What are the significant enhancements of GPT-4?

[gleem]

We can start the madness all over again ChatGPT has been replaced with ChatGPTplus ChatGPT based on GTP-4.

GPt-4 is the long-awaited upgrade of the popular GPT-3.5 and has several significant enhancements.

1. It can see and understand images: whatever that means.
2. It is harder to trick: now there's a downright challenge.
3. It has a longer memory: It can stay on topic for up to 64,000 words vs 8,000 for vers. 3.5
4. It is better in more languages.
5. It has multiple personalities: its verbosity, tone, and style are more readily changed.

For details see https://techcrunch.com/2023/03/14/5-ways-gpt-4-outsmarts-chatgpt/

 

[Greg Bernhardt]

The pace of generative AI has been dizzying.

 

[jedishrfu]

Woo Hoo! Time to play some more.

We are entering the era of vicarious programming where we ask the AI to write the program, poem, song, play, picture, or video of our dreams for us.

The next few years we could have immersive on-demand interactive movies where what you say generates new scenes ala like on Star Trek Next Gen where Data challenges the computer to make an original Sherlock Holmes story with Prof Moriarty as the chief villain.

https://villains.fandom.com/wiki/Professor_Moriarty_(Star_Trek)

 

[Filip Larsen]

Arstechnica also has write-up on it, noting its high performance on (selected) types of exams and test, and how it (apparently) is able to analyze and describe why something is funny, a task deemed very hard in 2012.

https://arstechnica.com/information...-gpt-4-its-next-generation-ai-language-model/

 

[Filip Larsen]

https://arstechnica.com/information...-see-whether-gpt-4-could-take-over-the-world/

As part of pre-release safety testing for its new GPT-4 AI model, launched Tuesday, OpenAI allowed an AI testing group to assess the potential risks of the model's emergent capabilities—including "power-seeking behavior," self-replication, and self-improvement.

Both reassuring and somewhat alarming I suppose. Personally, my greatest concern right now is merely the huge potential this technology has for misuse once (some) humans eye the opportunity, as they already has, to spin them a truckload of (more) wealth. However, if researchers eventually manage to remove or severely limit hallucination without hampering emergence of new capabilities in a future version, then I will move extinction back to the top of my worry list.

 

[Filip Larsen]

https://arstechnica.com/information...plugins-that-can-interface-ai-with-the-world/

On Thursday, OpenAI announced a plugin system for its ChatGPT AI assistant. The plugins give ChatGPT the ability to interact with the wider world through the Internet, including booking flights, ordering groceries, browsing the web, and more. Plugins are bits of code that tell ChatGPT how to use an external resource on the Internet.
[...]
For those who aren't aware: you write an OpenAPI manifest for your API, use human language descriptions for everything, and that's it. You let the model figure out how to auth, chain calls, process data in between, format it for viewing, etc. There's absolutely zero glue code.

(This is technically more a bit of news related to ChatGPT, which is the web/app interface to a specific GTP version such as GPT-4 which this thread is about).

 

[gleem]

No sooner is GPT4 released and there is a buzz going on about GPT5 reportedly to be finished training this December. So the pace continues. There is speculation that GPT5 may be approaching AGI. This would seem pretty optimistic to me. This is (probably) the reason that the letter asking for a moratorium on AI development be paused for six months. GPT4.5 is projected to be released in Sept or Oct.

OpenAI has become closed AI by not making available how it is improving the GPT model. Additionally, OpenAI is apparently now working with a Norwegian robotics company 1X presumably to give GPT something else to do. Some researchers have moved the arrival date of AGI from 2056 to 2035. Anyway, if the improvement from 4 to 5 is as great as from 3.5 to 4 we will have something to look forward to.

With such a concern about AGI, I wonder if it is attained will whoever develops it reveal its full capability at least initially, and if not how long might they keep the wraps on it?

I hope those that are testing the GPT versions keep the results so that they can compare them and help to make up their minds about the improvements.

