You have 4 summaries left

"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis

Investing in AI for Hard Tech, with Eric Vishria of Benchmark and Sergiy Nesterenko of Quilter

Thu Jun 13 2024
AIVenture CapitalCircuit Board DesignReinforcement LearningEthical Data Sourcing

Description

The episode covers investing venture capital into AI technology and its application to hardware problems. It explores the idea of using AI for circuit board layout and problem-solving, as well as the challenges and potential of reinforcement learning in circuit board design. The Brave Search API is discussed in relation to ethical data sourcing for AI models. Thoroughness in circuit board design and simulation methods are examined, along with diffusion models and challenges in circuit board design. The development of advanced AI technology and different categories of AI companies are explored, as well as the challenges in implementing AI solutions. The importance of working with infrastructure companies and the role of obsession in today's environment are highlighted.

Insights

Investing in AI Era

Benchmark's investment philosophy involves identifying companies with potential for long-term success in the AI era.

Disruptive Solutions

The concept of disruptive solutions in the market is explored, with an emphasis on starting at the lower end and expanding to serve a broader audience.

Reinforcement Learning in Circuit Board Design

Reinforcement learning allows for achieving superhuman levels of performance by training systems without human data input.

Ethical Data Sourcing

Using the BraveSearch API can lead to more ethical data sourcing and human representative datasets for AI models.

Thoroughness vs Speed in Design Process

Users have the choice to prioritize speed or thoroughness in the design process based on their needs and constraints.

Diffusion Models in Circuit Board Design

Diffusion models could be applied to circuit board design, similar to their use in other domains like proteins

Challenges in Circuit Board Design

Challenges in circuit board design include data quality and identifying actual working boards among many designs

Categories of AI Companies

There are different categories of AI companies, including foundational model companies and infrastructure companies like AI chip and systems providers.

AI Development Challenges

Foundational models in AI are expensive to develop and quickly depreciate, posing a challenge for venture capital investments.

Importance of Obsession

Obsession is considered crucial in today's fast-moving environment.

Chapters

  1. Investing Venture Capital into AI Technology and Hardware Problems
  2. AI for Circuit Board Layout and Problem-Solving Approach
  3. Data Challenges and Reinforcement Learning in Circuit Board Design
  4. Brave Search API and Ethical Data Sourcing for AI Models
  5. Thoroughness in Circuit Board Design and Simulation Methods
  6. Diffusion Models and Challenges in Circuit Board Design
  7. Developing Advanced AI Technology and Categories of AI Companies
  8. Categories of AI Development and Challenges in Implementation
  9. Working with Infrastructure Companies and Importance of Obsession
Summary
Transcript

Investing Venture Capital into AI Technology and Hardware Problems

00:00 - 07:38

  • Podcast focuses on investing venture capital into AI technology and its application to hardware problems
  • Discussion includes the vision for superhuman circuit board design, reinforcement learning approach, and societal impact
  • Benchmark's investment philosophy involves identifying companies with potential for long-term success in the AI era
  • Quilter stands out due to the increasing use of electronics in products, manual process of circuit board design, and unique approach not relying on copilots

AI for Circuit Board Layout and Problem-Solving Approach

07:17 - 14:03

  • The podcast discusses the idea that certain tasks, like circuit board layout, should be handled by AI rather than human assistance.
  • There is a focus on using the right kind of AI for specific problems, not just following popular trends like large language models.
  • The guest's experience and approach to problem-solving are highlighted as key factors in determining potential success in solving complex issues.
  • The concept of disruptive solutions in the market is explored, with an emphasis on starting at the lower end and expanding to serve a broader audience.

Data Challenges and Reinforcement Learning in Circuit Board Design

13:34 - 19:56

  • Deep learning approaches require a significant amount of data, but open source datasets in the field are limited and often locked behind companies like Apple or Google.
  • Using data for supervised learning can lead to board designs with unnecessary complexity and cost, while reinforcement learning offers the potential to surpass human capabilities.
  • Reinforcement learning allows for achieving superhuman levels of performance by training systems without human data input.
  • Sparse reward problem in circuit board design is addressed by breaking down the problem into smaller steps and using heuristics to guide decision-making.
  • Long-term goal is to have a single sparse reward indicating whether a board will work or not, but current focus is on automating processes based on existing human heuristics.

Brave Search API and Ethical Data Sourcing for AI Models

19:31 - 26:40

  • The Brave Search API offers affordable developer access to an independent index of the web with over 20 billion pages, free from big tech biases.
  • The Brave Search Index is built from real page visits by humans, refreshed daily for accurate and up-to-date information.
  • Using the BraveSearch API can lead to more ethical data sourcing and human representative datasets for AI models.
  • Discussion on reinforcement learning processes in language models and the reliability of human reward signals in training.
  • AI models are good at interpolation within known bounds, while humans excel at extrapolation and creativity.

Thoroughness in Circuit Board Design and Simulation Methods

26:11 - 32:30

  • The depth of search in a system like AlphaGo greatly impacts its performance, with more thorough searches leading to superhuman results.
  • Users have the choice to prioritize speed or thoroughness in the design process based on their needs and constraints.
  • The compute resources for designing circuit boards involve a mix of CPU and GPU cores, with a focus on training during production runs.
  • The company is expanding by addressing different physics considerations step by step, starting with low-speed boards before moving on to more complex simulations.
  • Various methods, such as finite difference time domain and FEM, are used to solve differential equations in electromagnetics simulations for circuit board designs.

Diffusion Models and Challenges in Circuit Board Design

32:15 - 38:43

  • Diffusion models could be applied to circuit board design, similar to their use in other domains like proteins
  • Challenges in circuit board design include data quality and identifying actual working boards among many designs
  • Manufacturers receive limited information about signal behavior on circuit boards, making evaluation challenging
  • Generating data points from verified physics principles is crucial for training diffusion models in circuit board design
  • Talent acquisition is a bottleneck in advancing towards superhuman circuit board designs
  • There are various checkpoints and steps in the process of improving circuit board designs incrementally

Developing Advanced AI Technology and Categories of AI Companies

38:19 - 45:01

  • The speaker discusses the checkpoints and challenges in developing advanced AI technology.
  • They believe that the problem of creating superhuman circuit board design is not unsolvable, though it will be difficult and require significant effort.
  • The focus of the company mentioned is on developing a compiler for electronics layout to simplify the process of creating circuit boards.
  • There are different categories of AI companies, including foundational model companies and infrastructure companies like AI chip and systems providers.

Categories of AI Development and Challenges in Implementation

44:43 - 51:40

  • There are three main categories in AI development: training, infrastructure enabling other advancements, and vertical applications like Quilter.
  • Foundational models in AI are expensive to develop and quickly depreciate, posing a challenge for venture capital investments.
  • Infrastructure companies have business opportunities but face uncertainty about the longevity of their solutions similar to early software development stages.
  • Vertical AI applications show rapid revenue growth, but concerns exist about barriers to entry and durability against competitors.
  • There is a gap between research-based AI code and production-ready scalable products that needs bridging for successful implementation.
  • The merging of DeepMind into Google reflects the shift towards engineering challenges in applying existing research to real-world applications.

Working with Infrastructure Companies and Importance of Obsession

51:16 - 52:39

  • The speakers are excited about working with companies in the infrastructure sector and meeting passionate individuals with long-term ideas.
  • They aim to be involved in the best companies at an early stage.
  • Obsession is considered crucial in today's fast-moving environment.
1