Can AI Predict the Future? with Ben Fielding, Co-Founder and CEO at Gensyn | EP 171
Summary
Ben Fielding’s core call is that frontier labs are trapped in a capital-intensive scaling race whose models are unlikely to repay their development costs through usage alone. Distillation and open-weight releases rapidly reproduce capabilities built with tens of billions of dollars, leaving “no protective moat.” Durable value should therefore accrue to products with large, paying, retained customer bases—not temporary benchmark leadership.
The binding constraint on AI has shifted from compute to timely, trustworthy information. In 2021, abundant internet data made compute the bottleneck; now laboratories can build enormous clusters but are “literally combing the planet” for new training material and reliable truth signals. Jastremski framed the escalating stakes as roughly $50–70 billion per gigawatt, with perhaps 20 gigawatts of new capacity this year, 30 next year and 50 in 2028.
Gensyn proved important pieces of decentralized training, but Fielding no longer believes distributed compute is the highest-leverage business. Its verification technology can reproduce and check machine-learning operations across NVIDIA, AMD, x86 and MacBook hardware, while RL Swarm reportedly coordinated about 40,000 devices. Yet centralized pre-training remains faster, compute supply has expanded, and no natural customer currently wants the radically different model architecture decentralized training would require.
The company’s “middle way” preserves decentralization’s benefits without forcing every computation onto a decentralized network. Open1B was trained efficiently in a centralized data center, but anyone can reproduce and audit every operation—not merely inspect open weights or a recipe. Fielding’s distinction is crucial: decentralize verification and trust where necessary, while allocating compute wherever it works best.
Fielding believes forecasting could become AI’s next major wave, despite receiving “probably 10%” of the attention devoted to LLMs. LLMs compress and reformat information but still require humans to judge consequential decisions; predictive systems could instead estimate their outcomes. His product metaphor is “a poker-like odds calculator for everything”—a machine that predicts the future better than people and continuously improves decisions.
The hard forecasting problem is collecting evidence before outcomes are known, then rewarding it after reality resolves. Historical reconstruction is contaminated by hindsight, editing and updated search results, so Gensyn’s Deli project uses an immutable blockchain record, permissionless markets and reliable, tested AI settlement to connect time-stamped claims with later outcomes. Evidence markets could additionally purchase the information behind a forecast, not merely observe a $10 wager.
Fielding sees crypto and AI converging only where decentralization purchases something indispensable. Prediction-market volume, token narratives or philosophical commitment are insufficient; “philosophy doesn’t pay bills.” The defensible intersection is global trust, automated settlement, micropayments and direct rewards for humans acting as information sensors—capabilities he argues cannot be reproduced by a centralized operator.
Deep dive
1. Fielding’s route to decentralization began with parallel search.
Fielding entered deep learning in 2015, when GPU acceleration was turning neural networks from theoretically scalable mechanisms into useful systems. His first paper described a camera-equipped 3D chatbot, but object recognition took roughly five minutes, so it discussed whatever had been in view five minutes earlier—a poor system that nevertheless showed where the technology could go.
His dissertation work moved from computer vision toward neural-network architecture search. Rather than having researchers manually discover one better stack of layers at a time, he used evolutionary and swarm algorithms to explore many candidate networks in parallel; the dissertation’s underlying conviction was that machine learning should be treated as a set of computational principles, not reduced to “backpropagation is always what we use.”
That framing led naturally to decentralization: geographically distributed machines offer the largest possible environment for parallel work. Fielding’s practical motivation was equally mundane—he had several GPUs under his desk, wanted more and found access difficult—revealing resource constraints as a barrier to research rather than merely an infrastructure inconvenience.
2. Horizontal scaling requires changing the algorithm, not enlarging the box.
Fielding described the familiar vertical path from a GeForce GTX 980 to a GTX 1080, then from one GPU to four and eventually to clustered nodes. Every step initially feels almost free because the algorithm stays unchanged, but memory, motherboard buses, interconnect bandwidth and cooling successively become the next constraint.
His preferred analogy is Google’s early transition from vertically scaling PageRank to MapReduce. Reformulating a task into separable pieces can make it “embarrassingly parallel,” allowing additional machines beside one another to scale the task instead of stacking ever-more-expensive capacity onto a single system.
