Andrew Ng launches $175M AI fund
techcrunch.com
techcrunch.com
Andrew has the reputation to do this and not be hated. But you know who did exactly this? Rocket internet, the clone-war masters.
What is he talking about here? First of all, this is just the basic lean startup methodology. Any startup (especially consumer ones) try to do that, i.e. A/B testing new features. Second, he may be referring to the case that once you have a big comprehensive dataset already gathered, then you can just iterate improving the model.
My (cynical) answer: good luck with that. The real issue with AI especially in mission-critical situations is data in edge cases. Even if you can improve on your train/test data, that barely translates to customer satisfaction unless the edge cases are treated well. And edge cases will be revealed very slowly by encountering data in the field. Just ask anybody working on voice-assistant agents (Alexa, Siri, Google one) or self-driving. It's not a fast loop by any means.
Of course, Ng probably knows all this which makes me feel even more bothered.
Ng's implied claim is that he has a head start on how to tackle the edge cases, along with other hard problems encountered in the field, based on his experience at Baidu. This is what got him the "easy" capital raise. If he really has an edge (no pun intended), then it makes sense for him to apply it while it's still relevant.
What I will bet money on though is that no massive success will come out of this. The key to AI is data & hardware (and not algorithm or methodology -- where everyone is essentially doing the same) and that's the domain of already giant tech companies. There can be a lot of good use cases in teaching far-off industries on how to use AI (manufacturing, etc.) but that's essentially a consulting business which will have a limited window. So lots of little successes is completely achievable; massive successes I hardly doubt.
True. And the key to keep (increasing) velocity is momentum. I believe the fundamentals to that is infrastructure. Once you get to the "flywheel stage" of an AI company (http://nicodjimenez.github.io/2018/01/25/stages.html), experimentation becomes so easy that new models can be rolled out nightly.
If that makes any sense at all, that is.
This is really a different kind of venture, and it will be interesting in a few years to see how they do.
Courses on Coursera aren't really taught that way. The course instructor(s) prepare the course materials (videos, presentations, quizzes, assignments, etc.) once only. Courses start on a regular schedule and the platform, rather than the instructors, takes you through the material in sequence. There are forums with moderators for peer-support, but you don't typically interact with the instructors as they're not really there.
I assume the OP is referring to Ng building the fifth of the five courses that comprise the specialisation.
Phil Libin is the ex-CEO of Evernote.
Maybe there's room for multiple players in this space, but I imagine the name recognition and bona fides of Andrew Ng is going to suck the air out of All Turtles.
I can imagine that Andrew focuses more on companies that have or have the potential to have significant AI themselves. I doubt that Andrew would consider a bot that uses the Google Vision API and Dialogflow an "AI" company. If that makes sense.
PS: I do think that the latter should be considered an AI company. Just like a company using AWS to host is called web company. In fact, probably a lot of companies will be using some sort of AI api pretty soon.
Predicting real estate opportunities (which houses are about to sell) predicting energy usage for thousands of sites in real-time, predicting churn, predicting Ad-Tech prices and fraud in real-time, predicting anomalies in seismic radar scans, customer segmentation and recommender engines.
Those are examples of real world paying customers using Deep Learning
Planet labs now have the ability to map the entire planet every day at a 3m resolution. So object detection applied on these images can be a quite efficient. I still wouldn't call this AI as we are talking mostly about supervised learning but it's a highly practical real world use case. CrowdAI and the like are on this path already.
It is battle tested by me and I can say it is surprisingly accurate.
We would not fit his definition of a vertical specific startup. (We have also been around a while though) The bulk of what we do is time series.
Applications we do for real paying customers include:
Detecting theft of power on the raw grid
Online payments fraud
Detecting people stealing from the telco network
Detecting faults in assembly line machines
Detecting computers about to fail
Detecting root cause of dropped calls
Kind of researchy, but we've also done robotics with RL to teach a robot to learn an obstacle course.
The problem is that though the above may currently be a profitable business, I can't see how you will generate a "moat" and mature into a monopoly or a big market share in a vertical.
Skymind is a weird mix of joint ventures in asia, US and Chinese investment. We're not looking for an exit anytime soon. We are looking to build a big business though.
The "moat" is a land and expand strategy. There is lockin with our tooling. It's a standard on prem play. We're hard to get rid of once you install us. We help in house teams compete with external vendors. Interest actually aligns there. Happy to elaborate a bit otherwise.
AI as we have it now is just a bunch of ML algos and *NNs that perform well in different niches and, well, after all cat pictures were categorized, what to do next?
So, in case Andrew is reading this post, I would like to have a chat with him about my research ideas and if they could be funded. Because if this happens, many more would follow and then we can see a real progres towards real/hard AI.
AI is on a different scale andmay have less to do with the actual implementations.
How do companies justify such a tiny edge as being worth so much? I wish Andrew Ng the best. I do. He's a class act. I enjoyed his course years ago, but came away a little disappointed. At the end, I realized there's no intelligence in the algorithms. It's all just applied statistics and a lot of sweat preparing training data.
Maybe I should just shut up and vie for some of the easy money sloshing around. I think I would just feel weird asking for a ton of cash when the end product is just a bit of recursive math.
We are on the very top of that hype curve, a few of years from now we'll forget how absolutely stupefied we were when ML gave us models that "can X better than humans" and the enthusiasm will give way to a feeling of "I overpaid for this". It happened before with databases way back when, then with specialist systems, then ...
The only difference is now the layperson hears about it with an astounding frequency. I don't know, maybe that will make things different but I can't see it having any other effect than making that 'disappointment crash' harder
The ML course though is great as a precursor though.