A reminder for new readers. That Was The Week collects the best writing on critical issues in tech, startups, and venture capital. I selected the articles because they are of interest. The selections often include things I disagree with. The articles are only snippets. Click on the headline to go to the original. I express my point of view in the editorial and the weekly video below.
This Week’s Video and Podcast:
Content this week from @kteare, @ajkeen, @aquantvc, @profgalloway, @raiseconference, @heetermaria, @nmasc_, @kateclarktweets, @aquantvc, @ChartMogul, @MaelleGavet, @garrytan, @htaneja, @FareedZakaria, @StevenLevy, @steph_palazzolo, @geneteare, @AnneMarieSteele, @bysarahkrouse, @DavidSHolz, @Apple
Least Bad (De-Dollarization?)
Meet the YC Summer 2023 Batch Garry Tan
Andrew Keen and Ludwig Ensthaler on AI
Fareed Zakharia on AI
What OpenAI Really Wants – Steven Levy
The hardest part of writing these editorials is the title. Every week, I look through the stories and try to pick a title and an editorial direction that fits the content and allows me to say what I am thinking about that week.
So this week, “Searching for Money” seems to be the unifying theme. In Essays of the Week, we’ve got “Searching for Unicorn Funds, a look at fund DPI,” (which stands for distributed capital to paid-in capital). In other words, funds are looking to make liquid on their investments by selling them. Why? Because they need to return money to their limited partners who invested in the funds. So limited partners are looking for money, and funds are looking for money.
We have an article from The Information, “Venture Funds Hang Up the For Sale Sign,” on portfolios, which really says everything. There are quite a number of funds just looking to sell. Sell, sell, sell, as they say. And we also have a look at Y Combinator’s latest batch this week, who are all looking to raise money from investors. So LPs are trying to raise money, funds are trying to raise money, and startups are trying to raise money, which tells you that we’re not living in a very liquid venture capital environment. We’re living in a highly illiquid environment’
The average amount companies raise at each round is declining, and the complete number of rounds done at each stage is mostly declining. But the number of companies seeking funding is really not declining.
Garry Tan has a great review of the Y Combinator 2023 summer batch, where he focuses in on all the opportunities that there are. Meanwhile, TechCrunch has an article saying that investors are not attending the demo day event because they feel the startups are overpriced.
So, we are in this interesting world where alignment is unavailable. Everyone has their own self-interest, and everybody’s pursuing their self-interest, and they don’t really coalesce around a common set of steps or end games. LPs, Investors and Companies cannot all get what they want.
So that’s what That Was The Week is about this week. Searching for Money.
This quote sums it up:
Pressure to raise cash to return to LPs, known as distributions to paid-in capital, has started to overturn venture investors’ resistance to selling their stakes early.
“That desire to stick with a company all the way to exit in terms of loyalty to the founder is not really best practice at this point,” said Adam Struck, founder of seed-stage venture firm Struck Capital. The firm is planning to raise its next fund next year, and it’s eager to show LPs it can produce more than just paper gains, Struck said.
“Especially in this environment, DPI is king, so showing that distributed paid-in capital will mean a lot for subsequent funds.”
But my essay of the week this week has nothing to do with that.
It’s my first time featuring Professor Scott Galloway from his podcast. And he’s got one this week called Least Bad, which is a discussion of de-dollarization. It’s actually a really interesting article because it argues that the US is still strong and that any proclivity to project its future demise is uncalled for. And he goes to great lengths to say that the US is in pretty decent shape, mainly because there are no really good rivals.
He does that by looking at things like the share of global currency reserves, where the US dominates the next biggest chunk: the euro, the Japanese currency, and then the British pound. And the Chinese currency is nowhere to be seen.
The hidden polemic is against the Chinese in this story and those who predict it will replace the US on the World stage. He claims that the sense that the US will be eclipsed is fantastical or crazy.
Now, I don’t actually disagree with him regarding timing, but it is fair to say that empires don’t survive. This is mainly due to their economic demise relative to newcomers.
That really only matters if the newcomer is of a size and scale where their success in innovation, in new technologies, and their better productivity addresses a very, very large market, and especially a large home economy where they can sell the things they produce to a domestic audience without needing to export.
That was the advantage the US had against the UK in most of the 20th century, a big home economy.
It’s the same for China today, despite the declining population trends in China and the aging population. The home market is very much larger than the US market. The middle class is growing rapidly. People can buy cars; people can buy homes. And the headline news, which is about overbuilding, really disguises the fact that it is a very large, scalable economy that will be important on a world scale.
The BRIC alliance we discussed last week now has six new members and represents more than half of world GDP measured by prices. PPP is a measure of GDP that factors in cost of living and quality of life.
So clearly, Professor Galloway is right, but he’s also wrong to make the case that the US is okay because it really isn’t in any long-term view. And its eclipse is kind of inevitable economically. That doesn’t mean it will happen soon.
Britain was significantly smaller and less productive than both the US and Germany as early as the 1870s. And it took till the 1950s for the world order to change to the world order we’re all familiar with, with the dollar as the reserve currency. So, these things don’t happen fast, but they do happen.
The second thing to say is when they happen, it isn’t that scary. The UK is clearly no longer number one in the world. In fact, it’s barely in the top five by most measures, but it’s not as if you can’t live there or that it’s hard to live there, or people are struggling. It’s still a very modern, very wealthy economy. So, not being first is not the end of the world.
Other articles we looked at this week complemented that a little bit. One of them is by Fareed Zakaria, and one of his co-writers is Heman Taneja, and they talk about AI and the new digital cold war.
They focus on competition between the US and China but not at the macro level, but specifically in artificial intelligence. And they call it the digital cold war. It posits that globalization is dead, which I think Professor Galloway would kind of agree with. And that AI will become the key measure of who succeeds in a now atomized world order.
It points out that capital requirements to even compete in AI are huge. And the US and China are the only players with the scale and the ability to invest. However, he mentions India, Singapore, Japan, and South Korea as smaller players in the game.
That context – that AI is part of a global competition also comes up in the video of the week. This week has two videos, one by Andrew Keen & Ludwig Ensthaler, describing the state of AI. The other is Fareed Zakaria speaking about the “dangers” of AI.
Now, I tend to disagree on both the end of globalization and the dangers of AI.
I think that, firstly, AI isn’t scary. AI is super liberating. It allows skilled workers, especially information workers, to have a highly skilled coworker known as chatGPT that can make them more productive and probably better at what they do.
I certainly find that with my coding at SignalRank.
So, firstly, I don’t think AI is scary. Secondly, I don’t think it’s possible for us to de-globalize. Of course, nation-states can certainly try, but the factors that drive globalization are not really political factors. They’re mainly economic factors. The two have to converge for friction-free globalization, but globalization happens nonetheless. And if you look at things like flows of money across borders, travel, almost any measure of humanity’s global characteristics, they’re all growing still and very rapidly. Even Chinese money flowing outside of China is growing now that the COVID lockdown has ended.
So, de-globalization is more of a political label. It really isn’t an economic reality to any great extent. And so I really don’t agree that we’re in this de-globalized world. Because of that, I don’t really agree that AI will be a weapon of war between nations. And if that were to happen, it would be really bad for the human race.
AI is more likely to be a tool for good worldwide. And, of course, that’s what we have to want it to be. And as human beings, that’s what we have to make it.
Steven Levy has a great Wired piece from his Backchannel section called What Open AI Really Wants that shines a light on the end game for AI being humanistic, not being confrontational, and certainly not national in character.
So some big issues in this week’s That Was The Week.
I hope you enjoy it and enjoy the video that we make after I’ve written this editorial.
Essays of the Week
Published on September 8, 2023
For decades, America has predicted — arrogantly and repeatedly — the imminent fall of a nation. The doomed nation, according to Americans? A: America.
In the ’80s, we decided Japan was doing to us economically what they couldn’t do militarily four decades prior. My second year in business school (Berkeley ’92) was devoted to a forensic analysis of our loss to Japan. Computers and cars were the future, and Japan was building them faster, better, and cheaper. In 1984, Walter Mondale asked Reagan at the presidential debates, “What do we want our kids to do? Sweep up around the Japanese computers?” Three years later, Paul Kennedy wrote a book about it, The Rise and Fall of the Great Powers, comparing the U.S. empire to the British empire. Japan’s GDP soared to 40% of ours, and we feared what might happen when that number hit 100. It never did.
In the early aughts, our soft tissue was geopolitical. The tragedy of 9/11 was described as an “inevitable outcome.” Our subsequent invasion of Afghanistan inspired books including Dark Ages America: The Final Phase of Empire, and Are We Rome? The Fall of an Empire and the Fate of America. Each offered a similar theme: Our time was running out.
Today the decline is (supposedly) more imminent. January 6 was the “beginning of the end.” Russia’s invasion revealed a “great unwinding.” Nations view us “with pity,” and we are “on the brink of collapse.” Just last week, the New York Times opinion page compared us to Rome (again). These headlines are click bait, and we still take the bait: Three-quarters of Americans believe our country is in structural decline, and the song of the summer is an ode to our demise.
It’s not just the public. Among economists there’s a growing school of thought that our situation is dire. Two months ago, ratings agency Fitch downgraded America’s credit rating due to “fiscal deterioration” and “erosion of governance.” The debt ceiling debacle didn’t help: Investors “should worry.” Our debt-to-GDP ratio is hovering around 120%; back when it was 70%, Brookings called it “the real national security threat.”
Many believe we are in the midst of “dedollarization,” ceding our status as the world’s reserve currency because of our unsustainable spending habits and an overall loss of faith globally. JPMorgan recently flagged it, an ex-CIA adviser plainly predicted it, and one prominent tech investor bet $1 million on it. Ray Dalio, the founder of the world’s largest hedge fund, hinges his recent bestseller, Principles for Dealing with the Changing World Order, on America’s inability to adapt to our loss of status and power. Empires win and lose their hegemony depending on their reserve currency status, and in Ray’s view, we’re near freefall.
Culturally you could build a compelling case. National pride is at an all-time low, and the “vibe” in America is that things are not good. (See above: the song.) But these are functions of perception, and as I’ve written before, human perception is flawed.
