Over the years I’ve become bearish on Stanford. The aesthetics of the place are off and the people you meet seem to optimize more for optics than depth. Not uniformly true obviously but you notice that no one from Harvard or MIT talks like this.

Thomson Reuters, Harvey, Legora, MikeOSS

.@harvey is hiring a social manager for up to $204K equity to be the new voice of harvey on social. apply here: jobs.ashbyhq.com/harvey/9ecf…

Anthropic xenophobia discourse working Malicious That’s why: opensourceaimustwin.com/?sha…

Our model is so dangerous, we're keeping your data, just in case

I miss Kate so badly that I physically hurt. It's been a week since we were together. She's still in Australia sorting her visa. I feel more stressed. My emotions are stunted. My thoughts are cloudy and my mood has dipped. It just hurts everywhere.

We just tell them the next model is too dangerous to release, and go raise a round...

Some background incase interesting: We landed PwC’s legal team as our second customer after A&O. This was right around the time GPT-4 launched. The killer feature at the time was the ability to query over multiple documents and receive an answer with inline citations. Later that year we built Vault (ability to upload a large amount of documents and query them) and knowledge sources (ability to search specific legal data sources). Around that time we started getting organic usage for PwC’s tax team. Turns out they also needed all of this functionality. We had also solved a lot of security / governance issues (dedicated capacity, BYOK, multiple workspaces in an org, retention, usage tracking, etc) that made it easier to roll out our product to adjacent verticals than use a horizontal solution. Future features like the ability to build custom workflows / agents led to expansion to other groups. We’ve consistently found that many of the features we build for law firms are useful to other professional service firms. All of them need to deal with the challenge of operating over client data. They need a secure way to share data with clients (shared spaces / datarooms), to run AI over that data in a secure way and with different configurations for different clients, as well as all of the governance to track their users, agents, client data, etc. Our focus is still law firms and in-house legal departments. The easiest way to see this is to look at how many lawyers we’ve hired across our EPD and GTM org (~200) and how many accountants (0). However, in the future we do think a lot of the infrastructure we are building for law firms will also be useful for other advisory firms like audit, tax, banking, consulting, etc. Each of these will need a platform that explicitly handles client data in a way clients / advisory firms are comfortable with (which likely won’t be horizontal products for the same reason you don’t use google drive as a dataroom).

The belief at the time was that model costs were halving every 6 months meaning tokens would get cheap and so application layer companies would need to find a way to charge for the value of tokens by selling the work / services. What actually happened is AI got much more expensive than people realized at the time. The shift from chat to agents led to an explosion in cost. One user could trigger hundreds of agents and each of those agents could trigger more agents. Agents started running longer and more autonomously. On top of that frontier models like Mythos are getting more expensive not less. If you look at what is happening in engineering, the coding companies are doing incredibly well selling tokens because engineers are consuming so many rather than needing to sell the value of the work. In fact, there is starting to be a massive demand for enterprise infrastructure to help large organizations track, manage and optimize their agents / tokens. Now the problem for application layer companies is how do you take that large token cost and convert it into something useful for your customers. A rough analogy is every company is about to get the ability to hire infinite employees. The main challenge is going to be figuring out how to manage those employees and make your business model work the same way it did with human employees. You can think about the previous generation of enterprise SaaS as building tools so that organizations could manage a large number of humans and make them productive. All of this is going to get rebuilt to support hybrid human / agent organizations. Some will be built by model / cloud, some by existing enterprise SaaS and some by new cos. For Harvey this means we don’t have to become a services company. The infrastructure for every law firm to deploy, train and manage a large number of agents is going to be so complex that model / cloud providers and law firms likely won’t build all of it. And AI is going to be expensive enough that we can capture something that looks like labor spend which is much closer to services without actually having to sell services.

Friends in SF : -$800k TC -Worried about green card -Bottom of the dating pool -Checking FIRE calculator everyday -Eyeing Fremont 2 bed 1 bath for $2.8 Mn -Envious of friends at OpenAI & Anthropic Friends in Bangalore : -$250k TC -Full paid home in whitefield with a staff of 5 -Apex of the dating pool -FIREd 5 years ago but still working for the love of the game -Envious of their IAS batchmates siphoning $20 Mn public funds