[seabass-labrax]

It's a risky investment strategy to put so much emphasis on AI startups. Not only do you have the already volatile nature of early-stage software companies (which YC is of course used to), but this is a bet on whether machine learning, chiefly LLMs, are going to continue to outperform other technologies and become sustainable to run.

There's no question in my mind that 'Open'AI is subsidizing the vast majority of LLM research and use today. If the efficacy of LLaMA and its derivatives actually start to approach GPT4+ in any meaningful way, there is quickly going to be a shortage of suitable compute that will completely dwarf the now-subsided Bitcoin mining craze. Plus, untainted training data is going to be harder to find amongst the text contaminated with mountains of early LLM drivel.

As a technology expert, I couldn't in all honesty say that I would want to have money in the YC fund right now. But if it pays off, it could be the biggest software windfall since social networking took off at the beginning of the 2010s.

[quadcore]

Friendly reminder: the strategy here is to invest in the founders, not the ideas. So great founders right now create AI startups. So either AI solves problems or they arent great founders.

[snowmaker]

That's exactly right. When people see YC funding a lot of AI startups, a lot of them think it must be because YC has some thesis about AI.

Actually, it says something much deeper about the world than whatever YC's partners' opinions are. YC funds founders, not ideas, so the reason that so many companies in S23 are AI startups is that that's what founders want to work on right now. It's an emergent phenomenon, like stock prices in the market.

One thing that's interesting is that there have been many hype cycles between 2006 and now (chatbots, several waves of crypto, VR, online-to-offline, etc). YC funded a few companies in each of those hype cycles but never anything like the current %.

[mtnGoat]

I would argue that just because they want to work on AI while it has a lot of hype, might not be a good thing. Folks should do what they are passionate about and interested in, not just raise money to turn a fume from the karat and greatest buzzword technology. Will they leave AI if the tides change?

[tinco]

I'm working on an idea that I'm submitting to YC this year. AI is revolutionizing the field I'm most passionate about, I feel like it would be a waste if I didn't even try to make a contribution at what's possibly the last time we get to make a significant contribution in this way.

I was hyped about Bitcoin when it came out, but the only truly revolutionary thing that happened after Bitcoin in that field was Ethereum, and there really wasn't anything to be hyped about besides that in the decade since even though people kept hyping it.

This hype is so different, it's as powerful as web 2.0 was. I felt like I missed out on building something epic then, and now I get another shot so I'm not gonna miss it. I bet there's thousands of people who feel this way.

[unusualmonkey]

That seems to beg the question.

Maybe YC isn't investing in the right founders, and that's why it has so many AI startups.

[nektro]

while there is certainly merit to the idea that they invest in founders over ideas, its scope is more to founders within an idea rather than wholly.

your statement fails to consider the possibility that the pool of applicants to YC are not great and the demographic of such founders are bad founders who are easily swayed by hype into fragile and failing topics such as "AI" and crypto.

in the situation where the applicant pool of ideas is in fact incredibly diverse, then it supports the thesis even moreso because it would show YC went with "AI" founders over other founders, regardless of their quality as potential business owners.

[alex_lav]

> So either AI solves problems or they arent great founders.

So being wrong one time means you're not a great founder?

[RexM]

It certainly feels like AI ideas are being targeted for funding.

https://www.ycombinator.com/blog/yc-ai/

https://news.ycombinator.com/item?id=36734110

[fijiaarone]

Great VCs are a gullible hype driven herd. Great founders are not.

[slg]

Jared made the comparison to S06 and it might be interesting to note that was the worst performing class in YC history in terms of percentage outcomes for each company. 73% of those companies are dead, 9% are still alive, and 18% got exits[1]. Although YC almost certainly cares more about money than percentage outcomes and one of those exits was Zynga buying OMGPop, so YC probably views that class as a success. Seems like they are taking on more risk for greater upside.

[1] - https://www.ycdb.co/

[hn_throwaway_99]

S06 was also the --first batch-- (Edit: I was mistaken, S05 was the first batch, but the rest still stands), well before YC was the "hit maker" its known as today, so it's not surprising that class didn't do as well as subsequent ones. The acceptance rate for startups into YC now is miniscule compared to what it was in 2006.

