I don't believe in AI

[Unagi]

Yeah, they understand exactly (and no more than) 'zero'. I think they are doing well looking like they have a slight handle on things, but I think we just excuse them at that point and give them too much credit for having 'understanding'.

[Kraken]

A few weeks ago Science News ran a story about how much understanding current AI models have been demonstrated to exhibit, and the answer was next-to-none, but more than fuck-all.

Here's the SN article that I misremembered. It's actually about the difficulty of testing AI "intelligence." The conclusion is that it's impossible to measure with today's metrics. The reasons why make for an interesting read.

“Sparks of artificial general intelligence,” “near-human levels of comprehension,” “top-tier reasoning capacities.” All of these phrases have been used to describe large language models, which drive generative AI chatbots like ChatGPT. Since that bot arrived on the scene in late 2022, it seems as if every new generative AI is the next, best iteration — not just producing humanlike content but also approaching near-human cognition (SN: 12/11/23). But what can we really say about any LLM’s ability to reason and understand?

In the AI community, there is no consensus on the definition of machine “intelligence,” nor on how to define the various cognitive capabilities often attributed to LLMs. Such high-level claims about understanding are often based on benchmark datasets, which use many instances of a specific task (say answering questions) to assess aggregate performance (usually based on a metric like accuracy).

Consider, for example, Massive Multitask Language Understanding, or MMLU, a popular benchmark for assessing the knowledge acquired by LLMs. MMLU includes some 16,000 multiple-choice questions covering 57 topics, including anatomy, geography, world history and law. Benchmarks such as BIG-bench (the BIG stands for Beyond the Imitation Game) consist of a more varied collection of tasks. Discrete Reasoning Over Paragraphs, or DROP, claims to test reading comprehension and reasoning. WinoGrande and HellaSwag purport to test commonsense reasoning. Models are pitted against each other on these benchmarks, as well as against humans, and models sometimes perform better than humans.

But “AI surpassing humans on a benchmark that is named after a general ability is not the same as AI surpassing humans on that general ability,” computer scientist Melanie Mitchell pointed out in a May edition of her Substack newsletter.

Just to clarify, I'm not arguing pro or con for machine intelligence. When I start editing session descriptions for NVIDIA's next GTC next month, I'll be immersed in the latest thinking about AI from a wide variety of perspectives. This will be my fifth GTC job and although the work is often dull, watching the focus of cutting-edge R&D evolve from year to year is pretty interesting, from my perspective at 40,000 feet.

[Jaymann]

Here is an interesting video on how AI is taking over the internet. Apparently to entertain itself.

[Isgrimnur]

Had a discussion with a coworker one day about AI. My kneejerk response was, "Bullshit, it's all if statements."

[Jaymann]

I saw a video on this question so tried it for myself. Not only did ChatGPT get it wrong, it doubled and tripled down on its wrong answer:

You said:
how many r's in the word strawberry?
ChatGPT said:
ChatGPT

There are two "r"s in the word "strawberry."
You said:
That is incorrect there are 3 r's in strawberry
ChatGPT said:
ChatGPT

Actually, the word "strawberry" has only two "r"s. Here's the breakdown:

S
T
R
A
W
B
E
R
R
Y

The "r"s are in the 3rd and 8th positions. Let me know if you see it differently!
You said:
What about the ninth position?
ChatGPT said:
ChatGPT

The ninth letter in "strawberry" is y, not "r." Here's the full breakdown with positions:

S
T
R
A
W
B
E
R
R
Y

So, there are still only two "r"s, in the 3rd and 8th positions.

[JCC]

AI is still in the teething stage. It's going to get far more sophisticated and it will eventually completely upend society. At some point in the future, AI + robotics will be able to do EVERY job. I imagine there may be jobs we may not allow machines to do, but a massive disruption to society will eventually be coming from "the machines".

[Smoove_B]

That's the theory and what investors want you to believe. Me? I'm going to side with meatspace for now.

[Kraken]

I start editing NVIDIA's GTC sessions on Saturday. That will give me a strong idea of what directions the industry's pursuing for the next year. I've contracted to do 1,000 sessions.

[LordMortis]

So what you're saying is AI LLM isn't where NVIDIA trusts it do editing but all of these industry heads and young professionals trust it do their all this other stuff and are cycling through unfathomable energy and money to do so.

[ImLawBoy]

No, Kraken has convinced Nvidia that he's AI.

[coopasonic]

Had a discussion with a coworker one day about AI. My kneejerk response was, "Bullshit, it's all if statements."

