Monday, August 29, 2016

Startups in Machine Learning & AI

Artificial intelligence is going to have a massive impact on multiple business verticals over time. The displacement of both blue collar and white collar work by machine learning is going to cause major societal displacements in the next 10-20 years[0].

While there is a lot of discussion in the popular press about general purpose AI (aka AGI - which is defined as a machine that can perform any intellectual task a person can), much less emphasis has been placed on near-term specific vertical markets or areas that AI and machine learning (ML) are likely to transform in the coming 5 years. In short, I think AGI is still 10+ years away, but vertical products driven by AI will be transformative in the coming years.

The areas listed below are underinvested by entrepreneurs and VCs. I believe some of the largest AI companies in the near term will emerge from these areas[1]. The two keys to success in all machine-learning driven areas will be (i) the ability to construct useful data sets with which to train models and having the staging ground to test these models recursively and with closed feedback loops and (ii) choosing a market where machine learning helps to create a product that the market needs. AI for AIs-sake alone is simply a technology looking for a problem - and those sorts of startups tend to fail[2].

Here are some areas where I think AI can create big companies in the next 5 years:

1. Hardware & ASICs. 
One likely outgrowth of the inevitable rise of self driving cars, and other markets using lots of machine learning, will be the need for more efficient hardware optimized to run ML models. There has been little investment by startups or venture capitalists in the underlying chip architectures to allow for running faster models or ML systems. Many companies that use specialized clusters for ML tend to use NVIDIA GPUs (graphic processing units), which have not been optimized specifically for machine learning. There is room for hardware innovation (ASICs or other approaches) in this area and for the equivalent of an ARM or Qualcomm to emerge. This will be driven not only by the overall growth in ML applications across various verticals, but also by the need for processors in self-driving cars and other hardware at scale. Perhaps the first $10 billion dollar plus AI company will be a chip company.[3]

Companies like Cerebras[4] and Nervana are currently working in this area.

2. Fin Tech. 
The rise of the robo-advisors like WealthFront and FutureAdvisor (acquired by BlackRock) shows that financial services companies are aware of machine driven portfolio management and trading. Machine learning will have a growing impact on how portfolio are constructed and traded, as well as insights extracted from various types of financial data.

Fin Tech applications of ML and big data will have at least 3 components:
a. Tools to make traders more efficient or glean unique insights. Companies like OmegaPoint are focused on making a “Next-generation Bloomberg” where machine learning models can be built on top of data feeds used for trading.
b. ML based portfolio management and trading. 
c. Applying machine learning models to understand and properly price financial products. 

Areas like insurance, mortgages, and derivatives may all benefit dramatically from the applications of artificial intelligence.  A startup or incumbent that will use new statistical approaches and ML to generate better priced mortgages will end up being a massive company.

3. Self-driving cars and trucks. 
Self driving cars are going to disrupt the multi-billion dollar transportation market. Major car and truck manufacturers already realize an existential crisis is coming their way. The pressure in this area on incumbents is exacerbated by Tesla, who may finally have a mass market car and is pushing hard on the self-driving side. Matters are made worse for incumbents by tech companies like Google and Baidu working on self-driving technology. Many car companies would prefer to go it alone and avoid dependence on these tech companies – hence the $1 billion acquisition of Cruise by GM and Uber paying 1% of its company for Otto. Acquisitions in self-driving cars and trucks will continue over the next 12-36 months.

 Self-driving technology will lead to the loss of millions of jobs and the transformation of society (should be largely deflationary force economically). While the industrial revolution occurred over a period of 150 years, I would not be surprised if there is massive job displacement over a 10-20 year period as multiple professions disappear and people are displaced. If opportunities are not created for this segment of society, we will see political upheaval.

4. Medicine. 
While most examples people tend to use talks about AI & robotics displacing blue-collar workers, I think AI will actually displace a much larger segment of white collar workers[5].

One area of big disruption will be medicine. ML will radically change how diseases are defined, diagnosed and treated. Vinod Khosla has some great thoughts in this area.