[slg]

>so it's not surprising that class didn't do as well as subsequent ones

But it also didn't do as well as the previous ones. I don't know when exactly YC got that "hit maker" reputation you are talking about, but S06 is an outlier regardless. The class directly before and directly after had failure rates of 29% and 38% compared to S06's 73%. The only classes coming close to the failure rates of S06 were in the lead up to the 2008 financial crisis.

[hn_throwaway_99]

Haven't done the math on this, but those early batches were so small - S06 only had 11 companies - so I think that statistically it's probably not that significant in any case, especially since there are huge grades of "success" even within the 3 categories of "alive", "dead", or "exit" on ycdb.co.

For example, in S06 there was one big exit (OMGpop), one company still alive and very well known today (Scribd), and one what I'd call "failure with an exit" - Xobni sold for a relatively low 60 million to Yahoo and was then shutdown.

Compare that with the batch directly after that, W07, which had 13 companies. Of those 13, it had one smash hit (Twitch) and one other relatively big success (Weebly). While it may have a lot fewer overall "dead" markers than S06, a lot of its exits look like acqui-hires or "OK buy our IP before we die" type exits. Better than an outright "dead" score I guess, but from the perspective of a VC like YC it hardly makes any difference. Point being I think it's a stretch to say there were really much significance in the "success difference" between S06 and W07 despite the differences in dead counts (with the possible exception that Twitch ended up being such a home run, but those are so rare they average 1 or fewer per batch anyway).

[hn_throwaway_99]

The underlying assumptions in this post don't really make sense for an organization like YC.

It's probably a good idea for YC to have some thesis about which technologies will be big, and "AI" is probably a good bet. The entire point of investing in so many AI startups is that I'm quite sure YC expects the vast, vast majority to fail (for various definitions of "fail"). But I think most people believe that there will be very few winners in the AI space (like pretty much all the other tech spaces over the past 20 years), so the biggest fears of someone like a YC is not getting in in those one or two AI companies that made it big.

[wokwokwok]

?

Concentration and diversification are fundamentally opposed in risk management.

It’s ludicrous to suggest that somehow YC is “immune” to the risks that are associated with concentration of investment in one area.

I don’t buy the “we pick founders not technology” argument.

Rubbish. It’s a hype bubble. You’re seeing that as emergent behaviour because it gets funding; there’s no amount of hand waving that makes any of the risks in the parent comment either a) wrong, or b) YC somehow magically exempt from.

Are you seriously suggesting that YC just doesn’t do risk management or cross-startup risk analysis?

I simply cannot believe that.

[version_five]

If this was NFTs, what would the winning strategy have been? Strike when the iron is hot and hype is at the maximum and hope to get some exits or subsequent rounds asap, or draw it out. From a portfolio perspective, it's just one batch, might as well go all in and maximally capitalize on hype, no?

[voytec]

NFT strategy was rather simple: hire so called celebrities to advertise the scam, profit from suckers, pull the rug.

[version_five]

Sure, and the faster you did that the more you would have made. I don't think AI is equivalent, the emperor has some clothes on here, they're just not quite as nice as he's been led to belive, but the same principle applies. Move fast, break stuff, sell while the cat's still in the bag.

[seabass-labrax]

These gigantic valuations for AI startups are only there because the startups aren't choosing to exit yet - big demand, low supply. Whilst acknowledging my lack of a crystal ball, I would imagine that once the few biggest companies have had their fill of acquiring AI startups for billions at a time, the remaining startups won't be so desirable any longer.

YC needs to make sure they actually have a market to sell these startups to before their valuation starts dropping again. Predicting that requires a certain kind of intuition, whereas 'traditional' software startups have a more predictable lifecycle. I'm not the 'hype type' myself - I don't really trust stocks without intrinsic value behind them :)

[seabass-labrax]

I'm really curious to know why my root comment saying 'this investment strategy is very risky and would make me nervous' has loads of upvotes, but my parent comment here - which I think has broadly the same message phrased in a different way - has three downvotes. Please enlighten me :)

[ignoramous]

> It's a risky investment strategy to put so much emphasis on AI startups.

At some level, YC is a marketplace for VCs. One'd think that this over-emphasis (it was crypto, b2b SaaS, mobile app startups before this) is YC working as it is supposed to be? http://paulgraham.com/herd.html / https://archive.is/vkzMd

[benreesman]

Hype cycles have downsides no one is arguing about, but they don’t seem avoidable so I try to focus on the fact a bunch of people are going to get subsidized long enough to pivot to something durable and positive.

