Synth
... agent learning, automated.
Opportunity
Automating knowledge work, and then automating all work, won’t just happen. Smart, hardworking teams will need to solve hard problems. Problems like:
- Supplying cheap general-purpose intelligence
- Designing effective interfaces (product UI) between intelligent systems and human users
- Teaching intelligent systems to act according to the priorities and constraints of those effective interfaces
We think a few hyperscalers (Google, Microsoft, AWS via lab proxies) will own the first problem, whereas a very large number of application startups will be necessary to “solve” the second problem - product vision will take a while to automate. We believe one company can win the third opportunity, which has enough “product complexity” (we’ll need to work hard to make it easy for product builders to configure their AI system) to avoid commoditization yet a homogenous enough customer profile to prevent market fragmentation.
Winning this intermediate layer will be the biggest opportunity in AI.
Wedge
We believe that labor automation software will take the form of compound AI systems - software with AI models interspersed. Software helps teams provide customers guarantees and reason clearly about changes to their product. It’s not going away. What will change, however, are the patterns these compound AI systems take on. The most useful mental model here is that of a graph.
- RAG and AI pipelines are small directed acyclic graphs.
- Agents are medium-sized directed graphs.
- ??? are large undirected graphs.
We don’t even have a name for the frontier category because it’s not feasible to build today.
It won’t be until Synth wins. When we win, though, this entirely new slice of applications will require software like ours to exist. The goal is not to just win RAG or even agents. The goal is to uniquely enable developers to build systems making hundreds or even thousands of AI calls to complete a unit of work.
Initial Product and Short-term Roadmap
Synth’s first product is software that logs the actions of small-to-medium sized digital control agents, identifies and communicates short-comings, and fine-tunes language models that will achieve better performance in that agent scaffolding.
This core pattern:
- Customer shares production artifacts
- Synth assesses system performance
- Synth reasons through, tests, and delivers system improvements
will form the basis for our roadmap.
Because agents, and compound AI systems generally, are product software developed by teams, Synth cannot deliver its promise to a customer overnight. If we make their base software A reliable enough to enable an improved version B and C - which we then improve - the time between A and B is “dead” for Synth. Hence we have a very strong incentive to help our customers ship fast, and specifically complete our “1 2 3” pattern as quickly and frictionlessly as possible.
Achieving that outcome isn’t just a matter of design, product, and UX work (although it is that, too) - it requires large engineering projects such as e.g. running customer software in secure sandboxed environments, training AI systems to submit pull-requests to client repositories, and simulating customer interactions. Before 2026, we’ll want to have earned the right to intervene “live” in customer’s production code. We’ll need to work for it.
Pricing
As the “1 2 3” pattern tightens, the amount of value Synth delivers for its customers will go up, and so will the price. Once software is written to tackle ambitious and valuable applications, and which can only exist with Synth, our marginal pricing power will go up substantially.
Presently we’re targeting $1k per month as a milestone price tag. By 2026 we’d like to target $10k per mo for larger clients.
We’re not starting this company to win $1k contracts, but much of the 2030 market doesn’t exist yet. We’re betting on a large market of firms willing to pay $100k-1m yearly for this solution in 2030 and if you join us you probably should too.
Team/Culture/Hiring
Building Synth will be hard. It will require a clear vision for a completely new category of AI software coupled with enough empathy to build a form factor that developers love using today. We’ll need to both engage in the open-ended research necessary to push forward the state of the art on multiple scientific axes and ship product exceptionally quickly. Exceptionally talented people will need to build a product whose full potential won’t be fully realized for years. We’ll need a particular kind of team.
Me
I’ve spent over a year building out real automation software at Basis and working with academics to research how to optimize AI agents. MIPRO is an agent-optimization algorithm I helped write that’s been used by dozens of researchers and technologists to build better agents. I have the research intuition necessary to know how to build agent-learning systems, what’s possible, and how to sequence experiments / bets. I know the painpoints agent devs face and what the dev cycles Synth will slot into look like. I did math and ml research at Yale. I’d like to think I learned a thing or two about startups after selling a tiny startup right out the gate to Halborn after college.
Team
Marcus Dominguez-Kuhne is on board as an experienced Applied Scientist. His background is Caltech, reinforcement learning PhD (dropped out), ML at Amazon. He’s leading our ML effort.
We’re going to look to hire a design-focused fullstack dev to own the frontend and a research engineer who loves distributed systems to own the systems underlying our backend and research codebases. Additional hires will come as needed and as earned. Broadly speaking, we want to hire exceptional thinkers to fully own and be responsible for entire functions. We want a frontend engineer who’s going to develop strong opinions about how the UI works in two years. Until there’s a “genius sized hole” in a function we won’t hire for it.
Traction
We validated the product with Ellipsis in early November, are working with Zenlytic, and have a few teams in late pipeline / trial phase.
Customers have high UX standards, and this is a hard product to build. We had to do a lot of work to make the UX and algorithm good enough to win those customers, and have plenty more ahead of us.
Next
Hire 2 engineers to own frontend/design and backend functions over the next 3 months. Iterate on the core product pattern to dominate our toe-hold segment (series seed-A startups building vertical SaaS agents) and prove out an offering ready for larger enterprise players.
At Series A Synth will be making the case to scale our offering across the enterprise. Let’s get to work.