Revisiting Old Startup Discussions

What worked? What didn't? What's next?

Silicon Valley doesn’t have a very long memory, which, while likely a net positive for innovation, means there’s alpha to be found in seeing which ideas stood up, fell down, or morphed into something entirely new.

Some of this was covered in Elad Gil’s “AI: Startup Vs Incumbent Value” and I think this piece also is a partial answer to another of Elad’s questions about why some people remain relevant throughout multiple platform shifts and eras while others don't.

Here’s some of what I found interesting:

• Neural Networks worked when they got big and will continue to be important

• YC felt much smaller (because it was) under pg than now

• Everyone knew about AI but the startups didn’t REALLY work until recently; also everyone gets paid way more now

• 2009 Marc Andreessen was absolutely correct from a VC perspective

• Valuations Got Way Bigger

• Mobile was mixed for predictions

• Thiel tried to find the successful weird companies and there weren’t that many of them

• Finance is mixed but quant got really big and there’s still opportunity in crypto trading

• VR/AR were big company things

• IoT was overhyped but Samsara is really big now and other companies use the tech

• Healthcare + Education improved (but more from the outside)

• 2016 Marc Andreessen was correct and then went two for three

• Synthetic Biology Mostly Flopped

• Agriculture was so-so

• Defense is still underrated

• Solar and batteries both took off but batteries were the business that worked

• Commodities startups didn’t work in the past, didn’t work recently, and I don’t think they will work in the future.

Neural Networks worked when they got big and will continue to be important

“Deep Neural Nets: 33 years ago and 33 years from now” by Andrej Karpathy is worth reading. In the past neural networks didn’t have big datasets or big compute, and weren’t very well optimized. Now they have big datasets and big compute, and are better optimized. In the future they will have even larger datasets (which could be interpreted in many ways), much more compute, and we will have even more optimizations.

A small side note, but it’s interesting that neural networks were seen even by the Google founders as academic toys that weren’t particularly useful. Statistical techniques such as Bayesian Filtering were used to combat email spam, but neural networks didn’t start picking up steam until around the 2007/2008-era, and until ImageNet didn’t break into the broader research consciousness.

Thus, question: What’s the ImageNet of today equivalent that we should be building?

YC felt much smaller (because it was) under pg than now

I’ve only ever spent my adult life in a Bay Area where YC is this famous institution, and while I’ve read the early pg essays I didn’t quite intuit how small/informal they were until I had read/saw the following:

Mr. Startup: Paul Graham, founder of YCombinator, and his “no asshole” rule

-> was an interview done about YC fourteen years ago. I did not know any of the hottest companies that they talked about. Also pg writes like he talks.

(Side note: Despite it frequently being mentioned around this time, DropBox didn’t quite reach the size of Coinbase, Stripe, Airbnb (roughly ~1 OOM less in market cap).)

I think reading Chaos Monkeys, which talked about pg, sama, and all of them around 2010 was interesting now that they’re seen much more as “tech legends” today. Even at a young age Sam was intensely well connected in the Valley with important people.

Everyone knew about AI but the startups didn’t REALLY work until recently; also everyone gets paid way more now

Everyone knowledgeable somewhat knew that AI was going to be a big deal in the early to mid 2010s. Sam Altman was talking about machine superintelligence in 2014/2015. The hype around self-driving was enormous, AI was (and still is) the topic of the hottest classes at Stanford, and then…

None of those startups really panned out in a spectacular way. I think Elad was correct in saying that AI then offered a 0.5-3x gain but not a 10x gain. I’d also add that in places where it did promise to offer a 10x gain (driving), the safety standard required to ease people’s anxiety made it a bad product to start with as compared to modern products like Cursor or ChatGPT which have much lower stakes for failure.

The salaries around deep learning weren’t that unusually extraordinary when OpenAI started, but kept climbing quite fast.