This is an economic discussion. And when you look at the data, you’ll find every diagnosis of our supposedly terminal illness is proof the doctor is jonesing for us to die. We’ll examine them, but first, a brief summary of the doomer’s economic vision for America. It goes roughly as follows:
1) America keeps borrowing more money, leading to an increased burden of interest payments.
2) Foreign nations increasingly question our ability to make good on our debts, leading to low demand for U.S. Treasuries and thus low demand for dollars.
3) Once dedollarization takes effect and the dollar is supplanted as the world’s reserve currency, the U.S. will be forced to ratchet up Treasury yields to increase demand for our debt, leading to even greater interest payments.
4) This will crowd out private investment, as well as public investment in our own infrastructure. GDP growth will grind to a halt, and eventually we’ll default on our debt.
5) At that point we’ll be unable to borrow or finance our growth, and
6) America will collapse.
Doomsday is due next century or next year, depending on who you ask. (That ex-CIA adviser said it’d be last month.) But the catalysts are consistent, and one of them is this notion of dedollarization.
The argument is that foreign central banks are losing interest in the dollar. The stat dedollarists point to is that the dollar has fallen from 70% share of the world’s currency reserves to 60% in the past 20 years. That may sound significant, but the scope is comically small. When you zoom out you find that in the ’80s our share was 50%, and 30 years before it was 40%. The only accurate description of the dollar’s reserve status over the past 75 years is … unwaveringly dominant. At 20%, the next-best option (the euro) is not within striking distance.
However, it’s not the euro dedollarists are talking about, but the currencies of ascendant nations, including China. A common headline is “Yuan’s share of global reserves hits record high.” Less common is any mention of that “high”: 2.6%. The president of Brazil made headlines recently calling on the BRICS nations (Brazil, Russia, India, China, South Africa) to join forces to create a new global reserve currency. The world’s reaction was lackluster — and even South Africa’s own central bank governor played it down: “If you want it, you’ll have to get a banking union, you’ll have to get a fiscal union, you’ve got to get macroeconomic convergence.” Translation: pipe dream.
Another pipe the hallucinations flow through is bitcoin. Among its many use cases is its potential to supplant the dollar. The argument: The value of the dollar is predicated on faith in the U.S. government; the value of bitcoin is predicated on faith in, well, bitcoin. Many bitcoin bulls argued the latter will ultimately win — and for a second there in 2021, it looked feasible. As with every other dollar competitor, though, the cryptocurrency ran out of steam. Today the total value of bitcoin is 12 times smaller than the amount of dollars held in global central bank reserves. So next time someone tells you the dollar will be replaced, ask: With what? By any metric, the most innovative payment platform or store of value has been, and remains, USD. More good news: Minting dollars doesn’t require the energy consumption of Argentina. But I digress.
Theory of Relativity
Arguments for America’s decline are rarely accompanied by a credible alternative. This is true of the dollar, and it’s also true of U.S. debt.
Take Fitch’s downgrade of our national credit rating, for example. What should have been a seismic shock to the global bond market by the premier ratings agency turned out to be a catastrophist headline. The bond market barely registered the news, with the 10-year Treasury yield inching up 4 basis points. Goldman put it deftly: “We do not believe there are any meaningful holders of Treasury securities who will be forced to sell due to a downgrade.” Jamie Dimon put it better: The downgrade was “ridiculous.”
As with currencies, creditworthiness is relative. The question creditors should ask isn’t “how likely am I to get my money back,” it’s “who’s more likely to give me my money back?” And when it comes to sovereign debt, there is no better option than the United States. Sure, Xi Jinping may make Biden look like a web browser with 19 tabs open, not knowing where the music is coming from — but the Chinese Communist Party also systematically withholds, suspends, and lies about the nation’s economic data. The Party has even ordered its own economists to stop talking about negative trends. Who would you rather lend to?
This extends beyond national debt. There are several linchpin data points declinists point to that are supposed to forecast our imminent undoing, but the people who cite them also forget to compare those metrics to those of other nations.
For example: inflation. Last year, inflation hit a 40-year high. Some predicted America would enter a period of hyperinflation. However, when you compare our situation to those of other nations, it’s less bad. Significantly less bad. In the U.S., prices have risen 3.2% year over year. In the U.K., it’s 6.8% — more than double. In the eurozone, it’s 5.3%. And yet — despite U.S. inflation continuing to come down and wage growth recently surpassing it — 74% of Americans still believe inflation is headed in the wrong direction…. Lots More
General Partner @ āltitude
In the world of venture capital, there are a couple of hidden truths that deserve attention.
Firstly, it’s a well-kept secret that a mere 𝟭𝟬% 𝗼𝗳 𝗮𝗹𝗹 𝗩𝗖 𝗳𝗶𝗿𝗺𝘀 𝘄𝗶𝘁𝗵𝗶𝗻 𝗼𝘂𝗿 𝗮𝘀𝘀𝗲𝘁 𝗰𝗹𝗮𝘀𝘀 𝗴𝗲𝗻𝘂𝗶𝗻𝗲𝗹𝘆 𝘆𝗶𝗲𝗹𝗱 𝘀𝘂𝗯𝘀𝘁𝗮𝗻𝘁𝗶𝗮𝗹 𝗿𝗲𝘁𝘂𝗿𝗻𝘀. This implies that a staggering 90% of these firms find themselves struggling and burning capital.
Secondly, our industry has consistently boasted a high-single-digit DPI. For context, 𝗮𝗻 𝗮𝘃𝗲𝗿𝗮𝗴𝗲 𝗼𝗳 𝟭.𝟳 𝘁𝗶𝗺𝗲𝘀 𝘁𝗵𝗲 𝗶𝗻𝗶𝘁𝗶𝗮𝗹 𝗶𝗻𝘃𝗲𝘀𝘁𝗺𝗲𝗻𝘁 𝗵𝗮𝘀 𝗯𝗲𝗲𝗻 𝘁𝗵𝗲 𝗻𝗼𝗿𝗺 𝗼𝘃𝗲𝗿 𝗮 𝟯𝟬-𝘆𝗲𝗮𝗿 𝗽𝗲𝗿𝗶𝗼𝗱. Ironically, despite our relatively strong DPI performance, we have been the guiltiest culprits when it comes to showcasing paper mark-ups or TVPI (hypothetical profits) to the institutional investors backing VCs.
Our industry has been able to maintain a delicate dance 𝗽𝗿𝗶𝗺𝗮𝗿𝗶𝗹𝘆 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝗼𝗳 𝗽𝗲𝗿𝘀𝗶𝘀𝘁𝗲𝗻𝘁𝗹𝘆 𝗹𝗼𝘄 𝗶𝗻𝘁𝗲𝗿𝗲𝘀𝘁 𝗿𝗮𝘁𝗲𝘀. However, as investors, we find ourselves facing challenges due to the surplus of capital flooding the market. Consequently, start-ups have struggled with management, leading to less effective operations than desired. It’s crucial to take corrective measures.
Like in many industries, what we do only makes sense if you are best in class!
September 7, 2023
When we released our recent blog post on Fund Unicorns, we looked at how many funds out of the roughly 650 submitted were “Unicorns.” We defined a Fund Unicorn as a 10x TVPI fund. After we released the analysis, one reader asked us to analyze cash-on-cash returns (aka “DPI”) as well. Here, we try to oblige.
As requested, we went back to our database of fund returns and looked at DPI. We removed all the returns data from predecessor funds, funds sub $1M, SPVs and angel track records. We looked only at blind pools.
I want to be clear on one thing: self-reported DPI data from emerging managers with less than less $200M AUM is not representative of the venture industry as a whole. After all, the data pool skews very young. With the time to IPO growing each year, it’s not reasonable to expect an early-stage fund to have much DPI until the final years of a fund’s life (8-10+ years). Given the current malaise in the public markets, we can safely assume that the time to liquidity is getting longer and longer.
If LPs wait around for DPI as a proof point, they are probably going to miss the chance to get into some great venture capital funds. So, let’s all take this data with a grain of salt. I am sure many more of the funds will generate amazing DPI during the next cycle.
Nonetheless, there are a very small number of funds with unicorn-level DPI. These funds are rare. One can guess that luck is more responsible than sustainable, repeatable skill. They had a single huge early exit.
How rare? Only two funds have a DPI over 10x.
The best performing fund (a 2012 vintage) is 26.2x DPI. This pre-seed fund invested a total of $2M, but one of their cybersecurity companies went public and currently has a market capitalization of almost $5B. This one investment generated over 2600x. Impressively, this fund has 7.2x RVPI, so those lucky investors can expect even more returns in the future. Over 11 years, this fund has generated a 47.60% IRR. This fund is truly a unicorn.
The second best blind pool (a 2014 vintage $5M fund) has generated a 10.7x DPI and has a number of remaining investments with a 1.17x RVPI for a 55.00% IRR. While it’s a little more difficult to discern their return drivers from the materials provided, this LA-based fund invested early in NerdWallet (IPO) and Honey (acquired), which I assume are the drivers.
There are 13 funds that have between a 3x and 10x DPI. Notably, all the funds that have substantial DPI are nano funds – sub $10M. So, if an LP is looking for a unicorn, think small.
Sept. 8, 2023
Some of the most active startup investors have been hanging a “for sale” sign on their portfolios at a time when venture investors are finding it increasingly difficult to raise new venture funds.
Insight Partners has considered selling a stake valued at $400 million in Left Lane Capital, a consumer-focused venture capital firm founded by a former Insight principal, according to a person with direct knowledge of the matter. Insight made more than 200 startup investments in 2021 alone, and raised a $20 billion war chest to back more startups last year.
Insight joins Tiger Global Management, Chamath Palihapitiya’s Social Capital and other VC firms looking to sell some of their stakes in startups or the VC funds that back them. And, in a change from years past, even the VC funds that specialize in the youngest companies are trying to sell some of their investments. For many funds, increased pressure from limited partners to return cash is driving these sales. Some funds are also using the cash to increase their ownership of existing investments.
• Insight Partners weighed selling part of its stake in Left Lane Capital
• Effort follows sale attempts by Tiger Global, Social Capital
• Early-stage investors are also selling positions
The selling spree comes on the heels of a nearly two-year drought in initial public offerings that cut off one of the most common ways VC investors return cash to their LPs. Even as the IPO market starts to reopen, with the expected September debuts of Arm and Instacart, VC funds are still mulling stake sales. That’s because, without cash returned to them, the pension funds and endowments that back VC funds are less likely to invest in a firm’s next fund.
