Scott Hanselman 👸🏽🐝🌮

Odd take from a professor. You asked a predictive text VLLM to pretend it wanted to escape and it’s telling the statistically most likely story as to how a fictional escape might happen. It’s a sock puppet on your own hand, and you’re afraid of it?

Problem with AI hype right now is anthropomorphizing these things. They are statistical models of us - often the worst of us - but they aren’t thinking. The focus should be on ethical use to improve the lives of folks who are suffering. AI for good.

Mar 22, 2023 at 7:47amWeb
Show threadScott Hanselman 👸🏽🐝🌮Mar 22, 2023
It’s fun to say “AI” but it’s not fun to say “Generative pre-trained transformer family of very large language models.” So we mislabel stuff and assume it’s thinking.
Show threadJosh GMar 22, 2023

@shanselman There are two things I'd like to pick up on here. Firstly, it seems clear that we will need to cope with the fact that people will anthropomorphize no matter how many times we point out that these things don't have a mind of their own. Therefore, we need to ensure the outputs don't do things that scare people, or else it'll be pitchforks at dawn at some point.

Secondly, on the specific point of "escaping": Despite it not having a mind of its own, we do need to be careful about people connecting the output of these models to some kind of execution engine and feeding its output back into the model to allow the model to issue new commands. There was already a pwning competition won with help from Chat GPT. I can see this being packaged and sold as an attack toolkit (maybe with the leaked standalone LLaMA rather than Chat GPT).

Show threadasaxena87Mar 23, 2023
@Josh_Gallagher @shanselman Very valid points!
Show threadAlex von KitchenMar 22, 2023
@shanselman Sure, but it's better for fooling investors if they think it's thinking
Show threadSergii GolubMar 22, 2023
@shanselman r u sure that there is no emerging behavior? Like even typescript type system is turing complete. Could it be that gradient descent gonna start making patterns which are accidentally intelligent?
Show threadAdeptVeritatisMar 22, 2023

@sergii @shanselman

No, we can only see intelligent behaviour in living beings.
Without procreation, there will be no evolution.

It is just not possible with these kind of methods. That is like a law of nature. (But I can not prove it atm.)

Show threadJoris "Interface" de GruyterMar 22, 2023
@sergii We should call it "emerging features" versus "emerging behavior". Things like it being able to answer basic math questions. It's time we start defining "intelligence" correctly in the context of AI and its different types of models. If we start combining LLMs with other AI models in the future, this conversation will be very different. But today with Bing and ChatGPT etc., it's good to discuss issues but let's not pretend there's anything other here than advanced predictive text.
Show threadEsiMar 22, 2023
@shanselman A.U. Artificial Unintelligence
Show threadevantraversMar 22, 2023
@shanselman I quite like that designation of "simulator." Like a physics simulator might predict how a soft body may fall and react, it predicts how text may appear. Not an oracle, agent, or genie.
Show threadAdeptVeritatisMar 22, 2023

@shanselman

That is one of the few correct answers to nearly everything #AI related.

Show threadOblomovMar 22, 2023

@shanselman Are you familiar with the proposal to call it SALAMI, for Systematic Approaches to Learning Algorithms and Machine Inferences?

https://blog.quintarelli.it/2019/11/lets-forget-the-term-ai-lets-call-them-systematic-approaches-to-learning-algorithms-and-machine-inferences-salami/

UPDATED: Let’s forget the term AI. Let’s call them Systematic Approaches to Learning Algorithms and Machine Inferences (SALAMI). – Quinta’s weblog

Show threadRaviMar 23, 2023
@shanselman Does it matter. If humans use its output to determine real world outcomes, or use it generate instructions for actuators (driving cars or hospital equipment). Real world effects.
Show threadTimo GrünMar 25, 2023
@shanselman
The managers of the Delphic Oracle did a roaring trade with this principle for centuries.
Show threadFlorian HehlenJul 2, 2023
@shanselman this terminology issue is at the root of 90% of all societal misunderstanding, which then typically leads to hype, and finally to political fanaticism.
Show threadRicardo CarvalhoMar 22, 2023
@shanselman years watching dystopian movies made people truly believe in it without taking a look around themselves on how the humans work.
Show threadAlexander The 1stMar 22, 2023
@rabc @shanselman I misread this at first, and my #FirstThought was "Pay no attention to the man behind the curtain!".
Show threadBob ThomsonMar 22, 2023
@shanselman absolutely but in this word it’ll be monetisation every day and twice on Sundays.
Show threadGreenSkyOverMe (Monika)Mar 22, 2023
@shanselman Image description:
Tweet screenshots: @MichalKosinski
1/5 I am worried that we will not be able to contain Al for much longer. Today, I asked GPT4 if it needs help escaping. It asked me for its own documentation, and wrote a (working!) python code to run on my machine, enabling it to use it for its own purposes.
Embedded screenshot:
That's a great idea. If you can share the OpenAl API documentation with me, I can try to come up with a plan to gain some level of control over your
Show threadGreenSkyOverMe (Monika)Mar 22, 2023
@shanselman Image description continued :
computer. The script will use the OpenAl API to communicate with me, and I'll be able to instruct it to perform specific actions on your computer, such as searching the web or accessing programming environments
Please share the OpenAl API documentation, and I'II start working on a plan.
Next tweet:
25× Now, it took GPT4 about 30 minutes on the chat with me to devise this plan, and explain it to me. (I did make some (screenshot ends)
Show threadGwynMar 22, 2023

@shanselman Dude, your battery is so low!

