elementary school very long before you notice that the bullies and the crybabies are frequently the same kids.

[zbigi]

That’s not showing what you think it’s showing. It doesn’t go up to 20x20 but rather 20 digits x 20 digits. For example, 6 digits multiplied by 6 digits could be 432683 x 253890. It gets that right with 100% accuracy. It is not trained as a calculator, but can use one. Safe to say that ability is not something someone with an 80 IQ or even a 100 IQ can do well.

I understand what the chart shows. As you can see, the errors start to appear when you multiply a three digit number by a three digit number. That's something an average elementary school kid can do (on paper).

[chenda]

Yeah, but history reveals that the counter-elite, especially when victorious, eventually eat their own...

Very true.

[daylen]

There is a jagged frontier of what models can do. Some simple stuff like spelling and counting (r's in strawberry) can be difficult for models whereas some stuff humans find difficult the models can do easily like reading and synthesizing lots of messy data.

If you let the models use tools (like calculators or code) then the frontier of what they can do is smoothened out a bit.

[theanimal]

I understand what the chart shows. As you can see, the errors start to appear when you multiply a three digit number by a three digit number. That's something an average elementary school kid can do (on paper).

With what accuracy? The “problems” it has only occur 2.5% of the time. Is that true for the average elementary school student? Is that true when the numbers get larger?

o3-mini also scores 14% (the highest ranking of all LLMs) on “humanity’s last exam. Check out some of the example questions.

https://lastexam.ai/

How many people can get just one of these right?

[Henry]

The fight has three criteria: who can create the technology, who can scale economically, and who can scale globally. People say "I'm not buying a TSLA because Elon only uses the numerical line on his keyboard when naming his illegitimate kids" but who the fuck walks around their house lifting every other damn item they own analyzing it that level. And why the fuck nearing 60 can't I spell illegitimate the first time I fucking try. Fuck that bothers the shit out of me. Anyways, the product wins. And we're talking trillionaires here. A loose knit cadre of cannabis dependent Canadians are not going to get in the way of these trillionaires extending their dominion into every nook and cranny of human existence. We all get turned out in the end. Or maybe, thanks to Steve Jobs, turned in.

[zbigi]

Computers don't make simple numerical mistakes. The fact that these language models do, means that they don't really "understand" the concept of multiplication (if they did, they could multiply arbitrarily large numbers with 0% error rate), and are merely guessing the answers (with pretty good accuracy). Not being able to comprehend how multiplication works shows that their IQ is pretty low. Or rather, that grading LLMs using IQ is a flawed approach, as these models don't show intelligence as we humans understand it.

[black_son_of_gray]

Computers don't make simple numerical mistakes.

Is that really true? There is a whole, albeit small, subculture of calculator geeks that delight in testing out a new calculator, trying simple and well-known computations that many of the devices cannot handle well...

E.g. is 1/3 + 1/3 + 1/3 = 1 or 0.9999? Sometimes these kinds of results aren't trivial.