For frontier builders and operators. Research, dealflow and rooms before they're consensus.
IN THIS ISSUE
Syndicate deal: Trossen Robotics SPV
Three FRC research rooms this week: robotics, cell biology, human preferences
Frontier Forecast: Ground Station Zero // 2030
// 01 FROM KRISTOPHER
 | From Kristopher |
Last week I was one of the hosts of the National Security + Space Innovation Summit, with VetsinTech and the Bay Area Council. Two days aboard the historic Klamath at Pier 9, during Fleet Week and SF Tech Week.
The whole ecosystem was in one room: military innovation units, national labs, investors, and veteran founders building defense and space tech.
The Bay Area Council Economic Institute previewed its first report on the region's defense tech economy. Two numbers stuck out. DIU, headquartered at Moffett Field, grew its funding roughly fivefold, from $191M in FY2023 to $983M in FY2024. And four of the ten most active venture investors in defense and aerospace are Bay Area firms (2025 space report).
The highlight: Artemis II pilot Victor Glover rode Navier's Quanta-D foiling vessel on the Bay with the Blue Angels overhead, the first test pilot outside Navier's core team. Navier is run by my friend Sampriti Bhattacharyya, and being part of it was special. Their full vessel reveal is November 4th: a hybrid electric, hyper long range vessel built to run manned and unmanned.
Friday night we crossed to the USS Hornet in Alameda for Forgepoint's annual Fighter Pilot Roll Call, which doubled as Victor's unofficial retirement party. I met him and his Artemis II crewmates Reid Wiseman and Christina Koch, and spent the night with Air Force, Space Force and Navy aviators. Something I'll never forget!
Some news: I'm now the San Francisco Bay Area Chapter Lead for VetsinTech. Want to get involved? Here's the VetsinTech events calendar.
// 02 DEAL FLOW // ACCREDITED ONLY
Deals: Trossen Robotics SPV
Physical AI needs hardware that works beyond the lab.
For 22 years, Trossen Robotics has built robotics hardware, largely bootstrapped. Its arms, teleop systems, and mobile platforms are designed for research and deployment.
Company-reported footprint: 400+ universities and colleges and 60+ Fortune 500 organizations.
My thesis: existing research customers could become much larger buyers as they scale into fleets. Picks and shovels for physical AI.
Planned Frontier Syndicate SPV on AngelList: a Rule 506(c) offering to invest in Trossen via a SAFE. Terms and memo via the SPV Access list.
For accredited investors only; verification required. Any investment is subject to the SPV’s definitive offering documents on AngelList. Early-stage investing is risky and illiquid; you could lose your entire investment.
// 03 INVITES
Upcoming
At Mission Robotics.
One question: before a robot acts, does it need richer senses or better reasoning? Joseph Lee (Stanford, Somil Bansal's lab) shows that steering a VLA's chain of thought cut collisions 30.4% in driving and barely moved manipulation. Shashank Tripathi (Max Planck Institute for Intelligent Systems) makes the case for contact as the signal for learning from human demos.
At Stanford University School of Medicine.
Ganesh Pusapati (Stanford, Rohatgi Lab) walks through his Nature Genetics study: about 150 previously unreported mutations in colorectal cancer's destruction complex, and how restoring it cut cancer-cell growth in the lab. Angela Oliveira Pisco (Chan Zuckerberg Initiative, Tabula Sapiens) asks what evidence we need before trusting a virtual cell to guide drug discovery.
At Array VC. Speakers TBA.
What should a model learn when human preferences and measured outcomes disagree? Reward models, preference ranking, post-training and evaluation for creative work, where quality is hard to define and people reasonably disagree. If you build systems that learn from human judgment, this is where reward models quietly learn the wrong proxy.
General Creative runs AI video ad contests: brands post a brief, creators compete, and only the winners get paid. They're building a model that predicts which ad wins, and they want a founding CTO with hands-on RLHF, reward modeling or RL experience to lead it, full time in SF with co-founder equity. Interested? Learn more
Coming up
Nov 4: FRC #30: Worlds for Agents: World Models and the Agents That Learn Inside Them
Nov 9: FRC Munich: chapter launch
Nov 18: FRC #32: AI for Financial Markets: Agents, Learning & Decision-Making
Dec 2: FRC #33 NeurIPS Preview
Dec 3: FRC Zurich: chapter launch
Dec 16: FRC #34 NeurIPS Recap
Full calendar: luma.com/frontiersyndicate
// PARTNER EVENTS
| ||
| ||
|
// 04 FROM THE RESEARCH ROOMS
From the Research Rooms: FRC #26, Agents in Production: The Data Layer
Oct 7, an official session of TECH WEEK by a16z. Huge thanks to Constructive (@constructive_io) and Dan Lynch for sponsoring, Pebblebed and Nicole Levin for hosting, and GMI Cloud, our FRC compute sponsor.
 | FROM OUR SPONSOR Dan Lynch, Constructive |
Dan Lynch's FRC #26 recap: agents now write the schema and the policies, so the security boundary has to live in the database. Your agents can stay stochastic. The database underneath them can't.
