Key Takeaways
- • Sequoia Capital partners Sonya Huang and Pat Grady argue that large language models such as OpenAI's GPT-3 are enabling machines to write, code, draw and create at credible and sometimes superhuman levels, a shift they label "Generative AI" as distinct from analytical AI.
- • Sequoia contends that knowledge and creative work spanning billions of workers could become at least 10% more efficient or creative, implying potential economic value in the trillions of dollars.
- • The firm identifies four waves of development, with the current phase (2022 onward) marked by cheaper compute, diffusion models, open beta and open-source model access, and the emergence of killer applications.
- • Sequoia highlights text as the most advanced modality, with code generation (GitHub Copilot generating nearly 40% of code in projects where it is installed), images, speech, and video and 3D models as fast-developing areas.
- • Applications flagged include copywriting, vertical-specific writing assistants, code generation, art generation, gaming, media and advertising, design, and social media, with most current apps delivered as plugins inside existing software ecosystems.
Summary
Sequoia Capital's September 19, 2022 essay "Generative AI: A Creative New World," written by partners Sonya Huang and Pat Grady, argues that a new class of large language models is enabling machines to produce original work rather than only analyzing existing data. The authors distinguish this "Generative AI" from "Analytical AI," the traditional machine learning used for tasks such as fraud detection, spam filtering, delivery time forecasting and content recommendation.
The firm contends that every industry requiring original output, including social media, gaming, advertising, architecture, coding, graphic design, product design, law, marketing and sales, is open to reinvention. Sequoia expects generative AI to reduce the marginal cost of creation and knowledge work toward zero. The authors estimate that the fields involved comprise billions of workers and that generative AI could make them at least 10% more efficient or creative, which they say has the potential to generate trillions of dollars of economic value.
Sequoia frames the development of the technology in four waves. Pre-2015 small models were strong at analytical tasks but not expressive enough for general-purpose generation. From 2015, the transformer architecture introduced in a Google Research paper, "Attention is All You Need," enabled scaling; the authors note that compute used to train these models rose six orders of magnitude between 2015 and 2020, with results surpassing human benchmarks in handwriting, speech and image recognition, reading comprehension and language understanding. OpenAI's GPT-3 is singled out as a major leap over GPT-2. The current wave, starting in 2022, features cheaper compute, techniques such as diffusion models, and access expanding from closed beta to open beta or open source. The authors say the application layer is now ripe for "killer apps," comparing the moment to the mobile inflection point.
On modalities, the essay describes text as the most advanced domain, though outputs are typically used for iteration or first drafts. Code generation is expected to have a near-term productivity impact, with GitHub Copilot cited as generating nearly 40% of code in projects where it is installed. Images have gone viral, speech synthesis is improving for utilitarian uses, and video and 3D models are advancing quickly. Sequoia lists application categories it finds attractive, including copywriting, vertical writing tools for areas such as legal contracts and screenwriting, art generation, gaming, media and advertising, and design.
On application architecture, Sequoia says generative apps sit as a UI layer and "little brain" atop large general-purpose "big brain" models, and can fine-tune those models with user data to improve performance and reduce cost. Today most such apps exist as plugins within existing software, while standalone web apps include Jasper and Copy.ai for copywriting, Runway for video editing, and Mem for note-taking.
The firm contends that every industry requiring original output, including social media, gaming, advertising, architecture, coding, graphic design, product design, law, marketing and sales, is open to reinvention. Sequoia expects generative AI to reduce the marginal cost of creation and knowledge work toward zero. The authors estimate that the fields involved comprise billions of workers and that generative AI could make them at least 10% more efficient or creative, which they say has the potential to generate trillions of dollars of economic value.
Sequoia frames the development of the technology in four waves. Pre-2015 small models were strong at analytical tasks but not expressive enough for general-purpose generation. From 2015, the transformer architecture introduced in a Google Research paper, "Attention is All You Need," enabled scaling; the authors note that compute used to train these models rose six orders of magnitude between 2015 and 2020, with results surpassing human benchmarks in handwriting, speech and image recognition, reading comprehension and language understanding. OpenAI's GPT-3 is singled out as a major leap over GPT-2. The current wave, starting in 2022, features cheaper compute, techniques such as diffusion models, and access expanding from closed beta to open beta or open source. The authors say the application layer is now ripe for "killer apps," comparing the moment to the mobile inflection point.
On modalities, the essay describes text as the most advanced domain, though outputs are typically used for iteration or first drafts. Code generation is expected to have a near-term productivity impact, with GitHub Copilot cited as generating nearly 40% of code in projects where it is installed. Images have gone viral, speech synthesis is improving for utilitarian uses, and video and 3D models are advancing quickly. Sequoia lists application categories it finds attractive, including copywriting, vertical writing tools for areas such as legal contracts and screenwriting, art generation, gaming, media and advertising, and design.
On application architecture, Sequoia says generative apps sit as a UI layer and "little brain" atop large general-purpose "big brain" models, and can fine-tune those models with user data to improve performance and reduce cost. Today most such apps exist as plugins within existing software, while standalone web apps include Jasper and Copy.ai for copywriting, Runway for video editing, and Mem for note-taking.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.