Fireside Chat, Interview, Panel
The Ultimate AI Roundtable: What Happens Now in AI, Why Google are Vulnerable | E1085
20VC with Harry StebbingsDes Traynor, Yann LeCun, Emad Mostaque, Jeff Seibert, Tomasz Tunguz, Douwe Kiela, Cris Valenzuela, Richard Socher, Harry Stebbings, Myles Grimshaw, Christian Lang
- A market consolidation is expected within three to five years, leaving only five or six foundation model companies (including NVIDIA, Google, Microsoft, OpenAI, Apple, and Emad's organization) while testing for other providers like Mistral, Grok, Anthropic, and Cohere increases.
- The foundational model layer faces commoditization where technology once deemed hard and expensive will become widely accessible, with no single model dominating and existing models expected to be obsolete within a year due to order-of-magnitude improvements in parameters and quality.
- Training efficiency gains will enable models to run on laptops, desktops with one or two GPUs, and eventually phones, though experts disagree on whether smaller models can handle all tasks or if size remains critical for specific objectives.
- Data quality will become the primary differentiator for fine-tuning at the next tier, distinct from the correlation between size and data at the base LLM layer, while open-source projects are predicted to succeed in basic infrastructure by leveraging global intelligence.
- Open source alternatives are projected to reach GPT-4 equivalence before the end of the year and will continue improving on-device, with university researchers relying on fully open-source foundational models due to limitations of closed APIs.
- Infrastructure market capitalization is concentrated in three businesses, creating razor-thin margins and direct price competition similar to AWS, whereas the application layer contains 100 businesses with significantly higher odds of success for investors.
- Business models will shift from selling software seats to selling work outcomes and consumption-based pricing, moving from co-pilot strategies to a control center paradigm where AI SLAs focus on efficiency targets rather than uptime.
- Within five to ten years, a labor shortage will necessitate a complete change in work experience as LLMs take over a significant volume of work, with technology creating as many jobs as it eliminates while increasing productivity.
- Apple is forecast to make massive strides in the next three to five years, potentially winning the interaction model across desktop and iOS by making Siri conversational, while running sufficiently large LLMs on devices without cloud dependency.
- Google faces an existential threat if AI replaces search, requiring the company to potentially destabilize itself by killing its "golden goose" through strategies like sponsored injections or giving away free Android phones to control the intent layer.
- Amazon is moving faster in engineering but has not fully transitioned research to engineering for all initiatives, with speculation that it may acquire Anthropic to integrate it into its EC2 cluster.
- Experts predict that AI will drive a renaissance for humanity by amplifying intelligence and creativity across text, art, music, and video, while arguing that regulating research or fearing uncontrolled catastrophe due to model size is unfounded.
- Users will increasingly demand open, vetted base infrastructure for intelligent assistants with human-level intelligence, and regulation will focus on critical decision-making products rather than slowing down research.
- Future intelligent assistants will possess more accumulated knowledge than most humans, and the opportunity for disruption lies in being orthogonal to incumbent co-pilot strategies rather than building similar tools for existing workflows.