Deep Learning ‘Hit a Wall’ Market Sits at 4% as GPT-5.6 and Claude Opus 5 Ship

Deep Learning “Hit a Wall” Market Sits at 4% Even as the Debate Reignites. A Manifold prediction market asking whether it will “seem as if deep learning hit a wall by EOY 2025” is drawing top-tier volume this week, yet traders are pricing the idea at just 4% as of August 19, 2026. The renewed interest tracks a fresh round of the scaling-plateau argument, colliding head-on with a summer of new frontier model launches.

The offbeat market everyone is suddenly watching

The contract in question lives on Manifold, the play-money forecasting platform, and was created by AI writer Zvi Mowshowitz. Its title is blunt: “By EOY 2026, will it seem as if deep learning hit a wall by EOY 2025?” The market has attracted roughly 200 traders and around Ṁ62k in cumulative volume, and it ranked among the most-active AI contracts on the platform over the past 24 hours. Because Manifold runs on mana rather than dollars, the “money” here is reputational, but the crowd’s conviction is real: the market has hovered near single digits for weeks, and it currently reads a 4% implied probability of YES. You can view the live contract on Manifold.

What the question is actually measuring

The wording is deliberately slippery, which is part of why it draws attention. Mowshowitz, who resolves the market, defines “hit a wall” as a scenario where the “rate of advancement of underlying capabilities has slowed greatly compared to the pace of 2021 to 2023.” Crucially, the market tracks perceived progress in raw AI capability, not economic impact, product adoption, or revenue. Resolution will lean on consensus opinion, using the best available polling if needed. So a YES outcome does not require deep learning to literally stop. It requires the broad expert conversation, by the end of 2026, to look back and agree that things clearly stalled around late 2025. If you are new to reading contracts like this, our explainer on how prediction markets work walks through how a subjective resolution question gets priced.

The odds, and what 4% implies

A 4% YES price means the market assigns a 96% chance that, by year end, the consensus will not read as “deep learning hit a wall.” In prediction-market terms, this is close to a settled view. For context on translating a percentage into a forecast, see our note on implied probability. A price this low is unusual for a question about something as contested as AI progress, and it tells you the crowd sees the “wall” framing as a minority interpretation rather than the likely narrative. The gap between how loud the debate is and how lopsided the odds are is exactly what makes the contract interesting to a finance-minded reader.

What is driving the fresh volume

The renewed betting maps onto a familiar collision. On one side, the scaling-skeptic case got a second wind. Cognitive scientist Gary Marcus, long the loudest voice arguing that pure large language models face diminishing returns, has repeatedly pressed his case in a widely shared post titled “CONFIRMED: LLMs have indeed reached a point of diminishing returns.” That framing, first sparked by reporting that pretraining gains were slowing, keeps resurfacing whenever a major model launch underwhelms. On the other side, the summer of 2026 delivered a dense cluster of releases that cut against the plateau story, which is precisely why the market keeps snapping back toward NO.

The “wall” case: diminishing returns and saturation

The bear case for capabilities is not fringe. Researchers at MIT’s FutureTech group documented what they called deep learning’s diminishing returns, showing how much more compute each incremental accuracy gain now costs. The core architecture, the transformer, has been largely standardized since 2020, with recent progress leaning on data quality, post-training, and reasoning tricks rather than new foundational ideas. Marcus’s original 2022 essay, “Deep Learning Is Hitting a Wall,” made the structural argument years before the current debate. Venture investors have echoed pieces of it: even scaling optimists concede that naive “bigger model, same recipe” pretraining is producing smaller jumps than it did in 2022. If the mainstream story by December centers on incremental releases, gated access, and a lack of AGI-style leaps, YES bettors argue the “wall” perception could firm up.

The counter-case: a crowded summer of releases

The problem for the YES side is the release calendar. OpenAI shipped its GPT-5.6 family, marketed on its own page as “frontier intelligence that scales with your ambition,” with a three-tier lineup (Sol, Terra, and Luna) that reached general availability in July and saw price cuts on the cheaper tiers on July 30, 2026. Anthropic shipped Claude Opus 5 on July 24, 2026. xAI followed with Grok 4.6 in mid-August. Open-weight challengers from Chinese labs and others kept pushing parameter counts and agentic benchmark scores higher. For a tech reader, the practical takeaway is that price-normalized capability, meaning what you get per dollar of inference, kept improving through mid-2026. Our side-by-side on Claude Opus 5 vs GPT-5.6 vs DeepSeek V4-Pro covers how tightly clustered the current frontier has become.

