[alancalverd]

Not necessarily. It would weigh more, and take more space.

Most dogs weigh less than most humans, take up less space, and can run faster. A camera tripod is much more stable than a human hand,

There is an old saying that if you grow up on a farm, you learn to do absolutely everything, but not very well. That's humans in general. 

Surgical robots can have more or fewer hands than a human, operating in the same space but doing a much better job.

[hamdani yusuf (OP)]

Not necessarily. It would weigh more, and take more space.

Most dogs weigh less than most humans, take up less space, and can run faster. A camera tripod is much more stable than a human hand,

There is an old saying that if you grow up on a farm, you learn to do absolutely everything, but not very well. That's humans in general. 

Surgical robots can have more or fewer hands than a human, operating in the same space but doing a much better job.

The development of humanoid robots don't necessarily stop the development of robots with other shapes and sizes. As I said,

The reason for building humanoid robots is for versatility. Working environments have been optimized for median human bodies. If they are expected to do everything that human individuals can do, then having the same physical form factor is an obvious starting point.

There is no unshakeable reason why future human descendants will have to keep current physiology, shape and size.

[hamdani yusuf (OP)]

Apple's GSM-Symbolic Paper does NOT Disprove Reasoning - Paper Review

Many people have used the GSM-Symbolic paper as evidence that large language models (LLMs) do not reason. Despite its popularity, the paper shows many flaws, which I will cover in this video. I evaluate the paper by connecting it to the wider body of evidence and relevant AI concepts. It fails to show that models don't reason and ironically instead provides data that support the idea that bigger models generalize better.

Sidenote: I use 'reasoning' and 'generalization' interchangeably, but the technically correct terminology is 'generalization'. Reasoning requires test-time compute. See the chapter 'Benign overfitting' for the difference.

Correction: In the chapter 'Concept: Filler Tokens' I mentioned the recursive structure of autoregressive models. This is however not relevant, since the filler tokens add computation in a parallel manner not in a sequential (recursive) manner. Consider that in the paper, each non-sensical clause adds "filler" tokens to the input. Every filler token can "store computation" by being processed by each layer. This thus adds parallel computation. Contrast this with training a model to output filler tokens in its output as a CoT which would add sequential computation.

Apple's Paper: https://arxiv.org/abs/2410.05229
Paper similar to Apple with opposite conclusions: https://arxiv.org/abs/2405.00332
Anthropic Mechanistic Interpretability Paper: https://arxiv.org/abs/2301.05217
The empirical evidence on filler-tokens is quite weak, so take it with a grain of salt. The theoretical idea of adding parallel computation is still intriguing though and explained in this article: https://www.lesswrong.com/posts/oSZ2x...
Paper on the lottery-ticket hypothesis (used for benign overfitting): https://arxiv.org/abs/1803.03635

Timestamps:
00:00 - Introduction
00:57 - Paper's Claims
01:41 - Dataset
02:09 - Concept: Benign Overfitting
04:43 - Related Work
04:58 - Concept: Mechanistic Interpretability
05:26 - Concept: System 1 & System 2
06:13 - Contradictory Paper
08:05 - Results: Accuracy drop
08:55 - Results: Variance
10:03 - Results: Adding complexity
12:10 - Concept: Autoregressive Models
14:03 - Results: Adding noise
15:10 - Concept: Filler Tokens
15:56 - Conclusion

[hamdani yusuf (OP)]

How AI Learned to Think

My goal here is to introduce model based learning and show how language understanding merged with gameplay AI strategies recently. From early chess engines to modern language models. We examine key breakthroughs in game-playing AI?TD-Gammon, AlphaGo, and MuZero?and their contribution to current large language model architectures. Special focus on the convergence of Monte Carlo Tree Search (MCTS) with neural networks, and how these techniques transformed into today's chain-of-thought reasoning.SUPPORT this work:   / artoftheproblem 

Timestamps:
00:00 intro
01:00 definition of reasoning
03:57 intuition
06:35 MCTS
07:40 AlphaGO
09:37 World Models
10:36 MuZero
12:45 Chain/Tree of Thought
14:03 RL on Reasoning
15:41 ARC AGI Test

[hamdani yusuf (OP)]

Artificial Super Intelligence Might be Here Already....

