David Luan: Why Nvidia Will Enter the Model Space & Models Will Enter the Chip Space | E1169

Intro (00:00:00)

• OpenAI realized that the next phase of AI after Transformer models would involve solving major unsolved scientific problems.
• There are still ways to improve model performance, which will require a significant amount of compute.
• David Luan is excited to be on the podcast and has watched previous episodes.
• David Luan has worked at several incredible companies, which have served as training grounds for him.

Lessons from Google Brain & Their Influence on Building Adept (00:01:03)

• Google Brain was dominant in AI research during the 2012-2018 period.
• Google Brain's success was due to its ability to attract top talent and provide an environment that encouraged pure bottom-up basic research.
• Pure bottom-up basic research involves hiring brilliant scientists, giving them the freedom to work on open technical problems in AI, and allowing them to collaborate and share ideas.
• This approach led to significant breakthroughs, such as the invention of the Transformer model.
• The Transformer model, invented at Google Brain, was a general-purpose model that could be applied to a wide range of machine learning tasks.
• Prior to the Transformer, different models were required for different tasks, such as image understanding, text generation, and playing games.
• The Transformer's versatility made it the foundation of modern AI and reduced the need for low-level breakthroughs in modeling.
• Researchers could now focus on using the Transformer to solve complex real-world problems.

Why It Took 6 Years for ChatGPT to Emerge After Transformers (00:05:06)

• The success of the show is attributed to the willingness to ask simple questions.
• ChatGPT is seen as a consumer breakthrough after the introduction of Transformers in 2017.
• The gap between the Transformer breakthrough and consumer adoption of ChatGPT is questioned.
• Language models have been gradually improving since the introduction of Transformers in 2017.
• GPT-2, released in 2019, demonstrated impressive generalist capabilities.
• Two key factors contributed to the viral success of ChatGPT:
  • Increasing intelligence of language models reaching a compelling level.
  • User-friendly packaging that allowed consumers to interact with the models.
• The GPT-3 API was available over a year before ChatGPT, but it lacked the consumer-friendly interface, hindering its widespread adoption.

Takeaways from OpenAI (00:06:49)

• OpenAI realized that the next phase of AI after Transformer was not about research paper writing but about solving major unsolved scientific problems.
• Instead of loose collections of researchers, OpenAI focused on building a culture around large groups of scientists working on specific real-world problems.
• This approach is different from the traditional academic curiosity-driven research and is similar to the Apollo project, where the goal is to solve a specific problem rather than just conducting research.
• Historically, scaling up language models has shown diminishing returns for every incremental GPU added.
• However, every doubling of GPUs has resulted in very predictable and consistent returns.
• To make a base language model predictably and consistently smarter, the amount of compute needs to be doubled.

The Key Bottleneck in AI Model Performance (00:09:57)

• Model scaling involves enlarging the model and having it gather data to enhance its intelligence.
• The focus is shifting from base model scaling to more efficient methods like simulation, synthetic data, and reinforcement learning.
• Models trained on data have limitations in discovering new knowledge or solving problems not in the training set.
• Chatbots and agents are diverging into distinct technologies with varying use cases and requirements.
• Hallucinations in chatbots and image generators can be beneficial for problem-solving, novelty, and creativity.
• Agents performing tasks like handling taxes or shipping containers should not hallucinate to avoid errors or problems.

Understanding Minimum Viable Capability Levels & Model Scale (00:16:06)

• AI systems have become less predictable as they grow larger and more complex.
• The capabilities of a model cannot be fully predicted in advance, even with estimates.
• Minimum viable capabilities are a function of model scale, meaning certain capabilities emerge at specific model sizes.
• It is difficult to predict the exact resource investment needed to achieve specific capabilities.
• Reasoning is a challenging problem in AI that requires new research.
• Pure model scaling alone does not provide solutions to reasoning.
• Reasoning involves composing existing thoughts to discover new ones, which cannot be achieved solely through internet data regurgitation.
• Solving reasoning requires providing models with access to theorem-proving environments, similar to how human mathematicians approach problem-solving.
• The general capability of reasoning needs to be solved at the model provider level to improve the model's ability to reason.

The Future of the Foundational Model Layer (00:20:17)

• The high costs involved in reasoning skills limit the number of long-term steady-state LLM providers to a maximum of 5-7.
• Solving reasoning involves training a base model, granting it access to various environments to solve complex problems, and incorporating human input.
• Memory in AI presents challenges in both short-term working memory and long-term memory, with progress made in the former but not the latter.
• Application developers should integrate long-term memory about user preferences as part of a larger system.
• LLMs are components of a larger software system, and the winning LLM providers will be those that existentially need to win, such as tier-one cloud providers.
• Nvidia should expand its chip business by offering LLM services, while major LLM providers are developing in-house chips for better margins.
• The interface of the LLM provides significant leverage downstream, leading to expected vertical integration pressure between model builders and chip makers.
• Chip makers need to own something at the model layer to avoid commoditization.
• Apple has an advantage in running smart models for free at the edge but may not be enough to compete with the smartest models.
• Apple will excel in private, fine-tuned models that don't require extensive reasoning and will run on the edge.
• OpenAI's GPT-40 represents a scientific advancement towards universal models that can process various inputs and generate any output.
• Apple's partnership with OpenAI suggests recognition of OpenAI's progress and hints at a future where foundational models become commoditized.
• Tier-one cloud providers will invest heavily in foundational models, while independent companies selling models to developers must either become first-party efforts of big clouds or build a substantial economic flywheel before commoditization.
• Adept's focus on selling end-user-facing agents to enterprises differentiates its business model from companies selling models to developers.
• OpenAI's Chat GPT provides a significant advantage in building an economic flywheel for independent foundational model companies.

