AI Academic Brain vs. AI Founder Brain
Academic-Brain vs. Founder-Brain
I see many academics try to start companies. While there are a few noteable examples of this working out: Coursera, Tableau, Google (though Larry/Sergey were still early in the PhD when they left), all of Chris Re's companies, etc. Those are the exemptions, not the rule. There are many examples of this not working out. This is because the habits, skills, and instincts engrained into great academic are almost antithetical, certainly orthogonal, to what would make you a great founder. Academics need to unlearn key instincts beat into them by their PIs, over years of bleary-eyed nights, to make it as founders. This is literally killing a part of their identity. Academics have the intelligence to pull it off, but most do not have the will. I break it down to a few key issues below:
1. Most of what Academics care about would be deemed "Vanity Metrics" to a common startup CEO
Academic-Brain is focused on # of citations and # of papers published (literally the only 2 measures in H-Index calculation); Founder-Brain is focused on usage, are you building something people want or care about. That requires focus on a single customer's problem and walking backwards and maniacally focused on that one singular customer and that person's problems. In a lab, a 3% lift on a static benchmark is exciting; in startups, who the hell cares? You need to be 10x better to get a company to switch to you. Academics are constantly optimized for Reviewer #2 and proving him/her wrong; in startups you optimize for your customer always. Will he/she pay for this? The researcher loves a new loss function; the founder loves watching a user share their screen and say, “Wow! This is so much better than before”.
2. Immutable datasets vs. Solve the problem by any means necessary
>99% of Academic benchmarks and papers are written and tested against immutable (unchangeable) datasets. This is great for science so its clear what is happening. However its impracticle for the real world. The easiest way to improve on a specific skill (like self-driving car, or robotic manipulation) is to capture more, diverse, clean data. The ROI on that beats always any architecture improvement out there. So just do it. Do the dumb thing to improve the product. Academics will consider this "cheating" which it is in academia, but in the real world, its the right answer! Every academic turned first time founder wants to start building training pipelines. This is usually the wrong first step. The right first step is almost ALWAYS HITL (Human in the loop) pipeline. You need the incoming live data stored remotely anyway. You need the labeling pipeline anyway. Your v1 model is going to have some % human validation anyway that will start at 10% going through HITL then will get to 0.1% or so. So just start it at 100% and get customer feedback, etc, right off the bat! You will find that once you do this, your understanding of the end product, labeling docs and how to solve the problem, domain shifts, etc etc will all be drastically different than what you expected. And you will pull in timeline by a year or more. Also you will allow your company to run in parrallel vs. in series bc now product can start customer interviews, sales can start demoing / selling (even if unprofitably), etc etc. All good things.
3. Focused on the Unimportant but Interesting vs. Important
In startups, you dont have time. You are dying. You are falling out of the sky while you are trying to assemble the plane. You dont have time to worry about why this tweak to the tokenizer results in an improvement, it just does, and thats good enough. Empericism is fine. The scent for what is important and what is not is completely different. Important = anything that moves the needle for the business: sign-ups, activation, retention, revenue, or referrals. Unimport = anything that is extremely interesting but doesnt move the needle, may make your brain tingle though! Proving a new lemma that your algorithm is globally optimal? Unimportant. Visiting 10 users to get feedback on a new launch? Important. Publishing. Unimportant. Pushing product. Important. Many academic startups die because they are run out of money focusing on the interesting but unimportant before they get to PMF and default alive.
4. Novelty
Joe Lonsdale has a great talk about MIT startup vs. Stanford startup. MIT startups are all about this new invention, the black box that does some magic. Stanford startups are all about a single customer problem, and how to solve it better and better over time. The former barely ever turns into a massive company. The latter almost always does if done right. This is because MIT is much more "Academic" and academica loves (LOVES!) novelty. You dont get published with a me-too paper.. BUT in startups not true! You can become a billionaire with me-too'ing. e.g. Building Uber 2.0 (lyft) or OpenAI 2.0 (Anthropic) or Instacart for Latam...