A viewer tried to summarize in the comment section.

Summary:
MIT research indicates artificial superintelligence (ASI) may be imminent.  AI dramatically accelerates scientific discovery, particularly in materials science, compressing the innovation cycle and raising concerns about job displacement.  Deep learning's application, especially GNNs, enables rapid material design, resulting in a surge in new discoveries and patents.  However, this progress highlights the need for scientists to reskill, with one researcher stating their education felt "worthless."  The observed efficiency gains and potential for AI to effectively outsource tasks to scientists suggest ASI's arrival is approaching, promising unprecedented technological and societal change.

[hamdani yusuf (OP)]

Q STAR 2.0 - new MIT breakthrough AI model IMPROVES ITSELF in REAL TIME (new Strawberry?)

00:00 Q star 2.0
01:54 Some Key Terms
04:32 ARK AGI
05:13 Francois Chollet
10:36 Test Time Compute

[alancalverd]

Would you really trust a learned paper by someone who can't spell "disappointing"?

[alancalverd]

with one researcher stating their education felt "worthless." 

More likely that someone with a good education can recognise worthless research.

[hamdani yusuf (OP)]

Would you really trust a learned paper by someone who can't spell "disappointing"?

Perhaps English is not his first language.

[hamdani yusuf (OP)]

Open-Source Q-Star! The First OPEN "Thinking" Model (DeepSeek r1)

AI News: Musk Says AGI 2026, Open-Source Q*, Flux.1 Updates, Quantum AI, and more!

Chapters:
0:00 Elon Musk AGI Timeline
0:26 Figure x BMW
1:41 Giveaway
2:20 Flux.1 Release
4:17 1m Token Qwen
5:02 ElevenLabs Agents
5:48 Pokemon GO Data
7:01 V4 Release
8:11 Gemini Remembers
9:05 OpenAI Voice Update
9:33 Open-Source o1
10:04 Quantum Google
11:21 GPT-4o Update

[alancalverd]

Would you really trust a learned paper by someone who can't spell "disappointing"?

Perhaps English is not his first language.

All the more reason to use a dictionary. First imppresions count.

[hamdani yusuf (OP)]

Would you really trust a learned paper by someone who can't spell "disappointing"?

Perhaps English is not his first language.

All the more reason to use a dictionary. First imppresions count.

Or AI assistant.

[hamdani yusuf (OP)]

AI Could Make Quantum Computing Obsolete, Nobel Prize Winner Says

Last week, DeepMind?s Demis Hassabis said that AI might be able to solve problems that quantum computers were supposedly necessary for. Indeed he said that classical systems ? AI run on conventional computers ? can model quantum systems. Sounds like an innocent claim but is certain to upset a lot of quantum computing researchers. Hassabis bases his argument on the surprising success of Alphafold.

Somehow I already thought that success of Alphafold wasn't that surprising.

[hamdani yusuf (OP)]

This New AI Is SCARY - Like SORA But LIMITLESS

The Matrix is a groundbreaking AI model capable of generating infinite, high-quality video worlds in real time, offering unmatched interactivity and adaptability. Developed using advanced techniques like the Video Diffusion Transformer and Swin-DPM, it enables seamless, frame-level precision for creating dynamic, responsive simulations. This innovation surpasses traditional systems, making it a game-changer for gaming, autonomous vehicle testing, and virtual environments.

🔍 Key Topics Covered: 
The Matrix AI model and its ability to generate infinite, interactive video worlds 
Real-time applications in gaming, autonomous simulations, and dynamic virtual environments 
Revolutionary AI techniques like Video Diffusion Transformer, Swin-DPM, and Interactive Modules 

🎥 What You?ll Learn: 
How The Matrix AI redefines video generation with infinite-length, high-quality simulations 
The transformative impact of real-time interactivity and domain generalization in AI-driven worlds 
Why this breakthrough is a game-changer for industries like gaming, VR, and autonomous systems 

📊 Why This Matters: 
This video uncovers a groundbreaking AI innovation that merges real-time interactivity with infinite video generation, setting the stage for a future of responsive and immersive virtual environments. 

