Why Markets for Ideas Keep Failing, and What Just Changed

People have been trying to build a market for ideas since at least 1986. The mechanism was never the problem. The buyer was.

A Sudoku in Barbados

On 26 February 2016, at the Financial Cryptography conference in Barbados, Gregory Maxwell bought the solution to a 16×16 Sudoku from Sean Bowe for 0.10 BTC, on the Bitcoin mainnet. Bowe published an encrypted solution with a zero-knowledge proof that it decrypted to a valid grid, and the payment was locked to the hash of the decryption key, so that, in Maxwell’s words, “when the seller collects his payment he is forced to reveal the information that the buyer needs.” Neither man had to trust the other.

That is the whole of the mechanism people keep proposing for a market in ideas, and it shipped ten years ago. The lineage runs further back. Brassard, Crépeau and Robert formalised all-or-nothing disclosure of secrets in 1986, a US patent with a 1999 priority date claims an idea marketplace built on nondisclosing synopses and escrow, and Cerezo Sánchez’s 2018 Raziel paper has a section titled “A Solution to Arrow’s Paradox”. A 2025 paper models a trusted execution environment holding an invention and releasing it against payment, and calls the result an “ironclad NDA”.

The market has still not appeared, and the best evidence for why comes from an exchange with no cryptographic problem to solve at all. IPXI traded standardised licence rights to granted patents, every one published, with clean title. In seven years it made three offerings. One priced, thinly. The flagship, 194 patents essential to 802.11n Wi-Fi against a market it sized at roughly eight billion chipsets, never priced at all, and the company shut six months later. Jorge Contreras’s post-mortem lists the causes: patent holders stayed away, the exchange had no record of enforcing anything, and the standard licence didn’t fit how deals actually got done. One of his interviewees put it in five words: “patent lawyers aren’t commodities traders.”

The parts that look hard from a distance were solved, several times, and the market kept not appearing. My claim is that the binding constraint was never on the seller’s side of the table. A purchase only happens when the value of the thing clears the whole cost of acquiring it: the price, the work of establishing what it’s worth to you, the work of making it fit, and the risk that it doesn’t. For most outside technical work that sum has never cleared, and the terms nobody could compress were the last two. Coding agents compress them. Not to zero, and the rest of this essay is fairly unsparing about that, but far enough that purchases which were never worth attempting might now be worth attempting. That is a narrower claim than it first sounds, and it is still the most interesting thing to have changed about this problem in forty years.

What Arrow actually said

Kenneth Arrow set out the problem in a 1962 NBER chapter, on page 615:

There is a fundamental paradox in the determination of demand for information; its value for the purchaser is not known until he has the information, but then he has in effect acquired it without cost.

In the same paper he named a second problem that the first tends to eclipse: “Any one purchaser can destroy the monopoly, since he can reproduce the information at little or no cost.” Valuation and appropriability. Every cryptographic escrow ever built addresses the first and leaves the second where Arrow found it. I’ll come back to that.

The paradox also bites less often than its fame suggests. Samuli Leppälä’s working paper on unprotected markets for information observes that “we can usually describe what particular information is about without giving it away”, and that buyers routinely price undisclosed information on the seller’s capability and reliability. It bites where a description that would let a buyer price the thing is close enough to the thing to be worth stealing: a trading signal, an exploit, a set of model weights, an algorithm whose value is a number on a benchmark. Define the class operationally, because that is what the mechanism needs. This essay is about work whose relevant properties can be tested before its contents are disclosed, and only that. A product architecture can be described in a sentence, and the sentence is the disclosure.

The conventional fixes don’t attack valuation at all. Patents, as Leppälä puts it following Gans and Stern, work “not by making the pre-disclosure valuation easier but by removing the disincentive to disclose the information.” Employment, the workaround the industry actually uses, routes around it the same way: you hire the person indefinitely and take whatever ideas come attached, with the CV as the filter that keeps that affordable. A zero-knowledge proof against a benchmark is the one workaround that meets valuation head-on. That’s real, it isn’t new, and on its own it has produced a Sudoku.

Grading was never the expensive part

The standard story says markets for ideas failed because evaluating a stranger’s idea cost a senior engineer’s afternoon, and a senior engineer’s afternoon is worth more than the expected value of the average unsolicited idea. The arithmetic is right. The history doesn’t support it.

For benchmarkable work, evaluation has been nearly free since 2006. Netflix scored every Prize submission automatically against a held-out set, and Kaggle has run a scoring server since 2010. Neither turned into a market for ideas. Kaggle, with 29 million registered users, paid out $3,668,000 across 68 competitions in 2025, under four times the single Netflix Prize of 2009. Sixteen years of free evaluation and the money never compounded.

What Netflix says it choked on wasn’t the grading. The Grand Prize ensemble beat the benchmark by roughly ten percent, and Netflix’s engineers wrote in 2012 that “the additional accuracy gains that we measured did not seem to justify the engineering effort needed to bring them into a production environment.” What they did use were two component algorithms from the 2007 Progress Prize, built for 100 million ratings and reworked by Netflix for more than five billion, still running five years later. The same sentence gives a second reason, and it deserves quoting because it is a separate failure mode: “Also, our focus on improving Netflix personalization had shifted to the next level by then.” The next section of that post is headed “From US DVDs to Global Streaming”. The problem moved while the prize was being competed for. So the market delivered a verified artefact against a target that had drifted, and the cost of adapting it exceeded what the drifted target was now worth. Netflix took the two pieces it could absorb and left the prize-winner on the shelf.

Bug bounties are the working market, and they’re large: HackerOne reported $81 million in payouts in the year to mid-2025, to people who needed no degree and no referral. The obvious reading of why is wrong, though. They are not cheap to absorb. HackerOne treats validation and remediation as separate obligations, remediation being “the duration it takes for the customer to fix a submitted vulnerability”, with a reported median of 37 days, and a suggested fix is explicitly optional in a report.

What the researcher supplies is the search, which is the expensive, high-variance half and the half a buyer can’t reliably do cheaper in-house. Fixing it afterwards is ordinary schedulable engineering. The market clears not because absorption is free but because a prevented breach is worth far more than a month of one team’s time, where Netflix’s ten percent was not worth what adapting it cost. Same arithmetic, opposite answer. And the researcher discloses in full before payment, trusting the programme’s reputation rather than any cryptography, which is why bug bounties sidestep Arrow rather than solving him.

The costs sat on either side of the benchmark, never inside it. Somebody senior had to build a test that measured what the buyer actually wanted (Netflix’s own post calls RMSE “a proxy question that was easier to evaluate and quantify”, and the proxy is what the winner optimised), and somebody senior had to make a stranger’s artefact fit a system it was never written for. Both are work a coding agent now does a serious fraction of. Describe a problem in prose and it drafts the held-out set and the scoring rule; when a submission clears the bar, it reworks the code for your scale, wraps it in your interfaces, and reports how much of the gain survives contact with the real system.

Careful about what that licenses, though. Absorption doesn’t become free, as anyone who has watched an agent confidently integrate something subtly wrong will tell you; correctness still has to be established, dependencies checked, deployment approved, the thing maintained. What moves is the threshold. A purchase that had to be worth a senior engineer’s week might now only need to be worth an afternoon, and far more outside work sits above that line than above the old one.

The grade was never the expensive part. The expensive parts were deciding what to measure and making the winner fit, and those are the two an agent can now take a real bite out of.

Why not just generate the ideas too?

If an agent can build a benchmark and integrate a solution, it can also generate ten thousand candidates and keep the best. That’s cheaper than paying a stranger and there’s nobody to negotiate with. It’s the strongest objection to everything above, and it lands further than I’d like.

Search-based systems have already beaten published records in exactly the class of work this market would trade. DeepMind’s AlphaDev found shorter branchless assembly routines for sorting three, four and five elements; reverse-engineered into C++, they were merged into LLVM’s libc++ in April 2022 and are still in the standard library, having displaced routines that had stood for over a decade. In 2025 AlphaEvolve found a way to multiply two 4×4 complex-valued matrices with 48 scalar multiplications instead of the 49 that two rounds of Strassen’s 1969 algorithm require, the first improvement in that setting in 56 years, since reproduced by mathematicians with no connection to DeepMind. Where the space is well defined and the score is cheap, search is a formidable competitor to any human seller.

The tempting reply is that models only interpolate and humans make the sideways jumps. I don’t think that reply survives its own examples, since AlphaEvolve is a search system built around a model and it left the space anyway. Arguing from a model’s training objective to the limits of a system built on top of it is a bet I wouldn’t take.

A word on “agent”, since I’ve been using it loosely and the industry uses it worse. What ships today delegates the action and keeps the objective human: you decide the problem is worth solving and what counts as solved, and the thing goes and does it. That is action autonomy, not objective autonomy, and the whole argument here depends on the distinction holding. The buyer still has to know which benchmark is the right benchmark.

The better answer is duller and doesn’t need machines to have a ceiling. Markets don’t exist because one side possesses a faculty the other lacks. They exist because suppliers differ. An outside seller may have data you don’t, may have spent four years on a problem adjacent to yours, may have a better search process, or may simply have solved this already for somebody else and be able to sell it again for less than your first attempt costs. That seller will use agents too, and should. The question a buyer faces is never “can a machine do this” but “is somebody else’s answer cheaper or better than my own agent’s first pass”, and for a nontrivial slice of problems it will be.

That is a smaller claim than human irreplaceability and a much harder one to refute.

What the thing would look like

A buyer posts a problem in prose. An agent turns it into a benchmark: a scoring rule, the conditions under which the score counts, and a test set that stays hidden. The rule is published, because a receipt reading “12% better than the agreed benchmark” has to name a benchmark somebody agreed to; the data behind it is not, or the seller optimises the answer sheet. Enforcing that is harder than stating it, since anyone who can submit repeatedly turns a fixed hidden set into a training target eventually, which is the oldest failure on Kaggle’s leaderboards.

A second tension is worth naming. A benchmark drawn from one company’s private system produces a receipt other buyers have no reason to trust, while one broad enough for several bidders establishes less about any single production environment. Buyer-specific tests make valuation easier and shrink the pool of bidders, and that trade decides whether this is a market or just a procurement channel.

The seller runs their work inside an enclave or a multi-party computation against the published benchmark and comes out with a proof of the score and nothing else. That is Maxwell’s Sudoku with a benchmark in place of a grid. Buyers bid on the receipt. When payment settles, the plaintext goes to the winner and the winner’s agent integrates it. The cryptography is an escrow, not a vault: secrecy lasts exactly as long as the bidding, and its job is to get the seller paid.

One thing Numerai got right belongs in any serious version. Its participants train on obfuscated data, submit only predictions, and stake NMR on them; correct predictions earn more, wrong ones are burned, and burned NMR “does not go to Numerai, to other users, or to anyone else.” Crucially the stake is settled against live scoring that arrives after the submission, not against the backtest that won the slot, and Numerai’s own documentation warns how easily validation metrics are overfitted. A stake with no later measurement to settle against is just a deposit. It works at scale: JPMorgan secured $500 million of capacity there in August 2025, against assets that have grown from roughly $60 million to $550 million.

One design note. Exclusive single-winner auctions and non-exclusive licences are a genuine choice rather than an obvious improvement. Selling to five buyers pays five times only if each still bids what they would have bid alone, and they won’t, because the advantage they’re buying is one their competitors will also hold. Whether settlement runs on a chain or through a broker with a bank account isn’t the interesting question. A boring escrow company could run this next quarter.

What doesn’t work yet

Six problems. The first two are the ones I’d attack.

Nobody has said who carries the integration risk. Payment settles at reveal. Integration happens afterwards, and that is exactly where the buyer learns how much of the advertised gain survives contact with a real system. The proof attests to the submitted artefact, not to the adapted one, so a seller can be entirely honest and the buyer still end up with two percent where the receipt said twelve. Escrowing part of the payment against a post-integration re-measurement is the obvious answer and it hands the buyer a lever to underpay by integrating badly. I don’t have a clean solution, and any real version needs one before it needs a token.

Verification at scale selects for luck, though not everywhere. Quantopian licensed the outputs of trading strategies it was not allowed to inspect, which is this proposal almost exactly, and shut its community platform in November 2020. As QuantRocket put it, “you can’t look at the source code, only license the algorithm’s outputs”, and once every proof is just a number, the top of a large enough distribution of numbers is noise. Cheap submission worsens it. Two cases differ sharply though: a Sudoku solution is valid or it isn’t, and no amount of sampling fakes it, while a trading signal picked from ten thousand backtests is mostly sampling. Severe for noisy predictive work, mild for deterministic correctness. Staking has the most evidence behind it and carries a cost this essay should own: a stake is a wealth filter, and the nineteen-year-old it keeps invoking is exactly who can’t post one.

The benchmark can be built wrong, and now it can be built wrong quickly. Netflix’s scoring rule measured the proxy and the winner optimised the proxy. The Heritage Health Prize offered $3 million for predicting hospitalisation days, and at its scheduled end in 2013 nobody had cleared the bar; secondary accounts blame the anonymisation needed to protect patient privacy. An agent that builds the benchmark from a prose description removes the cost of building it and does nothing about the judgement of whether it measures the right thing.

Arrow named two problems and this solves neither cleanly. What the escrow does is narrower than it looks: it establishes that a concealed artefact passes a specified test, and makes payment conditional on delivery. That is not valuation. Valuation is what the thing is worth to a particular buyer in a particular system, and this essay’s central example is a verified ten percent improvement that turned out to be worth less than the cost of having it. A benchmark result is an input to valuation, not a substitute for one. On the second problem, appropriability, the escrow does nothing: once the buyer holds the plaintext it can reproduce it forever at no cost. Which is also why the seller gets paid once, at reveal, and not in royalties. Royalties run on audit rights and courts, and an anonymous cross-border seller of fifteen lines can neither audit a hyperscaler nor sue one. Reveal is the last moment she holds anything, so it is the only moment enforcement is real.

The plumbing has known holes. Raziel rejected trusted execution environments outright, calling reliance on Intel SGX “a risky bet” and listing around thirty published attacks, while the 2025 NDAI paper builds on them anyway. Zero-knowledge contingent payments have their own documented failure: Campanelli and colleagues showed in 2017 that a buyer who controls the common parameters can extract partial information without paying. Both have proposed fixes and neither is finished.

It only reaches work whose properties can be tested before disclosure. The scope statement near the top is also the ceiling. An algorithm has a benchmark. A product architecture, a mental model, a new way of framing a problem, none of them do, and those are the ideas whose owners are worst served today. The philosopher whose framework gets scraped and regurgitated to millions is the person one most wants to pay, and she’s the person this mechanism can’t reach.

What survives those six is smaller than a manifesto and larger than nothing. Numerai shows a market in undisclosed work can clear at scale when the stake settles against later scoring. Bug bounties show strangers with no credentials will supply valuable, actionable work in volume when what they supply is the expensive half of the problem. What neither has had is a buyer that can draft the benchmark from a sentence and absorb the result without a senior engineer’s week. That is the new part, it is a change of degree rather than kind, and a change of degree is enough when the threshold is what decides whether a transaction happens at all.

Who sells into it

Strangers, mostly, and there are about to be more of them. Senior roles were 69.3% of software development postings in the first quarter of 2026 against 4.5% entry-level, where the economy-wide split runs nearer 14% and 46%. Whether AI caused that is genuinely contested and I’ll leave it contested here, but the entry-level rung of this industry is thinning while the people who would have climbed it keep arriving.

Kaggle, as the prize-market research puts it, functions as recruiting more than as a market, and bug bounties are a hiring pipeline for security engineers. So the receipt may matter more to a seller than the cheque: a verified, auctioned result is evidence that needs no degree and no referral. Don’t oversell that either. A credential is not an apprenticeship, and if the seller’s agent did much of the work it isn’t a reliable credential. What this offers is a way for somebody outside the system to be paid and be seen, which is less than a first rung and more than nothing.

(Why that rung is disappearing, and whether a receipt can stand in for one, is a longer argument than this essay can carry. It’s the subject of the next one.)

Request for startups

The pieces have been on the board for a long time. Escrow, verification and auctions were solved before most of us were working, and what held the market back was that buying outside work cost more than it was worth in all but a handful of cases. That threshold has moved. How far it has moved is an empirical question nobody has answered, and it is more useful to answer it than to argue about it.

So the thing to build is small and the thing to measure is specific. Take one vertical where the metric is uncontroversial, the way bug bounties started with “does this exploit reproduce”. Run perhaps twenty real purchases. Then report the numbers that decide whether any of this works: how much human review each purchase actually needed, how much of the benchmark gain survived integration, what share of sellers beat the buyer’s own agent given the same problem and the same afternoon, and what the whole thing cost against doing it in-house. If buying loses to generating on most of them, this essay is wrong in a way that will be obvious and cheap to discover. That is the best property it has.

The nineteen-year-old with the good idea is currently a rounding error in somebody’s efficiency calculation. I’d rather be able to buy from her.

References

Mankowitz, D.J., Michi, A., Zhernov, A., et al. “Faster sorting algorithms discovered using deep reinforcement learning.” Nature 618, pp. 257-263, 7 June 2023. Merge: LLVM Phabricator D118029, landed 8 April 2022.

Novikov, A., Vu, N., Eisenberger, M., et al. “AlphaEvolve: A coding agent for scientific and algorithmic discovery.” arXiv:2506.13131, June 2025. Independent reproduction: Dumas, J-G., Pernet, C., Sedoglavic, A., arXiv:2506.13242.

Arrow, Kenneth J. “Economic Welfare and the Allocation of Resources for Invention.” In The Rate and Direction of Inventive Activity: Economic and Social Factors, pp. 609-626. Princeton University Press for the National Bureau of Economic Research, 1962. Passage quoted from p. 615. https://www.nber.org/system/files/chapters/c2144/c2144.pdf

Leppälä, Samuli. “Arrow’s Paradox and Unprotected Markets for Information.” Cardiff Economics Working Paper No. E2013/2, Cardiff Business School. https://orca.cardiff.ac.uk/77948/1/e2013_2.pdf

Gans, J. and S. Stern. “The product market and the market for ‘ideas’: commercialization strategies for technology entrepreneurs.” Research Policy 32(2): 333-350, 2003. Cited via Leppälä.

Brassard, Gilles, Claude Crépeau and Jean-Marc Robert. “All-or-Nothing Disclosure of Secrets.” Advances in Cryptology, CRYPTO ’86, Lecture Notes in Computer Science vol. 263, pp. 234-238. Springer, 1986. DOI: 10.1007/3-540-47721-7_17

Powell, Michael D. (inventor); Pitchware Inc. (assignee). “Method and apparatus for a cryptographically assisted commercial network system designed to facilitate idea submission, purchase and licensing and innovation transfer.” US 7,043,454 B2. Priority 27 December 1999; granted 9 May 2006. https://patents.google.com/patent/US7043454B2/en

Maxwell, Gregory. “The first successful Zero-Knowledge Contingent Payment.” Bitcoin Core, 26 February 2016. https://bitcoincore.org/en/2016/02/26/zero-knowledge-contingent-payments-announcement/

Campanelli, Matteo, Rosario Gennaro, Steven Goldfeder and Luca Nizzardo. “Zero-Knowledge Contingent Payments Revisited: Attacks and Payments for Services.” ACM CCS 2017. IACR ePrint 2017/566. https://eprint.iacr.org/2017/566

Cerezo Sánchez, David. “Raziel: Private and Verifiable Smart Contracts on Blockchains.” arXiv:1807.09484, 2018 (v2 13 February 2020), sections 8.2 and 8.5. https://arxiv.org/abs/1807.09484

Stephenson, Matt, Andrew Miller, Xyn Sun, Bhargav Annem and Rohan Parikh. “NDAI.” arXiv:2502.07924, 11 February 2025. https://arxiv.org/abs/2502.07924

Contreras, Jorge L. “FRAND Market Failure: IPXI’s Standards-Essential Patent License Exchange.” Chicago-Kent Journal of Intellectual Property 15(2): 419-440, 2016. https://studentorgs.kentlaw.iit.edu/ckjip/wp-content/uploads/sites/4/2016/06/5-Contreras-FINAL.pdf

Amatriain, Xavier and Justin Basilico. “Netflix Recommendations: Beyond the 5 stars (Part 1).” Netflix Technology Blog, 6 April 2012. https://netflixtechblog.com/netflix-recommendations-beyond-the-5-stars-part-1-55838468f429

Heritage Provider Network. “HPN Announces Team POWERDOT Wins $500,000 As Current Leader.” PR Newswire, 3 June 2013. https://www.prnewswire.com/news-releases/hpn-announces-team-powerdot-wins-500000-as-current-leader-209924971.html

ML Contests. “The State of Machine Learning Competitions 2025.” https://mlcontests.com/state-of-machine-learning-competitions-2025/

HackerOne. “Hacker-Powered Security Report,” 9th edition, 2025. https://www.hackerone.com/report/hacker-powered-security

HackerOne. Press release, 1 October 2025. https://www.hackerone.com/press-release/hackerone-report-finds-210-spike-ai-vulnerability-reports-amid-rise-ai-autonomy

HackerOne. Press release, 27 May 2020. https://www.hackerone.com/press-release/hackers-earn-record-breaking-100-million-hackerone-0

QuantRocket. “3 Takeaways from Quantopian Shutting Down.” 4 November 2020. https://www.quantrocket.com/blog/quantopian-shutting-down/

Numerai. Tournament documentation, staking and burning. https://docs.numer.ai/tournament/learn

Numerai. “JPMorgan Secures $500m Capacity in Numerai Following Breakthrough Year.” 26 August 2025. https://blog.numer.ai/jpmorgan-secures-500m-capacity/

Numerai. “Numerai Raises $30M Series C At $500M Valuation.” 20 November 2025. https://blog.numer.ai/numerai-raises-30m-series-c-at-500m-valuation/

Aidala, Felix and Sneha Puri. “The Labor Market Is Tilting Toward Seniority.” Indeed Hiring Lab, 23 July 2026. https://hiringlab.indeed.com/2026/07/23/the-labor-market-is-tilting-toward-seniority/

Gallacher, Guillermo. “AI and Job Postings: From Destruction to Creation?” Indeed Hiring Lab, 8 July 2026. https://hiringlab.indeed.com/2026/07/08/ai-and-job-postings-from-destruction-to-creation/