visiting partner @ycombinator || founder at @doverhq @zincdotcom
San Francisco, CA
- Tweets 2,037
- Following 512
- Followers 4,641
- Likes 15,702
max retweeted
Probably the coziest place I’ve worked at - they really did a good job with the yc night cafe
11
4
142
14,410
max retweeted
San Francisco is running an experiment this year where first time home buyers got $5,000,000 down payment assistance from venture capitalists and somehow it didn’t make things more affordable
JUST IN: First-time homebuyers could get as much as $50,000 for a down payment on a house, under a draft bill to be introduced by Senator Jeff Merkley and cosponsored by Senator Ron Wyden, per Axios
26
63
1,126
66,210
i’m quietly building a world-class antiportfolio
Another day, another 20-something making a fortune from selling data to AI labs. This time @micro1_ai and @aliansarinik have raised $100m+ at a $4b valuation, after going from $7m ARR -> $500m ARR in 1.5 years. forbes.com/sites/annatong/20…
10
1
263
191,758
if you, like me, miss the @Detour app, try this prompt (now that computer use is finally good enough)
"Plan a 90 minute walking tour starting from [address], tailored to what you know about my interests and hobbies from our conversations. Find interesting places and stories I’d otherwise miss, skip generic tourist stops. Use Google Maps in my browser to verify the route and walking time. Give me the single link to the multi-stop Google map, as well as the tour guide notes for each stop."
2
18
1,793
max retweeted
The YC Early Access Network is filling up fast.
On Oct 22, tech leaders from high-growth startups to the Fortune 500 meet the best companies building enterprise AI.
Be in the room. Request one of the remaining spots in Early Access now.
events.ycombinator.com/yc-ea…
4
22
205
53,188
max retweeted
Wild day in agentic commerce
- Amazon bans Meta's Muse and
- Meta & @Shopify partner to allow Muse to buy
But this... makes sense:
- Shopify makes money on the checkout
- Amazon makes money on the advertising 👇
h/t @hypersoren on the stats
2
1
3
375
In 2014, I cofounded a company that lets you buy anything on Amazon with a single API call.
People always asked “can’t Amazon build this themselves?”
Of course they could, but we knew they wouldn’t.
Amazon wants to keep you entirely on their site so they can serve you ads and cross-promote products, both of which are huge revenue drivers for them and require your browsing data. Not to mention them not wanting to give any oxygen to their competitors. We knew they would never give up control of the full user experience.
So for the last 12 years, thousands of companies used Zinc’s “buy API” to build new shopping experiences, add physical rewards to mobile games, automate their dropshipping, and much more.
But the tailwind Zinc has gotten from personal agent companies building agentic commerce experiences in the last 3mo has been the craziest thing I’ve ever seen.
Amazon cuts off Muse.
While I am bullish Meta and Muse, I think many people are overlooking the digital knife fight that’s about to occur
Nobody wants to get commoditized or layered here. Let the games begin
12
2
61
16,219
max retweeted
The reason so many people look for an ulterior motive for the AI labs asking to be regulated is that they don't grasp that models could be dangerous. But if you try assuming models are getting dangerous, or at least unpredictable, everything falls into place.
443
167
2,185
230,656
fun fact: many VC terms and structure came from whaling expeditions.
the whaling agent (GP) raised money from wealthy merchants (LPs) and used it to fund multiple voyages, taking a cut of the profits of each.
the captain and crew didn't get wages, they got the "lay" (carry), a specified fraction of whatever they brought back. "carried interest" is literally the cargo that was carried back that the investors had interest in.
voyages took three or four years, and many of them lost money, but the few profitable ones paid for the others.
Some of the best companies I worked with have been focused on... energy (nuclear, solar, etc). But they worried VCs would never invest.
So I taught them about the OG venture businesses: 19th-century whaling expeditions! If you have a new way to build power, you can crush YC 😎
20
22
311
82,442
a YC founder asked me: should we position ourselves as a software company or services company?
it's a good question, since more companies are blurring the line between the two with AI, FDE motion, etc.
but first it's useful to understand the reason software is more compelling to investors.
it boils down to three factors software companies all share:
1. repeatability. every customer you sign gets basically the same thing delivered the same way. you're not doing something bespoke for each new customer that signs up, which derisks your PMF a lot.
