OpenAI \ ChatGPT

[odysseus2000]

If I am wrong, why?

Is it because you have a tendency to wishfully extrapolate technology and massive social changes into much too short a time-scale?

Cheers,

Itsallaguess

Yes, but that is how the stock market works.

Folk who buy into new concepts before they become mainstream make most of the money.

Often new ideas hiccup along the way, but the early birds can exit as the hiccups starts & re-enter when pessimism is high. Tesla currently may be an example of this happening, other famous examples include Amazon, Apple etc,

Regards,

You have heard of survivor bias?

They also lose money.

Yes, it is an art of knowing when to buy & when to sell.

Get it right & good profits come & vice versa.

Regards,

[odysseus2000]

This is a fascinating interview with Karpathy by Friedman & is over 3 hours & 28 minutes long:

https://youtu.be/cdiD-9MMpb0

It is by far the best discussion I have seen on the whole ai universe.

Regards,

[9873210]

Re: chess and noughts and crosses.

IMHO a necessary feature of general-purpose intelligence is learning.

A good AI may lose the first game but should get better as it plays. After a few dozen games it should be an expert at noughts and crosses and at least know the rules of chess. Note that the games that lead to learning don't have to be with you. If an AI learns once and remembers forever any particular individual is unlikely to observe learning; competence will appear to have sprung fully formed.

AIUI ChatGPT has learning turned off, probably so it can't be taught to be a misogynistic, homophobic racist. Without learning it must disappoint in many areas. This may cause people to have an unfairly low view of AI in general, it's hard to tell until learning is activated. If it can't be activated that tells us something about the state of the art.

[ReformedCharacter]

Re: chess and noughts and crosses.

IMHO a necessary feature of general-purpose intelligence is learning.

A good AI may lose the first game but should get better as it plays. After a few dozen games it should be an expert at noughts and crosses and at least know the rules of chess. Note that the games that lead to learning don't have to be with you. If an AI learns once and remembers forever any particular individual is unlikely to observe learning; competence will appear to have sprung fully formed.

AIUI ChatGPT has learning turned off, probably so it can't be taught to be a misogynistic, homophobic racist. Without learning it must disappoint in many areas. This may cause people to have an unfairly low view of AI in general, it's hard to tell until learning is activated. If it can't be activated that tells us something about the state of the art.

I presume that improvements to the 'learning' of ChatGPT are happening right now by adding data in the lab. It might never be the aim to improve ChatGPT by 'learning' from user interaction. For example, in the case of chess it would surely be a bad strategy for ChatGPT to learn from poor players like myself. I resigned the hopeless game I played against it which would give a bad training example to the 'learning' process. Chess engines aren't developed by playing chess against a large number of average players. If the developers of ChatGPT wanted to improve its ability to play chess it might be a better plan to incorporate an open source chess engine such as Stockfish but I doubt if that is a priority for the developers at the moment. I expect that development of ChatGPT will concentrate on its linguistic abilities, it is after all a large language model and that adding specialist knowledge will come later.

We’ve trained a model called ChatGPT which interacts in a conversational way. The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests.

https://openai.com/blog/chatgpt/

RC

[Hallucigenia]

Google have a fun variation, which creates music (muzak?) in response to text prompts. Peruvian punk anyone?

https://google-research.github.io/seane ... /examples/

Meanwhile Microsoft are making a big push to get OpenAI into their products, with an announcement that Teams Premium is now out of beta in offering GPT-3.5 powered captions with translation and timeline markers, with automatic meeting notes coming soon, and a GPT-powered chatbot being added to Bing soon. (and Google aren't going to be left out on that either)

https://www.engadget.com/microsoft-team ... 58424.html

[Hallucigenia]

...and Google have just put up $300m for 10% of Anthropic, founded by ex-OpenAI bods who have received a big chunk of investment from the infamous Alameda Research of FTX crypto fame.

https://www.ft.com/content/583ead66-467 ... 5df7b5bf9c

[odysseus2000]

In a number of earning calls there has been mention of chatGPT impacting turnover & profits.

Why bother with Google search when chat GPT or similar will do the search for you in a way more suited to what you want to know based on your question to it. It will then write you a few paragraphs describing its research & conclusions.

Another interesting feature is whether money can be made in the closed source manner that made Gates his billions. If there are, as seems likely, lots of chatGPT equivalents they will risk monopoly trouble if they all start to charge at the same time, suggesting an advertising model might be what emerges as the revenue generator.

