
Just saw a startup website comparing itself to 200+ alternatives. Bullish on their SEO/AEO. Less bullish on their pricing power.

Co-Founder/GP at Horizon Pre-seed/seed VC Fund: SendCutSend, Baseten, TypeSafe AI Angel: Palantir, Canva, BillionToOne

Just saw a startup website comparing itself to 200+ alternatives. Bullish on their SEO/AEO. Less bullish on their pricing power.

launched a model accidentally served trillions of tokens a day 29.4% of the fortune 500 showed up raised a really big series A from @a16z it's been 3 weeks we would like to sleep now

jev is the top model at 1k-10k context on openrouter!

Masterclass in strategy from @CompleteSkeptic. So many in tech/AI have forgotten what disruption actually looks like. If it doesn’t sound at least a little crazy at first, it probably isn’t disruption.
it is 100% a bet on jevons paradox! we do have a positive margin, but if current AI consumption stays stagnant, we will make very little revenue (every $100 we take from an LLM is $1 for us and less than half of LLM traffic is compatible with our interface) but if it's reliable and economical enough for automation, people who deploy it are creating value and incentivized to run each more things through their little value machines (this is all of software) and hopefully market grows even larger than it was before making intelligence so cheap

It's not disruptive if it's not dismissed by experts.
This is cool, but if your output domain is known in advance, why not just train a model to produce logprobs over enums?

TypeSafe just priced output tokens at zero. Input is $0.042 per million tokens, per the launch. Every yes or no from a language model arrives inside a sentence it had to write one token at a time, and you pay for the whole sentence. Pricing output at zero follows from a model that writes almost none of it. Diogo's launch post has the mechanism, and the waitlist. TypeSafe says Jev, its first public System One Model, answers many structured questions at once and returns a typed decision with a probability attached, so the calling code decides what happens next. So an LLM call becomes something your code can branch on directly. The launch example is support tickets. One long prompt splits into the judgments the app needs, is this urgent, is it billing, does it need a human? Each answer comes back typed with a confidence score, you pick a threshold, and anything under it goes to a person. TypeSafe's phrase for it is smart if statements. The confidence number is the piece I'd watch. Cheap tokens were never the whole blocker, since an output you have to double check still waits on a human before it touches production data. A probability that holds up the way TypeSafe describes would be the signal for when to let things run unattended. Once a judgment is cheap and legible it lands in form handlers, notification checks, and every small place nobody would have written a prompt for. Threshold tuning turns into a real engineering skill, and the teams who learn it early automate the boring half of their stack.
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x cheaper (w/ output tokens free) • Frontier composable intelligence optimized for decisions AFAICT the shortest path to AI-based economic revolution

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x cheaper (w/ output tokens free) • Frontier composable intelligence optimized for decisions AFAICT the shortest path to AI-based economic revolution

Oura is such a monster. -74% yoy rev growth, 24% cash flow at 1.5b+ run rate -Cohorts steadily improving with paid membership retention up from 81% to 87% in past 2 years -Median paid member wear device 23 hours/day! What a wonderfully healthy addiction (at least, that's what I tell myself, as an addicted user) -Monthly membership only $5.99...lots of room to raise that if they want. Bending Spoons would probably move decimal point over by one and see rev 3x overnight. -If you want one sign that there is a lot of juice to squeeze, annual discount on membership is a whopping $2 (thank goodness they have not let growth marketers loose on this business and I hope they never do) -Reminds me of Netflix which had to outmaneuver all kinds of bigger competitors, Oura has managed to scale in shadow of Apple and many other goliaths...truly exceptional execution Disclosure: I have shares (a startup I invested in sold to Oura) although I foolishly sold some last year.

I got addicted to @try_soma a few months ago in the beta, and it keeps getting better.
Soma is out of stealth see what your voice tells you about your mind

Great analysis of Palantir tho I think it understates how ridiculously unconventional it was. For ex: -No sales people for 15 years...took founder-led sales further than anyone could have imaged -Rejected defense playbook of using generals, etc in hard early days of winning US gov market -Rejected tech playbook of having partners do the dirty work of implementation -Had salary cap for first 10+ years, forcing everyone to take below-market cash comp -Massive title deflation ie "Director, Forward Deployed Engineers" had 5-10x the direct reports as the CEO...imo huge factor in keeping mission-first culture ...the list goes on. I don't think every unconventional choice worked, but the combo of extreme talent density + mission orientation got massive leverage from true first principle thinking. It's not a coincidence Palantir had no established VC backers (who would have tried to impose conventionality and stifle the company).
after 4 years of research, i wrote the world's definitive guide to what palantir actually is. thank you to @nikunj, @gokulr, @bennstancil, @jasnonaz, @dkrevitt, @_amankishore, @edraluk, @caelin_sutch, & others for reviewing this essay