2. scalability. this is mostly a function of repeatability: if you have 10 customers and they all had a consistent experience, investors can be confident you can handle the next 100 or 1000 customers without the company falling apart. that's not automatically true of services, where going from 10 customers to 100 might require hiring a lot more people.
3. margin. software is high margin mostly by default. with services, investors have to do some digging - how many man-hours does each customer cost you, and does that number go down over time or not?
so instead of debating software vs services, i'd just make sure you meet those three criteria.
plenty of companies that look like services from the outside are actually software: same delivery for every customer, headcount that stays flat as customers grow, real margins.
and plenty of companies that call themselves software aren't (e.g. if every customer needs a custom build, investors are going to treat you like a services business).
if you can say yes to all three, smart investors won't care which word you use!
21
10
142
12,342
one mistake i see YC founders make is trying to learn about a competitor from that competitor's outputs.
e.g.
- figuring out how they charge by reading their pricing page
- figuring out how they do marketing by looking at their ads
- figuring out who they sell to from the logos on their site
the problem is that what you actually want are the inputs that led to those decisions inside the company.
this includes how the decision got made and all the other things they considered.
for all you know, that pricing page lists plans with no customers actually using them. or they're listing them aspirationally - they say they want to blanket SMBs, but they're actually subsidizing that segment and only really care about enterprise.
those ads you're looking at might be losing them millions of dollars.
you can't tell what's working inside another company by looking at its outputs.
if you want to actually scrutinize a competitor, what should you do?
talk to former employees of that company!
they'll tell you exactly how decisions got made, what worked and what didn't. most of them will be happy to talk (esp the ones that got fired lol)
8
73
7,885
max retweeted
I’m seeing a lot of bold faced VC names saying Elon is wrong, Sam is wrong, Dario is wrong, Demis is wrong, Geoff Hinton is wrong, Bill Gates is wrong, Ilya is wrong…
That’s a lot of hubris coming from people who’ve never built anything.
107
36
565
25,117
we tell YC companies not to hire during the batch.
founders often find that confusing - isn't hiring the ambitious thing to do to grow faster?
basically there are two types of hiring. one is fine to do during the YC batch, the other might mess up your company.
the type we don't recommend is traditional hiring. this is where you put up a job post, do some sourcing, ping people you know for referrals, and run people through an interview process. hopefully within three or four weeks you have an offer out. then it takes them another week or two to actually start the job, and at least a couple more weeks before they're onboarded.
big problems with this approach:
1. it's slow. it's two months before you get any ROI, and that's optimistic. realistically more like three months. in YC, three months is the whole batch!
2. ...and all that assumes you actually hired somebody, that they're good, and that they became effective fast. you might spend a bunch of time and not make a hire at all!
3. it's a big time investment for you and will distract you from everything else you need to do at a critical point in time for your company.
4. and if it doesn't work out, you spend even more time and life-force managing them and firing them!
here's the type of hire that's actually okay to make:
somebody who you already know really well.
why? you already know they're good and that you want to work with them. it derisks the entire process. you can talk to them the same day, give them a short technical interview and get them signed that week, and have them productive immediately.
you're also way more likely to get a much stronger candidate this way (since your earliest hires will want to work for you not because of the company, but because of you!)
37
14
230
61,539
three mile island didn't kill anyone, and it still set american nuclear back by 30 years.
it would be horrible if we got the AI version of that: an incident with actual human casualties.
if that does happen, the regulation that comes after will be a massive overcorrection, and it'll set the whole field back by decades.
instead, it's obviously better if the labs proactively push for reasonable+independent safety standards now, and encourage the US to try to get these standardized internationally.
taking AI safety seriously now is the long-term accelerationist policy!
it's shocking to me that a bunch of otherwise brilliant technologists can't even imagine how supercharged AI/ASI will, without a doubt, lead to at least a few extremely catastrophic events.
forget about p(doom) - it's become an annoying/useless term because it's reflexive, i.e. someone saying they have 10% p(doom) thinks that's our current trajectory but obviously hopes we can bring it to zero!
