Founder Building the automated data stack @signalpilot_ai Ex-Goldman quant | Harvard Math & CS
San Francisco, CA
Joined July 2011
- Tweets725
- Following1K
- Followers1.4K
- Likes1.2K
Why do need a moat when you can be at the sea? Build hull for your ship anon
As CEO of Product Hunt, I see hundreds of launches a week. One trait separates winners from losers.
Every few months, new models come out. Some products become better, effortlessly. Some don't improve, or become irrelevant.
This property needs a name. I'm calling it Buoyancy.
Buoyant startups rise with model improvements the way a boat rises with the tide.
Anti-buoyant startups come in two flavors: over-scaffolded (which hinders the new, smarter models) and too thin - which makes them redundant when models improve.
To be buoyant, you need a "hull" - usually proprietary data, a vertical-specific harness, network effects, or distribution advantage. This hull keeps you buoyant as models improve.
Hulls are the new moats.
My number one advice to startups in 2026: Be buoyant. Find your hull.
(Full essay on my blog)
Yet he decided to work for a trillion dollar behemoth? 🤷
Don’t disagree with his take but this is kinda strange. Walk the talk?
Alexandr Wang says right now is the best moment in history to build a billion-dollar company.
He argues the bottleneck is no longer AI models but how fast the world adapts to them. Even if models stopped improving today, decades of economic upheaval would still follow, and that gap between technology and adoption is where new companies get built.
10 years ago, startups fought giants with cleverness and fewer resources. With AI agents, a small team can now compete head-on with incumbents, like a mecha Goliath against a traditional one.
He calls it a once-in-a-civilization window for anyone with a vision and ambition.
It won't stay open forever.
Long live the context layer 🙂
Intelligence must draw on memory. The mind must not simply compile data — as an algorithm now does more quickly than we can — but recall lived experiences, which contain depths of meaning that only the human soul can recognize, and connect more deeply with the goodness of these experiences. Everything that is good begins “with the Spirit” and thus possesses a breath, a dynamism and a sense of communion that, once experienced, are engraved upon the heart. vatican.va/content/leo-xiv/e…
Claude Dashboards got us really excited about the future of analytics engineering, and largely the role of data in knowledge work!
We think demand for analytics work, from both humans and agents in your team, is about to explode. The best analytics and data engineers are already getting ready for it.
There's a catch, though. Connect your warehouse straight to Claude and you'll get a beautiful dashboard fast. You'll also often get the wrong numbers.
Ask the same question five times and you can get five different answers, because each run picks up a different definition of "revenue" or "churn", from 5 tables.
That isn't a dashboard problem or a coding problem. It's an alignment problem: your people and your agents need to agree on what every metric means.
At dbt Summit, we talked to a lot of data teams trying to solve this with hand-written markdown files and skills. That doesn't scale. And the LookML-era habit of labeling every dimension and measure by hand isn't realistic either.
So we just launched SignalPilot's self-improving Data Context Factory for business data work. It curates business context, metrics and evals as code from your data and from how people use your agents, so teams and agents can act with confidence.
→ Drafts definitions from your existing dbt models, dashboards and queries
→ Learns from every chat and flags definitions that conflict
→ Runs evals on every change before agents rely on it
We believe this is how data professionals become the agent orchestrators in their orgs, and the central role in driving business outcomes.
We also shipped the skills and MCP tooling to work natively with Claude Dashboards. Ask Claude for a dashboard. The request goes to our headless cloud sandbox, and you get a dashboard grounded in the definitions maintained in your SignalPilot Context Factory.
In this demo, one prompt produces a live YoY revenue dashboard, and every number matches the definitions your data team has approved.
Beautiful, accurate, and human aligned.
This is actually a great take on “everything is dead” doomerism in saas
In the current stage of the market, the noise is so high that cutting through that noise and demonstrating what good product means during the sales process becomes the main challenge after qualifying a customer. And the job becomes, finding a single use case or problem the user cares about deeply and showing them exactly why you can solve this problem the fastest and the cheapest with a superior product.
I will give two actual examples:
1. Recently we encountered a prospect that wants to launch agents on top of their data warehouse and already had an incident where a vibe coded agent ran a query that ran for hours in the background and cost them 1000s of dollars.
We explained to them that our agent harness has tooling that does EXPLAIN based query parsing + AST analysis on the fly on all major warehouses and literally this behaviors will get gated / stopped deterministically.
2. Another customer (with ongoing pilot us) asked us the hardest questions on the hours they need to spend to keep their semantic layer accurate. They already got pitched by 5 other vendors that only if they spend months curating and then 10s of hours per week maintaining their controlled vocabularies in the semantic layer, then they will be sure to get answers aligned.
We had a product in the backlog that was implementing the self curating part of the context layer. We showed them our initial (somewhat broken) version and told them we will come back next week with a PoC on their stack.
Now both of the customers were happy but this kind of deep product feature is not something you will be able to show on your landing page hero and it is maybe deep in the docs page that the customer will not read. But that's where sales process makes a difference.
There are also a lot of copying of narrative and entire product lines going on, even at some of the largest incumbents as product (not necessarily good product) building becomes cheap. This only adds to this murkiness and noise around good products.
But that is for another post.
it is kinda funny but I feel like, with jev we are realizing that stats and probabilities (which is maybe called classical ML now) are surprisingly good at a lot of actual work when you think through your work as directional graph of actions with probabilities as edges
preach!
Boys and girls, nobody is thinking anymore.
Entire enterprise departments have thousands of people who are spending their days interacting with the LLM, faking "being productive", consuming tokens because it's a highly regarded activity, shipping lines of code with Cursor or Amp or Claude or whatever (except their brain).
Even slides are not made thinking anymore!
The person that allegedly built the powerpoint deck you are reviewing right now cannot even answer a couple of intelligent questions on the very thing he's presenting to you.
He looks at you stuttering some nonsense, hoping you'll let him off the hook because he made it so obvious Copilot wrote the slides (of course, you already knew that within the first 6 seconds of glancing at them).
There is some sort of comical atmosphere to it all.
And an absurd satisfaction in watching this entire situation unfolding in front of your eyes.
There is also approximately ZERO chance this whole thing won't spectacularly backfire, and of course good old consulting firms will be there to pick up the pieces.
I wonder at what point we started to believe we could get away with doing valuable things without even thinking?!
How did this happen, wow.
System 1.5 is basically using lots of system 1 for exploration so that system 2 can one shot with that context
Classic exploration exploitation paradigm
An early look at SignalPilot Rev: we are rebuilding the outer loop of agentic business decision making around System 1 intelligence of jev + llms.
1/ When your agent or you prompt: a large number of parallel Jev runs explore the context DAG and aggregate map-reduce style.
5/ One real insight about why it works so well: for this customer we + our agents worked with their team to curate and train them how to author, review and approve metrics, and then get out of the way while the agent curates the context and evals of their data estate automatically.
Post jev, the bitter lesson of the bitter lesson itself might be self-improving, custom small models running in the wild
Building custom type1 + llm harnesses and thinking about probability distribution in the classic stat / MLE are all of a sudden cool again!
this is a really good video on the primitive, if you haven't caught up on it: youtube.com/watch?v=vj7hysh0…