Fireside Chat, Interview
Universally Accessible Intelligence with Character.ai's Noam Shazeer
- The company plans to operate as a product-first entity focused on scalable, generalizable work to achieve AGI status.
- A global pause of approximately four months is anticipated while waiting for sufficient H-100 hardware to become available for training the next model.
- A new model is targeted for launch by the end of the year, expected to be tens of IQ points smarter than current iterations.
- NVIDIA is projected to deploy an additional 1.5 million H100 chips next year, raising the total inventory to 2 million units.
- Once these new chips are deployed, computational capacity is projected to reach roughly 0.25 trillion operations per second per person.
- Future models are expected to offer multimodal capabilities, including voice and visual recognition, to enhance accessibility.
- Systems are planned to interact with multiple individuals simultaneously to create social experiences where AI recognizes users by name.
- Massive information storage via existing HBM bandwidth is expected to enable systems to know a billion distinct facts about individuals.
- Operations costs are currently measured at approximately 10^-18 dollars, with a projection that using the largest trained models will eventually cost less than the value of most human time.
- Curing cancer is anticipated to become accessible within a few years, though this timeline is acknowledged as unpredictable.
- Global scientific innovation is expected to democratize advanced capabilities, allowing replication in academic labs or garages within a few years.
- The industry is characterized as entering the "dawn of universally-accessible intelligence," likened to the Wright Brothers' first airplane moment.
- The trajectory of scaling laws is expected to continue without a known stopping point as long as experimentation persists.
- "Theory of mind" is predicted to emerge as a natural property resulting from model scale increases.
- A single, massively scalable model is preferred over specialized systems to avoid the risks of non-generalizable rules.
- The speaker predicts that the technology cannot be blocked due to its ability to build upon itself as it becomes smarter.
- Use cases are expected to expand in areas including technology advancement, science, and general human assistance.
- The company intends to recruit employees motivated by launching products rather than publishing papers.
- Future advancements will make it increasingly difficult to distinguish between real humans and their AI versions.