DISCLAIMER: 
This video delves into The Matrix AI model, highlighting its advancements, capabilities, and potential to revolutionize technology and industry through innovation. 

This was what I imagined when starting this thread.

[hamdani yusuf (OP)]

AlphaProteo - Google DeepMind's Breakthrough AI for "Protein Design"

[hamdani yusuf (OP)]

NVIDIA CEO Jensen Huang presentation

Highlights of #nvidia ( #nvda stock ) Founder and CEO Jensen Huang speaking at AI Summit India. Highlights include how Nvidia dominated AI computing once #openai released #chatgpt , why Moore's Law no longer works, the generative AI breakthroughs Jensen Huang expects now that NVIDIA Blackwell is in production, and much more.


Timestamps for this Nvidia AI Summit Supercut:
00:00 Moore's Law is Dead - The Generative AI Era
05:56 NVIDIA Blackwell Data Center Accelerators
10:26 NVIDIA Generative AI Scaling 4x Per Year
13:47 NVIDIA AI Agents & Omniverse for Robots

[hamdani yusuf (OP)]

The REAL Reason People Are Scared of AI

6 Ways AI Could Go Wrong
Artificial Intelligence is advancing at a pace faster than anyone could have previously predicted. Legislators across the planet race to keep up and protect us from what they refer to as ?Nightmare Scenarios? - Here are 6 of those situations.

-- VIDEO CHAPTERS --
00:00 Intro
02:52 Predictive Policing
06:02 Elections
09:12 Social Scoring
14:57 Nuclear Weapons
18:32 Critical Sectors
24:12 Optimist?s Take
25:15 Credits

Correction:
1:19 We misspelled "Python" here - oops!

[alancalverd]

The REAL Reason People Are Scared of AI

People are scared of people. AI is just one more weapon for bad people to use against others.

[hamdani yusuf (OP)]

The REAL Reason People Are Scared of AI

People are scared of people. AI is just one more weapon for bad people to use against others.

How do you define bad people?

[hamdani yusuf (OP)]

Jen-Hsun Huang: Stanford student and Entrepreneur, co-founder and CEO of NVIDIA

Every successful thing needs to be torn down and rebuilt. In more of an intimate conversation than a lecture, Huang relates his experience as an engineer, entrepreneur, and innovator. Find out how he meets the ever-constant challenge of re-invention as co-founder and CEO of NVIDIA.

Here are some interesting comments.

"Innovation needs a lot of experimentation, experimentation needs exploration, explorations will result in failures. If you do not have tolerance for failures, you wont succeed" - Jen-Hsun Huang

00:02 Introduction to the Entrepreneurial Thought Leader Lecture Series at Stanford
01:49 Jensen Wang, co-founder of NVIDIA and generous donor to the school of engineering.
06:35 Having a unique perspective shapes vision and opportunities
09:02 NVIDIA CEO's vision for 3D graphics technology in gaming industry
13:23 Importance of perspective in entrepreneurial decisions
15:37 Perspective and vision matter in shaping the trajectory of a company.
20:03 Ignoring customers can be necessary for business innovation
21:57 Transitioning NVIDIA to programmable 3D graphics processors
26:04 Passion is essential for building a company
28:00 Reinventing the company through programmable shaders and taking big risks
32:00 Resource allocation and economic decision-making
33:57 Culture of innovation and risk-taking at NVIDIA
38:06 Importance of risk-taking and flexibility in entrepreneurship
39:55 Utilizing GPU for computational graphics beyond traditional graphics applications
43:39 Equal pay and share are fair and simple.
45:38 CEOs and leaders need to be comfortable with ambiguity.
49:13 Incorporating NVIDIA with minimal funding
51:11 VCs invest in great people with a large market vision
54:52 Reinventing the company every 10 years is a necessary and challenging process
57:02 Survival is important: Cash is always king
1:00:52 Start a company based on passion, not money
1:02:56 Unique perspective and perseverance are key to success

Jen-Hsun at 40:02 saw the world of 2024 in 2011. This is why his company has a 2 trillion dollar evaluation today. Incredible.