Artificial Intelligence thread

tokenanalyst

Brigadier
Registered Member

Integrated electromagnetic sensing system based on a deep-neural-network-intervened genetic algorithm.​

Abstract​

With the deepening integration of artificial intelligence (AI) and the Internet of Things (IoT) in daily life, electromagnetic sensing presents both attraction and increasing challenges, especially in the diversification, accuracy, and integration of sensing technologies. The remarkable ability of metasurfaces to manipulate electromagnetic waves offers promising solutions to these challenges. Herein, an integrated system for electromagnetic sensing and beam shaping is proposed. Improved genetic algorithms (GAs) are employed to design the metasurface with desired beams, while spatial electromagnetic signals sensitized by the metasurface are input into the GA enhanced by deep neural networks to sense the number of targets, their azimuths, and elevations. Subsequently, the metasurface device is designed as the hybrid mode combining tracking and avoidance in alignment with practical requirements and sensing outcomes. Simulation and experimental results validate the efficiency and accuracy of each module within the integrated system. Notably, the target sensing module demonstrates the capability to precisely sense more than 10 targets simultaneously, achieving an accuracy exceeding 98% and a minimum angular resolution of 0.5°. Our work opens, to our knowledge, a new avenue for electromagnetic sensing, and has tremendous application potential in smart cities, smart homes, autonomous driving, and secure.​

1738960074372.png

Please, Log in or Register to view URLs content!
 

AndrewS

Brigadier
Registered Member
5 hour Podcast with Semianalysis et al below.

DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast
Please, Log in or Register to view URLs content!


Key points for me are:

1. Reinforcement learning without human input is feasible, as proven by Deepseek. Therefore agents can train by themselves on verifiable tasks eg. booking an airplane ticket, buying stuff, etc. Tech companies expect 2-3 years to develop such autonomous agents.

2. Whoever figures out how to serve up advertisements on an LLM will make a fortune

3. For Inference, output tokens are typically 4x more expensive than input tokens. Outputs have to be processed serially with every increasing memory requirements. So for inference, total addressable memory is more important than FLOPS?

4. Nvidia cancelled orders for 2 million H20s (the China-specific GPU)

5. Stargate Phase 1 is:
a) $1 Bn for the datacentre
b) $5-6 Bn for the servers

6. Total Stargate buildout would be $50 Bn. Total cost of ownership would be $100 Bn

7. OpenAI is renting Stargate from Oracle

8. Semianalysis expect Chinese semiconductor industry to eventually catch up to world class (implied 10 year timeframe)

9. US needs to retain AI lead over the rest of the world, no matter what it takes

10. US would need 10 years and $1 Trillion to build out a world-class semiconductor capability domestically
 

GulfLander

Colonel
Registered Member
"[...]we're excited to introduce reasoning in Unsloth! DeepSeek’s R1 research revealed an “aha moment” where R1-Zero autonomously learned to allocate more thinking time without human feedback by using Group Relative Policy Optimization (GRPO).

We've enhanced the entire GRPO process, making it use 80% less VRAM than Hugging Face + FA2. This allows you to reproduce R1-Zero's "aha moment" on just 7GB of VRAM using Qwen2.5 (1.5B).[...]"
Please, Log in or Register to view URLs content!
 
Top