Character AI's 2021 seed deck surfaces in lawsuit, revealing Noam Shazeer's early vision and GPT-3 origins
Key Points
- Character AI's December 2021 seed deck credits Noam Shazeer's research as foundational to OpenAI's GPT-3, claiming OpenAI employees attended his 2018 talk and implemented his approach.
- Shazeer's pitch prioritizes transformer architecture and distributed computing over individual models, arguing the real value is in the ability to keep building models rather than any single one.
- Character AI planned $20–30 million in first-year spending to buy GPUs, train a GPT-3-sized model in three weeks, and launch products spanning companionship, writing assistance, and customer support.
Summary
Character AI's 2021 Seed Deck Reveals Noam Shazeer's Vision for LLM-Driven Conversational Products
A Character AI seed deck from December 2021—surfaced in a lawsuit—documents Noam Shazeer's early thesis on what would become the company's core product, along with his already-developed conviction that GPT-3's architecture owed substantially to his own research contributions.
The Shazeer Foundation
The deck positions Shazeer's trajectory as the linchpin: spell checker and "Did You Mean" at Google, AdSense targeting, large-scale ML infrastructure, and by 2016, a public prediction that large neural language models represented the future. His stated goal is to invent critical innovations and make them available to the world, grounded in specific technical bets on transformer architecture, distributed computing, sparsity optimization, and fast decoding for long sequences.
A notable framing appears in the deck: Shazeer argues the true value isn't in individual models—"the eggs"—but in the architecture itself, "the goose." The power to keep making models is what matters.
The GPT-3 Attribution
The deck explicitly credits Shazeer's work as foundational to OpenAI's GPT-3, citing multiple OpenAI employees who reportedly attended Shazeer's 2018 talk, took notes, and implemented the approach. The framing suggests OpenAI was "directionless" until that talk, and continued to monitor and implement Shazeer's papers afterward. The deck also references Kevin Lacker's 2021 assessment that GPT-3 couldn't quite pass a coding phone screen but was "getting closer"—GitHub Copilot, built on GPT-3, followed.
The Product Vision
Character AI's intended use cases span companionship, language practice, social situation rehearsal, assistive writing, life coaching, customer support, and personal assistance. The core mode: quality research engineers, intelligent products, high usage, and large compute cloud. The deck notes $1 million as the training cost for GPT-3, with a smarter model costing more.
Shazeer's operating plan is concrete: buy GPUs, evaluate make-versus-cloud, train a GPT-3-sized model in three weeks, with total first-year spend of $20–30 million.
Market Context
The deck was written one month after Shazeer left Google, with reported strong inbound interest from top engineers, angels, and early-stage VCs. By the time of this surface, Shazeer has since left Character AI and moved to OpenAI—a two-year run at Character, followed by a Google acquisition, then departure.
Every deal, every interview. 5 minutes.
TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.