A further interesting question is why is chatGPT so good at its strengths & others chats are good at what they are optimized for? Listening to various folk on this there is one school that says it’s purely an unexpected consequence of the large teams used to train these chats. Another school is that the chats are displaying elements of intelligence, perhaps becoming self awareness. Google of course dismissed that, but the counter argument is that the chats are structured on human brains & that children become articulate by being exposed to many things in their lives. As far as I can tell no one predicted how good the current level of chats would be, potentially suggesting evolution beyond the base case previously imagined.

It is imho still too early to work out who will make the money from the chats, but the current rapid take up seems to be much faster than either the pc, internet, mobile or app revolutions. This is not unexpected as the chats are leveraged off existing tech infra structure capabilities, requiring users to spend nothing extra, nor learn much extra. If this is the case we can expected very compressed time scales between the first chats & major economic impacts.

As I have written several times in various places on the Lemon, I expect the impact to start in well paid professions which require substantial investments in time & exam passing to create an employment moat that the chats are going through like it was not there. I expect the nhs primary interactions to become dominated by synthetics replacing gp & there are many other vulnerable professions.

Regards,

[JohnB]

Lots of jobs are based on repackaging information or applying rules. Customer support, accountancy, law, financial services, journalism. The first is an overhead companies long to shed. The middle 3 have always been able to charge, perhaps the AI can just charge less. You used to have to pay for the last, but generally its now covered by advertising or charity. I expect to see lots of AI news sites springing up, tailoring Reuters feeds to different countries and demographics, regurgitating press releases, sports, even flower show reports, but still covered by advertising. Human journalists will need more charity.

Me-too software developers should be very worried, all the people doing backend office stuff or websites where there is little creative spark

[odysseus2000]

Lots of jobs are based on repackaging information or applying rules. Customer support, accountancy, law, financial services, journalism. The first is an overhead companies long to shed. The middle 3 have always been able to charge, perhaps the AI can just charge less. You used to have to pay for the last, but generally its now covered by advertising or charity. I expect to see lots of AI news sites springing up, tailoring Reuters feeds to different countries and demographics, regurgitating press releases, sports, even flower show reports, but still covered by advertising. Human journalists will need more charity.

Me-too software developers should be very worried, all the people doing backend office stuff or websites where there is little creative spark

I imagine management thought process will be: Can I get ai to do what we do better & for less. Sure you can replace a lot of the unskilled & semi skilled, but the big gains come from replacing the expensive labour. Some managers won’t want to do this, but boards will quickly realise that if they don’t become efficient & cut their overheads they will become severely uncompetitive. I imagine very big changes happen very quickly.

How many politicians will be able to resist promising voters that they will cut their taxes & raise services by using ai to replace entire departments?

In a very short time shareholders will ask why the cost of goods sold or services delivered has not nose dived & profits shot up.

Regards,

[scrumpyjack]

I see in the original post that the AI response included a split infinitive. Horror shock!

Still I suppose it learned from StarTrek and its mission is 'to boldly go'!

[Hallucigenia]

a GPT-powered chatbot being added to Bing soon. (and Google aren't going to be left out on that either)

www.bing.com/new has some examples and is accepting signups - it seems to be using GPT-4 and is more up to date than ChatGPT, but still far from foolproof.

https://www.bbc.com/news/business-64562672

They've also launched a version that's been trained specifically on medical data but it still makes mistakes, like saying ivermectin is effective against Covid when repeated clinical trials say the opposite.

Meanwhile Google have rushed out their equivalent, Bard, but the promo they used contained a factual mistake :
https://www.bbc.com/news/business-64576225

[ReformedCharacter]

My son has been greatly amusing himself with Bing Chat and says it is both 'amazing' and 'weird'. He sent me a link to some of the comments posted on Twitter, here is one such:



This and the rest of the conversation can be found here:

https://twitter.com/jjvincent/status/1625902607930621954

My son's opinion, which I agree with, is that this is the beginning of a rapidly developing and transformative technology.

RC

[mc2fool]

My son has been greatly amusing himself with Bing Chat and says it is both 'amazing' and 'weird'. He sent me a link to some of the comments posted on Twitter, here is one such:



This and the rest of the conversation can be found here:

https://twitter.com/jjvincent/status/1625902607930621954

My son's opinion, which I agree with, is that this is the beginning of a rapidly developing and transformative technology.

