Proud of the @SnorkelAI team for all the growth (on every axis) over the past 12 months; already 8-digit revenue increasing 18x in one year is wild, and that number keeps rising. There are trillion dollar hardware companies and trillion dollar model companies. There will be trillion dollar data companies as well -- it's the other major AI ingredient. AI development at its core is teaching machines to do what humans do (pre-AGI) or want AI to do (post-AGI). Data = whatever you use to show the model what you want (training data) or evaluate whether it does what you want (evaluation data). The shape of that data has changed significantly since Snorkel spun out of Stanford -- we've gone from requiring 30s per datapoint (simple label on simple text or image) to 30h per datapoint in some cases (complex multi-turn, multi-player environments with tools, traces, tests, etc.). This shift toward more complex and higher quality data requires research expertise, not throwing bodies at the problem. AI is still in early innings. The winners in the data space will be the companies that create the compounding loops where AI and human work together and improve each other. The "10x (or 100x) data developer" is definitely a thing -- I know because I've seen it, and you can already start to see the results (18x YoY growth). This is what an AI-forward data company looks like.

Yesterday we released Headlong (headlong.ai/), a microharness that enables “persistent agents,” or self-guided agents that think continuously. A few observations from working with one the past few weeks (beyond what we included in the blog): Proactivity - In every work environment I’ve ever been a part of, proactivity is one of the most valued traits an employee can have. It’s hard for an agent to be proactive when it only “wakes up” when you or a pre-scheduled heartbeat pings it. There’s a special delight in reading a DM from your agent that starts with, “I’ve been thinking about our conversation on X, and I had an idea…” Continuous memory - There is no “session” with a persistent agent. The beginning of its memory is its very first after being created. A new conversation I start is just a fresh observation injected into its ever-growing mindlog. To avoid context rot, we use a context with exponentially decaying resolution (verbatim recent conversations, summarized older conversations that can be retrieved at full fidelity when needed). For really old stuff it does the equivalent of me pulling out my early childhood scrapbooks to remind itself of fine-grained details. I find myself talking differently to the agent and expecting more from it because I know it has access to everything we’ve ever discussed, not just the current chat or explicitly extracted “memory” files. Multi-player flavor - It’s a new feeling to know that the same agent that I’m talking to (with the same context, same memories, same personality, etc.) is simultaneously talking to other team members. I can give it my preferences (e.g. when you talk to me, use really simple/straightforward language; check with me before talking about this project with others; etc.), but the same is true for human employees. It feels like we’re molding a new co-worker together, that learns some things about the world and workplace from me, and some things from others. E.g. “I’m developing my point of view on X. Braden thinks Y and Andy thinks Z – what do you think?” Having only ever seriously worked with single-player agents, it’s a new experience that interactions like this are now possible. Actual agency - The persistent agent can see my message and decide that it’s not worth responding to (there’s nothing to say), or not urgent (finish what it’s working on). Is that useful? On the surface, no – why don’t we just ratchet up subagent parallel processing and let it do all the things all the time? For a traditional agent, that makes sense – it only does the things you tell it, so why not do all the things you tell it as fast as possible? But for a proactive agent that you encourage to have its own side quests and proactive recommendations, the number of tasks it could work on is now unbounded, so learning to prioritize is actually a valuable skill. Will persistent agents unlock new quantitative gains on major benchmarks? I don’t know. Probably not. I suspect that most benchmarks we have today measure somewhat orthogonal axes to what persistent agency unlocks, and this is still quite young as a project. But I do know empirically that working with a persistent agent feels different. We shipped the minimum viable version of it so more people can join in on the fun of exploring the ramifications. h/t to others on the Laude + MIT research collab: @andykonwinski, @NickAJalbert, @lateinteraction, @NoahZiems, @a1zhang