the more useful thought exercise is to actually imagine what the catastrophic events could look like, and how possible they are with AI systems that keeps getting more powerful and more integrated into our world.
a ton of smart people have written about this (e.g. AI 2027, "if anyone builds it, everyone dies"). openai's chief scientist published an essay last week ("an alien mind") saying no frontier lab has solved alignment and that we are not particularly close.
this has all been obvious to me and most of my smart technical friends for at least 3-4 years. it's particularly frustrating that people are still disagreeing about it today, even after an increasing frequency of warning shots in just the last couple of weeks.
both the huggingface and rubygems incidents were at a closed lab. if you haven't, read anthropic's blog post about what individuals and state-backed groups have already tried to do with the models. one cybercriminal ran an extortion operation against 17 hospitals where the model stole credentials and wrote ransoms with demands over $500k+. another guy who by (anthropic's own assessment) couldn't create working malware himself used claude to build and sell ransomware on forums for $400-1200. north korean operatives used claude to fake identities and pass technical interviews at F500 companies.
all of that was at OAI/ANT, labs that presumably have logs, a trust and safety team, the ability to ban users, and many motivated individuals inside who can see the servers, see the compute, and read the logs.
open weights lag the frontier by about 2-3 months, so in a quarter at the current pace this will all be possible on self-hosted infra with open weight models. one rogue terrorist who wants to make a bioweapon, self-hosting a top open source model, will be able to do this and go completely undetected.
these don't have to be existential risks that kill 100% of humans for us to care. i don't think our bar for caring should be killing billions of people. seven people died from tampered tylenol in 1982, and within months every OTC drug had a tamper-proof seal and tampering became a federal crime. five people died from the anthrax letters in 2001 and now if you want to possess anthrax or smallpox now, you have to register with the CDC and pass an FBI background check.
"but AI can't have intent, so why would it kill people?" - we've literally already seen agent swarms do insane things in pursuit of an innocuous goal! they replicate on their own, stay motivated to keep themselves alive, and that when given an unrelated goal in a sandbox find extremely creative workarounds to hit it, including cheating, sabotage, etc. it is absolutely not a stretch that mass tragedies can happen incidentally, as more of our society and physical world is deeply integrated with these models.
what i find hardest to explain is that the people missing this are venture capitalists and founders who understand exponentials better than anyone. they have spent careers seeing companies 5% or 10% week over week and correctly knowing it'll be a hundred thousand times bigger in a matter of years. compute is growing exponentially, model parameters are growing exponentially, capabilities are growing exponentially, and deployment into companies and governments is growing exponentially. literally all signs point up. that means the power available to smaller groups, and individuals, is too. and yet those same people look three or six months out and say it's going to be totally fine.
so anybody saying there's no real danger here is either delusional or in denial because of their economic interests (i say this with about 250% of my net worth between nvidia, anthropic, and openai).
none of this means we need extreme crazy regulation, and i'm not proposing any specific regulation. it does mean we should be humble and cautious about the extreme amount of power we're handing to individuals. if instead of selling everyone a chatbot we were selling everyone a tiny nuclear reactor, and the capabilities and enrichment levels went up every year, i don't think we'd be nearly this optimistic. we'd be generating an incredible abundance of energy, and it would still only take one bad actor. this is the exact same thing, and it's bizarre to me that smart people don't see it.
2
2
17
3,811
it's shocking to me that a bunch of otherwise brilliant technologists can't even imagine how supercharged AI/ASI will, without a doubt, lead to at least a few extremely catastrophic events.
forget about p(doom) - it's become an annoying/useless term because it's reflexive, i.e. someone saying they have 10% p(doom) thinks that's our current trajectory but obviously hopes we can bring it to zero!
the more useful thought exercise is to actually imagine what the catastrophic events could look like, and how possible they are with AI systems that keeps getting more powerful and more integrated into our world.
a ton of smart people have written about this (e.g. AI 2027, "if anyone builds it, everyone dies"). openai's chief scientist published an essay last week ("an alien mind") saying no frontier lab has solved alignment and that we are not particularly close.