FRC #26 at Pebblebed. Photo: Sumi Shan.
Context as Production Infrastructure: Governed Facts, Invalidation, and the Audit Trail
Enterprises hold agent context to the same bar as the rest of their data: provenance, access control, retention, audit, their own encryption keys.
A commercial graph database built for one big graph broke on millions of small, mostly idle ones, so Zep built Konig, where the graph is the unit of scale.
Stale facts get marked invalid, not deleted, so agents can reason about what changed. Today: 20M+ graphs and under 100 ms p95 retrieval.
Why it matters: agent memory is turning into a governed database problem, and the teams that treat it that way are the ones getting through enterprise security review.
Supporting Our AI Overlords: Redesigning Data Systems to Be Agent-First
Agents don't query like humans. They speculate: hundreds of exploratory probes, parallel plans, what-if branches.
Across 50 parallel attempts, often only 10-20% of plan subexpressions were unique. That is a huge opportunity for sharing work.
"You don't really need optimizers now, you need satisficers." Neon already sees ~20x more branches and ~50x more rollbacks from agents than from humans.
Why it matters: the dominant database workload is about to change shape, and whoever builds the agent-first data layer owns a new infrastructure category.
Pre-read: the CIDR 2026 paper, coauthored with Ion Stoica and Matei Zaharia.
// 05 FRONTIER FORECAST
Frontier Forecast
A Frontier Forecast takes signals that are already real and plays them forward into one plausible future, so you can see where a market is heading before it's consensus. I'm a certified futurist through the Institute for the Future, and I've used these same methods with governments and Fortune 500s. Each one ends with what to build, what to invest in, and what to watch.
Kristopher
It's 2030. A two-year-old startup gets a tasking request at 3 AM, typed in plain English. Within minutes, three of its satellites have re-pointed and imaged the area, one of them launched on 48 hours' notice the month before. An onboard model compares the new frames against yesterday's, flags what changed, and throws away the rest, so only a few megabytes ever leave orbit. A higher-resolution satellite gets cued for a closer look on its next pass. By breakfast, an agent has written the brief, attached the evidence and routed it to the operator. Nobody touched a ground station.
Signals
Rocket Lab launched for the Space Force's VICTUS HAZE mission just 16 hours 42 minutes after the Notice to Launch, then had its spacecraft chasing and photographing a target satellite in under 59 hours (Rocket Lab). The target was True Anomaly's Jackal, flown on a rideshare (Space Systems Command). Why it matters: hours-to-orbit is now a contract requirement, and the other spacecraft in the mission came from a startup.
Planet's Pelican-4 used onboard NVIDIA Jetson Orin modules to identify airplanes over Alice Springs from orbit, in real time. Planet's words: "the insight arrives before the image even downloads." Why it matters: the onboard model that flags change in the scenario has already flown.
Loft Orbital and NASA JPL ran Google's Gemma 3 on the YAM-9 satellite, tasked it "with a prompt instead of a rewrite," and got plain-language answers about its own imagery (IEEE Spectrum). Why it matters: a satellite taking intent in English and handing back prose is the first draft of the agent-written brief.
Drivers
Speed is becoming the spec. Proof: VICTUS HAZE required launch within 24 hours' notice, commissioning within 72 and on-orbit operations within 84, and Rocket Lab beat all three.
Inference is moving onto the spacecraft. Proof: Pelican-4's real-time aircraft detection on Jetson Orin.
Tasking is turning into language. Proof: YAM-9 took a prompt and returned plain-language answers about its own imagery.
Trend
Defense and space are merging into one market built on proliferated, software-defined constellations that think before they downlink.
From: a few exquisite satellites, raw data, analysts on the ground.
To: hundreds of small satellites, onboard models, briefs delivered by software.
If this future happens...
🛠️ Build: onboard autonomy and edge inference, the data layer that links constellations, tools that turn a prototype into a program of record faster.
💰 Invest: dual-use hardware with a real government customer, teams that ship software to orbit like it's a cloud deploy, picks and shovels for proliferated constellations.
👀 Watch: the next TacRS notice-to-launch record, how many imaging satellites ship with onboard GPUs, and the first government buyer that accepts a machine-written brief as a deliverable.
// 06 WORK WITH US
Work with Frontier Syndicate
Short on compute? We can help source H100s through B300s, plus colo and power.
Deals: early access to frontier SPVs. Memos hit the list first.
Speak at FRC: got a paper worth arguing about? Bring it to the room.
Sponsor FRC: put your team in front of the Bay's researchers, founders and funds.
// STAY IN THE ROOM
YouTube · Events · LinkedIn (FRC) · LinkedIn (Kristopher) · X (Kristopher) · X (Frontier Syndicate) · Instagram · thefrontiersyndicate.com
Know a builder who should be in these rooms? Forward this.