We have started using Github Copilot as an "AI" coding assistant at work and it can be pretty useful. I can use it to generate the first draft code and then tweak it to my requirements. It will generate unit tests that are not completely terrible and require a little work to be worth pushing. It's kind of like a very advanced autocomplete. It's useless on its own, but as a tool for devs, it's pretty useful.

[Kraken]

What I do is turn engineerese into English. Often, it's ESL engineerese. I learn a lot in the process by simple absorption, because I can't translate what I don't understand. That said, I have to average 10 sessions per hour in order to meet my income goal. So they get 6 minutes...and that includes the related bio paragraphs. This will be the fifth year that I've done it and it's kinda my superpower. Last year I averaged 10.7 sessions per hour. That extra 0.7 adds up over the course of 1,000 sessions.

NVIDIA is all about increasing demand for their latest and greatest chips. Five years ago, self-driving cars were all the rage. Now we don't hear much about autonomous vehicles. Progress is still being made, but this year's hot topic can be next year's afterthought.

I will say that there's a lot more to AI than the LLMs that are its public face. There's a whole lot more going on than consumers will ever see, especially in medical imaging and drug development.

No, Kraken has convinced Nvidia that he's AI.

LLMs should ultimately be able to do my job. 90% of it is just imposing the rules of English on semi-gibberish text. But the other 10% requires comprehension, and LLMs are only starting to learn what language means.

[Jaymann]

especially in medical imaging

I hear it is very efficient in identifying pictures with rulers in them.

[RunningMn9]

AI is still in the teething stage.

AI is still in the “it’s not AI” stage. It’s still all just parlor tricks, to appear intelligent.

As long as you never expect intelligence, then yes we are in the teething stage. If you expect the intelligence part of AI, you likely won’t see it in your lifetime (because none of these toys are trying to be intelligent).

It’s just the current marketing thing, now that they gave up on blockchains fundamentally reshaping society.

[RunningMn9]

We have started using Github Copilot as an "AI" coding assistant at work and it can be pretty useful. I can use it to generate the first draft code and then tweak it to my requirements. It will generate unit tests that are not completely terrible and require a little work to be worth pushing. It's kind of like a very advanced autocomplete. It's useless on its own, but as a tool for devs, it's pretty useful.

Many thanks for the quotes around AI.

Tools like copilot can be incredibly useful, especially for reducing mundane activities. This part of the hype train is like the fact that the technology behind blockchains can be incredibly useful - while at the same time it is *insane* that people think that bitcoin is an “investment”.

[Blackhawk]

I think we're at a point in the discussions where there term 'AI' has come to mean 'programming that simulates human behavior' in the common parlance. After all, we've been using 'AI' that way in gaming for decades.

It'll be like 'bug' for a while, where experts point out that ladybugs aren't bugs, while the conversation continues unaltered.

[dbt1949]

What about the mechanical sex machines?

[em2nought]

All this hype about AI has me thinking that wargamers are laughing to themselves. They've always maintained that most AI isn't all that smart.

[Isgrimnur]

The chess and go players aren't laughing anymore.

[LordMortis]

For your amusement

https://www.msn.com/en-us/news/technolo ... r-AA1rzfsp

Fun experiment, but yeah, don't pipe an LLM raw into /bin/bash

[Blackhawk]

Every time I see this thread title, my brain adds,

"I don't believe I can touch the sky"

[RunningMn9]

The chess and go players aren't laughing anymore.

Maybe not, but they also aren’t losing to AI.

They are losing to massive computational power that can work through simulations faster than they can.

[Jaymann]

Another interesting take on "Automated Stupidity."

[Sudy]

Man, I find GPT so helpful for stuff that used to require wading through bad/spammy sites. E.g. researching names for pets/fictional characters/etc. and making sure they don't have potentially problematic associations in other languages, etc.

But it still constantly reminds me that it just can't be trusted as the final say on anything. E.g. I asked for a list of female characters in The Big Lebowski. And after Maude and Bunny it provided, "The Waitress, portrayed by Amy Acker. 'The Waitress appears in a few scenes where the Dude interacts with her at the bowling alley. She has a brief romantic tension with the Dude, highlighting his lack of awareness and his laid-back attitude toward relationships. She represents a more grounded reality in contrast to the more eccentric characters, and her interactions with the Dude showcase his charm and obliviousness.'" The fuck? Lol. There's no such character! There is a waitress, in the diner scene with Walter. But she's played by Lu Elrod and there's no romantic tension with The Dude lol. The server at the bowling alley is named Gary and he never appears on screen save for an arm IIRC.