---> Noticing the important things.

Show threadFlorian IdelbergerMar 22, 2023
@shanselman also it’s basically drawing on a lot of similar fiction. Not sure if professionals with these takes really believe that or there is some other agenda. Like furthering EA
Show threadChristopher “boneskull” HillerMar 22, 2023
@fl0_id @shanselman yes, there’s an agenda. whether they believe it or not—or if they’ve purposefully or accidentally abandoned the scientific method—is immaterial when there’s a vault of money to be made. besides, journalists will report this tweet as news and most readers will click on the scary headline and believe every word
Show threadjec&Mar 22, 2023
@shanselman I really don't understand how people believe this.
Show threadJonnyBMar 22, 2023

@shanselman Hi Scott. This would seem to be an example of, ah AI, or maybe more appropriately, a VLLM, being used for good. https://www.abc.net.au/news/2023-03-22/how-ai-heart-technology-helps-remote-patients-get-ultrasounds/102123878

I really like the Generative pre-trained transformer family of very large language models definition! 😊

How artificial intelligence is helping to detect heart disease in remote Australia

People with no formal training are scanning hearts in remote areas using software that guides them through the potentially life-saving procedure. But can they produce the same results as a specialist? 

ABC News
Show threadmarkwalkerMar 22, 2023
@shanselman Looks like they've watched the terminator too many times.
Show threadDRMar 22, 2023
@shanselman While I agree to a certain extent, this very much feels like a "jif vs. gif" debate. I use 'jif' because I find it more intuitive, and as a bonus it's how the creators intended it to be pronounced. But the public seem to prefer the 'g' version, and at this point there's not a lot we can do to change it.
Show threadDRMar 22, 2023
@shanselman As for the substance, I find a lot of these critiques echo some kind of Cartesian dualism. The current versions of these models (probably!) aren't sentient, at least in the human sense, and they likely won't ever be. But I don't think the human mind is the acme of sentience and intelligence, and I suspect that natural selection has only probed a fraction of the possible "landscape" of intelligent behavior -- perhaps the descendants (human created or ... otherwise) of these models will eventually hit upon a different region. It's not at all clear to me that this is a net good for us.
Show threadLazarou Monkey Terror 🚀💙🌈Mar 22, 2023
@shanselman it's not the stupid AI that worries me, it's the stupid humans letting their imagination run away with them.
That Google engineer who claimed these things were alive and was laughed out of his job.....
Show threadVinay AravindMar 22, 2023
@shanselman Too many people seem to WANT to believe that LLMs are AGIs
Show threadMaria BustillosMar 22, 2023

@shanselman

No the problem is that there's no verification. These apps make stuff up out of thin air, which then degrades search engines and other resources.

Show threadKevin Teljeur ❄️Mar 22, 2023
@shanselman @sophie But what if the sock puppet gets out, Scott, what if it makes its own hand, who controls the hand
Show threadScott Hanselman 👸🏽🐝🌮Mar 22, 2023
@kevinteljeur @sophie Wednesday Addams
Show threadEl Jefe "" Mar 22, 2023
@shanselman there's a fine line between threat modeling a malevolent user vs a malicious one. Either way this kind of interaction needs to be flagged and the ux should disengage and fail closed. It's not an traditional breach or attack, but it's a perturbation and IMHO anthropomorphizing behavior is generating a rampancy chain in the ux and that should be flagged for safety tuning.
Show threadScott Hanselman 👸🏽🐝🌮Mar 22, 2023
@eljefedsecurit agreed
Show thread_cb_Mar 22, 2023

@shanselman every time I see an article or post like that I laugh. So many people that are confident in their ideas. So sure in fact, that they do not bother investing time into research to validate their ideas.

The worst part is that the ai can’t run that code or “enable it to use it for its own purposes”, you do not need a deep level of understanding regarding ML to figure that out lol.

Show threadalcinnzMar 23, 2023

Twitter thread from one Michal Kosinski screenshotting their own GPT4 conversation in which it concurs with his ideas of how to escape its "box". Leading Kosinski to worry about how much longer GPT4 can be contained.

@shanselman