 

[dlgoff]

... where Data challenges the computer to make an original Sherlock Holmes story with Prof Moriarty as the chief villain.

Loved that episode.

 

[TeethWhitener]

https://www.theverge.com/23560328/openai-gpt-4-rumor-release-date-sam-altman-interview
There’s been a viral image comparing the training data set size and/or parameter space size going around that you may have seen. According to this article, the CEO of OpenAI says it’s BS and he’s worried about the wave of overhype. This was 3 months ago, before GPT 4 was released, but I doubt it’s that far out of date (wrt the hype surrounding the new AI models).

 

[DrJohn]

Google the game Wordle if you don't know it.
Try to get a decent answer out of the chat machine on that! It was dreadfully useless when I tried that yesterday.

 

[gleem]

Google has joined the generative AI fray with an updated version of BARD, based on PaLM2. It is supposedly much improved over its original version inauspiciously launched back in March when Microsoft launched it chatbot based on GPT-4..
https://www.msn.com/en-us/news/tech...n&cvid=dc61b3903f4a46098be23e7ae7a5d186&ei=36

 

[Algr]

2. It is harder to trick: now there's a downright challenge.

Trick it into doing what? One of my biggest concerns is that people may block AI from telling us things we need to hear. Someone starts blocking flat earth nonsense, and the next thing you know, nuclear waste concerns are on the same block list.

 

[gleem]

Trick it into doing what?

See this link:

For details see https://techcrunch.com/2023/03/14/5-ways-gpt-4-outsmarts-chatgpt/

 

[Algr]

I read that link already. It doesn't say anything about what is considered malicious or bad, or who gets to decide what ideas go outside the guardrails.

2. GPT-4 is harder to trick​

For all that today’s chatbots get right, they tend to be easily led astray. A little coaxing can persuade them that they are simply explaining what a “bad AI” would do, or some other little fiction that lets the model say all kinds of weird and frankly unnerving things. People even collaborate on “jailbreak” prompts that quickly let ChatGPT and others out of their pens.GPT-4, on the other hand, has been trained on lots and lots of malicious prompts — which users helpfully gave OpenAI over the last year or two. With these in mind, the new model is much better than its predecessors on “factuality, steerability, and refusing to go outside of guardrails.”

The way OpenAI describes it, GPT-3.5 (which powered ChatGPT) was a “test run” of a new training architecture, and they applied the lessons from that to the new version, which was “unprecedentedly stable.” They also were better able to predict its capabilities, which makes for fewer surprises.

 

[mathman]

Gave following to chatGPT and got answer for ;inradius; of triangle, not radius of semicircle! Why?
====================================
Geometry problem: Semi-circle inside triangle: Triangle with known length sides a, b, c where a is the longest. Place inside the triangle a semi-circle with diameter resting on side a. What is semicircle radius of largest possible semi-circle in terms of side lengths? Position of diameter center along a?

 

[gleem]

ChatGPT has now been given a memory.

The new memory feature also means users won't need to reintroduce themselves whenever they start a chat, Vladimir Fomenko, the founder of Infatica, said in an email interview.

"This leads to seamless and continuous conversations, creating a sense of familiarity and trust as if conversing with an acquaintance who knows me well," he added. "The memory function empowers ChatGPT to comprehend better users' preferences, interests, and specific needs, resulting in tailored responses."

https://www.msn.com/en-us/money/oth...n&cvid=3fb5918813794b31a820cc6ee4f067db&ei=37

In addition, other upgrades include

The new memory feature isn't the only upgrade coming to ChatGPT. The bot is also getting support for multiple languages, emotional intelligence, integration with real-world data, visual and auditory understanding, code writing and debugging, personal assistant capabilities, and advanced context comprehension, analyst Er. Ankit Dhadwal pointed out in an email.

 

[Borg]

I would guess that it's been posted elsewhere but as of Sept. 25th, GPT-4 can now see, hear and speak.