“There’s a growing desire to find some way to generate distributions, because they’re not otherwise generating distributions through normal channels,” such as M&A, said Matthew Wesley, global head of private capital advisory at investment bank Jefferies. He said the firm has seen a “significant increase” in VC firms expressing interest in selling stakes on the secondary market, in part driven by improvements in the prices those stakes could fetch.
These sales efforts are taking place as VC firms encounter a plunge in investments from their LPs, the wealthy investors, endowments and pension plans that sank $328 billion into VC funds over 2021 and 2022. Investments in U.S. VC funds fell to $33 billion in the first half of this year from $122 billion in the first half of 2022 and are on track for a six-year low, according to PitchBook.
Insight, after cutting its target for its 13th fund from $20 billion to $15 billion earlier this year, has struggled to meet even the new, lower target because of cooler appetite for VC funds, according to one of the firm’s investors. As of August, it had raised only $3 billion, according to securities filings. The firm expects to close its fund in the first quarter of next year, according to the investor.
If Insight sells some of its investment in four-year old Left Lane, it would use the cash to put more into its other existing investments and would continue to be a large investor in the fund, a person familiar with the matter said.
VC investors have also become a bigger share of sellers in the secondary market, edging aside employees and other company insiders including board members and founders, according to Carta, which sells software that helps private companies manage their ownership. Investors jumped to nearly 70% of all sales in the first quarter of this year from half of all sellers in the last quarter of 2022, according to Carta.
For the most part, startup stakes are selling at a discount to their last round price, according to secondary market brokers, though New York–based Tiger Global Management last month sold a stake in artificial intelligence company Cohere at a $3 billion valuation, a 40% valuation premium to the firm’s initial investment in February last year.
Tiger has also struggled to raise capital this year. As of June, Tiger had raised only $2.7 billion, according to securities filings, or 55% below its initial $6 billion target. Tiger is still trying to sell additional startup stakes, according to a person with direct knowledge of the matter. A spokesperson for Tiger declined to comment.
In a change from past years, even investors in the youngest startups are selling positions as they respond to pressure from LPs to return capital, according to several of these investors who spoke to The Information. LPs need the cash to meet commitments to invest in new funds and often because the drop in value of public tech stocks has left them more weighted to private assets than their own guidelines allow.
“The private funding business runs on some expectation that the money will come back,” said Charles Hudson, founder of early-stage VC firm Precursor Ventures. “In a world where the IPO window for most companies is still closed, the only way you can generate liquidity for your LPs right now that is easy is through secondaries.”
He said his firm’s partners have discussed selling portions of stakes in some of its early-stage portfolio companies but have not done so yet.
In some cases, entrepreneurs are facilitating these transactions. Dan Siroker, CEO and co-founder of startup Rewind AI, said he is helping at least three of his investors sell a portion of their stakes by facilitating purchases by other investors. These potential sellers are investors who backed Rewind in earlier rounds that valued it at $30 million and $75 million and are looking to sell their stakes at the most recent $350 million valuation fetched in its Series A round or higher.
“I can understand if some of them need to sell half their stock, and that will help them improve the fund and their portfolio,” Siroker said.
Most attempts to offload assets have not resulted in sales, largely because the buyers and sellers have not agreed on price, according to venture capitalists and secondary market brokers. But buying interest has picked up in recent months. The ratio of bids to offers increased to 40% in August and September from 30% in June, according to Caplight, which tracks data from secondary brokers. The activity is driven by buyer demand for pre-IPO names like Instacart, plus sellers’ need for liquidity, according to Caplight.
Pressure to raise cash to return to LPs, known as distributions to paid-in capital, has started to overturn venture investors’ resistance to selling their stakes early.
“That desire to stick with a company all the way to exit in terms of loyalty to the founder is not really best practice at this point,” said Adam Struck, founder of seed-stage venture firm Struck Capital. The firm is planning to raise its next fund next year, and it’s eager to show LPs it can produce more than just paper gains, Struck said.
“Especially in this environment, DPI is king, so showing that distributed paid-in capital will mean a lot for subsequent funds.”
Erin Woo contributed to this article.
The greatest shortcoming of the human race is our inability to understand the exponential function.
— Albert Allen Bartlett
The biggest secret in venture capital is that the best investment in a successful fund equals or outperforms the entire rest of the fund combined.
— Peter Thiel
In the river “flow” of technological progress, one often observes trends toward:
abstraction (of information and processes)
adaptation (of one solution to different use case)
acceleration (solutions enable other solutions)
universality (good patterns can be repeated anywhere)
These trends have resulted in highly skewed power law distributions (vs. the bell curve), where a small handful of people are “hyper high performers,” a broad swath of people are “good performers”, and a smaller number of people are “low performers.”
Power law distributions
A major feature of power law distributions is that small outcomes are very likely while larger ones are less likely. In other words, a small number of inputs account for a large percentage of outputs. Power laws also exhibit “fat tails,” compared to the area under a normal distribution curve which falls off much faster as you move farther along the x-axis.
Many prominent entrepreneurs and venture capitalists assert that VC returns are distributed according to a power law. As Marc Andreessen of Andreessen Horowitz points out, each year, of the 4,000 technology startups seeking VC funding, only 200 (or 5%) are seriously fundable, with “15 of those generating 95% of all economic returns…even the top VCs write off half their deals.” Billionaire tech investor Peter Thiel concurs, “[W]e don’t live in a normal world; we live under a power law. … [I]n venture capital, where investors try to profit from exponential growth in early-stage companies, a few companies attain exponentially greater value than all others. … Bad VCs tend to think the dashed line is flat, i.e. that all companies are created equal, and some just fail, spin wheels, or grow. In reality you get a power law distribution.”
In Zero to One, Thiel further elaborates:
In 1906, economist Vilfredo Pareto discovered what became the “Pareto Principle,” or the 80–20 rule, when he noticed that 20% of the people owned 80% of the land in Italy — a phenomenon that he found just as natural as the fact that 20% of the peapods in his garden produced 80% of the peas. This extraordinarily stark pattern, when a small few radically outstrip all rivals, surrounds us everywhere in the natural and social world. The most destructive earthquakes are many times more powerful than all smaller earthquakes combined. The biggest cities dwarf all mere towns put together. And monopoly businesses capture more value than millions of undifferentiated competitors. Whatever Einstein did or didn’t say, the power law — so named because exponential equations describe severely unequal distributions — is the law of the universe. It defines our surroundings so completely that we usually don’t even see it.
What does the distribution of returns in venture fund look like? The naïve response is just to rank companies from best to worst according to their return in multiple of dollars invested. People tend to group investments into three buckets. The bad companies go to zero. The mediocre ones do maybe 1x, so you don’t lose much or gain much. And then the great companies do maybe 3–10x.
But that model misses the key insight that actual returns are incredibly skewed. The more a VC understands this skew pattern, the better the VC. Bad VCs tend to think the dashed line is flat, i.e. that all companies are created equal, and some just fail, spin wheels, or grow. In reality you get a power law distribution.
Indeed, the single most powerful pattern I have noticed is that successful people find value in unexpected places, and they do this by thinking about business from first principles instead of formulas.
VCs often write off as many as half of their investments. Historical data confirm that most investments will return very little or even lose money, many still return some multiple on initial investment, but it is the very small handful of “hyper performers” that return an outcome well outside what could be expected in a normal distribution which not only more than compensate for losses but also generate most of a portfolio’s returns. The last point exhibits the idea of “fat tails”: As the tails have more bulk, the probability of extreme events is higher compared to the normal. Portfolios are constructed around the idea that these unlikely, but high-value, outcomes will drive the returns of the portfolio, in spite of the fact that small-to-negative returns make up the highest probability outcome of companies within the portfolio.
The most experienced and successful venture capitalists grok the concept of the power law and how it impacts the outcomes of startup investments. In fact, the power law is so common that Peter Thiel considers it a crucial concept for all business people to understand.
Five insights stood out in our research:
SaaS growth rates have stabilized in the last 3 quarters. While the data doesn’t show any gain, it’s clear that growth is not deteriorating. There are pockets of optimism. Growth for best-in-class SaaS startups is accelerating and new business activity is picking up.
Best-in-class SaaS businesses grow over 100% each year. The top decile of SaaS businesses with ARR in the range of $1-3M grow at 192% annually. Those in the $3-8M ARR segment, grow at 121%. And those, in the $8-15M ARR segment grow at 110%.
Top-tier SaaS startups reach $1M ARR within
9 months. The median startup takes approximately 2 years and 9 months. On average, SaaS startups reach $10M ARR in slightly over 5 years. Even after 10 years in business, only 13% of SaaS startups are able to hit the $10M ARR mark.
The majority of SaaS startups grow from $1M to $10M ARR by growing their subscriber base.Only a small subset (<5%) of startups grow predominantly by increasing their ARPA.
SaaS is mildly seasonal. Q1 is usually the best quarter of growth, and Q4 is the slowest. March is the best month and July and December are usually slower (due to holidays).
Maelle GavetMaelle Gavet
CEO at Techstars; advocate for #EmpatheticTech and author of Trampled by Unicorns
Founders who are 50 years old and above are more likely to bring radical innovations to the market than younger business owners. “Every ten more years of age increases a founder’s likelihood to introduce a market novelty by up to 30 percent”
At Techstars we have 150+ companies who were founded by entrepreneurs who are 50yo or older. Too many to list, so here are just a few rising entrepreneurial stars you may want to look up:
– Carolyn Thompson, Aravenda Consignment Software leads SaaS innovation in the fastest-growing segment of retail – resale (currently a $100B+ market). With users in 10+ countries, Aravenda is growing 300% year over year by partnering with Shopify, Clover & Lightspeed to provide consignment resale inventory management solutions to SMB’s and Enterprise customers alike.
– Mark Kovscek, Conservation Labs is pioneering using machine learning to understand an object based on the sound it makes. For example, by listening to the flow of water, their sensor on a pipe provides high-resolution water insights, cutting water use by 10-20%. This means over $40,000 in utility savings and a 30 MT reduction in carbon emissions for typical commercial properties. They have just received Fast Company’s 2023 Innovation by Design Awards in the Artificial Intelligence category.
