Go Summarize

No Priors Ep. 32 | With NEAR’s Illia Polosukhin

998 views|8 months ago
💫 Short Summary

The video discusses the intersection of blockchain and AI technologies, focusing on democratizing web3 and handling content authenticity. It explores AI's role in running organizations, coordinating biotech research, and providing on-the-job feedback. The importance of identity verification, combating misinformation, and secure online interactions is emphasized. Future trends include decentralized training, data labeling marketplaces, and the integration of web3 and AI. Challenges in transitioning technologies, optimizing training datasets, and evolving AI architectures are also addressed. The video concludes with insights on technological lock-ins, historical industry shifts, and the uncertainty surrounding hardware accelerators.

✨ Highlights
📊 Transcript
Near blockchain aims to democratize web3 with over 25 million users.
00:07
Co-founder Ilya highlights the intersection of blockchain and AI technologies, emphasizing content authenticity and the alignment problem in AI.
Near initially focused on teaching machines to code but transitioned to blockchain due to challenges in paying developers.
The company's community of developers, mostly students, faced monetary control issues, prompting Near to explore blockchain solutions.
Near identified an opportunity to build a blockchain for payment solutions while anticipating AI growth.
The intersection of AI and blockchain in web3 for abstract web3 experiences.
04:36
Building a framework and platform for users to benefit without low-level implementations.
Potential overlap between AI and blockchain in creating marketplaces for data and compute models.
Highlighting the evolving role of AI agents with blockchain accounts as economic agents.
Emphasizing the ability of AI agents to communicate, pay others, and perform work independently.
AI technology used to run organizations with AI agents as CEOs.
06:00
Potential elimination of middle management by assigning specific tasks to AI agents.
Integration of blockchain and AI in biotech entities like cancer research for streamlining processes.
Coordination of experiments, recruiting candidates, and allocating funding through AI.
AI can identify the best person or lab for experiments, oversee delivery, and make unbiased decisions based on performance.
The concept of onboarding people through AI language models for on-the-job feedback is explored.
10:15
AI feedback is seen as potentially more objective and depersonalized, making employees more comfortable.
Potential applications of AI in areas like Dallas and blockchain are mentioned.
The importance of human alignment over AI alignment is emphasized, with a focus on historical roots of misinformation problems.
Addressing the need to address historical problems related to misinformation is highlighted.
Building a society to combat misinformation using AI and web3 technologies.
11:25
Web3 introduces cryptography to authenticate content and establish trust.
Unchained identity merges content and interactions to build reputation across communities.
Identity and reputation are crucial for understanding the credibility of shared information.
Implementing systematic approaches like SSL-like verification is essential for trust and authenticity online.
Importance of blockchain and cryptography in secure online interactions.
15:25
Emphasizes the need for identity verification on the blockchain to enhance security and usability.
Private keys as identity are considered too complex for mainstream adoption, leading to the development of more user-friendly identity systems.
Discussion on the potential for blockchain to revolutionize social interactions and communication.
Projects underway to expand the use of blockchain-based identity in various applications for a more extensive and user-friendly model for online interactions.
Emerging technologies like SSL and identity applications require widespread adoption to become standard options.
16:24
The potential for AI misuse to create fake content for political candidates is a growing concern.
The upcoming year may see an increase in experimental uses of technology, leading to targeted marketing and misinformation.
Law enforcement is already struggling with malicious tool usage, emphasizing the need to verify sources and content legitimacy.
Concerns regarding the lack of effective ways for law enforcement to combat audio and video fraud.
20:01
Criminals are using technology to impersonate others and engage in fraudulent activities.
Additional verification measures like embedding cryptography and implementing authenticated passes are needed.
The overlap between blockchain and AI is discussed, specifically in training models and utilizing GPU capacity.
The potential of training models in a distributed manner across a blockchain is highlighted as a future possibility.
Transitioning from proof of work to proof of stake poses challenges for older GPUs used for mining.
21:55
Core Weave stands out in the field due to their data center expertise and access to Nvidia technology.
Decentralized training is hindered by bandwidth limitations, impacting its effectiveness.
Inference, needing more compute power, presents economies of scale for broader utilization.
Privacy and scalability are crucial factors for inference, with potential solutions including existing hardware and new privacy methods like multi-party computation and zero knowledge proofs.
Benefits of decentralized web stream marketplace for data labeling.
25:12
Quality and knowledge assessment, along with an escrow model, are essential for fair pay and protection of workers.
Anyone, anywhere can get paid at any time, benefiting both workers and task providers.
The future of data labeling involves larger workforces, specific parameters, and customizable pricing.
Emphasis on quality control and domain knowledge in data labeling.
Challenges faced by companies when setting up subsidiaries in different countries.
27:12
Complexity and costs associated with hiring and training local workers are highlighted.
Contrasting traditional approach with Webster Marketplace, where students can work without a specific company contract.
Webster Marketplace allows for cross-validation tasks and incentivizes quality work and self-evaluation.
The platform is more effective and efficient in handling tasks.
Launch of a web 3 AI data marketplace project.
29:39
Potential for partnerships with existing web3 teams to improve user functionality.
Transformation of projects like Sweatcoin from web 2 to web 3 enables increased economic activity and ecosystem interactions.
Integration of web3 and AI set to replace traditional SAS models with customizable databases for different tools.
Excitement about the evolving landscape and mention of a company already exploring this space.
The future of SAS applications and business processes is discussed, highlighting a hybrid approach combining traditional UIs with programmatically interacting agents.
32:22
Dynamic UIs tailored to specific use cases are emphasized, enabling AI-driven project management and task delegation.
A shared environment for collaborative work on business processes is promoted, focusing on flexibility and customization.
Evolving model capacities within the Transformer structure are mentioned to enhance decision-making and processing abilities.
Emphasis on architecture simplicity and optimization in training datasets.
34:56
Importance of cleaner data and self-critique for improved content accuracy.
Use of language model for prediction tasks and training architecture for output decisions.
Emphasizing effectiveness of model in knowledge space search over extensive searches at inference time.
Focus on refining training processes and output feedback loops for optimal performance.
Challenges and Lock-in Effect in AI Architecture Optimization.
37:22
Current AI architecture heavily relies on GPUs, leading to a lock-in effect that may endure for the next five to ten years.
Emerging alternative architectures face hurdles in integrating with existing models like Transformers and achieving cost efficiency.
Companies will need to invest in developing new architectures to compete with GPU optimization.
Technological lock-ins persist until a major shift or plateau in optimization occurs, making it difficult to disrupt established systems.
Evolution of Hardware Technology in the 90s and Beyond.
40:00
The 90s were characterized by the dominance of Windows and Intel, known as 'Wintel', with optimized chips and systems reinforcing each other.
The rise of mobile devices in later years shifted focus to mobile-optimized chips before returning to PCs.
The demand for GPUs has created a supply imbalance, prompting companies to explore heterogeneous hardware.
The market is now filled with hardware accelerators optimized for specific architectures, leading to uncertainty in the industry's future direction.