Jack Morris on Finding the Next Big AI Breakthrough

Key Takeaways Copied to clipboard!

• Current AI model progress often feels incremental, with awe-inspiring breakthroughs being less frequent than in the past, though significant advancements like reinforcement learning (RL) still unlock future potential.  Copied to clipboard!
• Measuring AI progress is difficult, relying on both standardized benchmarks (like SweetBench or Math Olympiads) and subjective human evaluations (like ELO scores), but these metrics often fail to capture real-world utility or nuanced model characteristics.  Copied to clipboard!
• The career decisions of top AI researchers are a balance between pursuing frontier science (often in closed-source labs with superior infrastructure) and financial incentives, with the competition between US and Chinese labs being heavily influenced by access to proprietary data and compute resources.  Copied to clipboard!

Segments Expand All Collapse All

Initial AI Model Impressions

Copied to clipboard!

(00:01:37)

• Key Takeaway: GPT-5 does not feel like an amazing step function improvement over O3 models for daily use, suggesting progress is becoming incremental.
• Summary: The hosts note that GPT-5 does not immediately strike them as obviously superior to the O3 models for their uses. They feel the awe-inspiring breakthroughs in AI models might be behind them currently. Progress on models currently feels incremental despite significant resource expenditure.

Defining AI Researcher Role

Copied to clipboard!

(00:03:51)

• Key Takeaway: AI researchers today navigate a split between public academic research and private, non-publishing industry research labs.
• Summary: AI research is divided between public work done at academic institutions and private work conducted within major AI labs like OpenAI or Meta. Humans are still necessary to drive improvements, as current AI cannot yet fully self-improve to the point of exponential takeoff. Day-to-day research varies, covering hardware efficiency, data selection, and training algorithms.

Measuring AI Model Progress

Copied to clipboard!

(00:07:49)

• Key Takeaway: Formal AI model evaluation relies on testing against specific datasets (like SweetBench) or human preference ranking using ELO scores.
• Summary: Model evaluation primarily uses testing on datasets with known solutions, such as coding problems or math competitions like the International Math Olympiad. A second method involves ELO scores, where humans rank the outputs of two models against each other to create a comparative ladder. Benchmarks can be misleading, as models can score highly on specific tests while failing basic real-world logic questions.

Supervised vs. Reinforcement Learning

Copied to clipboard!

(00:13:12)

• Key Takeaway: The jump in capability seen in models like O3 was largely due to the successful implementation of reinforcement learning (RL) alongside traditional supervised learning.
• Summary: Supervised learning involves the computer copying existing text from the internet, which limits performance to emulation quality. Reinforcement learning involves the computer performing actions and receiving a numerical reward for success, which significantly improved models’ performance in areas like math. The reward mechanism is a numerical signal indicating a higher score or better outcome.

Pace of Breakthroughs and Agents

Copied to clipboard!

(00:14:50)

• Key Takeaway: AI progress is unpredictable, with expected breakthroughs like functional agents failing to materialize while unexpected areas like competitive math see rapid gains.
• Summary: The improvement cycle in AI feels cyclical, with progress appearing suddenly in unexpected domains. The anticipated ‘year of agents’ did not fully materialize because the problem of collecting comprehensive training data for real-world scenarios proved harder than anticipated. However, progress continues behind the scenes, supported by an ecosystem building specialized environments for RL training.

Data as a Differentiating Factor

Copied to clipboard!

(00:22:45)

• Key Takeaway: The next major advancements in AI models will likely stem from proprietary, private data sources that competitors cannot access, such as user conversations or specialized content.
• Summary: Publicly available datasets are largely exhausted for training current frontier models. Labs maintain advantages through private data access, such as Google’s YouTube data or Anthropic’s years-long effort scanning thousands of old books. This proprietary data accumulation creates a significant, hard-to-replicate head start for established labs.

Talent Motivation and Career Choices

Copied to clipboard!

(00:41:11)

• Key Takeaway: For top AI researchers, the marginal difference between extremely high salaries is less motivating than the desire to participate in the most significant scientific frontiers, often leading to prestige-driven career choices.
• Summary: Once salaries reach a very high baseline (e.g., $10 million vs. $20 million), the financial difference becomes negligible compared to the excitement of working on the next major scientific leap, such as achieving AGI. Titles and perceived importance within the organization also play a significant role in attracting and retaining top talent. Researchers must reckon with choosing between high-paying roles focused on commercial optimization versus mission-driven work on the scientific cutting edge.