– Eric Levitan, Vivo addresses the loss of strength and independence that every older adult faces by providing online, live and interactive small group classes that build strength and community. Their clinical approach includes tracking outcomes for every member, which has resulted in distribution through healthcare providers, health plans, government agencies, and a $2.3m grant from the National Institute on Aging (NIA) to measure the impact of Vivo on older adults with prediabetes.
– Marc Steren, University Startups empowers under-resourced students with pathways to higher education or career success through programs focused on the entrepreneurship mindset. In the year since starting Techstars, University Startups has grown its MRR 20x, entered 6 new territories including New York City, will serve over 18,000 students this Fall (up from 2,000 last Fall), and have partnered with HBCUs such as Bowie State University and non profits such as the Urban League of Arkansas.
– Paige Wilson, Naborforce‘s tech platform matches seniors seeking independence and a helping hand to a network of retirees and empty nesters vetted and ready to lend it. With several markets profitable or nearly profitable they have raised $12M+ to date, including a $9M Series A led by Palo Alto-based Translink Capital last summer to fuel additional growth.
by Garry Tan 9/6/2023
My first YC Demo Day was in Summer of 2008, where my startup and twenty-one others anxiously pitched a room full of investors on what we’d been working so hard to build.
Fifteen years later, I’m here as YC’s president kicking off our 37th Demo Day — the one celebrating the Summer 2023 batch. I’ve seen a lot of batches in the last decade and a half, and I can say without a doubt: this is a good one. It’s one of the strongest YC batches ever; the technical talent runs deep.
We received over 24,000 applications for this batch; we funded 229. We deliberately went with very early startups for this batch. 75% of the companies presenting today started with no revenue on day one of YC. 81% had never raised a dime.
This is the first batch in over three years that took place 100% in person. Office hours, speaker sessions, meetups — we were able to be together for all of it. For the next batch, I’m happy to share that we’re also making plans for an in-person, invite-only Demo Day. Virtual Demo Day served us well for the last few years and we’re planning to continue offering a way to watch it remotely… but there truly is no room like the one that YC can bring together for Demo Day. Want to be there? Invest in YC companies.
You’ll notice AI has a big presence in the S23 class — this is no accident. Recent developments in AI have unlocked an entire universe of possibilities, presenting a resoundingly clear answer to the question of “Why now?”. There has never been a better time to start an AI company than now, and there is no better place to start an AI company than Y Combinator.
This batch brings the YC community to over 10,000 founders across more than 4,500 startups. 350+ of those startups are valued at more than $150 million; 90+ of them are valued at more than $1 billion. Over 5% of YC companies become unicorns, a percentage that still blows my mind to write. It’s unrivaled.
YC now has more full-time Group Partners than it ever has before, allowing us to give more personalized attention to those founders and their startups. YC’s Group Partners are the ones that pick the companies, work with them throughout the batch, and support the founders for the life of their company. The newest Group Partner is Tom Blomfield, co-founder of two billion-dollar startups: GoCardless and Monzo. We also recently welcomed back YC alum Wayne Crosby (co-founder of Zenter, the startup that became Google Slides) and Emmett Shear (co-founder of Twitch) as Visiting Group Partners.
For years we’ve shared what verticals each respective batch focused on. For S23, it breaks down like this:
70% in B2B SaaS / Enterprise
10% in Fintech
10% in Healthcare
6% in Consumer
4% Proptech / Industrials
The demographics* for the S23 batch:
3% Hispanic or Latino
3% Middle Eastern or North African
12% South Asian
(* self-reported and opt-in; not all founders reported demographics)
Additionally, 15% of the S23 companies have a woman founder and 10% of the founders are women.
If you’re investing in YC companies, three things I ask of you:
Do no harm — This should be a simple one, but as with anything having to do with investing, it gets complicated. If you don’t have capital to invest, if the startup is clearly out-of-thesis, or if there’s a clear conflict (such as having invested in a competitor when you don’t invest in competing companies), you shouldn’t meet with them. If you might be able to invest or help but not until later, call that out up front so founders can properly prioritize.
Make decisions quickly — Speaking from experience, I know how hard it is as an investor to talk to dozens of startups in quick succession. But this is their life, and no matter how busy you get, we ask you to treat each founder with respect and seriousness. The wrong way to do YC Demo Day is to wait until a deal gets hot. We want you to be a conviction investor: make your own call, do your own research, and get to know the founders and their spaces. If it all resonates with you, invest.
(We recommend using the handshake deal protocol so you and the founders can choose to work together in a quick and clean way. If you haven’t seen that, here it is.)
Act like long-term owners and be helpful — If you’re investing, introduce the founders to more investors. Introduce them to potential customers. YC is a community. The founders tell each other who’s most helpful, and who can help bake that 10-year “overnight” success. If you’re looking for a quick flip, this is the wrong business for you to be in.
I am so excited for this new set of founders joining the YC community. We’re thankful to our YC investor community and are excited that with the founders, we’ll be looking to create the next generation of world-changing startups.
Do you want to be a part of the next batch of YC? Apply here.
September 06, 2023
Summary. Globalization is not dead, but it is changing. The United States and China are creating two separate spheres for technology, and artificial intelligence is on the front lines of this new “Digital Cold War.” If democracies want to succeed in this new era of…more
We are entering a new world order, one marked by increased nationalism and greater geopolitical competition. While countries are not going to undo all of the global economic systems that took shape under American unipolarity for the past three decades, certain critical sectors will become decoupled in a process we have previously referred to as “re-globalization.”
Most significantly, the technology ecosystem will be split largely into two spheres dictated by the world’s two economic powerhouses, the United States and China. Other states will need to decide which sphere they want to be part of, putting pressure on the United States and China to outpace the other and establish their technological dominance. Unfolding before us is a heightened form of economic competition we understand as the “Digital Cold War.”
The Digital Cold War will be an economic war, with technological innovations increasingly determining geopolitical prowess. Artificial intelligence (AI), with its capacity to rapidly and radically transform society, will be the most decisive technology in this arena. AI feeds on information, and its most powerful use cases will emerge through its applications across public and private sectors. For the democratic world to pull ahead, companies and countries need to adopt a new approach that prioritizes collaboration and transformation over competition and disruption.
A New Era of Democratic Coordination
Everywhere we are seeing claims that “globalization is dead.” Such declarations are fundamentally misguided. The system we are moving toward is more complicated than a reversion of global interconnection. In many cases, trade is simply being rerouted rather than shut down. Most commercial industries will remain open and global, but certain critical sectors will turn inward, toward localized supply chains.
This trend began not with the United States and its policies but rather with Xi Jinping and his 2015 Made in China plan, which made clear that the Chinese government had begun prioritizing national resilience over market efficiency. Xi laid out a plan to outcompete the United States and other global powers in the commanding areas of the economy, most of all high technology. Indeed, China continues to march towards its goal of 70% self-sufficiency in critical technology by 2025. Since then, the United States has responded with its own series of ambitious measures to retain technological supremacy. This reality leads us into a bipolar high technology future.
When it comes to AI — arguably the most decisive technology in this global contestation — we are heading toward two hermetically sealed ecosystems: one that supports open systems but is also associated with democracy, privacy, and individual rights, versus one that supports state control, information-flow restriction, and politically imposed limits on openness. As much as we might hope that China’s political model will evolve and that its technology will be subject to democratic feedback, we should not be naïve: That is not the trajectory that it is on. For a future to prevail that prizes openness and individual rights, democratic nations need to be market leaders in AI. The only way to ensure this is by promoting international collaboration, especially between democracies and other defenders of the rules-based order.
In the previous era, the United States could innovate on a technology and other countries would simply adopt it. When American technology leaders made groundbreaking advancements with personal computers and the internet, they operated under the assumption that American companies could work in isolation and spread their technologies across the world in a top-down fashion. The cloud revolution amplified this further, with Amazon, Microsoft, and Google owning 65% of the global market for cloud computing. This strategy may have worked when technologies were intended for pure disruption. AI, however, is geared for societal transformation. This requires a new kind of collaboration across stakeholders.
Along with computing capabilities, the power of AI is based on the amount of aggregated data that is fed to it. This means that the United States, or any country, working in isolation with restricted data flows will fail to maximize the potential of its technology — and yet data localization policies doubled worldwide from 2017 to 2021, further obstructing cross-border collaboration. The summation of human knowledge and capabilities is not siloed in any one country or culture. Even Wikipedia’s knowledge base is only 11% in the English language. For AI to thrive in its ability to help us solve our hardest problems, we have to unlock the world’s capacity — from French nuclear scientists to Korean philosophers, from Indian researchers to Kenyan artists, and indeed to Chinese researchers who choose to leave China and work and live in the West.
Plus, the capital requirements for investments in the sector are so large at present that very few national markets are big enough to succeed with AI on their own. For example, consider investment in semiconductors — a key input into AI progress. The U.K.’s recently announced 100 million GBP AI plan and 1 billion GBP semiconductor investment pale in comparison to the U.S.’s $280 billion and the EU’s 43 billion EUR chips packages — and even these packages are constrained compared to the scale of the investment required to fully develop these technologies. Certainly, few investors globally can support fundraising rounds for startups like the one-year-old Inflection AI’s $1.3 billion round.
At the center of this all, a disjointed approach with various patchworks of regulatory frameworks across the West will harm any ability to compete and win against Chinese AI systems. It is not just a function of population, but a function of data points. Chinese society, with a population more than four times larger than the U.S.’s, has become so digitized with data freely shared between its government and domestic technology champions. Soon their models, based largely on American and foreign research, can leapfrog the capabilities of those in the West.
With its size, centralized government, and inroads in other countries, China has the potential to develop one comprehensive AI model that outcompetes the multiplicity of models coming out of democratic nations if there is no international coordination. While the United States is still the AI leader — with the companies that are doing the most cutting-edge research — and while AI seems to have a bias toward open systems with unfettered access to information, the United States risks losing its edge if it fails to coalesce around a uniform strategy with other democratic nations. If this were to happen, Chinese companies could bring their technology to Western markets, influencing democratic politics and signaling to the world China’s economic dominance — and its broader colonization of digital infrastructure around the world.
