Person
1 episodes · 18 quoted claims
Sharma says a study by Nathan Lambert found that once a model reaches 30 billion parameters, the capital cost of pre-training it further stops making economic sense compared with optimizing it.
“once a model hits 30 billion parameters, the capex to actually train a model. And put, you know, billions of tokens into a a kind of pre run kind of doesn't economically make sense.”
For Microsoft's Dragon product for physicians, Sharma says tuning on 600 expert-annotated patient-physician interactions performed far better than synthetic fine-tuning.
“We saw a massive difference from when we used, you know, synthetic fine tuning to when we annotated 600 patient um physician interactions by experts.”
Sharma says Dragon's character acceptance rate rose from between 30 and 60 percent, depending on the run, to about 83 percent after tuning on expert-annotated data.
“I think we were sitting between 30 and 60 character acceptance rate, depending on the run, to something like 83.”
Sharma says GitHub Copilot uses an ensemble of models fine-tuned across about 30 countries' worth of languages for features such as next-edit suggestions and code completions.
“You mentioned Cursor GitHub has very similar features that we're using kind of as an ensemble of models that have been fine tuned across, you know, uh, 30 different countries”
Sharma gives an example of applying AI to an existing process by cutting fraud resolution time from 15 days to ten days.
“Or taking fraud down from 15 days to kind of cure to to ten days.”
Sharma says something like 70,000 enterprise AI tools launched in the year before the episode.
“There's something like 70 zero enterprise tools, like in the AI space launched last year.”
Sharma cites assigning a pull request to GitHub Copilot and creating an agentic sales development rep as examples of embodied agents already in use.
“You can assign a pull request to Copilot, you can, uh, create a software development rep that's agentic that can kind of do some of the lead generation and mining for you.”
Sharma says Microsoft's product teams historically planned in six-month semesters, which they now supplement with named 'seasons'.
“historically, at least in our product teams, had kind of semesters that they planned against. So think of that as every six months,”
Sharma says Microsoft's AI platform teams work in squads that set four-to-six-week goals beneath loose quarterly OKRs.
“teams are operating in squads and they're kind of setting out, you know, four to six week, um, goals that they're trying to go after”
Sharma says about 15,000 customers have built agents on Azure and that the number of agents is in the millions.
“I think the number of agents is actually like millions,”
Sharma says the agent product Manus broke through by using tool-calling loops that let agents handle longer-running tasks.
“Manus's breakthrough in the space was that they could, like, do these tool calling loops and have agents kind of do longer running tasks”
Sharma was employee number seven at Porch Group, one of the first companies she worked for.
“one of the 1st companies that I worked at was a company called Porch Group. I was employee seven.”
Sharma says the most important thing she did at Porch was build a platform matching home-service professionals with service types, zip codes and homeowners, which she links to its first $500 million valuation.
“create a matching platform that matched the 6 million professionals with the one 300 service types”
Sharma argues WhatsApp won through contact reach via phone numbers, reliability and privacy with end-to-end encryption rather than through features.
“WhatsApp didn't win because it had stickers or stories or dark mode.”
Sharma says she read about a London hospital improving pregnancy rates and cutting costs by using AI to match eggs and sperm.
“I was just reading about a hospital in London that's, you know, able to improve pregnancy rates by using AI to match, you know, eggs and sperms.”
Sharma says Stanford is a major customer of Microsoft's AI platform and is working on using AI for tumor reviews.
“Stanford's one of our big customers with the with the platform that I, uh, build, and they're working on using AI for tumor reviews.”
Sharma cites surveys indicating that about half of developers are now fine-tuning models.
“50 of developers, uh, according to surveys, are now fine tuning, uh, and we know fine tuning is good.”
Sharma says The Home Depot uses an inverted-pyramid culture in which store associates sit at the top and the stores act as headquarters.
“they have this inverted pyramid, where instead of having like executives at the top, the associates are at the top and the stores themselves are, uh, headquarters”