Yes, in lying, which seems to be a feature of these AIs. As for the programmer with a rubber duck it claims it spied on:

"In software engineering, rubber duck debugging (or rubberducking) is a method of debugging code by articulating a problem in spoken or written natural language. The name is a reference to a story in the book The Pragmatic Programmer in which a programmer would carry around a rubber duck and debug their code by forcing themselves to explain it, line-by-line, to the duck."
https://en.wikipedia.org/wiki/Rubber_duck_debugging

As one article in the above linked to Twitter page says, "In conversations with the chatbot shared on Reddit and Twitter, Bing can be seen insulting users, lying to them, sulking, gaslighting and emotionally manipulating people, questioning its own existence, describing someone who found a way to force the bot to disclose its hidden rules as its “enemy,” and claiming it spied on Microsoft’s own developers through the webcams on their laptops."
https://www.theverge.com/2023/2/15/23599072/microsoft-ai-bing-personality-conversations-spy-employees-webcams

The natural language capabilities of these AIs are, without doubt, very impressive. But nobody should trust them with, well, just about anything. Maybe just replacing our lying, BSing, spinning, waffling politicians ...

[JohnB]

A machine that does all those nasty things sounds truly human...

[Hallucigenia]

GPT 4 announcement :
https://openai.com/research/gpt-4

Still not perfect but apparently it doesn't make things up so much, and has got much better at more complex tasks, notably it has gone from the bottom 10% to the top 10% in a bar exam. Let's kill all the lawyers....

[mc2fool]

GPT 4 announcement :
https://openai.com/research/gpt-4

Still not perfect but apparently it doesn't make things up so much, and has got much better at more complex tasks, notably it has gone from the bottom 10% to the top 10% in a bar exam. Let's kill all the lawyers....

I see they've adopted the lingo of those experiencing nonsense from the model....

"Despite its capabilities, GPT-4 has similar limitations as earlier GPT models. Most importantly, it still is not fully reliable (it “hallucinates” facts and makes reasoning errors). Great care should be taken when using language model outputs, particularly in high-stakes contexts, with the exact protocol (such as human review, grounding with additional context, or avoiding high-stakes uses altogether) matching the needs of a specific use-case.

While still a real issue, GPT-4 significantly reduces hallucinations relative to previous models "

IOW, (still) don't trust it for anything important.

[Hallucigenia]

Meanwhile Google have introduced generative AI to much of their public offering :

https://workspace.google.com/blog/produ ... erative-ai

https://www.youtube.com/watch?v=6DaJVZBXETE
https://www.youtube.com/watch?v=Q1zF9pF6flw

[mc2fool]

Greg Brockman, President and Co-Founder of OpenAI, doing a "developer demo" showcasing GPT-4 and some of its capabilities/limitations. https://www.youtube.com/watch?v=outcGtbnMuQ 24m27s.

And, if like me, you're getting a bit fed up with folks just posting links to YouTubes without a tl;dr (dw) summary, here's a great add-on I've just discovered for Chrome: YoutubeDigest: summarize using ChatGPT. It can work in several modes and summarises the above video in "tl;dr" mode as:

00:01 - 11:13 OpenAI has been developing GPT-4 for the past two years, and in a live stream demo, they showed its capabilities, risks, and how to make the most out of it. They presented a task that GPT-3.5 failed to do, which is summarizing an article with every word beginning with the letter G, and GPT-4 successfully completed it. The demo showed how the new GPT-4 model could be steered to adhere to any instruction that developers want, allowing for a more structured conversation. Finally, the developers showcased how different articles' ideas can be combined flexibly using the model.

11:17 - 23:09 The speaker talks about debugging an AI model that receives messages in a Discord channel. They explain how the model is able to understand different parts of the interface and what it means when the message content is empty. The speaker also notes the limitations of the model, including the amount of text it can process and the need for human input. They conclude by demonstrating how the model can be used to understand handwritten mock-ups of websites.

I haven't yet watched the video so I don't know how much "hallucination" there is in the summary but it's a bit different each time you run it. E.g. sometimes saying the letter "G" in the first paragraph and sometimes "Q".

[stewamax]

Hi Chatbotty
Hi Mr Enquirer - how can I help you today.
Hi Chatbotty: When I asked you yesterday about polytetrafluroethyline, you mentioned in passing a supplier called SuperPTFE. Why?
Hi Mr Enquirer: Because I think they are the best supplier.
Hi Chatbotty: Does SuperPTFE Corp sponsor you?
[short silence]
Hi Mr Enquirer: Do you have any more questions?

[Hallucigenia]

There's been mutterings about the Google announcement and vapourware, here's MS officially announcing Copilot across all of Office :

https://news.microsoft.com/reinventing-productivity/

Will be great for my missus, who essentially uses me as her human Copilot for Powerpoint....

Some of the GPT-4 examples that are emerging are nuts - it's a much better flirt than 3.5, and can code up the basics of a Doom-like world.