Adept’s Focus for Vertical Integration for AI Agents (00:33:26)

• Adept is focused on building an AI agent that can handle arbitrary work tasks.
• Adept is not trying to train foundation models to sell to others, but rather building a vertically integrated stack from the user interface to the foundation modeling layer.
• Adept believes that vertical integration of model with use case is the only way to solve the generalization problem and handle the variability of agent requirements across industries.
• Adept's advantage is that it is building a system that can handle any workflow or task in any enterprise, rather than focusing on a particular narrow problem.
• Every enterprise workflow is an edge case, and controlling the entire stack is necessary to handle the variability and complexity of these workflows.

The Distinction Between RPA & Agents (00:35:53)

• RPA is suitable for repetitive tasks, while agents require constant thinking and planning to solve goals.
• Agents disrupt the business model of RPA players by demanding self-serve capabilities and addressing challenging use cases.
• The pricing model may shift from per-seat to consumption-based in some areas, but valuable knowledge work will not be priced that way due to AI agents empowering people to explore new possibilities and enhance their productivity.
• The objective is to develop AI systems that act as co-pilots or teammates.
• Unlike traditional workers, these AI systems should not charge based on their workload but rather on their ability to augment human capabilities and create new opportunities.

The Co-pilot Approach: Incumbent Strategy or Innovation Catalyst (00:40:24)

• Co-pilots can be a great incumbent strategy for companies to morph their existing software business model into AI.
• AI is not going to take all jobs, but it will augment human capabilities and make them more efficient.
• Co-pilot style approach is necessary for humans to effectively use AI systems.
• Collapsing the talent stack, where individuals have multiple skill sets and can simultaneously fulfill different roles, will make projects and teams more effective.
• Humans at work will become more like generalists, with larger scope over various functions, while supervising specialized AI co-pilots.

Enterprise AI Adoption Budgets: Experimental vs. Core (00:42:46)

• Most Enterprise AI adoption is still in the experimental budget phase.
• Many Enterprises still have a lot of on-prem infrastructure and workflows running on mainframes.
• Cloud technology, which is considered mature from a startup perspective, still doesn't have full adoption in Enterprises.
• There is a concern that AI adoption may follow a similar pattern to autonomous driving, with a period of plateauing progress.
• Unlike self-driving cars, where there was an initial breakthrough followed by incremental improvements, AI models are constantly improving with new scientific advancements.
• Breakthroughs like Universal multim lt40 are visible and continue to enhance the capabilities of AI models.
• AI models are already deployed and don't need to reach a specific level of reliability before deployment.
• AI services companies, which help in the implementation of AI in large Enterprises, are projected to be bigger than the model providers themselves in terms of revenue.
• This projection has been proven true, with some people who criticized the prediction later acknowledging their mistake when the revenue numbers were revealed.

AI Services Providers vs. Actual Providers (00:46:53)

• AI Services providers may not be the biggest players in the long term.
• Companies that turn use cases with product-market fit into repeatable products will likely be the real economic winners.
• Many AI services will eventually be turned into generalizable products.
• There is a concern that regulations on data usage and collection could hinder the progression of AI models.
• Regulatory capture is a concern, with lawmakers potentially favoring established companies and making it harder for new entrants to compete.
• This could lead to a concentration of power in the AI industry and stifle innovation.

Open vs. Closed AI Systems for Crucial Decision Making (00:49:32)

• Open AI systems often lag behind closed AI systems due to fewer resources and incentives, but they are crucial for the field to keep up with major players.
• The final step to achieving Artificial General Intelligence (AGI) involves finding the optimal interface between AI systems and humans.
• The current sequential approach to AI technology development is not ideal. Instead, we should consider how humans will use AI systems and design the entire solution end-to-end.
• More attention should be given to developing effective ways for humans to interact with, supervise, and correct AI systems as they become more intelligent.
• Human interaction with AI systems is richer and more effective than writing instructions, especially as systems become more intelligent.

Quick-Fire Round (00:54:18)

• David Luan thinks agents and chatbots will diverge into two distinct products: one for practical tasks and the other for therapeutic or entertainment purposes.
• The biggest misconception about AI in the next decade is that it will fully automate every human capability. Instead, AI will likely serve as a tool to enhance human intelligence.
• In five years, agents could become like a non-invasive brain-computer interface, allowing humans to think and reason at a higher level.
• One reason why this vision of agents might not happen is the resistance from incumbent software companies that bundle software into specific functional areas, preventing agents from bridging different domains.
• Luan believes that the potential market for agents is much larger than that of robotic process automation (RPA) because agents can address a wider range of tasks.