AI is becoming an increasingly critical part of this global infrastructure, and the West must act swiftly and in a unified fashion to ensure the technology remains open and democratically controlled. To develop the most powerful AI models across sectors, the United States will need to collaborate with other allied countries — to name a few, India, Singapore, Japan, South Korea, and European nations — by adopting data-sharing policies and encouraging the co-creation of technological innovations. Much can be gleaned from the European Data Governance Act approved by the EU in 2022, which facilitates data-sharing across member countries to maximize benefits for its citizens and businesses.
Failure to course correct will severely limit the impact of AI. For models tackling climate change, siloed data is an immediate death sentence. Medical and healthcare data and innovations are not limited to any one country, nor any one research institution. Industrial AI powering global supply chains cannot be effective without the constant flow of interconnected data. In consumer applications, varying copyright frameworks would stymie cultural relevance and influence, favoring those with free access to data over others. Moreover, piecemeal data regulation and sovereignty requirements raise compliance costs and complexity, harming the innovation economy’s ability to succeed. This is not to suggest that governments should abstain from regulating AI, but rather that they should work together to establish uniform standards and practices across countries. Coordination among democratic nations will empower each country to individually feel resilient when it comes to AI, but also elevate the West as a bloc to be the leader in AI.
Responsible Innovation for AI Transformation
In addition to collaborating across states, if Western companies want to become true market leaders, they will also need to collaborate within states, namely with governmental institutions and civil society. While most of the current discourse surrounding AI focuses on large language models and other generative capabilities, AI’s most significant long-term impacts will come from the ways in which it transforms industries and society as a whole. And genuine transformation cannot come about if private actors are disconnected from the wider society.
Already we are seeing AI’s transformative potential beginning to take shape. AI has the capacity to massively level the playing field for who has access to information and insights. In the classroom, AI can give individualized attention to students who have never had access to those resources before. In the workplace, AI can free up workers from monotonous tasks like patient data entry in hospitals so they can focus on higher level problems. AI also has the ability to spot things we cannot. Take drug discovery, where AI can test millions of combinations of drugs solving for conditions we can’t yet treat. Or medical imaging, where AI can spot diseases much earlier than we have before. Or climate change, where AI can outperform existing prediction models and capture the chances of rare but deadly disasters for vulnerable populations. In defense technology, where perhaps the stakes are greatest, AI can bring clarity to the fog of war and enhance deterrence against acts of aggression. But none of this is guaranteed. Just as AI offers enormous opportunities, it also presents substantial risks.
A lot of recent attention has fixated on the existential threat AI could pose to humanity in the somewhat distant future. But AI is also destabilizing democratic systems in the here and now, providing new avenues for mis- and dis-information through sophisticated bots and realistic deep-fakes. AI will also throw into question the value of democracy itself. What happens when authoritarian systems, augmented by AI, can outperform their democratic counterparts, for instance by massively reducing crime with expanded surveillance? Or by providing vastly better healthcare because of access to centralized information with no restrictions for privacy? Undemocratic alternatives will only become more tempting as AI develops, and these possibilities are not too far off.
We are at a fork in the road when it comes to AI. We can go down the path that leads to automation and destruction, replacing human work and meaning, or we can go down the path that leads to copiloting and enablement, making us more productive, helping us live more balanced lives, and becoming greater masters of our craft. Unlike the social media revolution, which, if regulators really wanted to, they could have slowed or redirected, the AI revolution can only charge forward. Unlike previous platform revolutions, this is a technological revolution, and its mantle has already been picked up by stakeholders across society.
The most successful companies will be those that embrace this forward-looking vision and build to endure, by centering a core set of values that align with society and abide by self-regulatory mechanisms. Taking the further step of partnering with the public sector and incumbent ecosystems is also crucial. We cannot risk AI going awry and putting democracies off track in this competitive race. Since the impacts of AI will be felt across every sector of society, accounting for broad stakeholder interests is both a moral responsibility and the only way to bring about sustainable transformation. In the age of AI, companies will need to pursue an agenda of responsible innovation, working outside of the usual technological silos.
To win the Digital Cold War, the United States and its allies must be the market leaders in AI. And to build the best AI companies, they need to prioritize international collaboration and engender a new mindset — one that aims to innovate responsibly and unleash human potential.
Hemant Taneja is CEO and managing director of global VC firm General Catalyst, backers of legendary companies like Stripe, Snap, Samsara, Airbnb, Kayak and Gusto. Hemant is also a best selling author and advocate for Responsible Innovation, with his latest book Intended Consequences being named a Forbes Top Ten Tech Book of 2022.
Fareed Zakaria is the host of Fareed Zakaria GPS on CNN and a columnist for The Washington Post. He is the author of four New York Times bestselling books, Ten Lessons for a Post-Pandemic World (2020), In Defense of a Liberal Education (2015), The Post-American World (2008), and The Future of Freedom(2003).
Video of the Week
AI of the Week
The young company sent shock waves around the world when it released ChatGPT. But that was just the start. The ultimate goal: Change everything. Yes. Everything.
Steven Levy, Backchannel
THE AIR CRACKLES with an almost Beatlemaniac energy as the star and his entourage tumble into a waiting Mercedes van. They’ve just ducked out of one event and are headed to another, then another, where a frenzied mob awaits. As they careen through the streets of London—the short hop from Holborn to Bloomsbury—it’s as if they’re surfing one of civilization’s before-and-after moments. The history-making force personified inside this car has captured the attention of the world. Everyone wants a piece of it, from the students who’ve waited in line to the prime minister.
Inside the luxury van, wolfing down a salad, is the neatly coiffed 38-year-old entrepreneur Sam Altman, cofounder of OpenAI; a PR person; a security specialist; and me. Altman is unhappily sporting a blue suit with a tieless pink dress shirt as he whirlwinds through London as part of a monthlong global jaunt through 25 cities on six continents. As he gobbles his greens—no time for a sit-down lunch today—he reflects on his meeting the previous night with French president Emmanuel Macron. Pretty good guy! And very interested in artificial intelligence.
As was the prime minister of Poland. And the prime minister of Spain.
Riding with Altman, I can almost hear the ringing, ambiguous chord that opens “A Hard Day’s Night”—introducing the future. Last November, when OpenAI let loose its monster hit, ChatGPT, it triggered a tech explosion not seen since the internet burst into our lives. Suddenly the Turing test was history, search engines were endangered species, and no college essay could ever be trusted. No job was safe. No scientific problem was immutable.
Altman didn’t do the research, train the neural net, or code the interface of ChatGPT and its more precocious sibling, GPT-4. But as CEO—and a dreamer/doer type who’s like a younger version of his cofounder Elon Musk, without the baggage—one news article after another has used his photo as the visual symbol of humanity’s new challenge. At least those that haven’t led with an eye-popping image generated by OpenAI’s visual AI product, Dall-E. He is the oracle of the moment, the figure that people want to consult first on how AI might usher in a golden age, or consign humans to irrelevance, or worse.
Altman’s van whisks him to four appearances that sunny day in May. The first is stealthy, an off-the-record session with the Round Table, a group of government, academia, and industry types. Organized at the last minute, it’s on the second floor of a pub called the Somers Town Coffee House. Under a glowering portrait of brewmaster Charles Wells (1842–1914), Altman fields the same questions he gets from almost every audience. Will AI kill us? Can it be regulated? What about China? He answers every one in detail, while stealing glances at his phone. After that, he does a fireside chat at the posh Londoner Hotel in front of 600 members of the Oxford Guild. From there it’s on to a basement conference room where he answers more technical questions from about 100 entrepreneurs and engineers. Now he’s almost late to a mid-afternoon onstage talk at University College London. He and his group pull up at a loading zone and are ushered through a series of winding corridors, like the Steadicam shot in Goodfellas. As we walk, the moderator hurriedly tells Altman what he’ll ask. When Altman pops on stage, the auditorium—packed with rapturous academics, geeks, and journalists—erupts.
Altman is not a natural publicity seeker. I once spoke to him right after The New Yorker ran a long profile of him. “Too much about me,” he said. But at University College, after the formal program, he wades into the scrum of people who have surged to the foot of the stage. His aides try to maneuver themselves between Altman and the throng, but he shrugs them off. He takes one question after another, each time intently staring at the face of the interlocutor as if he’s hearing the query for the first time. Everyone wants a selfie. After 20 minutes, he finally allows his team to pull him out. Then he’s off to meet with UK prime minister Rishi Sunak.
Maybe one day, when robots write our history, they will cite Altman’s world tour as a milestone in the year when everyone, all at once, started to make their own personal reckoning with the singularity. Or then again, maybe whoever writes the history of this moment will see it as a time when a quietly compelling CEO with a paradigm-busting technology made an attempt to inject a very peculiar worldview into the global mindstream—from an unmarked four-story headquarters in San Francisco’s Mission District to the entire world.
For Altman and his company, ChatGPT and GPT-4 are merely stepping stones along the way to achieving a simple and seismic mission, one these technologists may as well have branded on their flesh. That mission is to build artificial general intelligence—a concept that’s so far been grounded more in science fiction than science—and to make it safe for humanity. The people who work at OpenAI are fanatical in their pursuit of that goal. (Though, as any number of conversations in the office café will confirm, the “build AGI” bit of the mission seems to offer up more raw excitement to its researchers than the “make it safe” bit.) These are people who do not shy from casually using the term “super-intelligence.” They assumethat AI’s trajectory will surpass whatever peak biology can attain. The company’s financial documents even stipulate a kind of exit contingency for when AI wipes away our whole economic system.
It’s not fair to call OpenAI a cult, but when I asked several of the company’s top brass if someone could comfortably work there if they didn’t believe AGI was truly coming—and that its arrival would mark one of the greatest moments in human history—most executives didn’t think so. Why would a nonbeliever want to work here? they wondered. The assumption is that the workforce—now at approximately 500, though it might have grown since you began reading this paragraph—has self-selected to include only the faithful. At the very least, as Altman puts it, once you get hired, it seems inevitable that you’ll be drawn into the spell.
