HUGE AI NEWS: AGI Benchmark BROKEN ,OpenAIs Agents Leaked , Automated AI Research And More

00:00 - Breaking AI news you might have missed
00:23 - The rise of AI scientists: Game-changer or hype?
01:45 - AI flood incoming: Are we ready?
02:53 - OpenAI's secret project leaked?
06:05 - Text-to-video revolution on the horizon
07:51 - Benchmark bombshell: AGI closer than we thought?
12:19 - AGI Day drama: Experts clash on AI's future
15:41 - Is AI progress slowing down? Controversial claim
17:29 - Debunking the AI skeptics: What they're missing
19:38 - The real state of AI: Surprising insights
21:58 - DeepMind CEO drops major hints about next-gen AI
23:54 - OpenAI's roadmap decoded: What it means for you
26:12 - Mind-blowing conclusion on AI's true potential

Links

Welcome to my channel where i bring you the latest breakthroughs in AI. From deep learning to robotics, i cover it all. My videos offer valuable insights and perspectives that will expand your knowledge and understanding of this rapidly evolving field. Be sure to subscribe and stay updated on my latest videos.

Was there anything i missed?

#LLM #Largelanguagemodel #chatgpt
#AI
#ArtificialIntelligence
#MachineLearning
#DeepLearning
#NeuralNetworks
#Robotics
#DataScience

[vaendryl]

so much debate over what is and what isn't AGI.
so much effort put into making new benchmarks like that ARC thing.
as far as I'm concerned the answer is obvious, and we've done it before: video games.

video games are made for humans but quite accessible to AI models, relatively speaking.
they're meant to challenge someone on a human level and require quite a bit of insight understanding to pick up and play - and there are insane differences between genres.
the moment someone can come up with an AI that can pick up and play *ANY* game, and just .. figure it out as it goes along (more or less as could be expected of any human)
THEN I'll acknowledge that as true AGI.

[DynamicUnreal]

17:31 I could argue that every time the error rate gets cut in half on that graph, it’s exponential growth.

[TheAero]

actually looking at MMLU, it seems that the current closed source models are able to 95-96% solve it. ( as around 8% of questions are wrong )
Also open source is catching up to closed source, without needing as much parameters.
GPT-4 was estimated to have over 1T params. Launched around a year ago.
Our models right now catch up to it with lot less parameters + are open source. That is a significant jump in a year, for the competition. Let's not forget that companies focus on video and images more lately. And on image and video the jumps have been insane.

[PlanetLogical]

The paper site ArXiv is prounced like the word "archive". The X is the Greek letter chi.

[john]

It’s pretty audacious to say openAI have set us back considering the LLM is the only type of model that has made any progress towards AGI

[DanieleCorradetti-hnnm]

The real problem with the graph is that is not considering that the curve (evan an exponential curve!) cannot go above 100%. In the the green curve the guy was suggesting the model should get 120% right 😂😂😂 "If you know anything on how to aggregate data..." you know you cannot go above 100% right and passing from 90% to 99% it means to lower the error of 1 order of magnitude... (even so anyone knows these benchmark are flawed and to have more than 95% it's just not right)

[burninator]

Is there an AI tool that automatically takes out the “you knows” in a video? There should be…

[dsfadfafdsfadf]

it's impossible to see exponential improvements using a test score. the maximum you can get is 100%.
also, the LLM that gets 100% isn't twice as smart as the one that got 50%. it's orders of magnitude smarter.

[Lolatyou]

Date format should always be yyyy-mm-dd. This is the best format because it is sortable....
EU people are lost if they use day in the middle.. it's the worst possible way to format it.

[stillavantis]

A couple of things a lot of people don't understand about AGI is the power and materials needed.

The US does not have enough electricity to train/run an AGI system. Currently the US is in talks with Canada and Saudi Arabia to get us over the electrical limitations. With Canada being generally the best option because of their huge hydro power and the fact that they have similar security concerns.

Materially, no one has enough chips. Which is why Taiwan is under threat and the CHIPS act was passed. This problem as big as it is, isn't as big as the power problem. Both being solved within a decade or so.

In the near term what we will most likely see is a vast array of MI (machine intelligent) agents built for specific tasks all controlled by an Overseer and directed by an Oracle. This will get us pretty close to what we would think of as AGI.

The Oracle would designate what agents are necessary, pass the work instructions to the Overseer who would then spin up and use any manner of agents, be they generative, predictive, rule based, or any other program to complete the work. Finally handing everything back to the Oracle to present back to us, the user.

[wwkk]

The gary marcus graph is so embarrassing.

[nyyotam]

Oh, you understand 100% why Gary Marcus tries to downplay the huge advances in AI. Because we are headed directly at the singularity. So what's the use telling ppl the truth? You want to make ppl panic? What good would that do?

[davidevanoff]

What struck me as most significant in the Demis Hassabis interview is his remark is that the Gemini interface is 90% common internally and externally. And we know that Gemini is routing queries to the available models most expediently. That implies that Google is making the most use of reinforcement learning from human feedback. The whole world is collaborating actively to make Gemini better. Think about that -- it has millions of mentors.

[dr_harrington]

Researchers read research papers. When machines do research, there will be no more papers.

[kellymaxwell]

So, how close are we to AGI


making video games? Do we need AGI for that, and how close are we to achieving it? Can AI agents already reason, code, program,




script, and map? Can AI break down games, create art assets, and handle long-term planning? With better reasoning, could AI


eventually develop a game rather than just writing out ideas? Could


it also put those ideas into action? I wonder if ChatGPT-5 will be able to create games with agents or even remake old, closed-down


games like Dawn of the Dragons if all the artwork and data are available in the wiki.

[kokobanana]

Please try using less "you know" when you explain stuff.

[agi.kitchen]

@13:00 but would we know if OpenAI had switched out their neural nets or whatever to neurosymbolic? We connect via api, even in software dev not just the chats, we don’t see what’s behind that veil, you know? So it’s kind of an assumption that they’d not be testing stuff and they sure as all he;$& don’t HAAVE TO tell us

[peterhorton]

I don't see how a word probability calculator is going to get us to AGI. Help me see it.

[gingerhipster]

The Hipster Energy Team of non-materialist GPTs will soon be using this approach or something like it to expand automatically on Hipster Energy Science.

[LawJolla]

I'm in the skeptical club, but this video does a nice job showing flaws in the skeptic's specific argument. That said, there are similar arguments with various benchmarks that point to the same result.