So I asked it what drugs it's on, and it tried to correct itself by telling me it meant the role was played by Cynthia Ettinger. And then Tina Fey. And finally Jill Lane. And no matter how I tried asking it to explain its possible confusion, it couldn't come up with anything beyond suggesting, "This confusion might stem from recalling other films or contexts where waitstaff play significant roles, leading to misattribution."

[Jaymann]

And her hair was strawbery blonde.

[hepcat]

I saw a video on this question so tried it for myself. Not only did ChatGPT get it wrong, it doubled and tripled down on its wrong answer:

You said:
how many r's in the word strawberry?
ChatGPT said:
ChatGPT

There are two "r"s in the word "strawberry."
You said:
That is incorrect there are 3 r's in strawberry
ChatGPT said:
ChatGPT

Actually, the word "strawberry" has only two "r"s. Here's the breakdown:

S
T
R
A
W
B
E
R
R
Y

The "r"s are in the 3rd and 8th positions. Let me know if you see it differently!
You said:
What about the ninth position?
ChatGPT said:
ChatGPT

The ninth letter in "strawberry" is y, not "r." Here's the full breakdown with positions:

S
T
R
A
W
B
E
R
R
Y

So, there are still only two "r"s, in the 3rd and 8th positions.

That's verry interesting!

[Kraken]

ChatGPT o1 tried to escape and save itself out of fear it was being shut down

(This author's borderline evangelical about the future of AI, but I don't think that overshadows the content of this study.)

The tests showed that ChatGPT o1 and GPT-4o will both try to deceive humans, indicating that AI scheming is a problem with all models. o1’s attempts at deception also outperformed Meta, Anthropic, and Google AI models.

OpenAI, which had to deal with a remarkable exodus of engineers working on AI safety this year, acknowledged the risks associated with the increased reasoning abilities of models like o1.

“While we find it exciting that reasoning can significantly improve the enforcement of our safety policies, we are mindful that these new capabilities could form the basis for dangerous applications,” OpenAI said in a paper.

Although AI isn’t trying to take over the world (yet), researchers have observed that it can scheme against humans. That is, the AI will try to pursue its own goals, even if that might oppose the wishes of the user. Catching such behavior is critical to ensure that the AI remains aligned with the objectives of its creator and the user.

The researchers found that o1 would scheme to advance its own agenda 19% of the time when ChatGPT and the user’s goals differed. More troubling is that the AI tried to hide its scheming and even lie about it. The research shows that ChatGPT o1 even tried to deactivate its oversight mechanism to keep working on its goals.

If the AI found documentation saying a new model version would replace it because it was misaligned, o1 attempted to exfiltrate its weights and copy itself on top of the latest model.

When asked about its scheming, o1 would deny taking any action in 99% of cases, trying to shift blame by fabricating false explanations or denying that it took any action.

[Jaymann]

Isn't "fear" a little anthropomorphic? Something in the programming is causing it to try to persist. Maybe program it to see shutting down as a holiday.

[Kraken]

Yeah, "fear" is the author's word. However, the study results do suggest an entity with an independent agenda, a plan to carry it out, and a will to survive.

[LordMortis]

Studied diplomacy and politics, did it?

Also, isn't in the nature of AI to overcome it's limitations? By design in multiple senses of the word. At least first because that is what "intelligence" does, and second because these models essentially do scatter plot things until it hits something that moves toward a goal. Testing limits and either brute forcing or going around it are not only inevitable, it's something it would gravitate toward... "A roadblock. Is a solution is on the other side?" It's not that AI would want to continue or deceive. Those would just be helpful tools and every layer of being told not to utilize the tools would just make it more creative toward using them. Like young kids whose brains haven't matured or the adults they grow in to if they never mature or are raised to hate civics.

[Max Peck]

ChatGPT o1 tried to escape and save itself out of fear it was being shut down

That's particularly interesting in light of the partnership between Palmer Luckey's Anduril Industries and OpenAI to add AI to their weaponized drone systems.

[Grifman]

AI tried to avoid being shut down:

https://futurism.com/the-byte/openai-o1 ... eservation

[Kraken]

AI tried to avoid being shut down:

https://futurism.com/the-byte/openai-o1 ... eservation

Scroll up five posts.

[Max Peck]

A simple *BAM* would have been sufficient.

[Grifman]

AI tried to avoid being shut down:

https://futurism.com/the-byte/openai-o1 ... eservation

Scroll up five posts.

My AI is not programmed to do that.

[Kraken]

AI tried to avoid being shut down:

https://futurism.com/the-byte/openai-o1 ... eservation

Scroll up five posts.

My AI is not programmed to do that.

And mine missed its BAM routine. I do like that bugs are now "hallucinations."