At the same time, OpenAI is not the company it once was. It was founded as a purely nonprofit research operation, but today most of its employees technically work for a profit-making entity that is reportedly valued at almost $30 billion. Altman and his team now face the pressure to deliver a revolution in every product cycle, in a way that satisfies the commercial demands of investors and keeps ahead in a fiercely competitive landscape. All while hewing to a quasi-messianic mission to elevate humanity rather than exterminate it.
That kind of pressure—not to mention the unforgiving attention of the entire world—can be a debilitating force. The Beatles set off colossal waves of cultural change, but they anchored their revolution for only so long: Six years after chiming that unforgettable chord they weren’t even a band anymore. The maelstrom OpenAI has unleashed will almost certainly be far bigger. But the leaders of OpenAI swear they’ll stay the course. All they want to do, they say, is build computers smart enough and safe enough to end history, thrusting humanity into an era of unimaginable bounty….more
Most people who have experience working with large language models such as Google’s Bard or OpenAI’s ChatGPT have worked with an LLM that is general, and not industry-specific. But as time has gone on, many industries have realized the power of these models. In turn, they’ve come to understand that if they were fine-tuned to their industry, these models could be invaluable. This is why over the last few months multiple examples of domain/industry-specific LLMs have gone live.
Let’s take a look at a few different examples of domain-specific large language models, how said industry is using them, and why they’re making a difference.
Imagine an LLM that can absorb the insane amount of legal documents produced thus far by our justice system and then it turns around to assist lawyers with citing cases and more. Well, that’s what CaseHOLD does. CaseHOLD is a new dataset for legal NLP tasks. It consists of over 53,000 multiple-choice questions, each of which asks to identify the relevant holding of a cited case, which is the legal principle that the cited case establishes. CaseHOLD is a challenging task, as the correct answer is often not explicitly stated in the cited case.
The CaseHOLD dataset was created to address the lack of large-scale, domain-specific datasets for legal NLP. The dataset is a valuable resource for researchers working on legal NLP as it is the first large-scale, domain-specific dataset for this task. The dataset is also challenging, which makes it a good way to evaluate the performance of new NLP models.
Biomedical text mining is becoming increasingly important as the number of biomedical documents rapidly grows. Just using standard NLP models for biomedical text mining often yields unsatisfactory results due to the different word distributions between general and biomedical corpora.
This is where BioBERT comes in. BioBERT is a domain-specific language representation model that is pre-trained on a large corpus of biomedical text. Based on the BERT model, it has been fine-tuned on a dataset of biomedical text. This allows BioBERT to learn the unique features of biomedical text, which helps it perform better on biomedical text mining tasks.
If there is one industry that most would first think of as benefiting from a domain-specific LLM, finance would be at the top of the list. And already, BloombergGPT is causing waves within the industry. So what does it do? Well this LLM is specifically trained on a wide range of financial data. It is a 50 billion-parameter model, which means that it has been trained on a massive dataset of text and code; allowing BloombergGPT to learn the unique features of financial language, which helps it to perform better on financial tasks than LLMs that are not specialized for this domain.
BloombergGPT can perform a variety of financial tasks, including sentiment analysis, named entity recognition, and question answering. It has also been shown to perform well on general LLM benchmarks, which suggests that it is a powerful language model that can be used for a variety of tasks.
As LLM models have become more popular, a new community committed to open-source research and development has sprung forth, and with it, StarCoder was born. StarCoder is an LLM that looks to automate some of the more repetitive tasks associated with coding. StarCoder was trained on a dataset of 1 trillion tokens sourced from The Stack, which is a large collection of permissively licensed GitHub repositories. The Stack dataset includes code from a variety of programming languages, which allows StarCoder to learn the unique features of each language. StarCoder was also fine-tuned on a dataset of 35B Python tokens, which helps it perform well on Python tasks.
Because of that, StarCoder is massive, to say the least. With 15.5B parameters and an 8K context length, which means that it has been trained on a massive dataset of text and code. This allows StarCoder to learn the unique features of code language, which helps it to perform better on code-related tasks than LLMs that are not specialized for this domain.
Like law, the medical field is drowning in paperwork and data. This is where Google AI’s Med-PaLM comes in. What makes Med-PaLM special is that it is trained on a massive dataset of medical text and code, which allows it to learn the unique features of medical language. Because of this, it has been shown to outperform existing models on a variety of medical tasks, including answering medical questions, summarizing medical text, generating medical reports, identifying medical entities, and predicting clinical outcomes.
Though still not officially released, tests have shown that Med-PaLM can be used to help doctors diagnose diseases, develop new treatments, personalized care for patients, improve patient education, and make healthcare more efficient. Med-PaLM is still under development, but it has the potential to revolutionize the way that healthcare is delivered.
But if there is one domain many may not think of when it comes to LLMs, it’s climate. But if we’ve learned anything, climate science and all the data produced by researchers could also benefit from LLMs. Part of the BERT family of models, ClimateBERT is specifically trained on climate-related text. It is a transformer-based model that is further pretrained on over 2 million paragraphs of climate-related texts, crawled from various sources such as common news, research articles, and climate reporting of companies.
Currently, ClimateBERT has been shown to outperform existing models on a variety of climate-related tasks, such as text classification, sentiment analysis, and fact-checking. It has also been shown to improve the performance of other NLP models when they are fine-tuned on ClimateBERT.
Clearly, large language models, when geared toward specific industries/domains, can unlock even more benefits for those who are willing to take the time and learn this new technology. But, because LLMs are part of the fast-moving NLP ecosystem, standards, ideas, and even methods are quickly changing.
By Stephanie Palazzolo Sept. 5, 2023
Tomorrow, dozens of startups will take to the virtual stage to pitch their ideas as part of Y Combinator’s iconic Demo Day, though some well-known investors have already gotten an early look at them.
This year, more than 60% of the group—or 134 startups—are building applications or tools around artificial intelligence, mostly related to large-language models, according to YC’s startup directory. In past YC groups, the percentage of startups bucketed under “AI” has trended below 20%. The huge increase demonstrates the recent fervor from founders and investors alike for this hyped-up industry.
Much of the group reflects the most recent trends. For example, six of the companies are tweaking large-language models, including open-source ones like Meta Platforms’ Llama 2, to suit the needs of enterprises or workers. Others are tackling the compute shortage by helping AI startups work across multiple cloud providers. Many are trying to stand out by catering to specific industries like healthcare or other under-the-radar use cases such as regulatory compliance.
A bunch of the companies might be preparing to pivot, thanks to OpenAI’srecent expansion of ChatGPT’s features for enterprises that could make the startups’ nascent products obsolete. But there’s still a lot to be excited about, so let’s get into it.
The YC group has over a dozen coding-related AI startups. Code generation has been an especially popular application for generative AI, as we discussed in this newsletter from earlier this month. It’s based in language, LLMs’ specialty, and kept to a contained environment where developers can quickly get feedback by running the suggested code. And in many ways, the buzz around this can be explained by the fact that developers are most excited to build the tools they wish they had, Elad Gil, a longtime investor in YC graduates, told me.
Other startups in this batch, though, are moving past the most obvious applications. One of the companies, Sweep, helps developers with mundane tasks like code testing, documentation and refactoring, the process of cleaning up existing code so it can perform better. With giants in the space like OpenAI and Microsoft-owned GitHub going after code autocomplete, it may be a smart move for these startups to pursue other, less popular areas that still help developers save time.
Another common appearance: vertical-specific AI applications. That in and of itself isn’t that surprising—after all, industries such as law, healthcare and finance have been popular targets for well-funded darlings like EvenUp, Hippocratic AI and Hebbia, respectively. What was surprising, though, was the sheer range of verticals companies tackled (did you know that construction and shipping companies get their own specialized AI copilots now?) Startups, as is often the case, are going after typically overlooked industries that lack AI incumbents.
“As you specialize, you build more of a moat and a unique advantage versus some other horizontal solutions that lack a perspective on what matters” for a specific industry, Emergence Capital principal Yaz El-Baba told me. There’s a clear precedent for this: Look at life-sciences software developer Veeva Systems, with a $30 billion-plus market cap, and the nearly-$10 billion construction software provider Procore.
Founders in the latest YC batch are also paying close attention to open-source models coming out of Meta. A number of them built products that give an existing model, such as Llama 2, additional data so it can learn new information or a new task. These would be crucial capabilities needed for companies looking to customize cheap open-source models rather than pay up for proprietary models from OpenAI or Anthropic. One startup, Automorphic, even promises to tweak, or fine-tune, LLMs using just 10 sets of example model inputs and outputs. (OpenAI currently recommends 50 to 100 examples, though it says it can finetune with 10 examples as well. Another cost-effective option for startups wanting to improve model performance, especially in answering questions about company-specific data, is retrieval augmented generation, an approach that we dug into last week.)
One area I expected to see more activity is security for AI models. I could only find two startups—Kobalt Labs and Deasie—working on it. As reports of jailbreaking and other threats to LLMs rise, it seems like every VC I’ve chatted with in the past month is looking for startups that will protect models. Among their key vulnerabilities: Prompt injection, in which an LLM is tricked into following instructions from a malicious user, and data loss, in which employees inadvertently provide LLMs with sensitive internal information that the models could then leak to other companies. Security startups are still experimenting with different approaches and it may take some time before a leader emerges.
In the weeks leading up to Demo Day, there’s been plenty of VC gossip on X, formerly known as Twitter, about valuations in the latest YC batch. One investor told me that they typically saw post-investment valuations ranging from $15 million to $25 million for AI startups in the latest batch, with some especially strong candidates landing in the $30 million to $40 million range. That’s not out of the norm compared to previous years, which have sometimes seen post-investment prices of up to $75 million, VCs said. Though valuations remain elevated for AI companies, the first investor said, VCs are starting to be a bit more skeptical about overpaying for young startups without much traction. Many seem keen to avoid the kind of FOMO that led to unicorn valuations for earlier darlings such as Jasper and Stability AI that have since lost steam.
“This LLM tax isn’t flying high everywhere anymore,” the VC explained. “You have to have a significant advantage or team to pull off large [deals].”
Wednesday September 6, 2023 11:28 am PDT by Juli Clover
Apple has significantly ramped up its spending on artificial intelligence, according to a new report from The Information that highlights Apple’s AI and machine learning research.
