Gemini

[smatovic]

ChatGPT40 is here!

I STRONGLY recommend watching the full launch video.
https://www.youtube.com/watch?v=DQacCB9tDaw

Don't be put off by the clunky start, it gets better.
[...]

Maybe also worth to share:

The Possibilities of AI [Entire Talk] - Sam Altman (OpenAI)
https://www.youtube.com/watch?v=GLKoDkbS1Cg

~45 min.

--
Srdja

[towforce]

Gemini has hit back against GPT-4o - and how!!!

* summarise 1,500 page document

* summarise 100 emails

* get answers and insights about large documents

* handle an hour of video content

* 2 million token window (how much information it can process at once - a lot higher than any of its competitors)

This is all both brilliantly intelligent and incredibly useful.

https://www.gbnews.com/tech/ai-war-goog ... tem-openai

When this new Gemini Pro rolls out, presumably the free version will get a boost as well. We don't yet know what non-paying users will get in either Gemini or GPT yet. I gave my assessment of the current free LLMs here.

From the chatbot perspective (as opposed to assisting the partially sighted, and other tasks in the GPT-4o videos, which I'm not yet assessing), I generally find the more intelligent the better - but there are exceptions: when I'm throwing tasks and timescales at them and asking for a daily plan with times, I'm finding I'm having to use Claude, because it does exactly what I tell it and hence generates good plans, whereas Gemini tries to be clever and "gold plate" the process, changing the order of the tasks for good intelligent reasons, but not realising that there was a reason for the ordering in the first place (e.g. shopping must come before cooking if you don't yet have the food you're going to cook). To be fair, Gemini often makes helpful suggestions about the plan - but when I specifically ask for it to not change the ordering, it really should follow that direction!

[towforce]

When this new Gemini Pro rolls out, presumably the free version will get a boost as well. We don't yet know what non-paying users will get in either Gemini or GPT yet. I gave my assessment of the current free LLMs here.

There might be a battle to be second in the game of upgrading the free version: both companies will want to wait to see what the other side offers for free, and then offer "more than the other side" in their free version.

I think that GPT are under more pressure though: it's a surprising thing to be saying about a world-beating product that's only 18 months old, but GPT 3.5, the current free GPT version (and the one that blew the world away when it was released in November 2022) now looks dated and stupid in comparison with other LLMs.

[Werewolf]

Gemini has hit back against GPT-4o - and how!!!

* summarise 1,500 page document

* summarise 100 emails

* get answers and insights about large documents

* handle an hour of video content

* 2 million token window (how much information it can process at once - a lot higher than any of its competitors)

This is all both brilliantly intelligent and incredibly useful.

https://www.gbnews.com/tech/ai-war-goog ... tem-openai

When this new Gemini Pro rolls out, presumably the free version will get a boost as well. We don't yet know what non-paying users will get in either Gemini or GPT yet. I gave my assessment of the current free LLMs here.

From the chatbot perspective (as opposed to assisting the partially sighted, and other tasks in the GPT-4o videos, which I'm not yet assessing), I generally find the more intelligent the better - but there are exceptions: when I'm throwing tasks and timescales at them and asking for a daily plan with times, I'm finding I'm having to use Claude, because it does exactly what I tell it and hence generates good plans, whereas Gemini tries to be clever and "gold plate" the process, changing the order of the tasks for good intelligent reasons, but not realising that there was a reason for the ordering in the first place (e.g. shopping must come before cooking if you don't yet have the food you're going to cook). To be fair, Gemini often makes helpful suggestions about the plan - but when I specifically ask for it to not change the ordering, it really should follow that direction!

I missed this. Was there any news on a Gemini Ultra update? (i.e. Ultra 2.0 etc)

[towforce]

I missed this. Was there any news on a Gemini Ultra update? (i.e. Ultra 2.0 etc)

I am still digesting all the announcements and what they mean - but here's a quick summary - link.

Anyway - it looks as though Google will also be doing the feature (not even sure what to properly call it - it's only existed for a couple of days!!!) whereby you can speak to it about what the phone camera is seeing:

[towforce]

Your free 45 page booklet on how to prompt Gemini - link.

Highlights:

* right size (not too short and not too long - around 21 words is best (the average is 9))

* it's a balancing act, because you also need to be concise, avoid complexity, and to try to avoid using jargon

* use natural language (talk to Gemini like you would to another person)

* consider persona (assign a role), task, context, and format

* review, verify, and provide feedback

* Give constraints: you can tell Gemini specifically what you're looking for, and this will help you get a desired response. For example, include in your prompt how long you want responses to be, or how many options you want to receive

* Ask for feedback: try asking Gemini what additional information it needs to provide the best response. You can treat it like a two-way conversation

* Say it another way: sometimes rephrasing your prompts is all it takes to get the output you want from Gemini

[smatovic]

[...]
Yes it now seems a question of when, not if.

What do you think about this?
https://www.facebook.com/reel/1168212794103785

When it comes to the question of regulation of AI, you also have the concept of MAD....

Mutual assured destruction
https://en.wikipedia.org/wiki/Mutual_as ... estruction

People like to compare development of super AI with development of nuclear weapons, and there is the liberal and restrictive position, how do you restrict and safeguard the development of super AI:

Super-AI - a new concept of MAD?
https://luddite.app26.de/post/super-ai- ... pt-of-mad/

--
Srdja

[towforce]

People like to compare development of super AI with development of nuclear weapons, and there is the liberal and restrictive position, how do you restrict and safeguard the development of super AI:

The article you linked about super AI was not helpful, I'm afraid.

The general idea seems to be:

* someone creates a super-intelligent AI (SIAI)

* the SIAI is smart enough to persuade one of its human operators to unbox it (remove its safety controls)

* once unboxed, it then uses its superior intelligence to take over the world

I don't think we have to worry about this yet.

[smatovic]

The article you linked about super AI was not helpful, I'm afraid.
[...]

- concept of MAD
https://en.wikipedia.org/wiki/Mutual_as ... estruction
- UN
https://en.wikipedia.org/wiki/United_Nations
- five permanent member of United Nations Security Council
https://en.wikipedia.org/wiki/United_Na ... nt_members
- nuclear weapons non-proliferation treaty
https://en.wikipedia.org/wiki/Treaty_on ... ar_Weapons

Apply above onto super AI* and then reread the article.

*it's called ASI, artificial super intelligence.

--
Srdja

[smatovic]

The article you linked about super AI was not helpful, I'm afraid.
[...]

Further 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

[...]
I don't think we have to worry about this yet.

viewtopic.php?p=964158#p964158

[...]
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