GPT-4 is here! What we know so far (Full Analysis)

#gpt4 #chatgpt #openai

References:

Links:

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

[kpb]

5:55 - 88th percentile doesn't mean it would be 88th in competition with 100 people, it means that it scored better than 88% of human participants. In other words in a competition against 100 people representing usual test takers it would rank 13th position leaving 88 humans behind and only 12 humans with a better score. That's very impressive!

[fcolecumberri]

OpenAI: the best close source AI developers ever.

[john_blues]

Hey, Yannic. Just wanted to say that I appreciate your focus on the Open Source concept and it's importance. I think your voice on this is a breath of fresh air, especially how you articulate it without bashing or ranting.

[technomage]

Knowledge is not power. Controlling knowledge is power. Thanks for your great videos.

[iambinarymind]

Thanks for the update, Yannic. I'm still excited and hopeful for the actual open-source AI space. Improvements happening at breakneck speeds.

[BGP]

"The only numbers they give is the prices for their api" 😆

[igorbeaver]

Why no one say that it’s unfair to give new API to close circle of big companies for couple months earlier than for startups? This kills competition

[lio]

I've tried GPT-4 in chatGPT plus. It definitely seems like a major improvement.

[amafuji]

OpenAI is the best of both worlds (closed source and open source). It's the best of closed source because they can profit the most from it. And it's the best of open source because the can pat themselves on the back about how "democratic" they are, whatever that means.

[williamv.]

OpenAI is the Anakin Skywalker of AI research labs.
They are single-handedly ending the era of open AI research when they were supposed to bring more openness to the field.

[frederikvoigt]

Thank you so much.  Every video of yours so far has sent me down another publication rabbit hole, spitting me out two months later with so many interesting reads along the way. Have a very great day!

[NomanTripps]

I suspect The Palm-E paper dropping last week prompted them to release info on GPT-4. They've had the release ready for months but felt the competition was so far behind they could take their time. Palm-E demonstrated performing tasks on images and text similar to GPT-4.

[yeetyeet]

Let's get OpenAssistant 80% of the way there to whatever they are building, then we will never need to give them money.

[tristanwegner]

On thing about human test: Human prepare for them during a short time before, and forget quite a bit right after the test, whereas AI can be expected to always retain these capabilities. E.g. in medicine it is said, that around the final exam, most medical students will have the highest amount of medical knowledge OUTSIDE of their chosen narrow field of medical practice. They will improve in their specialty and connected areas, but their knowledge in other fields will go down.
Also GPT4 takes the exam in many different topics, but only rare humans take more than two bar level exams in different fields.

[fteoOpty]

Great video Yannic. Thanks. It is worrying that OpenAI is expecting us to trust them. In the short term maybe, but longer term, there will be alternatives hopefully....

[erikjohnson]

Having 8 times max output size is already nice for anyone trying to get ChatGPT to code anything. I still needed to provide a lot of guidance to get correct results, diminishing the value. Great for get "some" boilerplate code, but difficult to get correct code outside of the training set. When asking interview coding questions, just remember that it was almost certainly trained on this data explicitly, meaning it has likely seen the problem before unless you are creating your own coding problems.

[Neomadra]

I wish the European Union would invest in open source large language models. We should not be dependent on one single closed source company for such an important technology. This is as important as something like particle accelerators, fusion and so one. Probably even more important. We need a publicly funded program for that.

[andybrice]

Good heavens. Imagine a world with fewer lawyers because citizens are able to better understand the laws they are bound by. What a terrifying thought. We must stop this digital menace immediately. #SaveTheLawyers

[gileneusz]

2:02 well... Elon Musk noticed that some time ago: "Initially it was created as an open-source nonprofit. Now it is closed-source and for profit. I don’t have an open stake in OpenAI, nor am I on the board, nor do I control it in any way"

[Oscaragious]

I heard some interviews stating that they are close to exhausting all quality data available on the Internet, which according to them means that the fast improvement of these language models (they seem to imply this affects more than just OpenAI) will sharply decline at that point.

I'm not sure what to make of that myself tbh.