Apprente acquired by McDonald's
2019-09-15T17:44:56Z (to long-term-world-improvement@googlegroups.com):
Apprente, where I am currently a Member of Technical Staff, has been acquired by McDonald's to automate ordering at the restaurant drive-thru:
https://bloomberg.com/news/articles/2019-09-10/mcdonald-s-buys-startup-to-add-automated-drive-thru-ordering <https://www.bloomberg.com/news/articles/2019-09-10/mcdonald-s-buys-startup-to-add-automated-drive-thru-ordering>
I'll now be joining McDonald's as a founding member of the new McD Tech Labs, which will operate out of Mountain View near our current offices on Castro St. Now that our company has been acquired, I hope to have some more time to spend on this list, although my primary job will continue to be deploying and scaling our AI system. In particular, I'd like to dig more into analysis of what exactly explains Thiel's and Cowen's observations of Western stagnation. What happened? Why? Zero to One has a few generalizations and case studies, and I think they're essentially accurate, but the book doesn't really do any novel data collection. Papers like the "Eroom's Law" findings in drug discovery are similar ( http://maodl.com/wp-content/uploads/2012/10/JackScannellNatureMar2012.pdf)
- observe a trend, kick around a few hypotheses and case studies, and call
for a new bureaucracy to investigate it further. I don't think that will happen anytime soon, so I guess I'll have to try myself, Lord have mercy. Why do the clinical trials take longer? Is it just "regulation"? Which regulations?
On the conversational AI side, although I'm very happy our company succeeded, my experience there makes me more sympathetic to Gary Marcus's arguments about the fundamental limitations of deep neural networks (and training with backpropagation on i.i.d. data). To a significant extent, this training paradigm was created by the availability of cheap compute when you do massively parallel linear algebra on GPUs (and now TPUs/ASICs). After many years of stagnation, we're now seeing steady drops in the cost-per-core for multicore x86 CPUs, so I'm now excited to think about what new techniques might come out of that type of hardware.
--
- Alyssa
2019-09-15T20:03:25Z (to goldhaber.ben@gmail.com):
Thanks Ben! Unfortunately the full analysis is still confidential, but basically just several years of trying to solve parts of the fast-food ordering problem, seeing that nothing out there fit our requirements, trying to work around or design a new system, and seeing how much effort / how many layers were needed to handle everything a customer could ask for with high reliability. Serious commercial applications are awesome as benchmarks, since you know that the problem hasn't been made artificially easier to "fit" AI limitations, and that the problem wasn't "solved" in a brittle way that breaks under slightly different assumptions. Even just ordering a burger is much more complex than a standard "classification task", since if you try to analyze it that way, you'll find that the space of possible orders makes the number of classification buckets exponentially large.
On Sun, Sep 15, 2019, 1:56 PM Ben Goldhaber <goldhaber.ben@gmail.com> wrote: