Problem
Some deals would leave both sides better off, and still they never happen. To prove that what you have is worth paying for, you usually have to show it. And the moment you've shown it, the other side no longer needs to buy it.
Proving knowledge and giving it away are the same act. This has a ton of cascading effects. Most of the data companies collect is never used. The large majority of patents sit idle. On top of that is a mountain of research and spare capacity nobody can put to work.
AKA Information Asymmetry: This is the oldest problem and most important problem in human history.
Why
The goal of life is to be useful. Our goal is to create the most net positive impact in the world that we can, using our resources, abilities, and intuition.
From first principles, creating value means improving what people can experience, accomplish, or become. Net value accounts for what producing that improvement costs.
Net value = benefits created − resources consumed − harms caused
Our impact is the difference between what happens because we act and what would have happened otherwise, including what those resources could have accomplished elsewhere.
Our net impact = net value with our actions − net value under the best available alternative
To maximize that difference, we seek to do valuable things that others are not doing, or valuable things that others cannot do—whichever comes first.
Do what others are not doing → discover what others have not discovered → develop capabilities others do not have
Develop capabilities others do not have → do what others cannot do → bring otherwise inaccessible outcomes into existence
Problems bounded by social constraints can contain the most alpha: convention, incentives, and failures of coordination can leave physically possible, valuable outcomes unexplored.
Removing those constraints can unlock value across entire industries, because the same barrier may be holding back thousands of people and companies at once.
Information asymmetry is one of humanity’s oldest problems—and one of the greatest barriers to reaching escape velocity. Knowledge, talent, and resources remain disconnected, creating artificial constraints on what we can build and become.
Our goal is to remove those constraints and bring entirely new possibilities into existence: ideas, products, and software the world has never seen. We intend to reshape reality by questioning inherited limits and relentlessly experimenting.
We will draw on every great force available to us: human capital, energy, culture, ancient techniques, and emerging technologies. No discipline is beneath our attention. No assumption is beyond examination.
That means working on everything that precedes an outcome, as well as the outcome itself: allocating talent, starting companies and institutions, directing capital, applying energy, building tools, and connecting people with the knowledge and resources they need to act.
Knowledge + talent + capital + energy + tools + coordination → capacity to create
We begin with physics: what is possible, what it requires, and what is preventing it from happening. Our decisions should directly address those requirements.
Desired reality → physical requirements → missing capabilities or resources → actions that supply them
We then use statistics to reason about which actions are most likely to create the greatest additional value:
Expected net impact = Σ [probability of an outcome × additional net value of that outcome]
Priority ≈ expected net impact ÷ scarce resources required
These are decision frameworks, not claims that every benefit or probability can be measured precisely. Intuition helps us identify possibilities; experiments help us distinguish promising possibilities from mistaken beliefs.
Intuition → hypothesis → experiment → evidence → updated judgment → next action
We want to work on important, deeply ambiguous problems that nobody is working on, and do everything in our power to make progress. In those problems, every step can require something that has never been done before: a new technique, a new allocation of talent, a new institution, or a new way for people to coordinate. Each step can reveal further opportunities.
Unexplored territory → new experiments → new knowledge → new capabilities → new possibilities
Novelty alone does not guarantee impact. The alpha comes from finding valuable possibilities that others have overlooked and developing the ability to realize them.
Our strategy is to stack and compound leverage until we can bring specific, valuable realities into existence faster than anybody else—and, eventually, pursue outcomes that nobody else is positioned to create.
Resources → action → learning → capability → greater leverage → more ambitious action
Compounding leverage → faster realization of valuable outcomes + expansion of achievable outcomes
Elon Musk illustrates the mechanism we mean: accumulated capital, talent, infrastructure, experience, and compounding bets can create an unusual ability to pursue outcomes that others cannot readily reproduce. In our framework, the net positive contribution is the additional value those outcomes create, including how much sooner they arrive, after accounting for costs and harms.
Exceptional net contribution = valuable outcomes that otherwise would not happen + value of bringing outcomes forward − additional costs and harms
Leverage matters because it expands the difference we can make. Its value depends on what we use it to accomplish.
Our central bet is that removing information and coordination failures can create the greatest net value we can contribute. Solving these problems can help every company turn more of its knowledge, talent, and resources into useful outcomes.
Better information → better allocation → more effective action → greater value creation
This is a positive-sum ambition. We want to uncover idle value: knowledge that is unused, talent that is misallocated, resources that are disconnected, and possibilities that nobody has attempted. The aim is for the value created to be additional, expanding what is possible for everyone.
Our total contribution = direct additional value we create + additional value we enable others to create
Progress will not always be visible. But our goal is to rewrite the laws by which humanity exists. The least we can do is try everything.
Problems:
- You can't safely look. To ask the world "who has what I need?", you have to describe what you're working on and that description is the exact thing you can't afford to reveal. Your question is your position or gives up leverage. A biotech publicly hunting for one specific patent is announcing where its program is stuck. So the smart move is not to look.
- You can't safely reveal. Even when two sides find each other, proving your side means showing it, and showing it removes the reason to pay.
This characterizes search and reveal into two groups. Companies willing to announce what they are looking for, creating a directional search, and companies holding assets but not announcing anything (double blind). The double blind companies are not searching for anything in particular.
Directional search:
We know what one party has/needs, and read public trails and turn it into a clear map: who likely holds what idle asset, and who's likely stuck on what. Design good discovery agents. People already do this for go to market, outreach, etc.
Double blind search
Once the public map is good, a user takes a position they'd never announce and matches it privately against that map, using our tooling, on their own side. Or, depending on where agent security research is, a user uses a secure agent network that flags mutually beneficial deals depending on surfacing criteria.
Both of these solutions call for nonexistent agent to agent infra. First, let's analyze the ideal A2A infrastructure. Depending on how enthusiastic you are on harness innovation, you may think that it is too early to design a world for agent to agent interaction. We assume it is not too early. Let’s assume all (a lot of) companies in the future will have company brains (agents or a consortium of agents) that manage context across multiple dimensions of the company. Assuming this, it's reasonable to want a network where these agents can communicate with each other. Bottlenecks towards this future:
- How do you control the IP that goes into the agent (open, potentially solveable)
- How do you control the IP that goes out of the agent (open, potentially solveable)
- How do you protect against prompt injections, etc?
- How do you make the agent act faithfully with its actions?
- How do you control agent interaction in an ideal way in a multiple agent system (open, potentially solvable)
These questions, along with months of interactions with data companies and frontier labs put us on the path forward
Existing Solutions
Why every existing fix falls short (based on conversations)
- The human broker had to be told everything, which makes them the leak, and could only juggle a few relationships at once.
- Data marketplaces (Snowflake, Databricks, AWS Data Exchange) only work if the owner is willing to list what they have (directional, active search)
- Data "clean rooms" let two parties compute over data without exposing raw records but only after they've already found each other and agreed to work together
- Patent and IP marketplaces are public listing models where the good stuff never gets posted (same disclosure reason)
- Expert networks (GLG, Tegus) match knowledge to need using people which is bandwidth-limited and the network sees everything that passes through
- The "trust our secret-protecting box" model relocates it onto a vendor nobody has a reason to trust with their most sensitive position
To solve these problems, we need to work our way up to “trust our secret-protecting box” by creating directional solutions and becoming a best agent to agent negotiation control for non-secrets
The closest incumbent is the boutique merchant bank, which does exactly this with humans who must be told the secret and who scale one hire at a time.
The target:
Stuck for a fixable reason; Both sides would win, and the only thing in the way is that finding and proving the deal would force an exposure neither can afford.
Where the stuck things live: data (companies sit on huge piles they never touch, surveys put dark, never-used company data around 55–68%), patents and inventions (an estimated 75–95% sit idle; only about 5–7% ever become a product), research and know-how (labs quietly re-derive what other labs already have), and spare capacity (idle compute, unused factory lines, money on a balance sheet).
The Path Forward
- Hand broker deals within immediate network to see if we can bridge gaps
- What we learned: People REALLY don’t know what they want; Startups, research labs, individuals are far more receptive to this than big companies we talked to
- Create a loose network to test data flow
- What we learned: Some knowledge has high willingness to flow (ex: marketing strats, warm intros, meeting notes), but data is still hard to find sellers and buyers
- Work closely with select groups to build them autonomous discovery agents
- Currently with: Researchers, labs, startups
- Conduct research and design good agent to agent benchmarks
- Post-train and run RLVR for trust, discovery, and cooperation
- Create autonomous discovery agents → simulation infra
The public-search engine: A model of every business built from its public exhaust, turned into a live, legible map of idle holdings and latent needs. An inferred graph of the whole market.
The private overlay: Tooling that lets a user match their own confidential position against that map, on their side, and act.
Judgment about which matches are worth acting on:
S = V − c − E : do the deal iff S > 0
- S = surplus created
- V = asset's value to buyer B
- c = seller A's cost to transfer it
- E = exclusivity: what A loses by no longer being the only holder (0 for a dead asset; huge for leverage assets, which pushes S negative)
Why it doesn't happen today: discovering S > 0 means exposing yourself first. A party searches only if p x (their share of S) − L > 0, where L = leak cost of revealing your position and p = odds a match exists.
- Money: take fee f from the surplus, with f < S. Surface only if S > f > 0
- Beachhead: dormant assets, where E ≈ 0, so S ≈ V − c
How we make money:
- A slice of the value we create, taken when a deal that otherwise wouldn't have happened actually closes
- Membership to the simulation network
Risks (from people we’ve talked to):
- Is public inference actually good enough?
- Agent to agent evals are hard to make
- The truly trace-less deal stays out of reach. We do unlock double-blind search (see above), but only as far as public search can generate a hunch about who to check.
- "See who has what" invites comparison to ordinary sales-intelligence tools. Our defense is that we surface idle assets and latent needs those tools can't see, and that the money is tied to closed deals, not to access
- We could misread how this market matures and build the wrong layer for it
- High ticket companies won’t feel comfortable putting data into agents for a while
- Deal making is very white-glove, relationship driven, not automatable
- People are not sure how many deals are stuck/how much value-to-be-generated there is
Example Use Cases
Startups
Ramp → Rippling
- Ramp's early B2B outbound playbook is settled internal knowledge; a company still scaling that motion would pay to skip the months of rediscovery Ramp already paid for
Brex → a Series A startup chasing an enterprise CFO
- Brex has a live relationship with a finance chief the startup has cold-emailed for months; the intro costs Brex nothing and unblocks the other entirely
Vercel → Retool
- Vercel built internal load-testing tooling for its edge infra that Retool is about to build from scratch; licensing what already works beats rebuilding it.
Research labs
A Stanford NLP lab → Anthropic
- A grad student's unpublished result is exactly what one Anthropic training problem is stuck on; neither knows the other exists, and the student can't advertise it without scooping themselves
EleutherAI → Allen AI (AI2)
- EleutherAI needs a trusted evaluation set to prove a model; AI2 has that exact benchmark data sitting idle after a prior release
Two labs on the same problem: a DeepMind team → a Meta FAIR team
- Both chase the same result from different directions, each years from where the other already is; surfacing the overlap without either revealing its approach halves the total work
Flapping Airplanes → Argonne National Laboratory
- Argonne's synchrotron generates dense, highly-structured measurement data as pure byproduct of experiments. Its non-redundant physical signal, the kind a sample-efficient method extracts far more from than a scale-first one
Enterprises / data holders
Bloomberg → Citadel
- Citadel needs one specific historical slice of Bloomberg's data but disclosing which slice reveals the strategy; the request stays private and only the data moves
CoreWeave → Mistral
- Mistral's training runs are spiky, so owning GPUs wastes money on idle days; CoreWeave's reserved capacity is cheapest on exactly those off-peak windows, turning its idle time into Mistral's burst compute
Waymo → Physical Intelligence
- Physical Intelligence needs long-tail real-world motion data it can't collect at scale; Waymo has years of labeled driving edge cases sitting inside one product line.
Recursion → a small oncology biotech
- The biotech's shelved compound may hit a target already screened in Recursion's phenotypic data; the answer is dead weight to Recursion and decisive to them
Healthcare / bio
Mayo Clinic → OpenEvidence
- OpenEvidence needs outcome data on a rare condition no biobank holds; rather than search for data that doesn't exist, the match is the offer that makes Mayo willing to generate it, HIPAA intact
Tempus → Isomorphic Labs
- Tempus surfaces genomic variants it can't assign function to; Isomorphic's structure prediction does exactly that, giving both a variant-to-function pipeline neither runs alone
Finance
Blackstone → Carlyle
- Blackstone wants an asset Carlyle holds, but expressing interest first loses leverage and revealing financials weakens Carlyle; their agents explore every version of the deal in a space neither principal sees, and it surfaces only if one works
Research
We created software and research. I had the honor to work with some of the best research minds, advisors and young IOI/IMO talent.
Our journey started from one idea: align the right forces in the world, and you can create something like magic.
This is where the core whitepapers come from, and where they live:
The figures above come from public research and industry surveys (IBM, Splunk, Seagate, WIPO, and academic work on patents and dark data). The underlying ideas trace to Kenneth Arrow on the information paradox, Alvin Roth on what makes a market work, and the broader economics of information.