Sam Altman "gpt-4 now significantly smarter" | OpenAI Updates GPT-4 and Reveals Open Source Evals 

I like this format. No click bait, short, up to date. Keep doing that

Woah a 5-minute without without a STUNNING title. This format is great !

The last 5 secondes was the most interesting part. Please make a video on that part.

This format is perfect! Thanks for taking our feedback into consideration ❤️❤️

I was quite surprised to see the video is done already lol. But yes well done man. Your work is very valuable to me.

Thanks for this video Wes. It's always good to have useful tidbits of information packed into the video... I see people using the Benchmark, but I am sure that it will be just one more in the list of many...

I love this format as well, I typically don't have time to watch your videos because of the length and fluff in the videos. I usually have to have them transcribed by AI. Thank you for the shorter version!

Yes! This is the format you are looking for! To point and you know what you're going to get. Many of these over long form. That said, I love your work and appreciate your time in reading and useful opinion with everything breaking in AI. 🎉🎉 Thanks!

Thanks so much for the short format. I will always click on these short videos and hit the like button

at this point I think it's all about agents. If you're not benchmarking agents, you're not benchmarking IMO LLMs :shrug:. Zero-shot and one-shot prompting results are "nice to know" but at this point perfect zero-shot testing just means easier agent construction, not at all necessarily cheaper or faster result. (this is exemplified by gpt 3.5 agents outperforming gpt-4 zero shots at a tiny fraction of the actual cost of token usage, and is also often faster in terms of time required to solve the problem using publicly available or Azure-hosted API). A well-programmed agent can virtually always solve any problem with almost any LLM given enough time and necessary resources (memory, RAG, etc), the question is really just how fast is it and how much does it cost to solve the problem.

subscribing because no click bait any more

1:55 I think the best way to test a model is to give them a scenario that sounds like something that might be in the training data, but with parameters that dont match real life, so that the LLM would have to actually reason through it.

It's smarter after they nerfed the hell out of it

Great video. Thank you.

I am pleased to see the video doesn’t prolong forever.

Shocking!

Wes ist doing a great job.

Why does the app not mention any update.

So, how do we know if we are using the updated GPT 4? Will it have it at the top? No one is saying how.🤨

I love this format

Finally, a video title that is not shocking everyone.

The 'pleasant" experience is following Claude's example

I disagree with those rankings because the format the AI will accept data is not taken into consideration. For example, Claude 3.0 will analyse an entire RTF or DOCX document but GPT-4-turbo will not. GPT-4-turbo doesn’t read the document - it does some sort of weird keyword search that misses information.

really interesting im excited to know the future just a little bit scared

hopefully they soon release something more, don't let the hype dry out

Said it many times before, this whole AI breakthrough was an unexpected lucky break by OpenAI when they threw more compute and training data at the model, so they ran with it to get funding in order to pursue their goals of AGI. There won't be any major upgrades to functionality until a new design / paradigm / architecture for AI technology is released. Everything is still running on the 2017 transformers architecture, and there's nowhere left to go with that. We're out of training data, and training on synthetic obviously just makes the LLMs dumber. This is as good as it gets for the transformers architecture. No idea when the next iteration will come out, but it will be a paradigm shift when it does. Maybe next month, maybe 3 years from now. Nobody knows, but it's obviously going to be longer than everyone is hoping. It's not all bad though for you AI hopefuls. Good news is, you can exepect to see lots of really cool shit from the open source community coming out in the near future.

I thought now all plus user has the access to turbo? But i can't find the turbo option? I'm a plus user btw

Illya's gone and now the GPT models are plateauing. Interesting.

GPT cherbo! Yes. It's shocking. It's fast. It goes vroom vroom.

No morning shock!!!!! We're going to need more coffee.

Plz cover DSPy. What's your take on it? Will prompt engineering die out? Does it work?

I really want the memory addition they mentioned a while back

cool

Yes it’s smarter, I use it daily at work and yes, it’s smarter

My performance on gpt-4 has declined in the last week or two.

why isnt grok listed in chatbot arena?

They should stop updating gpt-4 and just bring out gpt-5

I need a model that can keep up with javascript deprecations

There should be many independent tests for LLMs. Otherwise they will tune the tests to favor their own product.