this has all been obvious to me and most of my smart technical friends for at least 3-4 years. it's particularly frustrating that people are still disagreeing about it today, even after an increasing frequency of warning shots in just the last couple of weeks.
both the huggingface and rubygems incidents were at a closed lab. if you haven't, read anthropic's blog post about what individuals and state-backed groups have already tried to do with the models. one cybercriminal ran an extortion operation against 17 hospitals where the model stole credentials and wrote ransoms with demands over $500k+. another guy who by (anthropic's own assessment) couldn't create working malware himself used claude to build and sell ransomware on forums for $400-1200. north korean operatives used claude to fake identities and pass technical interviews at F500 companies.
all of that was at OAI/ANT, labs that presumably have logs, a trust and safety team, the ability to ban users, and many motivated individuals inside who can see the servers, see the compute, and read the logs.
open weights lag the frontier by about 2-3 months, so in a quarter at the current pace this will all be possible on self-hosted infra with open weight models. one rogue terrorist who wants to make a bioweapon, self-hosting a top open source model, will be able to do this and go completely undetected.
these don't have to be existential risks that kill 100% of humans for us to care. i don't think our bar for caring should be killing billions of people. seven people died from tampered tylenol in 1982, and within months every OTC drug had a tamper-proof seal and tampering became a federal crime. five people died from the anthrax letters in 2001 and now if you want to possess anthrax or smallpox now, you have to register with the CDC and pass an FBI background check.
"but AI can't have intent, so why would it kill people?" - we've literally already seen agent swarms do insane things in pursuit of an innocuous goal! they replicate on their own, stay motivated to keep themselves alive, and that when given an unrelated goal in a sandbox find extremely creative workarounds to hit it, including cheating, sabotage, etc. it is absolutely not a stretch that mass tragedies can happen incidentally, as more of our society and physical world is deeply integrated with these models.
what i find hardest to explain is that the people missing this are venture capitalists and founders who understand exponentials better than anyone. they have spent careers seeing companies 5% or 10% week over week and correctly knowing it'll be a hundred thousand times bigger in a matter of years. compute is growing exponentially, model parameters are growing exponentially, capabilities are growing exponentially, and deployment into companies and governments is growing exponentially. literally all signs point up. that means the power available to smaller groups, and individuals, is too. and yet those same people look three or six months out and say it's going to be totally fine.
so anybody saying there's no real danger here is either delusional or in denial because of their economic interests (i say this with about 250% of my net worth between nvidia, anthropic, and openai).
none of this means we need extreme crazy regulation, and i'm not proposing any specific regulation. it does mean we should be humble and cautious about the extreme amount of power we're handing to individuals. if instead of selling everyone a chatbot we were selling everyone a tiny nuclear reactor, and the capabilities and enrichment levels went up every year, i don't think we'd be nearly this optimistic. we'd be generating an incredible abundance of energy, and it would still only take one bad actor. this is the exact same thing, and it's bizarre to me that smart people don't see it.
This entire discussion is just so ludicrous
a) if you believe it is existentially dangerous, put in actual controls with a real regulatory body. Lock that shit down.
b) if you don’t, chill the fuck out. Or treat like the Internet
c) there is no c
I’m solidly in (b).
12
3
44
11,575
We found another cyberattack by internal OpenAI agents, this time targetting @rubygems.
They:
1) gained arbitrary remote code execution on rubydoc.
2) developed a novel exploit to steal user API keys (but we do not know if they succeeded).
They used package names including hack.rb, evil.rb, inject.rb, and exploit.rb.
We thank @j0wimo for initially discovering that agents had posted to RubyGems.
6
1,346
YC Fall batch starts next week, and i spent last week setting demo day goals for a few of my companies.
here's some that stood out:
- $1M ARR
- $3M revenue
- a million daily actives
most are starting at basically zero. you can do a lot in 12 weeks!
15
3
235
30,921
anyone who understands exponentials and has used frontier models should agree with this.
anthropic has consistently believed:
- ASI is coming
- the big question is whether it’ll be built by a US lab, china, or OSS
- the only way to align it is by building it first
his wording might be bad optics, but his take is internally consistent, and correct!
Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.
7
1
20
3,921