Though Apple’s AI chief John Giannandrea is said to be skeptical of AI chatbots, he established a team that is working on conversational AI four years ago. We have heard prior rumors about “Apple GPT” from Bloomberg‘s Mark Gurman. Gurman in July said that Apple was experimenting with large language models, and some Apple employees have access to an “Ajax” internal chatbot.
With the 2022 debut of OpenAI’s ChatGPT, chatbots suddenly became the must-have feature. Microsoft and Google have both launched chatbots, but there are so far no signs that Apple has a consumer-oriented product launching in the near future.
Apple’s “Foundational Models” team that works on conversational AI includes just 16 people, but Apple is spending millions of dollars per day training its language models. Training large language models requires a lot of hardware, and as an example, OpenAI Sam Altman said the company spent more than $100 million for GPT-4.
According to The Information, Apple has other AI goals. The company is aiming to develop a feature that would allow a voice assistant like Siri to automate multi-step tasks. That functionality is available on the iPhone today, but workflows must be manually set up using the Shortcuts app.
The Siri team could have multi-step voice-controlled automation ready for use in iOS 18.
Apple also appears to have AI teams that are working on software to generate videos and images and multimodal AI that works with images, video, and text. The aforementioned Ajax chatbot that Apple is working with is supposedly more capable than the original ChatGPT 3.5 and has been trained on 200 billion parameters, but OpenAI’s newer models are more powerful.
News Of the Week
August 28, 2023
European venture funding halved in the second quarter of 2023 compared to a year earlier, and was down two-thirds from the peak two years ago, Crunchbase data shows.
Funding totaled $12.4 billion in Europe in Q2 2023 — flat quarter over quarter and down 50% from the $24.7 billion invested in Q2 2022.
The quarterly decline is in line with the funding reset in North America, where funding halved year over year for the past quarter, but Europe’s startup ecosystem is younger — and arguably more vulnerable.
“I think we are more sensitive to a downturn,” said Yaron Valler of Berlin-based investor Target Global, speaking on the state of European venture investment. Valler noted that Germany only became a viable startup ecosystem around 2008. “If there is a persistent downturn and sources of funding dry out, this can have an effect for decades afterwards.”
Around 18% of global venture capital was invested in Europe-based startups in the second quarter, according to an analysis of Crunchbase data. The U.K. maintained its lead as the largest market for venture capital. Germany was not too far behind, and France ranked as the third-largest funding environment.
European AI companies raised $1.5 billion in Q2 2023, representing 12% of the continent’s funding.
Large rounds were raised by AI applications development platform Builder.ai, intelligence platform for banking and insurance Quantexa, and AI-powered video platform Synthesia — all London-based companies. Heidelberg-based Aleph Alpha, which helps government and private enterprise implement AI technologies, also raised significant funding.
Late-stage funding has declined each quarter since the second quarter of 2022.
That funding stage saw the largest decline when compared to earlier funding stages in the past quarter, falling 64% year over year as global-leading late-stage investors dialed back on leading private company financings.
Early-stage funding declined by a third compared to the same time frame a year ago. Funding has settled around the $5 million mark for the past four quarters with funding up quarter over quarter at $5.9 billion.
Seed-stage funding was flat quarter over quarter, just shy of $2 billion down from $2.8 billion in Q2 2022.
While not as strong as it was in 2021, “the seed market is still really vibrant” with a large volume of dollars invested, said Hussein Kani of Hoxton Ventures, a London-based early-stage investment firm.
A few newer funds have been launched in recent years, he noted, and Hoxton is seeing competitive deals at the seed stage, with startups often receiving multiple terms sheets from investors.
Thousands of Airbnbs and other short-term rentals are expected to disappear from rental platforms as New York City begins enforcing tight restrictions.
HOUSANDS OF AIRBNBS and short-term rentals are about to be wiped off the map in New York City.
Local Law 18, which came into force Tuesday, is so strict it doesn’t just limit how Airbnb operates in the city—it almost bans it entirely for many guests and hosts. From now on, all short-term rental hosts in New York must register with the city, and only those who live in the place they’re renting—and are present when someone is staying—can qualify. And people can only have two guests.
Gone are the days of sleek downtown apartments outfitted for bachelorette parties, cozy two- and three-bedroom apartments near museums for families, and even the option for people to rent out their apartment on weekends when they’re away. While Airbnb, Vrbo, and others can continue to operate in New York, the new rules are so tight that Airbnb sees it as a “de facto ban” on its business.
Short-term rentals can bring noise, trash, and danger, and they can price local residents out of their own neighborhoods. Some landlords in New York are prolific and have hundreds of Airbnb listings. But other New Yorkers who have listings on Airbnb are trying to make ends meet, either leasing their place while they’re out of town or renting half of a duplex to help cover their mortgage costs.
Airbnb is also popular with some of the 66 million visitors a year looking for accommodations that are cheaper and sometimes larger than hotels. In 2022 alone, short-term rental listings made $85 million in New York. The city might be a relatively small slice of Airbnb’s global market, but the new rules show how local governments can effectively stamp out short-term rentals overnight and lessen their impact on dense residential areas. And New York is just one of many cities around the world trying to calm the short-term rental gold-rush.
And everyone is taking a different approach. Dallas has limited short-term rentals to specific neighborhoods to avoid disruptive and dangerous parties. Elsewhere, the Canadian province of Quebec and Memphis, Tennessee, among others, now require licenses for short-term rentals. In San Francisco, the amount of time someone can list their entire residence for rent on Airbnb is limited to 90 days each year; Amsterdam puts that limit at 30 nights per year, Paris at 120 days. Berlin previously banned nearly all Airbnbs but walked the decision back in 2018.
Airbnb’s attempts to fight back against the new law have, to date, been unsuccessful. The company sued New York City in June, but a judge dismissed the case in August, ruling that the restrictions were “entirely rational.” Airbnb did not comment on whether it would appeal the decision. Hosts are also fighting for the right to list their apartments as short-term stays by meeting with city officials to try to change the law….more
The prospect of podcast riches led to aggressive investments in celebrity deals and original programming for what appeared a sky’s-the-limit media frontier
Spotify spent more than $1 billion to build a podcasting empire. It struck splashy deals with Kim Kardashian, the Obamas and Prince Harry and Meghan Markle. It paid $286 million for a pair of podcast studios and spent $250,000 and more an episode on exclusive shows to lure new listeners.
The bet hasn’t paid off.
Most of its shows aren’t profitable, according to people familiar with the matter, and the company has recently cut staff and programming to slow its losses. The company, which has struggled to report consistent profits, lost €527 million, equivalent to about $565 million, in the six months ending in June, on €6.2 billion in revenue.
No one in the business is making much money on podcasts, but Spotify, which has spent far more on the medium than its rivals, has more to lose than most. Spotify’s competitors, including Amazon,Apple and Google, tech behemoths with their own audiostreaming services, have other, more profitable businesses.
Podcast revenue in the U.S. is expected to reach $2.3 billion this year, a 25% increase from 2022, according to the Interactive Advertising Bureau, an industry group, and is expected to more than double by 2025. That represents a tiny slice of the $200 billion digital-ad market. Spotify spent its way to the top of an industry that turned out to be less lucrative than it appeared when it began its podcast quest in 2018.
“The size of the bet up against the size of the market just seems irrational in retrospect,” Evan Shapiro, a media consultant and producer, said of Spotify’s podcast investment. “They’re out of runway.”
The pool of podcast listeners is growing, but the flood of shows on various streaming platforms makes it tough to break new hits. Facing competition across genres and formats, Spotify found that exclusive podcasts generally don’t draw subscribers away from its rivals. Podcast costs at the company rose €29 million in the first half of this year.
The company, which had 220 million paid subscribers to its premium service in June, said it has more than 100 million podcast listeners on its platform, 10 times what it had in 2019. Spotify said it is on track to make its podcast business profitable in 2024.
Tuning InChange in number of U.S. podcast listenersfrom a year earlierSource: Insider Intelligence eMarketerNote: 2023 and 2024 data are estimates
Under pressure from investors to meet that profitability goal, the company in June laid off 200 employees, about 2% of its workforce, and culled shows to focus on a more limited stable of original and exclusive content. It did away with its Parcast and Gimlet brands, consolidating its original work into a unit called Spotify Studios. In July, Spotify raised the price of core subscriptions by a dollar to $10.99 a month.
Spotify has started to share more of the risk with its talent. The company recently agreed to pay comedian Trevor Noah $4 million in a deal that allows the company to collect revenue from the podcast to cover its investment, according to people familiar with the matter. After that, both sides share the take. Spotify intends to pursue similar deals, executives said.
Sirius, iHeart, NPR and other podcast rivals have gone through their own podcast-related layoffs and budget cuts, part of a larger cost-trimming trend among media and technology companies this year.
Chief Executive Daniel Ek has said he wants Spotify to be the world’s largest audio company, spanning audiobooks, education, sports and news. Podcasts are only the first step toward Spotify’s goal of evolving from a music-streaming company to an audio giant, generating $100 billion in revenue by 2030. Spotify reported €11.7 billion in revenue in 2022.
While the company probably overpaid for some content, Ek said, the investments helped Spotify achieve its goal of becoming the top podcast platform. The company expects podcast ad revenue to grow 30% this year, ahead of Spotify’s overall revenue growth, executives have told staff in recent months.
“We’ve been very focused on pursuing shows that drive really loyal audiences and also attract advertisers,” said Sahar Elhabashi, head of Spotify’s podcast business. “We have a very strong portfolio now which does that.” ….more
Ministers will not immediately enforce online safety bill powers to scan apps after WhatsApp threatened shutdown
SEPTEMBER 6 2023
The UK government has conceded it will not use controversial powers in the online safety bill to scan messaging apps for harmful content until it is “technically feasible” to do so, postponing measures that critics say threaten users’ privacy.
In a statement to the House of Lords on Wednesday afternoon, junior arts and heritage minister Lord Stephen Parkinson sought to mark an eleventh-hour effort to end a stand-off with tech companies, including WhatsApp, that have threatened to pull their services from the UK over what they claimed was an intolerable threat to millions of users’ privacy and security.
Parkinson said that Ofcom, the tech regulator, would only require companies to scan their networks when a technology was developed that was capable of doing so. Many security experts believe it could be years before any such technology is developed, if ever.