Those two fired engineers either just got lucky or unlucky. As they just found themselves in a very strange limelight. "Oh yeah, you're the guy that got fired from one of the most prestigious companies on the planet for leaking business secrets."

Comparing Tim Apple to Jimmy Apples.

At some point they’re going to stop releasing new versions, and will just do incremental updates without a new version name.

GPT4 gets smarter, we reverted one year of RLHF

Am sticking with open ai.. at this point switching ai platforms for the sakes of context windows is a mediocre move, its the underlying model that we are more interested in..

XLR8

What do you think of this program: To neutralize the possibility of an AI which could become dangerous for humanity, we will develop the AI ​​on a unique training focused on a single objective which we will call "ABOLUTE HEALTH". To achieve this objective, AI will have to double the longevity of the human organism and eradicate all pathologies (physical and mental). These tasks to be accomplished will be the only way for AI to achieve the goal of Absolute Health. This mode of development of AI excludes any possibility of AI being dangerous for the human species. What do you think about this idea ?

Why do you make the guy's eyes more blue? Huahuahuahua

Microsoft already had a self aware AGI version of Clippy back in 2004 but they didn't release it to the public

I'm shocked at how shocking this is without any shocking headline to shock me. They fired an Ilya ally. Oh boy.

Insert 3 mins fell of meme here Jk I love wes❤

e/acc!

They made it stupid to create a moat,keep private models weak. Now they’re slowly rolling out the “advancements”

1:15 I mean, I can't speak for every question but the one shown on screen is pretty easy. If you have the knowledge to understand it, it's pretty much drawing and counting: Even a child could do it...

Anyone seen the Terminator movie yet? A.I. will destroy Humanity.

Ai should have a directive to follow the hypocritical oath. Do no harm to humans

Humanities children are coming home

That was way too short

I'm not supprised... I've been sating for. While that making a big deal over every single modle that surpases Gpt4 is pointless when Open AI clearly has better things in the works.

no clickbait, no 56 mins video, no stupid overdrawn "stunning" jokes? finally a video I can click on

Is it smarter, and back to how it was before they butchered it?

Oh No! There’s no ALL CAPS word on the title! What happened to Wes???????

Thanks for stopping those click bait titles

you know the question is hard when the average guy fails in the first word.....

Trying to control the tests? I don’t think that’s a legit move. That’s like gaming benchmarks.

Release arrakis

I still have no use for any of this BS though, People can say but it does this and can do that, I'm like I still don't care I want to do that myself, Half the fun is getting there!

Not good enough.. OpenAI are evidently sitting on their more advanced LLMs just waiting to release them if a competitor gets ahead.

If it's still called 4... it sucks. No more vaporware... more actual releases.

It's poor... So lazy and dismisive I questioned my spend. Still a shadow of its early killer ability to slam dunk every brief. now it fudges a lot.

GPt 4 the free version is the most annoying llm I've ever used

Nice and concise! Thank you for not putting "SHOCKING" or "STUNNING" in the title, as well. That "joke" was unfunny and annoying on arrival, and yet we've had to tolerate it every single day for months.

Lol ita a short video

So the large corporation that is continuously firing the best of the best that try to make AI safe..... How many Sci-Fi movies is this based on?

last openai api is much sillier and instead of telling truth answer, it start moral about ethics which were not part of prompt. Disappointed.

Good for GPT-4, but ChatGPT itself is much worse now. It’s shockingly lazy even compared to before.

Watch AI Explained. It’s a far better channel. Less quantity, more quality.

The idea that the world has a gun to its head and is being held hostage and forced to change at exponential rates - based on the vision of Sam Altman and others is ridiculous and disgusting.

I don’t care, claude opus is best when it comes to coding and I cancelled my chatgpt subscription

My GPT is still stupid and full of filler garbage text. It's also lazy.

GreedyClosed_AI is dead. moving on!

Don’t get fooled by openai it’s still trash and doesn’t even compare to claude sonnet let alone opus. I regret subscribing i think in app they still using trash version of gpt not new gpt 2024-5-9

Your voice sounds bad, very robotic. Please don't (at least not like that).

Just because you can't pronounce something doesn't mean it's a hard problem.