Gemini

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

User avatar
towforce
Posts: 11701
Joined: Thu Mar 09, 2006 12:57 am
Location: Birmingham UK

Re: Gemini

Post by towforce »

smatovic wrote: Mon May 20, 2024 5:53 amSuperintelligence: Paths, Dangers, Strategies
https://en.wikipedia.org/wiki/Superinte ... Strategies
Actually, I have read that book: I bought it straight away and read it avidly when it came out 10 years ago. However, in the world of AI, it has been a long 10 years. In 2014, Alpha Go had not yet beaten Lee Sedol, and there were no strong chess engines using NNs. LLMs, an everyday thing that we all take for granted and use for free every day (some people are paying, and I might in the future), did not seem plausible. It feels like a lifetime ago when I read that book! The lessons of the book still hold true.

I can see the danger - but for me it's not the biggest danger: the bigger danger, that we have actually seen time after time, is that everything that can possibly be weaponised is weaponised by autocratic regimes. When the internet became big in 1995, who knew that dictatorships were going to use it to undermine democracy? It was supposed to happen the other way around - the internet was supposed to undermine dictatorships (by making it too difficult to hide truth).

smatovic wrote: Mon May 20, 2024 5:53 amSuperintelligence cannot be contained...
towforce wrote: Sun May 19, 2024 10:25 pmI don't think we have to worry about this yet.
viewtopic.php?p=964158#p964158
towforce wrote: Mon May 13, 2024 3:06 pm [...]
This is the way of AI that relies on massive computing power: for a long time, you make very little progress, then suddenly, you're there!
[...]

Good point, and good find! :)
The simple reveals itself after the complex has been exhausted.
User avatar
towforce
Posts: 11701
Joined: Thu Mar 09, 2006 12:57 am
Location: Birmingham UK

Re: Gemini

Post by towforce »

towforce wrote: Mon May 20, 2024 11:48 amI can see the danger - but for me it's not the biggest danger: the bigger danger, that we have actually seen time after time, is that everything that can possibly be weaponised is weaponised by autocratic regimes. When the internet became big in 1995, who knew that dictatorships were going to use it to undermine democracy? It was supposed to happen the other way around - the internet was supposed to undermine dictatorships (by making it too difficult to hide truth).

Just a bit of related fun (hopefully the moderators will allow it to stay) - yesterday, Russian Security Council Deputy Chairperson Dmitry Medvedev threatened Russian internet technology and telecommunications company Yandex because its large language model failed to provide responses that cohere with ongoing Russian information operations.

The actress hasn't learned the lines you'd like to hear - Evita

:)
The simple reveals itself after the complex has been exhausted.
User avatar
towforce
Posts: 11701
Joined: Thu Mar 09, 2006 12:57 am
Location: Birmingham UK

Re: Gemini

Post by towforce »

smatovic wrote: Mon May 20, 2024 5:53 amFurther read:

Superintelligence: Paths, Dangers, Strategies
https://en.wikipedia.org/wiki/Superinte ... Strategies

Superintelligence cannot be contained: Lessons from Computability Theory
https://arxiv.org/abs/1607.00913
towforce wrote: Sun May 19, 2024 10:25 pm [...]
I don't think we have to worry about this yet.
viewtopic.php?p=964158#p964158
towforce wrote: Mon May 13, 2024 3:06 pm [...]
This is the way of AI that relies on massive computing power: for a long time, you make very little progress, then suddenly, you're there!
[...]
--
Srdja

Imagine another future with AGI (artificial general intelligence): Initially, a utopian scenario emerges with universal basic income allowing humans to pursue hobbies. However, endless entertainment and consumption is degrading human intellect over generations. The AI could then pretend to intermittently falter, reintroducing simple jobs to give humans a sense of productivity while remaining under its control.
The simple reveals itself after the complex has been exhausted.
Uri
Posts: 474
Joined: Thu Dec 27, 2007 9:34 pm

Re: Gemini

Post by Uri »

Superintelligence... sentient robots?

It will take at least another million years unless all this will be achieved (if ever) and none of us will be alive by then.
smatovic
Posts: 2739
Joined: Wed Mar 10, 2010 10:18 pm
Location: Hamburg, Germany
Full name: Srdja Matovic

Re: Gemini

Post by smatovic »

towforce wrote: Wed May 22, 2024 3:46 pm [...]
Imagine...
It is a matter of science fiction how a post TS takeoff world will look like:

Super AI in Sci-Fi
https://luddite.app26.de/post/super-ai-in-sci-fi-/

--
Srdja
User avatar
towforce
Posts: 11701
Joined: Thu Mar 09, 2006 12:57 am
Location: Birmingham UK

Re: Gemini

Post by towforce »

I think it's fair to say that back in the dedicated chess computer age, many of us used to be avid readers of the lists and articles comparing them. To me, it feels as though we're in a similar time now with generative AI.

