Thank you! By that we mean, all processing, i.e. data, forward run, evaluator run is done locally in, let's say, Jupyter Notebook. While it's running locally, it sends all the run data / stats to the server for storage and progress tracking.
We love it because we tried putting things into the UI, but found it to be much more limiting rather that letting users design evals and run them however they want.
It's an old trick in generative models, I've been using it since 2015: https://www.gwern.net/RNN-metadata When you have categorical or other metadata, instead of trying to find some way to hardwire it into the NN by having a special one-hot vector or something, you simply inline it into the dataset itself, as a text prefix, and then let the model figure it out. If it's at all good, like a char-RNN, it'll learn what the metadata is and how to use it. So you get a very easy generic approach to encoding any metadata, which lets you extend it indefinitely without retraining from scratch (reusing models not trained with it in the first place, like OA's GPT-2-1.5b), while still controlling generation. Particularly with GPT-2, you see this used for (among others) Grover and CTRL, in addition to my own poetry/music/SubSim models.
Nice hustle! How did you decide to start in CS research? Do you attend a math & sciences HS, or have family members in academics? Or perhaps just from reading the internet?
OpenAI mostly. Started following along what they're doing and eventually spiraled into going to one of their meetups and then attending ICML this summer.
I'm not particularly a genius, but deeply passionate and curious on what I do. As a result, doing research or an internship was more fun than most other things I'd be doing over the summer/fall.
Didn't grow up in the best of situations, but learned how lucky I was to live in the Bay, giving me access to meet people IRL and grow from there.
I’m rather confused why doing this as a consumer product in Assistant makes more sense than an enterprise product. There is so much room for improvement in call centers, and would be a huge market opportunity.
This is hardly even a time saver. What would be helpful is if this is also able to stay on hold for you — probably an easy task.
> Fox didn't rule out selling this kind of tech to call centers, but that's another potential down-the-road addition. "There are companies that do that well," he said. "There are very big and established companies that provide software for call centers. It's not the core of our business. We'll see as we go whether we think if we can help there. But it's not a business that we're in today, and it's a pretty well-served business."
Do you mean a call center bot on the answer side? I would think that's massively harder to do in a way that's noticeably better then the phone tree systems that already exist. Everything simple is already covered by simpler automated systems, and everything that isn't simple is impossibly hard for this level of tech.
Now if Google could make a bot that can call the service line for a company, jump through whatever hoops are needed to get a person on the phone, then ring the call through to me when they got one, that I would pay money for.
Enterprises are more demanding and have a lower tolerance for error. I have no doubt Duplex will be rolled out to enterprise once the technology is ready and has been thoroughly tested on consumers.
After your startup and Google experience, you were hired as an SWE1? Also, curious if the experience of the startup or subsequent integration was perhaps more valuable than the opportunity cost.
I never went to Google, I left the startup a year before that happened. That startup was pretty much my first job out of college and I was there for about two years.
In that specific case, I believe the startup was worth the opportunity cost. I was coming from San Antonio, not exactly a tech hotbed, and from a small liberal arts school nobody's heard of. I'd interviewed at a bunch of companies out the Bay and was rejected by all of them. A friend was an investor in TechStars' seed fund for their San Antonio program and he put me in touch with the founders of one of their companies. Those guys took a chance on me and hired me on a contract to solve some scaling issues while they were still in the program.
Once they graduated and raised a series A, they brought me on full-time. I worked remote for another six months after they moved the company up to Boulder, eventually following myself. They let me go after about two years but during that time I gained a ton of StackOverflow reputation for Scala and Akka, which led me to one of Twitter's open source advocates who made the intro for Twitter's Boulder office. That was back in 2015. About a year later the company sort-of sold to Google, who then fired almost everyone and re-sold the IP (or something, it was weird and I wasn't there, but my investor friend gave me some of the details).
So, in my specific case, the startup was worth the opportunity cost because my opportunity cost wasn't that high. I didn't have a six-figure SV job as a backup, but I was able to leverage the risk I took into that sort of job.
> Google is so strong with ML today, and not stopping, I don't know what will allow smaller competitors to stay differentiated. I just had a call with a G recruiter, and pretty much 100% of the roles they have in all their European offices are ML-related roles, on most products. And they hire without ML experience, which I think is a sign of how much they are investing in ML.
Weird. I'm having the opposite experience, where I'm not having luck finding an ML team, even with research experience. What teams/products are hiring for ML-related roles without experience?