What the data trackers say

Independent measurement complicates the “wall” thesis further. Research group Epoch AI, which maintains a running benchmark of frontier systems, published an analysis asking whether AI capabilities have accelerated. Its Capabilities Index showed state-of-the-art models improving at roughly 15.5 points per year, described as about double the pace of the pre-2024 period, with three of four tracked metrics showing evidence of acceleration driven by reasoning models. That does not settle the argument, since Epoch measures benchmark performance rather than the fuzzy “feels like a wall” sentiment the Manifold market resolves on. But it is hard to square a hard-plateau narrative with a tracker showing gains speeding up. This tension, loud skeptics versus rising benchmarks, is the whole trade.

Who and what is involved: the money and the messengers

The debate has heavyweight participants on both sides. Marcus anchors the skeptic camp. Mowshowitz, the market’s creator, is a prominent AI commentator whose newsletters track capability progress closely, and his own view has leaned toward continued rapid advancement, consistent with the low YES price. On the investor side, a16z’s Marc Andreessen has spent 2026 arguing that the real story is collapsing inference cost rather than stalled capability, framing AI as undergoing “extreme deflation” in his firm’s 2026 outlook. That cost-collapse view is itself a partial concession to the skeptics: if the frontier were leaping, labs would compete on raw capability, not price. The market sits at the intersection of these arguments, and the crowd has clearly sided with the “still progressing, just cheaper and more incremental” reading.

Timeline: 2026 frontier releases versus the wall narrative

What to watch next

Three things could move this contract before it resolves. First, a genuinely disappointing flagship launch in the fall, where a headline model fails to beat its predecessor by a clear margin, would hand the skeptics a concrete data point. Second, shifts in expert sentiment surveys, since the market resolves on consensus, mean a few influential voices flipping could matter more than any single benchmark. Third, access and regulation: if frontier variants get increasingly gated from public use, ordinary users may perceive fewer visible leaps even as internal capability rises, which is a perception effect the “will it seem” wording is unusually sensitive to. For readers tracking how these AI contracts trade week to week, our roundups on the Manifold AI markets and the PostTrainBench human-baseline odds follow the same crowd.

Frequently asked questions

Does 4% mean deep learning is definitely still improving? No. It means traders think it is very unlikely the year-end consensus will describe late-2025 as a wall. The market measures perception of progress, not progress itself, and it can be wrong.

Is real money at stake? Not in this case. Manifold uses play-money mana, so the roughly Ṁ62k in volume is reputational. Dollar-denominated platforms like Kalshi and Polymarket work differently; see our comparison of Kalshi vs Polymarket.

Who decides the outcome? The market creator, Zvi Mowshowitz, resolves it based on his read of expert consensus at the end of 2026, using polling if needed.

Why do smart people disagree so sharply? Because “hit a wall” is subjective. Skeptics point to expensive, incremental gains; optimists point to reasoning models, falling costs, and benchmark trackers showing acceleration. Both can cite real data.

The Bottom Line

Sources

• Manifold — By EOY 2026, will it seem as if deep learning hit a wall by EOY 2025?
• Gary Marcus — CONFIRMED: LLMs have indeed reached a point of diminishing returns
• OpenAI — GPT-5.6: Frontier intelligence that scales with your ambition
• Epoch AI — Have AI Capabilities Accelerated?
• MIT FutureTech — Deep Learning’s Diminishing Returns
• Nautilus — Deep Learning Is Hitting a Wall
• Andreessen Horowitz — Marc Andreessen’s 2026 Outlook

Prediction markets carry risk and are not investment or betting advice. Odds move constantly and can be wrong. Availability is restricted by jurisdiction: Polymarket is not available to US persons, while Kalshi is a CFTC-regulated US exchange; Manifold runs on 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.

Morgan Reeves

Prediction Markets Editor

Morgan Reeves covers prediction markets, sports analytics, and decentralized forecasting platforms. Background in quantitative finance and behavioral economics. Tracks Kalshi, Polymarket, PredictIt, and emerging US-regulated event-contract platforms since 2022.

View all articles