Roon’s “great minds discuss flops” tweet is now a prediction market, and traders just piled in to bet against it. A small Manifold market asking whether the OpenAI staffer’s 2023 line about compute, data, and architecture will “age well” by 2027 sits at 9% YES as of September 15, 2026, after roughly three-quarters of its lifetime volume traded in a single day. The sudden activity says less about a fresh news event and more about how the 2026 scaling debate has turned against the tweet’s core claim.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the market is actually asking
The market lives on Manifold under the title “Will Roon’s Tweet about various great minds discussing flops age well by 2027?” It was created by writer and AI commentator Zvi Mowshowitz, and it references a specific December 17, 2023 post from the pseudonymous account Roon (@tszzl): “Great minds discuss flops; average minds discuss data; small minds discuss architecture.”
The framing is deliberately provocative. FLOPs, or floating-point operations, are a rough proxy for the raw compute used to train and run a model. The tweet ranks that compute-first view above data quality and above model architecture. Eliezer Yudkowsky, a prominent AI-risk writer, replied that the statement “will not age well,” and the market exists to settle that disagreement. Per the market terms, resolution runs through a Twitter poll near the deadline asking whether the line aged well, resolving strictly YES or NO with, in the creator’s words, “no aging mid.”
The odds as of September 15, 2026, and what they imply
As of September 15, 2026, the market prices YES at 9% on Manifold. That is a lopsided read: traders are collectively assigning roughly a 91% chance that when the poll runs, the crowd will judge Roon’s compute-first hierarchy to have aged badly. In prediction-market terms, the 9% figure is the implied probability of YES, not a guarantee and not a payout ratio.
Two caveats matter before anyone reads too much into the number. First, Manifold runs on play money called Mana, so the “money” here is reputational rather than financial. Second, this is a thin market: total volume sits near Ṁ7,500 across about 51 trades and 34 holders, with only around Ṁ1,000 actively at stake. A handful of confident traders can move a market that small, which is precisely what appears to have happened.
Why the money moved this week
The eye-catching signal is the pace, not the size. Roughly 76% of the market’s all-time volume changed hands within a 24-hour window heading into September 15, 2026, one of the higher single-day readings on the platform for a niche question like this. That is the “sudden bet” behind the story.
There is no single, cleanly documented headline from the last one to two weeks that repudiates or vindicates the 2023 tweet. Based on the available reporting, the most supported explanation is structural rather than event-driven: the broader 2026 conversation about AI scaling has shifted decisively toward data, inference-time compute, and architecture, the exact factors Roon’s tweet ranked as “average” and “small” minds’ concerns. In a thin market, a shift in the ambient narrative is enough to pull a cluster of traders to the same side at once. Where a specific cause cannot be verified, that ongoing debate is the explanation the evidence best supports, and it should be read as context, not as a confirmed trigger.
The tweet, decoded: flops vs data vs architecture
Roon’s line compresses a real technical argument. The “great minds discuss flops” camp holds that AI progress is mostly a function of scale: build bigger models, train them on more compute, and capability follows almost as a matter of course. In that view, clever architecture tweaks and careful data curation are second-order details that scale eventually washes out.
The counter-argument, the one Yudkowsky gestured at, is that the decisive gains increasingly come from how models are built and what they are fed, not from brute-force compute alone. By 2026 that counter-argument has more evidence behind it than it did in 2023, which is the crux of why the market prices the tweet to age poorly.
The people and platforms involved
Roon (@tszzl) is a pseudonymous OpenAI technical staffer known for widely-read commentary on model capabilities, AI safety, and scaling. He remains active and influential in 2026, which keeps his older posts in circulation and under scrutiny. His employer, OpenAI, has seen a run of high-profile departures through 2026, as covered by CNBC, though none of that is directly tied to this market.
Zvi Mowshowitz, a rationalist writer who tracks AI developments closely, created and will resolve the market. Eliezer Yudkowsky supplies the market’s anchor with his “this will not age well” reply. For readers new to the account, LessWrong’s “Read the Roon” is a useful primer on why his posts get this level of attention. The venue itself, Manifold, is a play-money forecasting platform that lets users spin up markets on almost any question, which is how a three-year-old tweet ends up with an order book.
The 2026 scaling debate that hangs over the bet
The reason the tweet looks shaky in 2026 is that the field spent the year moving away from a pure compute-first story. Naive pretraining scaling, meaning a bigger model on the same recipe, has shown diminishing returns, a pattern some researchers call the scaling wall. High-quality public training data has largely been exhausted, pushing labs toward synthetic data and curation. And the most talked-about advances, reasoning models and test-time compute, are about how a model spends compute at inference and how it is structured, not simply how many FLOPs went into pretraining. Recent work such as the arXiv paper “Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs” folds architecture directly into the scaling math, and general explainers like this 2026 overview of LLM scaling laws document the shift.
| Date | Development | Relevance to the bet |
|---|---|---|
| Dec 17, 2023 | Roon posts “great minds discuss flops; average minds discuss data; small minds discuss architecture” | The claim being judged |
| Dec 2023 | Yudkowsky replies “this will not age well” | Sets up the disagreement the market resolves |
| 2024-2025 | Reasoning models and inference-time compute move to center stage | Elevates “architecture,” which the tweet ranked lowest |
| 2026 | Data exhaustion and “scaling wall” talk spread; MoE and inference-efficiency research proliferate | Undercuts the pure compute-first thesis |
| Sep 15, 2026 | Manifold market at 9% YES after a one-day volume spike | Traders price the tweet to age poorly |
None of this settles the question in a lab. Compute still matters enormously, and data-center buildouts are the physical backbone of the field. But the market is not asking whether FLOPs matter; it is asking whether ranking data and architecture beneath compute will look wise by 2027, and the momentum runs the other way. For a related read on how forecasters are pricing the broader “did deep learning hit a wall” question, see our coverage of the deep-learning wall Manifold market.
What a tech and finance reader should take from a play-money market
A 9% price on a play-money question is not a trade recommendation, and it should not be confused with the regulated, cash-settled contracts on venues like Kalshi. What it offers is a compact readout of expert-adjacent sentiment. The people who trade AI questions on Manifold tend to be close to the field, so a lopsided price is a signal about where informed opinion sits, weighted by conviction rather than by dollars.
The finance-relevant lesson is about market structure. Thin markets are noisy: with only a few thousand units of volume, a single day of activity can dominate the tape and produce a “surge” headline that reflects a small number of participants. That dynamic shows up across forecasting platforms and is worth remembering whenever a market’s 24-hour volume dwarfs its history. Tech Insider has tracked the same pattern in other niche Manifold questions, from whether AI can play random games by 2028 to whether an AI can do magic.
What to watch next
Three things determine where this goes. First, resolution mechanics: because the market settles on a Twitter poll with no partial credit, the outcome hinges on how the question is framed and who happens to vote near the 2027 deadline, which adds real uncertainty on top of the underlying debate. Second, the technical trajectory: any high-profile 2026 or 2027 result that credits raw compute for a capability jump would support YES, while another wave of architecture- or data-driven gains would push it lower. Third, Roon himself, whose continued visibility keeps the tweet in the conversation and the market liquid. For a sense of how quickly AI-adjacent forecasts can swing on sentiment alone, compare the Fed rate-move markets, where headline risk reprices odds in hours.
Frequently asked questions
Is real money at stake in this market? No. Manifold uses play money called Mana, so positions are reputational. That is different from cash-settled exchanges such as Kalshi, which is regulated by the CFTC.
What does 9% YES mean here? It is the implied probability that a future Twitter poll will judge Roon’s tweet to have aged well by 2027. Traders currently see that as unlikely, implying roughly a 91% chance the tweet is judged to have aged poorly.
Did a specific event this week move the market? No clearly documented single event has been verified. The most supported explanation is the broader 2026 shift in the scaling debate toward data, inference-time compute, and architecture, amplified by how thin the market is.
Who decides whether the tweet “aged well”? Market creator Zvi Mowshowitz resolves it based on a Twitter poll near the 2027 deadline, strictly YES or NO.
Why does a 2023 tweet still matter in 2026? Roon remains an influential voice on AI scaling, and the compute-versus-architecture question his tweet stakes out is one of the central debates in the field.
The Bottom Line
Sources
- Manifold — Will Roon’s Tweet about various great minds discussing flops age well by 2027?
- X (@tszzl) — Roon’s original “great minds discuss flops” tweet, Dec 17, 2023
- LessWrong — Read the Roon
- arXiv — Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs
- Build Fast with AI — LLM Scaling Laws Explained (2026)
- CNBC — OpenAI talent exodus raises ‘huge red flag’ ahead of IPO
Prediction markets carry risk and are not investment or betting advice. Market availability is restricted by jurisdiction: Polymarket is not available to US persons, while Kalshi is a CFTC-regulated US exchange. Manifold operates with play-money Mana. Age limits of 18+ or 21+ apply as applicable. If gambling is a problem for you or someone you know, call 1-800-GAMBLER for confidential help.
Related Coverage
- Traders Pile Into Manifold Market on China Seizing Taiwan's Outlying Islands
- Superconductor Bet Heats Up: Manifold Traders Give a Room-Temperature Breakthrough 15% Odds Before 2031
- Manifold Traders Push AI Masters Minecraft by 2035 Market to 95%
- New Zealand Cannabis Decriminalisation Odds Hit 1% as Traders Swarm Manifold Market
- Kalshi Traders Bet on 10% U.S. Unemployment by 2030 as AI Job Cuts Rattle the Market
