New AI Agent, GPT-5 Not That Good? 100 Billion Humanoid Robots, Mixutre Of AGENTS And More

Links From Todays Video:

00:20 HumanPlus
06:49 Elon Musk Robots
11:48 New AI Agent Demo
16:08 OpenAI NSA
20:04 GPT-5 Not that good?
23:55 Mixture of agents

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

[TheAiGrid]

00:20 HumanPlus
06:49 Elon Musk Robots
11:48 New AI Agent Demo
16:08 OpenAI NSA
20:04 GPT-5 Not that good?
23:55 Mixture of agents

[vmooreg]

I wish the standardization of timestamps were a thing.

[MrErick]

I think the main mistake with people expecting gpt5 is in the fact they expect the model to come as a chatbot similar to gpt4, but I think k gpt5 is so advanced in its capabilities that it will not make sense releasing it as such, it will probably take the form of an agent that can do a million things and integrates seamlessly into our lives. And thats were the surprise comes from, people expecting a nex « model » when it actually comes in the form of an agent - laymans will see it close to a simulated human/ assistant / companion that stays alongside you ans helps you wuth everything. And this will feel like we’ve actually stepped into a new era again.

[robertmazurowski]

Regarding Jace, I do not believe any company saying they automate a process based on a prompt and several questions, without guidance of the model.

I just won an international Gen AI hackathon with a project that does way more than jace, but I do not claim it does it automatically, you need to write each step in a special natural language programming language.

It will be able to do what Jace "does", once models will become better.

Now I know my idea has potential.

[charbelbejjani]

Sam Altman said in his last Lex Friedman interview that he "expect[s] the delta between GPT 5 and 4 to be the same than the one between 4 and 3" (pretty sure that is almost his exact words)
And as you said we also have Microsoft saying that the models they saw were at the PhD level or something like this.

So yeah I def think OpenAI has something impressive internally. I don' think it's AGI or something, but a good step ahead of GPT-4.

[fortium]

When optimus does construction work, auto-repairs, farming and landscaping it will be sufficiently useful.

[gubzs]

People are massively underestimating the power of automation. Thought experiment, work with me here:
-Human labor in the first world is worth a bare minimum of $10/hr.
-So, a $50k robot only has to be able to do 5, 000 hours of work in its lifetime before it pays off its $50k cost.
-But let's deduct 50% of that productivity for electrical cost, downtime, and maintenance, and assuming it's less productive than a human per hour.
-5000 x 150% = 7, 500 hours to pay itself off.
-It can work 24/7 when operational, remember, we deducted 50% of that already for maintenance and downtime
-That's only *312 DAYS* of ownership before it breaks even, and then completely outcompetes a human from that point onward.

This example ignored human costs like PTO, insurance, HR expenses, payroll costs, etc. Repetitive labor is absolutely done for, and productivity for these tasks is going to go through the roof.

[LailaSharshar]

Maybe Mira is saying that they are working with many different models in he labs, not all state of the art and why we got a 4o, working laterally by optimizing existing models based on feedback, not just new models.

[magicsmoke]

"girlfriend robot, or husband robot"...
The family unit as we know it will die very quickly.

[vmooreg]

Didn’t gpt5 training begin just recently? I think people’s expectations are unrealistic at this point. Be patient.

[Aggielife]

Robots will need breaks for charging and maintenance! Remember that!

[moonsonate]

00:01 Stanford University and Google Deep Mind collaborating on humanoid robot for autonomous tasks
02:09 Unitary robot lacks degrees of freedom compared to humans
06:01 Future of humanoid robots and AI development
07:50 Humanoid robots, like Optimus, could be widely adopted in households and industries.
11:29 New AI Agent, Jace, Demonstrates Impressive Capabilities
13:47 GPT-5 has limitations but could revolutionize AI
17:46 Discussion on open AI hiring ex-NSA surveillance official
19:31 Discussion on the advancement of AGI technology and its implications.
23:13 GPT-5 capabilities uncertain until release
25:02 Combining open source models improved responses
Crafted by Merlin AI.

[TheRealUsername]

Those who stated they have AGI, even ASI internally... Their secretive policy invites speculation and for some, extreme assumptions and hallucinations especially those who doesn't know the state of research in ML.

[veganforlife]

Elon should assign a dedicated sexbot group. That way, his female employees would not get put in that awkward position of wondering, if they decline his creepy advances, will they get fired.