Machine learning contains similarly parallelizable problems, architecture search being one Fielding knows directly. He argues that distributed data centers could collectively hold models 10, 100 or 1,000 times larger than any single facility, but realizing that upside requires new algorithms and a willingness to solve problems absent from centralized clusters.
Jastremski’s observation—worth keeping—is that modern racks, NVLink and 400-gigabit, 800-gigabit or even terabit-class interconnects were designed precisely to make a cluster behave like one machine. Fielding reiterated that vertical scaling delivers faster immediate returns; horizontal alternatives become attractive only when declining efficiency makes architectural change unavoidable.
3. Verifiable execution matters more than verifying the GPU’s identity.
Gensyn initially asked whether idle consumer GPUs around the world could be coordinated. Jastremski illustrated the trust issue with a claimed B300 versus a Mac mini; Fielding’s answer was that neither payments nor an economic network can rest on untrusted execution. The customer must know the requested work happened, regardless of the provider’s hardware specification.
His preferred trust primitive therefore proves the operations, not the hardware specification, reputation or service-level agreement. Periodically checking whether a provider owns a claimed GPU merely creates “whack-a-mole” incentives to game the test; proving exact execution removes the value of that deception while leaving providers free to optimize their infrastructure.
Gensyn says it can reproduce machine-learning operations across NVIDIA GPUs, AMD hardware, x86 machines and MacBook processors. That exact-reproduction foundation supports optimistic proofs and, more slowly, zero-knowledge proofs—the technical result that made genuinely permissionless compute possible even though the company later changed its commercial focus.
4. Centralized compute won the near-term market, forcing Gensyn to pivot.
Fielding’s change of mind is explicit: decentralized training worked technically, but centralized AI attracted resources so quickly that demonstrating a decisive advantage became extraordinarily difficult. Even where frontier systems face diminishing returns, laboratories can push through barriers by spending at a scale decentralized competitors have not matched.
The compute market also improved after Gensyn’s 2021 thesis. More chips, more neocloud providers and increased competition eased the resource bottleneck, while most GPU consumption became concentrated in centralized hyperscalers performing specific tasks.
A decentralized compute network therefore faces a missing-customer problem. Anyone building one must also become the research laboratory designing a fundamentally different model, yet Fielding has not seen a decentralized AI company raise the tens of billions of dollars he believes that frontier competition would require; hobbyists, small GPU users and some sovereign workloads form niches, not the scale Gensyn sought.
Pre-training should consequently remain centralized while it is materially faster and cheaper. Gensyn’s Open1B embodies the “middle way”: it was trained in a data center, yet every training operation can be independently repeated and checked, offering more than open weights or an open recipe while decentralizing audit rather than execution.
5. RL Swarm worked, but distributed post-training still needs the right product.
Fielding sees reinforcement-learning post-training as the strongest remaining opportunity for decentralized compute. A product could train from consenting customers’ activity on their own devices, turning a user base into both an information source and a distributed learning network.
RL Swarm demonstrated the research possibility with roughly 40,000 devices jointly training models on one task. The unresolved constraint is product design: tools such as coding assistants or open-source projects such as Hermes kernels might benefit, but the application and learning system must be built together.
That combination “seems a bit like a miracle,” because one company must solve a product problem and a research problem simultaneously. Gensyn decided it was not best positioned to do both and redirected its infrastructure toward the resource that decentralization could unlock more convincingly: information.
6. Open models turn research leadership into a perishable advantage.
Fielding divides machine learning’s constraints into compute, data and truth or labeling. Compute dominated when the internet supplied more text than researchers could process; self-supervised next-token prediction then eased the human-labeling bottleneck, while reinforcement learning created labels through reward functions rather than armies of annotators.
The constraint has now moved again. Laboratories can add compute faster than they can find useful information, leading them to scour the world for data—Fielding cited reports of books being physically cut apart and scanned—while rule-based environments such as Lean can serve as a reward function.
Distillation makes any capability exposed through an accessible model easier for competitors to reproduce than it was to create. Fielding expects open source to remain a strategic equalizer: lagging laboratories benefit by releasing models that reduce a leader’s advantage, although he suspects today’s open-source advocates might reverse positions if they become the frontier leader.