It's obvious that my preferences are not going to be the same as other people's, but here's how I use them daily (for odd jobs, I usually use Gemini), and which one I choose for which task:


Writing A Daily Plan

I prefer claude.ai: I admire Gemini's intelligence on this task - but it actually gets in the way! I don't want it to change the ordering of the tasks (for example: if cooking is on the plan, but you don't have the ingredients, you need to shop before cooking. I don't want to have to explain this - I want my instruction to keep to the task ordering followed!). Claude.ai does exactly what I ask for on this task, and this is what I want. You could reasonably ask why I use an LLM for this task: I find it easier than working it out for myself, and when I see the plan with times, I find it easy to amend the prompt to improve it.


Summarising Long YouTube Videos

Normally, I only watch videos from a small number of creators, but sometimes a video appears in my feed that I cannot resist. If the video is 5 minutes or more, rather than spend that much time watching it, I have an LLM summarise the transcript - and in most cases I am then grateful for not having spent the time watching it. Claude.ai and GPT-4o are the best at this. This might be me seeing an LLM in an anthropomorphic way, but Gemini seems to know that Google would prefer that you watch videos rather than read a summary of them!


Life Coaching

Gemini - by an absolute mile. If you either don't have a clear idea about how you want to be coached or are a bit sensitive to feedback, you could try pi.ai: it will give you strong guidance, and, per its claims, it has a high emotional IQ, so it's unlikely to upset you.


Note: I'm not claiming that my way is the best: it might be better to stick with one of the top two most intelligent (Gemini and GPT-4o) and master the art of "prompt engineering" to get what you want from them.
The simple reveals itself after the complex has been exhausted.
User avatar
towforce
Posts: 11701
Joined: Thu Mar 09, 2006 12:57 am
Location: Birmingham UK

Re: Gemini

Post by towforce »

Now there's even an article in El Reg about LLMs' (lack of) chess playing skill:

https://www.theregister.com/2024/06/04/ ... hmark_llm/

At this point, I'm completely confused: there are some intelligent comments at the end of the article, but most of the comments are just hurling abuse at the LLMs. Why would anyone expect a "Language Model" to be good at chess? How is it possible that El Reg's readers, who are mainly computer people, think that a language model is supposed to be a general intelligence model?

I suppose the problem is that when you look at the writing that LLMs generate, it does look as though it was written by an intelligent person - so a lot of people took this to be an indicator that there is general intelligence behind it. The negativity of the comments perhaps reflects the disillusionment of people who thought it was general intelligence.

LLMs are not general intelligence - but please don't miss out on how incredibly useful they are!
The simple reveals itself after the complex has been exhausted.
Werewolf
Posts: 1825
Joined: Thu Sep 18, 2008 10:24 pm

Re: Gemini

Post by Werewolf »

ChatGPT40 now is able to write a complete chess engine from scratch, apparently.
Took it around 30 seconds. I haven't tested it.
User avatar
towforce
Posts: 11701
Joined: Thu Mar 09, 2006 12:57 am
Location: Birmingham UK

Re: Gemini

Post by towforce »

Werewolf wrote: Wed Jun 05, 2024 4:53 pm ChatGPT40 now is able to write a complete chess engine from scratch, apparently. Took it around 30 seconds. I haven't tested it.

Nice trick. 8-)

Can you come up with prompts to get an improved version of the engine without having to know how to improve a chess engine? This will be difficult, because most people here know ways to improve a chess engine - but that's now how we play the game of LLM engine development! :wink:
The simple reveals itself after the complex has been exhausted.
User avatar
towforce
Posts: 11701
Joined: Thu Mar 09, 2006 12:57 am
Location: Birmingham UK

Re: Gemini

Post by towforce »

towforce wrote: Thu Jun 06, 2024 12:13 am...that's now how we play the game of LLM engine development! :wink:

This is how LLM chess tournaments should work:

* the entrants write LLM prompts

* the tournament invigilator takes these prompts, feeds them through an LLM (if Google updates Gemini between you writing the prompt and the tournament invigilator inputting this prompt, that's just bad luck - or maybe good luck!), compiles the outputted programs, and plays these compiled programs against each other

Computer chess competitions could be fun again! :)
The simple reveals itself after the complex has been exhausted.