Person
1 episodes · 19 quoted claims
Weights & Biases founders reached Toyota Research Institute through relationships built at CrowdFlower/Figure Eight.
“through some relationships that we had built at at CrowdFlower Figure eight to be able to, like, talk to a team like Toyota Research Institute.”
Chris Van Pelt studied art and computer science in college and links his interest in tools to his art background.
“I studied art and computer science in in college, and like, as an artist, tools are like, fundamental, right?”
A T group is a session where about ten startup leaders gather over two days in a facilitated discussion.
“ten random startup leaders are gonna go on a little field trip, uh, Thursday and Friday this week.”
Chris Van Pelt says the Ruby programming language is what got him into programming.
“haven't done, I haven't done like Ruby in a long time. But I love, uh, Japan and Japanese culture, and Ruby's what got me into this.”
Weights & Biases used MySQL because Google Cloud offered managed MySQL but not managed Postgres at the time.
“We used MySQL as the like data store for for everything because at the time we were on Google Cloud. And all they had was a managed MySQL offering, they didn't have managed Postgres.”
Chris Van Pelt says Semantic UI is no longer updated and the Weights & Biases team is replacing it with a new UI framework.
“Semantic UI isn't really being updated anymore. Now. It's like all over the code base, so the team's like creating a new UI framework.”
Chris Van Pelt said Weights & Biases had grown to about 200 people working on several parts of the MLOps pipeline.
“we've grown now. We're a company of 200 people. We have lots of engineers working on, um, lots of different aspects of the the ML Ops pipeline”
OpenAI turned to Weights & Biases when it had a problem because the company had kept in contact with its researchers.
“ultimately, because we had been talking to them and they were aware of us. When they had a problem, they thought, Hey, maybe Weights and Biases could help us with this problem”
Weights & Biases sets top two to three goals every two weeks and shares them with the company, rolling them into quarterly goals.
“So, um, you know, every two weeks we sit down and say, Okay, what are what are like top two to three goals for these two weeks? And those goals are then, you know, shared with the the company.”
At his previous company, Chris Van Pelt's team worked with the US Department of Defense and held a meeting at the Pentagon.
“At our previous company, we, uh, were working with the Department of Defense, so we had a meeting in the Pentagon.”
Chris Van Pelt says ChatGPT's release changed the AI landscape because of its interface and RLHF tuning, not because the underlying models were new.
“It's like the interface and the RLHF tuning to to make it like a real joy to to interact with.”
Chris Van Pelt said about 70 percent of a recent Y Combinator class was working on generative AI, often treating models as just another API.
“a lot of entrepreneurs. Like 70 of the YC class is doing some gen AI thing, uh, they're coming and using these things, seeing it just as another API they can hit and use.”
Chris Van Pelt advises that if an LLM is used to evaluate another LLM's output, the evaluating model should be a different one.
“you can use an LLM to measure an LLM. But I would, I would think you'd wanna use a different LLM than the one like outputting stuff.”
Chris Van Pelt cites Llama.cpp and Whisper.cpp as projects enabling language models to run locally, including in a browser via WebGL.
“With projects like, uh, Llama.cpp and Whisper.cpp, um, you can also, you can, like, run it in the browser using, uh, WebGL.”
Chris Van Pelt suggests enterprises can obtain a private GPT instance through an enterprise contract with Microsoft.
“Or you just get that enterprise contract with with Microsoft to have, like, a private, um uh, instance of of GPT.”
Chris Van Pelt said ChatGPT turned a designer's English description of a loading icon animation into CSS3 transforms.
“I could just ask ChatGpt and it perfectly translated that English into CSS3 transforms.”
Chris Van Pelt says an LLM given access to a database schema can generate and run SQL to answer business questions asked in English.
“if you've set things up right and and given the LLM access to the schema and where all the data is, it can generate and then run, you know, SQL and actually give you, give you charts or or information about things.”
Chris Van Pelt said GPT-2 seemed cool but toy-like to him, and he did not initially see its practical uses.
“I thought it was cool, but I also thought, like, all right, well, what can you do with this?”
Chris Van Pelt compares language as an interface to the transformer architecture, which he says made it possible to parallelize bigger compute.
“much like the transformer architecture became a really powerful way to parallelize, uh, bigger and bigger compute.”