Even after you strip off the 80% “and then LangChain connects to a different API” stuff, this technology is still massive and it’s here to stay.

Tuning up a bunch of bright founders on some really highly leveraged ways of looking at problems? At some VC’s expense?

It could be worse.

[yuvadam]

Bitcoin hash rate is steadily growing no matter which time frame you look at [1], what do you mean by "craze"?

[1] - https://www.blockchain.com/explorer/charts/hash-rate

[somsak2]

I mean, not "no matter which time frame you look at." Look at the drop in May 2021, from which the rate didn't recover until 2022. Similar situation for October 2018 - June 2019.

[baobabKoodaa]

The drop in what? Bitcoin has not been mined with GPUs in a looong time.

[baobabKoodaa]

I think they meant the GPU mining craze around 2017, which was unrelated to Bitcoin.

[steveBK123]

On the other hand, there's probably some sort of reasonable business model or two that will come out of the AI/LLM boom. So casting a wide net and hoping your accelerator will catch one of them is a decent play.

[Izikiel43]

I guess what you are saying is buy calls for Nvidia

[seabass-labrax]

Quite possibly; I'm not sure why calls would be better than straight-up buying the stock, but I am certain that Nvidia is going to have a pretty good time of the LLM hype.

Nvidia seem to me to have been consistently competent in improving their product lines. Perhaps my only criticism would be of their artificial hampering of some of their products, which they do in order to appease certain PC gamers by excluding the gaming market from usual supply/demand effects. This has the knock-on effect of severely reducing the compatibility of the hardware with Linux, which of course is by far the most sensible option for cloud computing hosts right now.

[__jonas]

> Perhaps my only criticism would be of their artificial hampering of some of their products, which they do in order to appease certain PC gamers by excluding the gaming market from usual supply/demand effects

Can you elaborate what you mean by this? I thought Nvidia is outpricing their PC gamer customers because of their focus on AI? I’m curious how they have hampered their products in your opinion.

[seabass-labrax]

Here's one reference: https://www.phoronix.com/news/NVIDIA-Lock-Broken

My reading of the saga is that the cryptocurrency mining bubble was causing huge demands for Nvidia GPUs as parallel processors (with ASICS for mining only starting to appear at this point). This meant that GPUs were being priced-out of the market for most PC gamers, as it was always profitable for cryptocurrency miners to buy more of them.

Nvidia introduced a lock on 'non-professional' cards which prevented alternative firmware from being loaded on them, effectively splitting the market into two, one for 'gamers' and the other for 'miners'. That lock interferes with the open source driver for Linux, Nouveau, meaning that normal things like CUDA (and not least running games!) couldn't be done on Linux without a labyrinthine network of proprietary drivers that, for instance, couldn't be updated automatically.

It seems as if Nvidia succeeded, because from my passive awareness of the PC gaming world, people seem very happy with their new GPUs. But for those using Linux for servers, at home or for AI research, it has been an ongoing nightmare.

[latchkey]

That lock was broken with the NVIDIA leak.

H100's were not used for mining.

Regardless, in general, there is little demand for GPUs for mining any longer after ETH switched to PoS.

[coffeebeqn]

Nvidia is specifically not on the LLM train - almost all possible forms of AI can use parallel compute chips to great effect. Their own AI products like DLSS have nothing to do with LLMs

[seanhunter]

DLSS sure, but they have products targeted squarely at model training eg the H100. https://www.nvidia.com/en-gb/data-center/dgx-h100/?ncid=pa-s...

[whatwhaaaaat]

I do not believe this is correct. Sure you can run llama or something locally on a cpu or your Apple arm but it’s not even close to the speed of running on a Nvidia card.

Do you think nvidias trillion dollar valuation is because entire governments are buying Nvidia cards to run stable diffusion?!

[tayo42]

Don't calls increase potential gains through leverage? With the usual high risk downside of the option ending up worthless.

[steveBK123]

Precisely - when there's a gold rush, sell shovels.

[epolanski]

Please don't turn HN in low effort WSB-like comments.

[r_singh]

The adjacent possible theory comes to mind