[Chuck_Hooks]

How do hundreds of millions of unemployed people afford these C3POs?

[edgardsimon]

maybe its because the difference of technology deployment on the market was bigger with any other revolutionary tech before and chat gpt version have big progression but it doesn't close the gap between what we got and what they had in past

[marcfruchtman]

O wow... Thanks for showing this. So, This "Shadowing" tech is amazing for "quickly" learning basic techniques! Combined with remote manipulation for fine control, it will revolutionize robotics.
Regarding the robot estimate of 1 to 1 or even 2 or 3 to 1 human, I think that estimate is totally on par. However, I don't think it will happen for the whole world as quickly... due to initial costs, sort of like adopting computers in the early days, when they were much more expensive. So, yea, the first few years will see fairly high costs in 70 to 100k range, until one or two "local" manufacturers can pump them out at lower prices.

[TiagoTiagoT]

Regarding AI capabilities, i think we're being gaslighted in both directions; lots of the stuff that is promised is underdelivered, and at the same time lots of the understating is being done qualm fears of the population and let the companies keep the most advanced stuff to themselves hidden in plain sight after having already demonstrated it's exists.

[ThomasTomiczek]

Robot: That is not a new way of teleoperation by definition, but a nice way to have something SHOWN to him. That said, the base training of 40 hours of human motions is comically low to start with. But it is those small steps that make us soon have robots that are amazing at home - just some years, as the hardware and the training clusters get a TON faster - fast.

Elon: Elon is obviously correct here - timelines are complex unless you are surreally conservative. But it is quite obvious that one day we do a ton of robots - I personally see Tesla rebuilding most factories for robots (they change the way they build cars now so that they only need about half the floor space - modules that then get put together at the end) internally - like 50% of the floor space - and then I see in 2030 that they run up 1 million robots... PER DAY. There will be a brutal demand explosion the moment they can do ANYTHING and there will be epic waiting lines. The endgame is definitely robots doing it all - and while one can argue it took decades to get us to where we are - we started VERY low, and now we STILL get a LOT more capable with hardware and algorithms every 18 months. And on the software side I read yesterday paper that makes transformers a LOT faster - likely 20 to 30 times for large layers, by eliminating the matrix multiplication. We do not need AGI in the robot, btw - imagine the robot being smart, but not AGI, but the HOUSE (or apartment) AI (which may live in the cloud) do the complex coordination. It will not be B2B - household work, elderly care, nurse at home.... here you go.
Also, now hitting another separation from my partner - I can tell you that soon people will ""marry" an AI - not for sex, not even necessarily for companionship, but it will be there for you, help you, organize your life etc.. Like you say "can you get me a holiday next week in Turkey" and it knows your favourite hotel, does all the reservations etc. Imagine a trusted Butler always on your side. THAT is the future. People will be raised by AI/Robots - at least partially, because it will always be there. Teachers will get replaced - home schooling with AI...

Jace: It does not look impressive, but it is amazing. Even if you need a lot of backend information that you insert - like instructions how to deal with specific things - that is the future. it will flow into the main AI system. It is not totally new - I know others handling and planning multi hundred step processes. Remember, the current version of Jace is the worst it will ever be. Same as Devin. Things will change fast and brutal.

Re OpenAI Models: Remember that GPT-5 has JUST STARTED TRAINING. What they have are only architectural demonstration so far - small to see whether it works. The last model is GPT-4 which is based likely on the GPT-4 architecture and main data set + some images and definitely a lot of Audio, likely as "ok, let's try that out first" approach, and that is public. It makes sense that they do not have GPT-5 finished yet. I just read in an interview where someone said (in April) that he has seen GPT-5 and it is not that impressive. That contradicts the timeline that is public - so either he saw a demo model of GPT-5 (a small validator that IS not that impressive but there to test that it works) or GPT-4o which.... is not GPT-5. I would say thy really have an idea how GPT-5 will look, but it is not finished training. By their history they say that they can evaluate capability by MUCH smaller models (1000:1) to check the architecture. We haven o plateau now because we have no new model - as I Said, GPT-4o is more an extension of the 4 architecture (definitely a total retraining as it has a different tokenizer) built also to test the integration of audio.

[swaydaygaming]

@19:35 Nolan is def gonna make the next movie about it in the next couple decades lol