Aidan Gomez: What No One Understands About Foundation Models | E1191

Intro (0s)

• There is no market for outdated AI models. (2s)
• Larger and more computationally expensive models tend to perform better. (7s)
• The AI landscape will likely feature a mix of specialized and general-purpose models. (12s)
• There is a risk in becoming overly reliant on a single cloud provider. (21s)

Childhood & Backround (45s)

• He was raised in rural Ontario, Canada, on a 100-acre forested property, which while a quintessential Canadian experience, provided limited access to technology. (1m1s)
• His interest in technology stemmed from his love of gaming and the challenges he faced with slow dial-up internet speeds, leading him to learn coding to improve his online experience. (1m24s)
• He believes gaming fosters resilience by encouraging players to persevere through repetitive tasks to achieve goals and promotes a growth mindset by allowing for second attempts and opportunities to improve. (2m44s)

Is More Compute the Only Path to Better Performance? (4m29s)

• Increasing the size and computational power of a model is a reliable but inefficient way to improve its performance. (5m0s)
• While scaling up model size has led to rapid advancements in model quality, smaller models have demonstrated comparable performance, highlighting the potential of alternative optimization approaches. (5m45s)
• The future of AI models will likely involve a combination of large, general-purpose models for prototyping and smaller, specialized models for specific tasks. (7m56s)

Can Anyone Afford to Stay in the AI Race Besides Tech Giants? (8m7s)

• The cost of maintaining a position in the AI race is exorbitant, potentially requiring a company to be the size of Microsoft, Amazon, Google, or Facebook. (8m29s)
• Most of the major gains in the open-source AI space have come from data improvements, such as better scraping algorithms and the use of synthetic data. (9m8s)
• There is a lack of training data that demonstrates reasoning, as the internet mostly consists of the output of reasoning processes rather than the processes themselves. (11m21s)

Is AI Heading Toward a Race to the Bottom? (13m44s)

• There is a concern that the value of AI models is diminishing due to price dumping and companies offering models for free. (14m1s)
• Selling models alone is expected to be a low-margin business in the short term due to price competition, but there is potential for higher margins in the long term. (15m41s)
• A significant portion of spending in the AI industry is now allocated to chips, with companies engaging with various chip providers like Nvidia, AMD, and Google to meet customer demand for platform diversity. (16m6s)

Will Companies Keep Building Their Own Chips? (16m55s)

• Companies are currently building their own chips due to the high profit margins and limited options in the market. (17m9s)
• The shortage of chips is decreasing, and more options are becoming available, particularly for inference, which already has a diverse range of choices. (17m42s)
• While Nvidia has been dominant in training large models, Google's TPUs have emerged as a viable alternative, and AMD's Tranium is expected to become a competitive option soon. (18m28s)

Is Model Progression Outpacing Compute Advancement? (18m30s)

• Concern is raised that the advancement of models is significantly outpacing the development of data centers and compute capabilities. (18m37s)
• There is a question of whether the newest models will be run on older hardware due to this potential misalignment. (18m48s)
• It is stated that building proprietary data centers is currently not economically viable, but could be considered if it becomes more cost-effective or if access to specific desirable chips is limited. (19m10s)

Early Challenges in Accessing Compute Chips (19m41s)

• Before the widespread popularity of AI, access to a significant number of computer chips was a challenge. (19m43s)
• The speaker did anticipate the eventual surge in AI's popularity, but the speed and timing of its rise were unexpected. (20m0s)
• The release of ChatGPT marked a turning point in AI adoption, making the technology directly accessible to users and eliminating the need for complex explanations. (21m14s)

Are We Underestimating the Short-Term Impact of AI Advancements? (23m48s)

• It is becoming increasingly difficult and expensive to improve AI models because they are becoming so sophisticated that it requires more specialized knowledge to find and correct errors. (24m14s)
• The cost of computing power is decreasing rapidly, which is allowing for the development of much larger models. (26m53s)
• A "flop" is a unit of compute, specifically a floating point operation, which represents one clock cycle of a computer's processor. (26m25s)

Is It Too Late for Startups to Enter the AI Model Space? (27m6s)

• The cost of building large language models is decreasing by a factor of 10 or 100 each year due to improvements in data and compute. (27m21s)
• Lower costs make it easier to build the previous year's models, but there is no market demand for outdated technology. (27m30s)
• Rapid technological advancements quickly make previous generations of large language models obsolete. (27m47s)

AI Development: The Exponential Rise in Costs (27m55s)

• The cost of developing AI models is increasing exponentially, with each new generation costing significantly more than the last. (28m7s)
• While the improvements between generations may not be noticeable to the average user, they are significant in specialized fields like medicine, mathematics, and physics. (29m19s)
• Continued investment in AI development is crucial for progress, even if the benefits are not immediately apparent to everyone. (30m37s)

Will Cloud Giants Continue Acquiring Smaller AI Model Providers? (30m40s)

• The increasing prevalence of cloud computing suggests that major cloud providers will likely acquire smaller AI model providers in the coming years. (31m7s)
• Raising capital from cloud providers presents a different financial model compared to traditional venture capital, potentially creating challenges for companies seeking to maintain their independence. (32m9s)
• While acknowledging the pressure to align with revenue multiples, there is confidence in the strength of the market for AI models, despite current pricing pressures and the availability of free models. (33m54s)

Is OpenAI Prioritizing AGI Over Practical Products? (35m10s)

• OpenAI is currently focused on building consumer products, particularly ChatGPT, which has become highly successful. (35m42s)
• There is a growing trend of businesses wanting to implement AI solutions quickly, driven by a fear of falling behind competitors. (41m55s)
• OpenAI's most successful bet was its commitment to the scaling hypothesis, which posits that increasing the size and training data of AI models will lead to continuous improvements. (48m9s)

What's the Biggest Overlooked Factor in AI's Future? (48m29s)

• The most significant overlooked aspect in AI's future is the development of reasoners, planners, and models capable of attempting tasks, learning from failures, and executing long-term tasks. (48m57s)
• These capabilities are currently absent in existing AI technology, but researchers have shifted their focus to incorporating them, and they are expected to be ready for production soon. (49m20s)
• The integration of these capabilities will significantly enhance AI's capabilities, leading to the emergence of new products and transforming existing sectors like social media. (49m39s)

Concerns About a Future Where AI Replaces Human Interaction (50m9s)

• There is concern that children will grow up speaking to AI more than humans. (50m20s)
• Some jobs, such as customer service, may be replaced by AI, but overall, AI will lead to job growth, not displacement. (52m57s)
• AI can be used to handle difficult tasks, such as customer support calls where customers are angry, allowing humans to focus on more positive interactions. (54m1s)

What Will AI Do in Three Years That It Doesn't Do Today? (54m20s)

• Significant advancements in robotics, particularly in cost reduction and the development of more robust models, are anticipated, potentially leading to breakthroughs in the field. (54m30s)
• Traditional robotics software was inflexible, requiring specific programming for each task and environment. (54m51s)
• Foundation models and language models have enabled the creation of improved robotic planners that are more adaptable and capable of more natural reasoning, suggesting a future with more versatile and affordable humanoid robots. (55m13s)

Quick-Fire Round (55m48s)

• The most significant change in perspective over the past year has been a greater appreciation for the importance of data quality in building AI models. (55m57s)
• Approximately $1 billion has been raised in funding. (56m40s)
• A key goal for the future is to leverage AI technology to drive productivity and enhance human effectiveness. (1h1m56s)