Chat GPT

[Carl Sagan]

Now GPT4 has been released to supersede ChatGPT, things move on again.

A team from Microsoft Research has released an academic paper entitled "Sparks of Artificial General Intelligence" (pdf link https://arxiv.org/pdf/2303.12712.pdf)

AGI or artificial general intelligence is the concept of an AI being more like a Human and good at everything rather than have a specific narrow band of expertise. I think the researchers writing the paper have been caught up by the hype, in the way they talk about GPT4 understanding things. My take is it "understands" nothing. They acknowledge they are pushng boundaries by their use of the term. An interesting paragraph from the paper reads:

"A question that might be lingering on many readers’ mind is whether GPT-4 truly understands all these
concepts, or whether it just became much better than previous models at improvising on the fly, without any
real or deep understanding. We hope that after reading this paper the question should almost flip, and that
one might be left wondering how much more there is to true understanding than on-the-fly improvisation.
Can one reasonably say that a system that passes exams for software engineering candidates (Figure 1.5) is
not really intelligent? Perhaps the only real test of understanding is whether one can produce new knowledge,
such as proving new mathematical theorems, a feat that currently remains out of reach for LLMs [Large Language Models]."

But that doesn't stop AI becoming more powerful and either more useful or more dangerous or both. 

[Bob The Badger]

I have an unusual writing style tending to blend humour with swearing and occasionally adding marketing/psychology value to my subscribers.

On Wednesday I decided to send out a newsletter that was 100% ChatGPT4 other than editing out its annoying habit of saying we rather than I.

I did give it a fairly in-depth prompt explaining my style and the topic.

I then sent it to about 2,500 subscribers and posted it in my FB group which has 7k members.

Nobody realised I hadn't written it including a good friend who was staggered.

I couldn't have done that n v3.5.

It is a bit scary now because what is version 5 going to do?

Or 25 ffs?

[Bob The Badger]

17 minutes ago, Bob The Badger said:

I have an unusual writing style tending to blend humour with swearing and occasionally adding marketing/psychology value to my subscribers.

On Wednesday I decided to send out a newsletter that was 100% ChatGPT4 other than editing out its annoying habit of saying we rather than I.

I did give it a fairly in-depth prompt explaining my style and the topic.

I then sent it to about 2,500 subscribers and posted it in my FB group which has 7k members.

Nobody realised I hadn't written it including a good friend who was staggered.

I couldn't have done that n v3.5.

It is a bit scary now because what is version 5 going to do?

Or 25 ffs?

Further to this, I didn't expect it to swear and it didn't. But sometimes I don't and that wouldn't have been obvious.

And for anybody who wants to play about with it, I am using the paid version at $20 per month. That's the only way you can use 4 at the moment.

I also just got the go-ahead to use Bard and will check that out over the weekend.

[EtoileSportiveDeDerby]

On 24/03/2023 at 18:44, Bob The Badger said:

Further to this, I didn't expect it to swear and it didn't. But sometimes I don't and that wouldn't have been obvious.

And for anybody who wants to play about with it, I am using the paid version at $20 per month. That's the only way you can use 4 at the moment.

I also just got the go-ahead to use Bard and will check that out over the weekend.

How is Bard performing?

[Carl Sagan]

On 29/03/2023 at 21:04, EtoileSportiveDeDerby said:

How is Bard performing?

For those who don't know, Bard is Google incorporating one of these "large language models" into its search technology. I'd say Bard is the tamest version of a large language model, which fits with Google being the dominant, trusted earch engine online. It has to navigate a fine line between incorporating sufficient technology to look as if it's keeping up without producing too much output that's wrong or speculative. The end result is servicable, but somewhat bland. But it is amazing how normal it seems to be having these interactions with a chatbit. Here's a Bard example I did to show what I mean:

[Carl Sagan]

I've been looking at a project with Salman Khan, that's recently become public. For those who don't know, the Khan Academy has an incredible mission to provide free high-quality education to any child in the world, as a great leveller and one of the best ways we can ensure social mobility, globally. It's mainly funded by Elon, as one of his myriad schemes to make the world a better place.

Sal does this through the Khan Academy, with over a 100 million registered users. He's been working behind the scenes with OpenAI over the last few months to incorporate GPT4 into the tutoring, with a view to every child having a personal AI coach and tutor who gets to know them, and helps with their weaknesses and strengths. At the same time, every teacher has an army of AI teaching assistants to help prepare lesson plans and grade work.

My take is that this could be an early killer app that really could transform the future. Here he is talking about it:

 

[EtoileSportiveDeDerby]

11 hours ago, Carl Sagan said:

For those who don't know, Bard is Google incorporating one of these "large language models" into its search technology. I'd say Bard is the tamest version of a large language model, which fits with Google being the dominant, trusted earch engine online. It has to navigate a fine line between incorporating sufficient technology to look as if it's keeping up without producing too much output that's wrong or speculative. The end result is servicable, but somewhat bland. But it is amazing how normal it seems to be having these interactions with a chatbit. Here's a Bard example I did to show what I mean:

I got access yesterday. Very quick to compute answers. It won't be long before  someone combines answers from several models.