At Rise of AI Summit 2020 in Berlin, Daniel Jeffries, Chief Technical Evangelist at Pachyderm, says he wants to focus only on the positive changes AI will bring in the next decade, after previously covering 5, 50 and 500-year horizons. He argues intelligence has driven every past leap in human history, from the agricultural revolution to the information age, and that AI needs a shared, Linux-like infrastructure stack before it can spread beyond big tech, an effort he connects to his own work with a 15-company Artificial Intelligence Infrastructure Alliance. He predicts AI will transform the arts, citing OpenAI's Jukebox recreating voices like Frank Sinatra, and will cut blockbuster film costs by blending digital and real actors. In healthcare, he traces rapid gains in disease detection, from Google's skin cancer algorithm reaching 76 percent accuracy in 2017 to 96 percent by 2019, and a Google research team's 99 percent accurate breast cancer detector called LYNA. He also describes an Oak Ridge supercomputer analysis that reframed COVID-19 as a vascular disease via the "bradykinin hypothesis" and identified nine existing drugs with possible efficacy.
Key takeaways
- Jeffries argues AI needs a shared, Linux-like infrastructure stack before capability spreads from big tech to smaller companies and individuals.
- He credits Google's skin cancer detection algorithm with improving from 76 percent accuracy in 2017 to 96 percent by 2019.
- He cites Google's LYNA system reaching 99 percent accuracy detecting breast cancer, outperforming radiologists on the hardest slides.
- He credits an Oak Ridge supercomputer analysis with producing the "bradykinin hypothesis," reframing COVID-19 as a vascular disease and flagging nine existing drugs.
- He says he worries less about AI superintelligence than about narrow AI controlled by selfish humans, citing loan and hiring algorithms as risk areas.
Detailed notes
Intelligence as the engine of every historical leap
Jeffries says he is deliberately focusing only on the positive next decade of AI, after discussing darker and more distant scenarios in past talks. He frames human history as a series of intelligence explosions: the agricultural revolution 12,000 years ago, a second, British agricultural revolution that ended fallow land and triggered specialization, the scientific revolution 400 years ago, the industrial revolution 200 years ago and the information age 50 years ago, arguing AI is now driving the next such leap because intelligence underlies every other advance.
A Linux for AI
He argues AI is still largely confined to big tech and R&D departments because it lacks a shared infrastructure stack comparable to Linux, and describes his own involvement in an Artificial Intelligence Infrastructure Alliance of 15 companies working on this. He predicts that once a canonical stack exists, capability will cascade from big tech to Fortune 500 companies and then to individuals, citing WhatsApp reaching 400 million people with only 35 engineers as an example of what shared infrastructure enables, and predicts roughly 95 percent of today's manual data-science work, like data labeling, will be automated.
AI in the arts
Jeffries points to OpenAI's Jukebox project recreating voices such as Frank Sinatra and Ella Fitzgerald, and research separating drums and vocals from tracks to recreate individual parts. He expects video game and movie technology to converge as real-time graphics approach photorealism, letting AI-driven digital actors, layered with real performers, cut blockbuster costs from around 400 million dollars to closer to 20 million. Using his discovery of Ella Fitzgerald's "lost Berlin tapes" as an anecdote, he imagines combining recreated voices, such as Fitzgerald's joy with Amy Winehouse's sadness, using emotional meta-tags layered onto musical notes.
Healthcare: disease detection, vaccines and prioritized care
He traces fast progress in disease detection: Google's skin cancer algorithm reached 76 percent accuracy in 2017, 86 percent in 2018 and 96 percent by 2019, and a Google research team's LYNA lymph node assistant reached 99 percent accuracy detecting breast cancer, outperforming radiologists on the hardest slides. He cites the Center for New Data tracking COVID-19 super-spreader events via GPS and machine learning, and points to Pfizer and BioNTech's vaccine, which reached 90 percent efficacy in trials with AI central to the work, alongside 34 other drugs then in human trials and 145 more in labs, calling this pace far faster than the mumps vaccine's previous four-year record from outbreak to vaccine. He credits an Oak Ridge National Lab supercomputer analysis of 27,000 GPUs for producing the "bradykinin hypothesis," reframing COVID-19 as a vascular rather than purely respiratory disease and flagging nine existing drugs with possible efficacy, and describes Boston University research predicting heart disease and diabetes hospitalization risk with 82-83 percent accuracy a year in advance.
Q&A highlights
Asked about resistance to AI-driven change, Jeffries says progress and resistance are both necessary forces in society, comparing music-industry pushback to AI in the arts to the earlier fight against Napster, and warns against regulating AI "before it's out of the womb." Asked whether he worries about being replaced by a virtual version of himself, he says it is technically plausible via brain-mapping connectome projects but says he would welcome talking to a "virtual Dan Jeffries." He adds he worries less about AI superintelligence than about narrow AI controlled by selfish humans, citing loan and hiring algorithms as areas needing human-in-the-loop oversight.
Quotes
The history of mankind is the history of intelligence, all of our innovations come from our mind.
Daniel Jeffries
I worry more about narrow intelligence controlled by selfish, egotistical, violent human beings than I worry about AI superintelligence taking over the world.
Daniel Jeffries
Questions answered in this talk
What accuracy has Google's AI reached for skin cancer detection?
According to Daniel Jeffries at Rise of AI Summit 2020, Google's algorithm detected skin cancer at 76 percent accuracy in 2017, improving to 86 percent in 2018 and 96 percent by 2019.
How did AI contribute to understanding COVID-19?
Jeffries describes an Oak Ridge National Lab supercomputer, using 27,000 GPUs, that analyzed billions of combinations and produced the "bradykinin hypothesis," reframing COVID-19 as a vascular rather than purely respiratory disease and identifying nine existing drugs with possible efficacy.
Why does Daniel Jeffries say AI needs a "Linux for AI"?
Jeffries argues at Rise of AI Summit 2020 that AI is still mostly confined to big tech because it lacks a shared infrastructure stack; once one exists, comparable to how Linux spread software development, capability will cascade to Fortune 500 companies and then to individuals.