I haven’t been buying the "SaaSpocalypse," but Q1’s nosediving SaaS valuations gave me pause. After a week in SF last month sampling the AI zeitgeist, I have a better feel for where the software sector is heading. It’s the SaaS-to-inference transition, and it’s good. My long-standing view has been that AI is a net positive for the software industry. It radically raises the ceiling for what software products can do. It should dramatically expand the market opportunity for software, just like the on-prem-to-cloud transition did back in the day. Yet many have been freaking out. After all, haven’t SaaS switching costs come down dramatically in SaaS, threatening one of the pillars of the business model? Yes, there’s no doubt that the “cement around the ankles” of legacy SaaS has weakened. At the same time, most legacy SaaS companies have barely scratched the surface of AI innovation while maintaining their historically high retention. This is how it played out in the last major transition: on-prem-to-cloud. Many legacy players (pathetically) ignored cloud innovation for 5-10 years (or longer) and still kept their customers. It turns out that technology is stickier than most in the tech industry believe. Take a look at Bending Spoons, which IPO’d off the back of buying crappy legacy products and jacking up prices because users didn’t want to give up their AOL email or Evernote notes. Tech industry people are not like this. They tend to be part of the very small minority of early adopters. Most people aren’t like this. Neither are most organizations. Legacy software isn’t going to disappear. But if pre-AI software companies don’t embrace AI innovation, their customers will be much less forgiving than on-prem customers 10-20 years ago. AI capabilities are too potent and obviously beneficial. What does embracing AI innovation look like? It means layering intelligent actions into all software. Historically, great software has helped users follow the right workflow. Now, great software must do the workflow by triggering agents to take actions. In other words, inference. The great news for everyone is that this opens the door to consumption-based pricing models that can scale exponentially. For legacy players and startups alike, delivering amazing AI-powered, agentic features is the way to get on the vertical-growth train. Remarkably, the door is still open for legacy players. Intercom’s 3.6b exit to Salesforce is a great example. Of course, new pricing models mean new margin structures. Just as SaaS had lower gross margins than legacy on-prem, expect consumption-priced inference to have lower gross margins. This is OK! We’ve already seen massive wins for inference-selling startups with negative gross margins, like Cursor. Legacy SaaS companies need to find religion on this. Dropping margins is never easy. Lock up the finance team if you have to. The priority is delivering AI-powered value for customers. Everything else is just details.

Last month, @baseten announced a $1.5B Series F at a $13B valuation. I've known co-founders @tuhinone and @saltyph for over a decade. I’ve learned a lot watching them build Baseten, but the biggest takeaway for me is the value of deep conviction–even if it means being dangerously early to a market. I met Tuhin and Phil at their previous startup, where I angel invested in their pre-seed round. A few years later, I spent a lot of time helping them navigate the sales process after they got an offer from a strategic buyer. Unlike the typical founder who dreads working for a buyer, Tuhin was excited to join their team. He realized it could be a great learning experience. This intense thirst for learning is one of those founder traits that’s hard to teach but incredibly valuable. In 2019, they started a new company. I angel invested in their first round. That was Baseten. Their thesis was built on a problem they'd lived firsthand: productionizing sophisticated ML models. The typical Series A-to-D startup then had a small data science team. These were smart, math-oriented people who could analyze data and build ML algorithms. But getting actual business value from said models required engineering resources. Tuhin, Phil, and their other co-founders, @amiruci and @defpan, identified this gap and aimed their company directly at it. While this was years before the launch of ChatGPT and interest in ML models morphed into the AI tsunami, they presciently saw a market opportunity. That said, they were very early. Over the next few years, SaaS mania ensued. Startups with very incremental SaaS products could fetch unicorn valuations off of a few million in ARR. Those products had little use for the sophisticated models Baseten’s product was anticipating. So the Baseten team had to bide their time. (Easier said than done for a startup!) Crucially, they were 10/10 in 2 key areas: 1. Fundraising: Tuhin is a brilliant storyteller and has a knack for building high-trust relationships with VCs. He kept the company perpetually well-funded by raising multiple pre-traction rounds from top-tier VCs. 2. Conviction: It would have been easy for them to pivot. They were great builders and could have quickly hacked together a SaaS product that would’ve whet the appetite of SaaS-crazy VCs in 2021/22. But they didn’t. They had deep personal experience suffering through productionizing models and could see that momentum was about to build for products with sophisticated (read: AI) models. After ChatGPT launched in late ‘22, the market rapidly coalesced. Startups began to see opportunities to build AI products that were essentially “productionizing sophisticated ML models,” just as the Baseten founders had originally envisioned. By 2024, the AI tsunami was emergent, and Baseten couldn’t have been better positioned. All told, Baseten is a great example of the dangers of being too early to a market. But also the incredible advantages. I’m very glad to have been along for the ride.

