If you're reading this, you're probably aware that Hatcher+ is using deep learning methods to evaluate and score business plans and associated data with a view to enabling the inclusion of an unbiased "voice" in the human-led investment committee process that happens within an accelerator or VC firm.
As my partner Dan Hoogterp likes to say, we're not using deep learning in order to take over the decision-making process - we're using it to provide a useful analytical "assist" in the deal evaluation process. As with the opinions voiced by different human analysts at a VC firm, the partners are free to disregard the analysis output by our algorithms. And if they are people we're co-investing with, we will back those human decisions, regardless of what our data analysis indicates. The analysis is there to provide a voice at the table - not to negate years of human experience and common sense observations.
So now that we've cleared that up, I'd like to turn to another question that we get asked relatively often: what would happen if someone used AI to write a business plan, and then input that plan into your system? I thought it would be interesting to explore this.
The first and most obvious observation I would make about the problem set is simply this: business plans require a real problem at their core to engage the reader - a real problem, real people (founders), and real data. Any attempt to write a business plan would need to be able to talk intelligently about an engaging issue, present accurate data on that issue, and perhaps reference some real LinkedIn profiles, as a minimum. It would need to be engaging.
Which is why, to assist in the writing of my AI-powered business plan, I chose for my subject the mission of "removing plastic from the ocean." This is something I'm passionate about (is any thinking person not worried about this?) - but have zero expert knowledge of. I don't know exactly how much plastic there is in the ocean, or how much gets added annually. I have little to no idea of the methods I would propose to start cleaning things up.
For my plan, I chose HuggingFace's web-based "Write With Transformer" tool (note: a "Transformer" is simply a deep learning tool that is designed to handle ordered sequences of data, such as natural language, and do various tasks such as machine translation and text summarisation - probably the most simple example of which is an individual Google search result.)
And here's what I did: I wrote an initial paragraph - and then spent the next fifteen minutes simply hitting the "Tab" key and selecting from the available options HuggingFace presented me with. The Transformer did the rest - including adding the reference data (which isn't contextually correct, but it's kind of impressive.)
The result that I managed to create in this 15 minute session is shown in the screen-capture above - with the machine-generated text highlighted. The single correction (in white) was only necessary because I (the human) made a bad selection - the machine suggested "petroleum products" as the best solution and I chose to ignore it to see what would happen - silly me.
So how did the Transformer do? You can make that assessment yourself. Personally, I think the first few machine-generated paragraphs would probably not pass muster with most humans in the business of reading business plans (but on the other hand, in terms of grammar and content, I've read far worse...) The fact that there is no mention of the team of founders post the first para raises an eyebrow. But I must confess to a couple of surprises. The second-to-last paragraph in which "three methods for reducing the use of plastic bags" are described was produced entirely by HuggingFace. It contains a sensible, logical list of options. If a human were to have written this, I might have even added "thoughtful" as a complimentary adjective.
This exercise shows, at least to me, that just as deep learning tools are becoming increasingly successful at producing fine art, fake news, and robots capable of advanced gymnastics, generation of a readable business plan will not pose a significant challenge to AI practitioners. And while we can take some solace in the fact that the very same deep learning tools are increasingly being utilised to spot these non-human creations, when an advanced deep learning system like Alpha Zero plays a traditional game, like chess, against another advanced deep learning system, the result is inevitably this: both tools get smarter.
[Note: In the link above, Alpha Zero plays itself in what one Grand Master has described as "The Perfect Game." I'll leave the philosophical question as to whether Alpha Zero got smarter as a result of playing this game against itself to the actual philosophers that may be reading this.]
As the mission statements of business plans and the technological challenges they articulate become more complex, and move further and further away from the personal experiences of venture partners, deep learning will become more and more necessary - both as a tool for evaluating false or complex claims targeting complex environments and opportunities, and as a methodology for evaluating what will be required operationally to enable the best chances of achieving the stated mission.
Our five to ten year goal at Hatcher+ is: continuously develop our data sets, and our deep learning skills and models, so we can assist humans to make better decisions in the face of these challenges. We're three years into this mission and it feels like we're making good progress.
At least that's what my human intuition tells me.
John is a serial entrepreneur and investor, and the Founding Partner of Hatcher+, a next-generation, data-driven venture firm that utilises a massive global database in combination with AI and machine learning-based technologies to identify early-stage opportunities in partnership with leading accelerators and investors worldwide.
The Best Investor Deck Ever|
The Most Unrewarding, Misunderstood, Underrated Job on the Planet|
AI vs. Human Reasoning in Venture|
The Power of Venture|
The First 100 Are Funded!|
|John Sharp (247)|
|Wissam Otaky (11)|
|Dan Hoogterp (4)|
|Amit Kaistha (3)|
|Manav Narang (3)|
|Darren Thang (1)|
|Stephen Fisher (1)|
|The 210 Steps Needed to Set Up a VC Fund|
|Asset allocation in Venture?|
|The Olympics and Startups….aren’t they so similar?|
|Innovation transcends boundaries… does your early stage investment in Venture too?|
|Hatcher+ and Australian Wholesale Investor Launch $10M Early Stage Fund|
|Accessing Day Zero Venture Deals - A Checklist for Family Offices|
|Weeks, Days, Hours, Minutes, Milliseconds|
|Lessons from SEALS for Startups|
|Spray and Pray - What The Data Says|
|How To Generate a 1207x Return|
|AI vs. Human Reasoning in Venture|
|Investors Invest in the Top 1%. These 10 Steps Could Help You Make The Cut|
|Why Unicorns are a 1 in 10,000 Bet|
|Science Says: Milestones Are Bad For Founders (And VCs)|
|The Gender Gap Could Soon Be History. Thank Biology.|
|Y Combinator is Pulling Back. They Should be Ramping Up.|
|Can a Robot Write a Business Plan?|
|What's Your TQ?|
|My Five Years in Space|
|Triple Your Bottom Line|
|AI, Alpha Zero, Hatcher+, Transformer, HuggingFace, deep learning, business plan, venture capital, VC|