The lasting moat, in his view, is “a huge customer base” that repeatedly pays for a useful product. He reads OpenAI and Anthropic’s expansion beyond research as recognition of this reality and rejects the idea that Meta lost merely because Llama 4 disappointed: Meta still owns an enormous, deeply embedded user ecosystem.
7. Neocloud economics can trap companies in yesterday’s bottleneck.
Fielding accepts that neocloud operators can exploit temporary compute-market inefficiencies and that one might become a major winner. His objection is strategic: at bottom, the provider remains “a merchant of commodities,” exposed to scale effects that should progressively normalize pricing.
Gensyn considered using a semi-centralized compute marketplace to generate revenue while funding research, then declined. Once revenue arrives, investors ask how to expand that revenue and the organization optimizes around serving it; five years later, the company may discover it is excellent at a segment it never truly wanted.
Fielding’s metaphor is “falling into a swamp.” He is interested in transformational mechanisms that scale machine learning, not an attractive bridge business that makes future pivots harder, particularly as product distribution becomes more important than incremental research advantages.
8. Forecasting could move AI from information retrieval to judgment.
Fielding characterizes today’s LLM as an extraordinarily capable librarian: it compresses huge quantities of multimodal information, restructures them and returns something persuasive. But consequential outputs still demand human verification because the model lacks a sufficient mechanism for understanding causal consequences in the world.
Forecasting receives “probably 10%” of LLMs’ attention, yet he sees it in roughly the position next-token prediction occupied in 2021. The scaling ingredients already exist; the remaining obstacles are engineering and resources, creating an opportunity to transform an academic forecasting task into general-purpose “intelligence for decision-making and judgment.”
His sharpest analogy is online poker, where software continuously displays the probability of every hand, versus live poker, where the player must calculate alone. A sufficiently scaled forecasting system would become “a poker-like odds calculator for everything”—not merely predicting events, but showing the likely consequences of everyday choices.
9. Deli turns time-stamped human evidence into a forecasting flywheel.
Forecasting data cannot safely be reconstructed after the event. A researcher searching one month later cannot know whether a page existed in its original form beforehand or was edited by the outcome; the only reliable method is to capture information contemporaneously, preserve it and score it when the future arrives.
Jastremski likens this to reinforcement learning with an unusually long reward cycle. The system records successive claims about how one event might affect another, waits for the terminal outcome, then returns to the sequence with a reward.
Gensyn’s Deli is a permissionless information market in which anyone can create a market for a future event, participants can stake money on their beliefs, and a reliable, tested AI model automatically resolves it. The immutable ledger establishes what was claimed when, while the wager provides a signal: a financially rational person risking $10 presumably believes they possess relevant information.
Conventional prediction markets often focus on trading volume; Jastremski noted that commissions make this commercially natural, while Fielding argues that some volume incentives actively degrade information aggregation. Gensyn’s proposed evidence markets instead purchase the information behind a forecast and reward it if it proves right, turning people into compensated “sensors in the world” rather than unpaid training-data suppliers.
10. Crypto and AI converge where trust cannot be centralized away.
Fielding distinguishes philosophical support for decentralization from a viable business. Decentralized systems are usually more expensive, so a founder must know exactly “what the hell are you buying”; companies that adopted crypto for attention or capital predictably leave when “philosophy doesn’t pay bills.”
Gensyn’s criterion is pragmatic necessity. Global market settlement, immutable claims and trust among unknown participants require decentralization in a way that ordinary enterprise adoption often does not; wallets, micropayments and on-chain payment flows are valuable additional benefits, but trust is the decisive purchase.
The resulting system could return model-training economics to users. Someone who supplies information that later proves correct could receive part of the reinforcement-learning value directly, creating a large off-policy training loop in which humans and machines both contribute and share rewards.
Fielding’s closing prediction is categorical even if the implementation remains difficult: a “box” that forecasts the future better than people “can exist.” Just as a 2021 chatbot evolved into a tool performing substantial intellectual work, he believes LLMs will become components of a larger predictive system—and forecasting will be AI’s next major wave.