Interview, Fireside Chat, Conference Presentation
Josh Woodward: Google Labs is Rapidly Building AI Products from 0-to-1
- Prompt engineering is expected to shift from paragraph-level text to multimodal context delivery via picture, voice, or video for most users, while specialized engineers may continue writing long prompts; this transition is described as occurring "pretty soon."
- Google Labs prioritizes a rapid development cycle, aiming to move projects from concept to user hands within a "50 to 100 days" timeframe to maintain a fast-moving culture during the AI platform shift.
- Success metrics for Google Labs teams include reaching "10,000 weekly active users" as a significant milestone, a threshold distinct from the scale required for larger Google groups.
- Generative video models like Veo are transitioning from a phase where results are "almost possible" to "possible and let's talk about it now," with a predicted historical cost reduction trend of "97 times" over the last year for text models serving as a comparable curve for video.
- Efficiency in video generation is forecast to improve such that the "cherry pick rate" drops to "one time," and physics simulation is considered "close" to being solved, leaving the application layer and workflow rethinking as "wide open" opportunities.
- A new "AI camera" interface is anticipated to allow for the generation of infinite scenes and global modifications, such as changing a character's clothing color across a whole film.
- Video generation business models are expected to innovate beyond simple subscriptions toward "pay per output" or "auction type models" within a timeframe of "quarters," reflecting a shift toward steerable entertainment and a "curator" class of content creators.
- The pace of AI innovation is predicted to continue accelerating without slowing, opening "30 doors" into "adjacent possibles," with the distinction between movies, games, and 3D world-building expected to blur.
- Tools like Veo2 have accelerated the "3D angle" by enabling the generation of entire product catalogs from "two or three" photos, contrasting with previous requirements of six.
- Computer agents like Mariner are predicted to evolve from single-session tasks to handling "an infinite number" of agents running "in the background," with initial killer use cases in the enterprise sector targeting high-toil activities in call centers and sales teams.
- Mariner is forecast to require "four, five, six revs" within a single year, following an iteration cycle of "every month or two," while addressing unsolved challenges regarding precise screen navigation and the appropriate level of human involvement.
- By the "2028" capsule scenario, knowledge is envisioned as "infinitely remixable," allowing any input to be transformed into any output, alongside the strategic leveraging of "long context" or "infinite context" capabilities.
- For 2025, Google Labs is focusing hiring and bets on coding, video, and agent capabilities involving thinking and reasoning models, with 25% of current code written by AI and major leaps expected in self-correcting and self-healing models this year.
- Pre-training is viewed as potentially "hitting a wall," with the industry focus shifting to inference time compute and values regarding whether tools eliminate people or amplify human creativity.