Beyond Bounding Boxes: High-Skill Data for Frontier AI
The Startup Powering The Data Behind AGI
More from Gradient Dissent: Conversations on AI
AI's Role in Drug Discovery: Overcoming Biology's Talent Bottleneck
AI revolutionizes disease cures by overcoming biology's talent bottleneck, not just replicating intelligence. Specialized AI, not general models, unlocks scientific discovery and development velocity.
View Episode Notes →
AI-First Autonomous Driving: End-to-End Learning Beats Infrastructure
Wayve's AI-first approach to autonomous driving learns to generalize across diverse environments, bypassing the costly infrastructure and mapping of traditional methods.
View Episode Notes →
Embracing Complexity and Discomfort for Durable System Advantage
Build resilient systems by prioritizing durable architectural choices over immediate fixes, unlocking long-term competitive advantages that naive approaches cannot replicate.
View Episode Notes →Resources
Books
- "The Art of Computer Programming" by Donald Knuth - Mentioned as an example of a foundational work in computer science, implying a similar depth of understanding is needed for advanced AI data research.
Videos & Documentaries
- None
Research & Studies
- Llama 2 Paper - Referenced for its findings on the effectiveness of RLHF over SFT data in training language models.
Tools & Software
- Crowdflower - Mentioned as a previous company the host started in the data collection business.
Articles & Papers
- None
People Mentioned
- Donald Knuth (Author of "The Art of Computer Programming") - Cited as a figure associated with foundational computer science knowledge.
- Terence Tao - Mentioned as an example of someone whose skill level would not differ in drawing a bounding box, highlighting the low complexity of such tasks.
- Albert Einstein - Mentioned as an example of someone whose skill level would not differ in drawing a bounding box, highlighting the low complexity of such tasks.
- Hemingway - Mentioned as an example of a great writer who did not have a PhD.
- Emily Dickinson - Mentioned as an example of a great poet who did not have a PhD.
Organizations & Institutions
- Twitter - Mentioned as a previous employer of the guest where he encountered data collection issues.
- Google - Mentioned as a company where the guest experienced similar data collection problems.
- Facebook - Mentioned as a company where the guest experienced similar data collection problems.
- Airbnb - Mentioned as an early customer of Surge.
- Meta - Mentioned in the context of potentially making Llama open-source models closed-source.
- OpenAI - Mentioned as a major AI lab with a distinct culture and model.
- Anthropic - Mentioned as a major AI lab with a distinct culture and model.
- Elon Musk's X - Mentioned as a major AI lab with a distinct culture and model.
- Stanford University - Mentioned as an example of where a professor might be working on frontier research.
Courses & Educational Resources
- None
Websites & Online Resources
- LLM Arena - Discussed as a popular leaderboard for LLM models that the guest believes has negatively impacted the industry due to flawed evaluation methods.
Other Resources
- Lidar model - Used as an example of a specific application where understanding underlying principles helps in data collection.
- Cloud Code - Mentioned as an example of a coding collaborator.
- AWS - Mentioned as a service that could go down in a simulated RL environment.
- Gmail - Mentioned as a communication tool within a simulated RL environment.
- Slack - Mentioned as a communication tool within a simulated RL environment.
- Jira tickets - Mentioned as a tool within a simulated RL environment.
- Github PRs - Mentioned as a tool within a simulated RL environment.
- Codebases - Mentioned as a component within a simulated RL environment.
- Sat style math problems - Used as an example of narrow, synthetic problems that models can become good at, but at the expense of broader capabilities.
- RL environments - Discussed as a new method for data collection, involving simulated universes with complex tasks.