I expect this rise in compensation will continue and that top researchers/executives in AI will be paid hundreds of millions to potentially billions in the future (if you count the equity from starting new labs, they already are) for their work.

2009 Marc Andreessen was absolutely correct from a VC perspective

“Andreessen's faith in the power of the web isn't terribly surprising, given his front-row seat at the outset of the Internet revolution. But Andreessen is an online absolutist. "No clean tech, no rocket ships, no electric cars. No China or India," he says, explaining Andreessen Horowitz's investment strategy, which differs from a lot of established VC funds that are fervidly pursuing all those areas. Andreessen's unwavering view is that the Internet will soon take over all aspects of our lives. Online services won't merely supplement your TV viewing or newspaper reading, but will replace those activities altogether. Amazon's Kindle is okay, but it would be so much better with access to the Net.”

He nailed it. Being all-in on software in 2009 was the right move for a VC to make. (The rocket ships and electric cars were an Elon outlier phenomena.)

This also reminded me of the piece “Grading Extropian Predictions” by Max Tabarrok, an economics PhD student at Harvard:

“Measuring by what’s resolved in 2023, I think that a normal, centrist computer science professor would have outperformed most of the Extropians [predictions of the future] here. They would be just as good on Moore’s law and the growth of online content but they wouldn’t be so pollyanna-ish about cryonics and free-market politics.”

Valuations got way bigger

The Road To $5 Billion Is A Long One. On 10/21/12, software companies started in 2002 or later that had a >$5 billion market cap were Twitter, Skype, Workday, LinkedIn, and Facebook. Only Facebook and LinkedIn had greater than $10 billion market caps.

Valuations are way bigger now. Will they continue to go up? Maybe. Lots of it came from fast growth which gave crazy multiples to Big Tech, and if that growth slows down, the valuations will come down, but if AI really, really works, then I wouldn’t be surprised by 10T+ mega-cap companies.

Mobile was mixed for predictions

I’ve grown up in a world with smartphones all around me, so it’s fascinating to read smart people talk about mobile when it was this very new thing.

They were right that it was going to be very big and important.

I think some folks overestimated how much startups would be able to gain on this platform shift from incumbents (there wasn’t a new mobile CRM it was just “Salesforce, but now on mobile”).

I think they overrated how important games/initial apps like Angry Birds were going to be. (They were important but not what drove many billions in revenue)

They were right that Twitter was going to be a really big deal. (And I also learned that there were other similar services to Twitter at the time; I did not know that!)

(Side Note: The Essay “Twitter is to Facebook as Google was to Yahoo!” by Elad Gil could be extended to give a rationale for why he invested in Happenstance as Happenstance:LinkedIn in this analogy, where people search is the most important part of LinkedIn and a startup will do that better. What other analogies exist like this that should be built but aren’t being built today?).

Thiel tried to find the successful weird companies and there weren’t that many of them

Peter Thiel Is Taking a Break From Democracy:

“It was harder than it looked,” Thiel said. “I’m not actually involved in enough companies that are growing a lot, that are taking our civilization to the next level.”

“Because you couldn’t find those companies?” I asked.

“I couldn’t find them,” he said. “I couldn’t get enough of them to work.”

I’m optimistic that this is starting to change, but I think these types of companies either require folks to already have money or to be very intentional.

Anduril was an incubation, Saronic was an incubation, Palantir was an incubation, I suppose SpaceX was an incubation in a sense from Musk (who already had money), OpenAI could be viewed as an incubation if you count it as a YC project.

Finance is mixed but quant got really big and there’s still opportunity in crypto trading

Peter Thiel started Clarium Capital and that did well at first and then blew up.

Leopold Aschenbrenner started Situational Awareness, LP (SEC filing here), and it’s too early to pass judgement on.

Ken Griffin’s Citadel has emerged as a pretty big winner in the space, and is now VC-backed.

Even though Alameda Research wasn’t technically a hedge fund in the strictest sense (it would better be described as a quantitative trading firm) it did quite well due to the inefficiencies of the crypto markets.

