STOP. Using AI Right now

Share

Summary

In this video, Felix shares his journey of building a supercomputer, running protein folding simulations for scientific research, and diving deep into self-hosting AI. He discusses the challenges and triumphs of maximizing his computer's capabilities to run large AI models, and critically examines the broader implications of AI, focusing on privacy and control.

Highlights

Future AI Plans and Sponsor: NordVPN
00:23:48

Felix expresses his enjoyment in running and training AI, hinting at training his own AI model to be shared later. He then introduces NordVPN as another sponsor, advocating for its essential role in online privacy and security. He shares anecdotes about the risks of unprotected internet use and emphasizes NordVPN's ability to safeguard against hacks and unlock the full internet. He offers a special deal for viewers using his link, including an additional four months free and a 30-day money-back guarantee, also mentioning their Marvel collaboration.

Bifurcation and Protein Folding Simulations
00:00:00

Felix confesses to bifurcating his PCIe lanes again to fit more GPUs, ultimately reaching ten GPUs. He explains his initial goal to max out his computer. Despite the power consumption concerns, he undervolts his RTX 4090. He then reveals that he uses his beastly computer for protein folding simulations, contributing computing power to scientific research to understand diseases. He encourages viewers to join his team for this initiative, highlighting its legitimacy and impact on charity leaderboards.

Sponsor: SURFshark for eSIMs
00:04:03

Felix introduces SURFshark as a sponsor, promoting their eSIM service for seamless internet connectivity while traveling. He praises its convenience, affordability, and privacy-respecting features, especially under the Nord company. He offers a 15% discount using the code 'PewDiePie' at checkout and encourages viewers to prepare for their next trip with SURFshark.

Navigating the AI Landscape and Maxing Out Models
00:05:56

Felix ventures into running AI on his multi-GPU setup, aiming to run the biggest AI model possible. He expresses his distaste for AI-generated images and the overuse of AI as a buzzword. He discusses the misinformation surrounding AI self-hosting and his quest to run a 120 billion parameter model (Llama 70B initially, then a larger open-source model released by OpenAI). He successfully runs a 245B model, pushing his computer's limits and even increasing token capacity significantly, demonstrating its practical use despite the challenges and the unexpected 'too much power' feeling.

Building a Personal AI and Web UI
00:13:00

Felix shares his experience learning to program and building a custom web UI for his self-hosted AI. He showcases its speed, audio capabilities, and added features like search, memory, and RAG (Retrieval Augmented Generation) for accurate information. He emphasizes the importance of privacy-respecting, free options for self-hosting AI, criticizing the reliance on APIs and advocating for solving problems independently. He also jokingly admits to 'stealing' code from a Chinese laboratory for deep research capabilities.

The Importance of AI Privacy and Data Control
00:16:54

Felix highlights the critical aspect of memory in AI and the illusion of privacy with cloud-based AI services. He reveals how even deleted chats can still be used for training, emphasizing the need for self-hosting to maintain control over personal data. He illustrates this by demonstrating his AI recalling his personal details from his own Google data, highlighting the discomfort of AI knowing too much and the contrasting comfort when the data is locally controlled.

The AI Council and Swarm Experiment
00:19:28

Felix introduces his 'AI Council,' where eight GPUs run the same AI model with different personalities, voting on responses to queries. He humorously recounts replacing underperforming council members and their eventual collusion against him. He then experiments with a 'swarm' of 64 tiny AI models running simultaneously, realizing the strength of smaller models with added tools like search and RAG for fast information retrieval. He concludes that a beast computer isn't necessary for running AI effectively, but rather a proper toolset.

Recently Summarized Articles

Loading...