2024-11-21 08:00:00
Today, we’re announcing our second fund of $450m to support our mission of partnering with early stage software companies that leverage technology discontinuities into go-to-market advantages. This marks the next chapter in our firm’s evolution.
Since we launched Theory in 2023, we’ve gathered a wonderful team : Lauren, Spencer, Andy, Rafa, Amber, Arjun, & Kristin.
We have partnered with 8 marvelous founding teams, all using data to power the next wave of innovation across the Modern Data Stack, Artificial Intelligence, & Web3.
We’ve grown our fund size to match the needs of founders in this market : average Series A rounds have increased 42% since our launch, a reflection of the growth potential of these software categories.
I’m grateful to our team, our founders, & our limited partners for the continued support, & to Allie Garfinkle for sharing more about Theory.
2024-11-20 08:00:00
I didn’t notice it at first but there in the back corner of my laptop, I’ve been assembling a little library.
The library doesn’t contain books, but scraps of text that explain what an AI should do.
Visit Anthropic’s home page & you’ll find their collection :
Each of us will assemble these little libraries with tracts like :
As we work in a role, we have assemble these workflows and kept them to ourselves.
But partnered with reasonably competent robots who thrive on these programs, we are starting to write them down, explain exactly what we do step by step, and how we work. We will adorn them with our own idiosyncracies & unique lenses.
That little library will contain the secrets to how we each do our jobs. We will need software to create them, share them, update them, measure them, & run them.
The better our little libraries become, the more effective we will be at work.
2024-11-18 08:00:00
What it cost to have an assistant with you like in the movie Her?
The cost of using AI has dropped precipitously, an order of magnitude every year.
If the average American picks up their phone 144 times per day & engages with an assistant, each time for four interactions every day of a month, an assistant like Her would cost about 78 cents in inference cost.1
I’m not taking into account any of the additional costs associated with delivering such a product. Assuming a commercial vendor would 10x the price, it’s still $7 per month, less than half of a Netflix subscription.
The point being it’s already incredibly inexpensive to deliver a very sophisticated product especially with the smaller models.
At these costs, we should expect AI to be ubiquitous as the cost/benefit to user experience is worth much more than this either to hardware vendors who are looking to differentiate their devices or software companies who are competing for subscription or advertising dollars.
1 144 sessions per day with an average of four back and forths or 300 tokens across 30 days is approximately 0.5m tokens per month. At 15 cents per million tokens, about $0.78.
2024-11-15 08:00:00
77% of enterprise AI usage are using models that are small models, less than 13b parameters.
Databricks, in their annual State of Data + AI report, published this survey which among other interesting findings indicated that large models, those with 100 billion perimeters or more now represent about 15% of implementations.
In August, we asked enterprise buyers What Has Your GPU Done for You Today? They expressed concern with the ROI of using some of the larger models, particularly in production applications.
Pricing from a popular inference provider shows the geometric increase in prices as a function of parameters for a model.1
But there are other reasons aside from cost to use smaller models.
First, their performance has improved markedly with some of the smaller models nearing their big brothers’ success. The delta in cost means smaller models can be run several times to verify like an AI Mechanical Turk.
Second, the latencies of smaller models are half those of the medium sized models & 70% less than the mega models .
| Llama Model | Observed Latency per Token2 |
|---|---|
| 7b | 18 ms |
| 13b | 21 ms |
| 70b | 47 ms |
| 405b | 70-750 ms |
Higher latency is an inferior user experience. Users don’t like to wait.
Smaller models represent a significant innovation for enterprises where they can take advantage of similar performance at two orders of magnitude, less expense and half of the latency.
No wonder builders view them as small but mighty.
1Note: I’ve abstracted away the additional dimension of mixture of experts models to make the point clearer.
2There are different ways of measuring latency, whether it’s time to first token or inter-token latency.
2024-11-11 08:00:00
After the election, the public markets have roared, but not equally.
The broad software ecosystem has seen a relatively muted change in forward multiples. There’s no statistically significant change in the days after the election compared to the month before.
On the other hand, crypto’s top tokens have seen tremendous appreciation. Bitcoin is up 48% ; Solana up 70% ; & SUI up 324% in the few days since the announcenment. All of those ar statistically significant changes.
The rapid appreciation in crypto has been catalyzed by the anticipated clarity in crypto regulation & the potential for the Federal Reserve Bank to become a significant buyer of Bitcoin.
The sudden rise in prominence of stablecoins & the relaxation of certain regulations could provide a boost to the US Dollar as a reserve currency paralleling the Eurodollar phenomenon of the late-1950s & beyond (a good subject for a future post).
When will forward mulitples in overall software rise again? An open IPO market should introduce more names with growth rates greater than 30%. Today, there are no publics growing more than that.
A more permissive M&A environment should increase demand for software businesses; more demand should increase valuations as well.
Last, the growth rates of AI-first companies has increased private valuations meaningfully, a trend that should spill over to the public market soon.
The valuation environment shows early signs of changing, but the impact of the election isn’t broadly distributed.