For my first 5 years of angel investing, I ranked everything. Founders, market, traction, team. But when I looked back on the data, it was all noise. So instead of fancy quantitative systems, I get much more personal mileage from a few simple heuristics today. One of my favorites is to ask myself: how am I going to feel when the competition arrives? I've tried other frameworks. Peter Thiel wants to invest in monopolies. Jason Lemkin from SaaStr says he invests in CEOs who are better than him. As the sort-of CEO of a little company for a few years, that one doesn't really work for me. A lot of what passes for wisdom here is moderately helpful at best, pseudoscientific cliché at worst. When something is called part art or science, I think they mean the science isn’t that good. Or they misunderstand the meaning of science. For me, I’d say my science of evaluating founders is a work in progress. But one thing I know for certain is that if a founder has a high-quality, earned insight into a market opportunity, there will be many other startups funded to target the same market in the next 6-12 months. Sequoia, Kleiner, and Andreessen are likely each going to back one, and some of them will attract great talent. In short, the competition is coming. And if I’m wrong about that, it just means the market wasn't there. Now, I've had angel investments in companies where someone came to the market, and I started to worry. I've been in that category more than I'd like. And then I've had others–like BillionToOne or SendCutSend or Baseten–where competitors start to flood the market, and I couldn’t wait to watch them take each one behind the woodshed. I wasn’t a Bulls fan in the 90s, but that’s probably what it felt like. The reason I don’t worry is always the same: they’re amazing at execution and they’re building great businesses that will compound for a long time. Oguzhan and @jimbelosic and @tuhinone are all maniacs in this department. And because they know what great looks like here, they’ve built teams full of people similarly gifted at executing on a mission. After years of rankings that were nothing more than noise, the difference between “Oh boy…” and “This is going to be fun” is what I've come to trust.

Founder trauma as an investment signal is a little overdone. In fact, I'm not so sure it was ever that useful. Bill Gates, Larry Page, and Mark Zuckerberg all had pretty good upbringings. I want to tread lightly here. We all have different experiences, and I'm not dismissing anyone's. But you arguably can’t compare one person’s trauma to another's. Take any person on the planet and ask what's the worst thing that ever happened to them. There are some people whose “worst thing” is unimaginably bad. There are others whose “worst thing,” comparatively speaking, is like a walk in the park. But it was still the worst thing that happened to them, and it was traumatic by their standards. It’s all relative, which means it’s very challenging to differentiate. What I do think you need to look for are absolute signs of outliers, and they can come in different ways. If you're a basketball scout and you see someone who's 7’5”, there’s no question. That’s about as legible as it gets. But in the context of early-stage investing, I'm more interested in less legible signals. Like extreme determination & unique technical breadth. One of our portfolio companies, @raven_computer, is a smart glasses company. The founder, @tomthecarrot, is a great example of what we look for. When he was 10, he programmed the viral whack-a-mole app “Bustin Jieber.” At 12, he delivered a TED talk on programming that has accumulated over 11m views on YouTube. But his life was completely changed when Google Glass came out a few years later. He became totally obsessed with smart glasses, AR, and the idea of spatial computing. He went on to Georgia Tech, where he became a Thiel fellow before dropping out to start an (you guessed it) AR company. He raised a bit of money but didn't have much success. Fortunately, it’s been a very different story with Raven. His personality is very different from that of some other founders. But to me, the signal is his utter determination to birth a revolutionary smart glasses product and a new computing paradigm. He's also a polymath and largely self-taught. He can write software backwards and forwards and talk about waveguides and optics or thermal management and quantum leakage. Constructing a truly breakthrough wearable device & surrounding ecosystem requires extreme depth across a lot of areas, and he has just that. Granted, I don't have depth in any of those areas. But I think I'm pretty good at picking up on someone who is really legit across them. And often I'll get reinforcement talking to other world-class experts. Ultimately, I don’t find trauma narratives particularly instructive. I'd rather find out what a founder was obsessed with before anyone told them it could be the seed of a business.