I suspect crypto hedge funds/financial institutions will continue to do quite well because that industry is still very inefficient and seen as somewhat scammy by “traditional finance” so they’ll overlook it more than they should.

VR/AR were big company things

The one really successful VR/AR startup was Oculus which got sold to Facebook, now Meta, and they’ve turned it into a somewhat successful platform.

I think the Meta Ray-bands are being underrated by most people who are terminally online (young tech folks) from seeing empirical adoption by my family and others in the Gen X generation.

The lesson here is probably that doing consumer electronics is just hard in general, which is why nothing else really worked so far, as there’s not really many other consumer electronics success stories (Humane and Magic Leap shut down, we’ll see how Rabbit/Friend/Truffle/etc. do in the coming years).

Funnily enough, Starlink is a referendum on consumer electronics in that you absolutely can make them in the US, but it seems like so far only Elon companies can do it.

IoT was overhyped but Samsara is really big now and other companies use the tech

IoT was really, really hyped but Samsara is now a 20B+ company, so, good for them.

Big Tech integrated some of it into their platforms, but it’s not really mentioned in the same vein as AI these days.

However, one could argue that a core concept in IoT, sensor fusion, is key to Anduril’s success and thus they could be viewed as an IoT company in a sense.

Healthcare + Education improved (but more from the outside)

I lump these two in with each other because they tend to be the prototypical “good for the world, also big percentage of our GDP (healthcare) and lives (education)” but which haven’t had a truly outstanding 100B+ outcome in space.

The same rhetoric about why these companies are important is mostly the same rhetoric used today as well, which does worry me somewhat. (If in the past we didn’t achieve these 100B outcomes using that rhetoric, why will that rhetoric work today? What have we learned in the past 10-15 years?)

For education, there was the MOOC craze in the early 2010s and that didn’t pan out quite the way folks were expecting.

I think Primer and microschools will be interesting from an education perspective thanks to new regulations, but from an educational perspective I think YouTube/Twitter/Social media got very underrated in terms of providing content and LLMs are getting overhyped. (Not because they’re useless, no, but because they’re too useful. I think many people don’t use these systems to learn more, they use them to turn their brains off more. This is one of the failure modes of weird nerds thinking about the impact of the internet and now AI: yes, all of the information in the world and now brainpower in the world is at your fingertips, but that doesn’t necessarily lead to a wiser, smarter population.)

For healthcare, there’s a lot more, such as Oscar Health, Flatiron Health, Ro, etc. but none of them are at the 100B scale.

I agree with Will Manidis that lots of the reform will come from Incumbents adding AI rather than new startups, although there will be new startups in healthcare that are doing well.

However, the interesting insight might be that the best way to make education and healthcare is not to start an education or healthcare company. YouTube/Social Media/etc. have been enormously impactful for education despite not being created for that purpose, and when it comes to nutritional advice, diagnostics, mental health, and so on many people are turning to LLMs for support. (I am not a medical professional so I cannot endorse or deny whether one should do this, but that is anecdotally what I see many doing.)

2016 Marc Andreessen was correct and then went two for three

“If I was 18 and going to college right now, I would do computer science again in a heartbeat. And then I would either focus on: #1 distributed systems and the broad domain of cryptocurrency; #2 AI & machine learning; or #3 intersection of biology and computer science—so genomics and synthetic biology. Personally I would do one of those three areas, and the hard part for me would be picking between them because I think all three of those are going to have transformative work happen in the next 20 years. They’re going to change a lot about how the world works.”

I think Marc was absolutely correct about the computer science major, and then for what to focus on I think he got ⅔ right. While I am bullish on longevity and biotech beginning to have more breakthroughs in the next decade, looking back 9 years ago I think crypto and AI were the much better bets from a “maximize market capitalization” career perspective.