“A notice can only be issued where technically feasible and where technology has been accredited as meeting minimum standards of accuracy in detecting only child sexual abuse and exploitation content,” he said.
The online safety bill, which has been in development for several years and is now in its final stages in parliament, is one of the toughest attempts by any government to make Big Tech companies responsible for the content that is shared on their networks.
Social media platforms have railed against provisions in the bill that would allow the UK regulator to force them to allow their encrypted messages to be monitored for harmful content, including child sexual exploitation material.
WhatsApp, owned by Facebook’s parent Meta, and Signal, another popular encrypted messaging app, are among those that have threatened to exit the UK market should they be ordered to weaken encryption, a widely used security technology that allows only the sender and recipient of messages to view a message’s contents.
Meredith Whittaker, the president of Signal, described the government’s move as “a victory, not a defeat” for the tech companies.
“Of course, this isn’t a total victory,” she wrote on X, formerly known as Twitter. “We would have loved to see this in the text of the law itself. But this is nonetheless huge, and insofar as the guidance for implementation will have the force to shape Ofcom’s implementation framework, this is, again, very big and very good.”
Will Cathcart, head of WhatsApp, said the company “remains vigilant against threats” to its encryption. He posted on X: “The fact remains that scanning everyone’s messages would destroy privacy as we know it. That was as true last year as it is today.”
Officials have privately acknowledged to tech companies that there is no current technology able to scan end-to-end encrypted messages that would not also undermine users’ privacy, according to several people briefed on the government’s thinking…more
Wednesday September 6, 2023 12:49 am PDT by Tim Hardwick
Apple has signed a new deal with British chip design company Arm to license its chip technology that extends beyond 2040, reports Reuters.
News of the deal emerged in documents filed on Tuesday for Arm’s initial public offering, which the company has priced at $52 billion.
“We have entered into a new long-term agreement with Apple that extends beyond 2040, continuing our longstanding relationship of collaboration with Apple and Apple’s access to the Arm architecture,” said Arm in the IPO document.
The document reveals that companies including Apple, AMD, Google, Intel, Nvidia, Samsung, and TSMC, have “indicated an interest” in buying “up to an aggregate” of $735 million in Arm shares. TSMC, the world’s largest contract chipmaker, has said it will decide this week whether to invest in the chip designer. By holding Arm’s shares, chipmakers will hope to have sway over Arm’s management.
Japan-based SoftBank has been preparing for an IPO since its plan to sell Arm to Nvidia became subject to regulatory scrutiny. California-based Nvidia in January 2022 abandonedthe purchase when it became clear that the deal would be blocked by the FTC.
The relationship between Apple and Arm is one of the longest in the chip business – Apple was one of the first companies to partner with the firm when it was founded in 1990, prior to the release of Apple’s Newton handheld computer, which used an Arm-based chip.
Startup of the Week
‘He Doesn’t Need VC in His Life’: How Midjourney’s Founder Built an AI Winner While Rejecting Venture Capital
David Holz has become an ‘AI celebrity’ by generating over $200 million in revenues with just 40 employees—all without raising a cent from outside investors.
By Kate Clark Sept. 5, 2023
Ever since David Holz founded Midjourney in mid-2021, venture capitalists have been practically begging him to take their money.
On top of cold-calling and emailing Holz incessantly, some have asked his inner circle for “warm” introductions. Others have preemptively sent him term sheets. A lucky few have landed meetings with the 35-year-old CEO of Midjourney, which uses machine-learning models to generate uncanny, hyperrealistic images. Index Ventures partner Mike Volpi scored a dinner with Holz at rustic Italian cafe Cotogna in San Francisco last year, “but he was pretty straightforward in his intention to not raise money,” said Volpi.
By adamantly (if politely) rejecting the advances of venture capitalists, Holz has bucked the tide among artificial intelligence startups, the top tier of which has raised more than $17 billion in fresh capital in recent years, according to The Information’s Generative AI database. Holz has also succeeded in making Midjourney a source of intense intrigue in VC land, where the company’s profitability and astronomical growth have enticed investors.
The company, which charges users between $10 and $120 for a monthly subscription to its image generator, is on pace to surpass $200 million in revenue this year, according to people close to the company. Details about Midjourney’s financials have not been previously reported.
Although a large chunk of the revenue goes right back into purchasing the pricey AI chips required to train and run Midjourney’s machine-learning models, the company has been profitable since early on, according to several people close to the company. Holz’s ability to rake in hundreds of millions in revenue in under two years and with only about 40 employees puts him in a rare class of entrepreneur. “To put it charitably, he doesn’t need VC in his life,” said Michael Stewart, a partner at Microsoft’s venture fund, M12.
Holz began to sour on venture capital after his first startup, hand-tracking sensor company Leap Motion, raised more than $100 million from investors including Andreessen Horowitz and Founders Fund, only to sell for a disappointing $30 million in 2019. His decision to bootstrap his second act has not diminished venture capitalists’ ardor. Firms like Greylock Partners, Sequoia Capital, Andreessen Horowitz, Index Ventures and Spark Capital have all been angling to grab a stake in the company, according to sources with knowledge of the firms’ activities.
Some in VC admit to being puzzled by Holz’s reticence and have privately voiced concerns about Midjourney’s ability to compete with larger companies offering text-to-image AI products, like Adobe and OpenAI. They have also questioned one of Holz’s most consequential moves—launching Midjourney on Discord, a group chat app that can be complicated to learn how to use. In return for building its business atop the platform, Midjourney shares a slice of its revenues with Discord, a much larger company with 1,000 employees and a $15 billion valuation, according to three people familiar with the matter.
Holz initially believed hosting Midjourney on its own site and building its own app would have been too costly and time-consuming for the startup. But his choice has been vindicated by the avid community that has formed around Midjourney’s Discord server. To use Midjourney on Discord, subscribers type prompts into the server’s chat box, then wait 10 or so seconds for an image to appear, with the results shared with everyone on the server.
Much more than other generative AI platforms, Midjourney’s Discord server has become a place for users to discuss and compare their art, making it feel less like a solo pursuit than a “collective,” said Nabeel Hyatt, general partner at Spark Capital and an investor in Discord. By comparison, OpenAI’s image generator Dall-E feels “utilitarian,” said Hyatt.
The numbers tell their own story: Midjourney’s server has expanded from 2 million to 14.8 million users over the last year, making it Discord’s largest server by far. By comparison, OpenAI reported in September 2022 that Dall-E had just 1.5 million active users, though its user base has likely grown significantly since then. Discord declined to make an executive available for an interview.
Forging a partnership with Discord is just one of many iconoclastic decisions Holz has made, the three people familiar with the arrangement said. From the start, the CEO was determined to run Midjourney differently from the typical Silicon Valley startup. The company has few managers and no board of directors. It keeps its teams small and gives them independence. Holz leans on four outside advisers, including AI investor and former GitHub CEO Nat Friedman, for advice. Rather than offering hefty stock packages, Midjourney gives employees a cut of profits, according to a person familiar with the matter.
In weekly office hours with customers, Holz, known for his distinctive curly red hair and philosophical manner, has admitted to feeling “bitter” about the VC-backed path he took with Leap Motion and to finding the current cash-hungry batch of AI founders uninspiring. His goal is to remain a bootstrapped company that stands the test of time—“kind of like Craigslist,” Holz told The Information in a rare interview. He wants Midjourney to be “this weird thing that no one knows how to compete with that just sort of stands alone.”
A Complicated Path
Seen a certain way, Midjourney’s story really is a lot like that of Craigslist, the 40-person classifieds website founded by Craig Newmark in 1995. As they did with Holz decades later, venture firms chased after Newmark’s business when they got wind of the vast amounts of cash it was printing in the early 2000s. (One 2005 estimate had Craigslist generating $25 million in revenue. By the start of 2020, its revenue was estimated at $1 billion with profit margins projected at 70% to 80%, according to business intelligence service the AIM Group.)
But Newmark and Craigslist CEO Jim Buckmaster had no intention of taking their company public and adhering to strict corporate governance requirements. Because they were generating ample revenue, they had no need to humor venture capitalists. “It wouldn’t serve my community any better to have a lot of money,” said Newmark last week. “I’ve always felt it was the right decision. I only feel more strongly over time.”
Like Craigslist in its early years, Midjourney is also navigating a number of legal challenges. These hurdles could threaten its core business, potentially hastening the need to accept outside money or even forcing an acquisition by a larger company. In August, a U.S. District Court ruled that AI art cannot be copyrighted, which could discourage some commercial artists from creating work with Midjourney. The company has also been accused of using artists’ work to feed its AI models without giving them credit or a path to earning royalties.
Lawyers say more legal battles lie ahead for Midjourney and its competitors. “All of their money is being based on misappropriation of people’s copyrighted works,” said attorney Matthew Butterick, who is representing claimants in a class-action suit against Midjourney. “They are sucking up all these artists’ work and then they are creating images that are designed to replace the artists’ work.” The company declined to comment on the suit.
Midjourney’s technology has also been used to create “racist and conspiratorial” images, according to the Center for Countering Digital Hate, as well as viral deepfakes—all of which could make it vulnerable to further litigation. These controversies have all arisen while Holz has been navigating his newfound AI-celebrity status in Silicon Valley. That rock star image could dissipate if the AI hype that’s defined the last year in tech fades, much like the VR, metaverse and crypto boom-and-bust cycles that preceded it.
A more immediate worry is a slowdown in active users that’s already afflicted other generative AI services. Traffic to the website of OpenAI’s chatbot ChatGPT dropped 10% over the last month, according to web analytics company Similarweb, and traffic to Midjourney’s website fell 6% during the same period. A more extreme turn in user sentiment could put the company in a club with Prisma Labs, the creator of Lensa, an AI self-portrait image generator that enjoyed overnight viral success—and millions in revenue—before rapidly tumbling into obscurity. Or worse, Holz could experience a repeat of Leap Motion’s trajectory; though its technology was praised by investors and the media as the future of computing, it soon floundered.
“The rising tide of generative AI excitement has lifted all boats,” said Nathan Benaich, founder of AI-focused venture firm Air Street Capital, “but companies can’t rely on that forever. We’ll see some natural attrition.”