In 2016, I became @BillionToOneInc's first outside investor when I wrote a check into their pre-seed round. I invested (and followed in every round until the IPO) because I saw exceptional potential in the founders. But my conviction had little to do with their science. I'd been interested in the intersection of software and bio for a while. A couple years earlier, I'd looked at Benchling but took too long to pull the trigger. Big mistake. So when B2O showed up in TechCrunch’s write-up of YC’s inaugural YC Fellowship program, I was already primed. I met the founders, Oguzhan Atay and David Tsao, for coffee in SF. I knew almost nothing about biotech diagnostics, so I asked newbie questions and paid close attention to how they answered them. In healthcare, you need to be able to communicate with people across different levels of background knowledge. These 2 could do that. By the end of the meeting, I felt like I really understood what they were building and why it had the potential to win a big market. I wrote a very small check initially and helped them informally as they kept raising. I doubled my investment a month later after observing impressive intensity and coachability. I wrote another check a year after that as they raised a small pre-seed extension. It was bigger than before, but it still wasn’t a huge position at that point. I had a bit of an impostor complex investing as a biotech novice. But imposter syndrome be damned, I ended up writing 7 more checks into the business and made it my single biggest position. The biggest signal of their potential was their almost fanatical level of execution. Every quarter, they sent out detailed investor updates. They always hit their goals, and usually extra ones. When they hit setbacks–no startup is immune to these slings and arrows–they always confronted them directly and overcame them. I can speak for other early investors: none of us ever saw another company with more consistently stellar performance. I take most satisfaction in the number of follow-on checks I wrote. It’s one thing to write a small check backing impressive founders doing something you can’t understand. I also love that I was able to put together 2 partnerships and convince 16 friends & family members to invest in between the A and the B. Some investors will never look at a round extension or a bridge round. Not me. I should be honest about the temptation to be overly deterministic about all of this. I made many other investments in 2016, all of which I was excited about. A few more became unicorns, and several others built strong businesses. But this is the best one, bar none. It’s easy to over-learn from low-sample-size successes. So I don’t try to look for clones of B2O. But my understanding of the B2O founder’s absolutely remarkable technical and operational brilliance informs my approach today. It’s a great example of the extremely high bar that founders must clear if they want to build a truly generational company.

The SaaS-to-tokens transition has begun. The on-prem to SaaS transition took 10-20 years. This one will go much, much faster. It will also expand the TAM for the best companies. @usepylon is nailing it. Love the analysis & transparency from @marty_kausas.
Pylon's long-term competitor isn't Zendesk. It's Anthropic and OpenAI. Here's why... All companies will go through these phases: 𝟭: 𝗧𝗲𝗮𝗺𝘀 𝗮𝗱𝗼𝗽𝘁 𝗔𝗜. They get access to Claude/Codex and start building skills and using it for everyday work. 𝟮: 𝗘𝘅𝗲𝗰𝘀 𝗿𝗲𝗮𝗹𝗶𝘇𝗲 𝗔𝗜 𝗮𝗱𝗼𝗽𝘁𝗶𝗼𝗻 𝗶𝘀 𝗲𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲. As an example, our support team was spending $1.5k/person/month on custom Claude skills we had developed to help them triage, investigate, etc. 𝟯: 𝗖𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀 𝗶𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁 𝘁𝗼𝗸𝗲𝗻 𝗯𝘂𝗱𝗴𝗲𝘁𝘀. For example a support team might be given that $1.5k/person/month as a budget for Claude/Codex use. Across 10 people that's $180k/yr. 𝟰: 𝗣𝘆𝗹𝗼𝗻 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝘀 𝗔𝗜 𝘁𝗵𝗮𝘁 𝗰𝗼𝗺𝗽𝗲𝘁𝗲𝘀 𝘄𝗶𝘁𝗵 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 / 𝗢𝗽𝗲𝗻𝗔𝗜 𝗳𝗼𝗿 𝘁𝗵𝗮𝘁 𝘁𝗼𝗸𝗲𝗻 𝗯𝘂𝗱𝗴𝗲𝘁. Once the $180k token budget is there, Pylon will compete for it. $180k/yr is ~10x what that same support team might be paying for their ticketing system. The SaaS/Zendesk layer is tiny in comparison. We're releasing this agentic customer support product on July 15th. -- Notes -- - The token budget concept dramatically changes the market size for software companies that compete for it. Anthropic / OpenAI are paving the way for new use cases, but those implementations (currently via skills) will be unoptimized and expensive variants of products waiting to be developed. - Purpose built software (e.g. Pylon for customer support) will be able to build those same use cases and have them be faster / cheaper / better. - This idea applies broadly to B2B software companies.