Machine learning has enormous potential to augment or replace major components of medical care. Imagine if literally anyone with a smart phone had access to the equivalent of the world's best specialist doctors, at low cost, from anywhere.

a. Diagnosis. 
Ever have the terrible experience of waiting for a doctor for 45 minutes, being seen for 5 minutes, and then told a simple solution to your medical problem? The poor customer service and high cost of medicine would not be tolerated in any truly competitive industry.

Moving diagnosis from people to machines will likely start on a per-disease vertical basis - for example IBM Watson has claimed progress in oncology. Similarly, there has been some fascinating work on diagnosis of depression and other mental health disorders using deep learning. The limiter is often the data sets available, and the ability to have a closed feedback loop on diagnosis and outcome.

One approach to accelerating machine learning application in medicine would be to buy out an existing radiology center or clinic. This center could be wired up to optimize for data generation, which would be used to train a machine learning model to diagnose and treat a patient. By launching the model side-by-side with standard care in this clinic, you would have both a feedback loop for the model as well as ease regulatory and patient care concerns.

In parallel, diagnostic tests themselves will improve with the use of ML models. For example, using ML to determine if a variant call on a DNA sequencer is correct, or what cell type you see on a FACS sorter. At my company, Color Genomics, we have started to apply machine learning to genomics in multiple ways.

b. Treatment. 
Similar to diagnosis, machine learning should be able to help select the right treatment options for patients. The biggest limiter here is probably (again) access to data).

c. Continuous monitoring and analysis. 
One way to increase available data for machine learning models is through new consumer-centric technologies for ongoing monitoring. Self-monitoring of health data has a niche, passionate community behind it[6]. In parallel, companies like Cardiogram are putting more power in consumers hands by continuously monitoring pulse and other data. The Freestyle Libre is being adopted by a number of Silicon Valley denizen to self-monitor glucose levels. This may change medical practice by empowering “citizen science” and having more people be proactive about their own healthcare and outcomes.

5. Education. 
I have never invested in an education startup as the education as a market in the US is pretty terrible from a technology perspective. That said I am pessimistically hopefully that an adaptive learning company will emerge that tailors online content and coursework based on the student. This should accelerate learning abilities in both the developed and developing world.

6. Other areas.
A number of areas are not mentioned in this post (manufacturing, advertising etc.) that have been, or will be dramatically changed by machine learning. As an entrepreneur and investor, I am personally most interested in the above right now but many other opportunities abound. Entrepreneurs need to ask themselves how machine learning makes a product 10X better in a market, rather then just think ML itself has value in its own right. That is the key to building a big company around AI today.

Data limitations.
Fundamentally, many of the limitations to vertical AI are data set dependent. Many advanced ML applications in finance, medicine, and other areas would exist today if the data sets for these areas were broadly available. Large companies (e.g. Google, IBM) as well as startups will generate useful data sets by either buying access or partnering on solutions. Data is going to transform multiple industries. This will be both a deflationary and democratizing force (same standard of medical care in poor as well as rich countries), but will also displace massive portions of the developed world workforce. The ultimate impact of machine learning in the next 5-10 years will be both to broaden access to key information (e.g. medical diagnosis) as well as to displace millions of people. More on this in another post soon.

[0] More on this in another post I am working on.

[1] Once, of course, we have true AGI, the world will change rapidly, but that is still enough time away that exact timeframes are mere speculation.

[2] Or get acquired by one of Google, Facebook, Uber, etc. There are going to be tons of acquisitions in this market in the coming years as companies try to beef up their talent in this area.

[3] One could argue that Google search / ads is really the world most successful vertical application of ML and it is already well above $10 billion in both market cap and revenue. So, I mean the next startup to succeed :)

[4] Thanks to Andrew Feldman for quick read of this blog.

[5] More on this in a future post.

[6] Most world-changing products start off looking like a toy, and self-monitoring of health information will likely fit this trend.

End of Cycle?
3 Types Of Platform Companies
Defensibility and Lock-In: Uber and Lyft
Uber And Disruption
Who Cares If Its Been Tried Before?
The Road To $5 Billion Is A Long One
How To Win As Second Mover
End Of Silicon Valley
Social Products
Hot Markets For 2015