Synthetic Biology Startups Mostly Flopped

We have Solugen and…that seems to be it for durable unicorns if you don’t count successes in pharma, which I don’t really count as a synthetic biology win so much as another pharma win.

Solugen is also interesting because much of their success I think didn’t come from synthetic biology necessarily but a lot of Sean Liu (their former CFO, now a partner at Founders Fund, wonderful guy) being a financial wizard with funding, their reliance on heterogeneous catalysis and non-synthetic biological methods, and an intense focus on the business/commercialization side of things. (Banana Capital did a good podcast with the founders about this)

Agriculture Startups Were So-So

David Friedberg’s “The Climate Corporation,” selling to Monsanto in 2013 for 1.1 billion, was the first AgTech unicorn. And while there are more unicorns now, the largest one has a valuation of around 4 billion and there’s a handful of other unicorns.

That’s an incredible accomplishment, but considering all of the hype around vertical agriculture, robotics, etc. this space I don’t think quite lived up to all of its aspirations. However, it has built durable value, so it wasn’t a 99% flop like synthetic biology, and new companies like Rainmaker and others have promise.

Defense Is Still Underrated.

While some say that defense has manufacturing problems (prototypes and fancy websites are easy but production is hard), I think it’s still early enough that you don’t have to call it a flop yet. Anduril and Saronic raised funding to do build-outs, so we’ll see how it goes.

But, from my perspective, what’s really powerful about defense is that you can actually manufacture things in America with great margins and some level of protection without having to deal with the heavy regulations of the FDA. (I am not an FDA hater, I appreciate the work they do, but the iteration cycles are very long because of the FDA in the medical device space compared to the defense space.)

Your competitors will naturally have to be American, and if they’re older companies then while getting the sales motion for the federal (and especially local) government will be tough at first, technologically you will be ahead and modernized.

I think as well for building hard-tech companies having funding from defense is still underrated. SpaceX lived off of NASA contracts, Radiant Nuclear is taking army contracts for accelerated regulation, and I think more founders are going to be working with the DoD to bring sci-fi products to life.

Solar and batteries both took off but batteries were the business that worked

Solar and batteries ramped up on an exponential curve in terms of production and came down the exponential curve in terms of cost. Elon was right about that.

Thiel was right that the solar manufacturing and installation businesses were terrible investments despite all of the exponential growth and publicization of “global warming” from Al Gore.

Grid batteries seem to be making Tesla Energy a lot of money but despite the ramp up in manufacturing capabilities there will be a lot of competition from China and others for Grid batteries.

I think Base Power Company and Tesla’s Powerwall will do very well, however, as home batteries have additional moats beyond just raw technical performance.

“But isn’t that just solar installation all over again?”

No, it’s really backup power, so it’s competing with generators, not other sources of energy generation like natural gas, and can also do arbitrage, which solar panels can’t.

Commodities startups didn’t work in the past, didn’t work recently, and I don’t think they will work in the future.

In the early days of “greentech” a la Sequoia in the 00s, none of them panned out except for Tesla, which was not a commodities startup.

In the late 2010s we had/have the Breakthrough Energy Ventures program (I worked for a few companies doing that program) and none of those really panned out with the exception of Commonwealth Fusion Systems, but that’s such an unusual situation that I do not count it.

Austin Vernon recently wrote a great piece called “Startup Strategy for Commodity Products” where he noted that Elon’s companies (and other successful hard-tech companies) tended to take commodity materials and turn those into products that were very valuable in terms of $/ton. (Tens of thousands of dollars for cars and rockets versus hundreds to thousands of dollars for cement, steel, etc.)

Perhaps tariffs or further national sentiment will raise the probability of outsized success for commodities startups, but even if you look at Elon, the greatest modern hard-tech entrepreneur, he’s not doing commodities startups. (Anticipated response: Lithium refining and such is internal within Tesla, and as a stand-alone startup likely not terribly valuable in terms of market capitalization compared to, say, a Stripe or AirBnB)