Eugene Cheah

Eugene Cheah is the cofounder and CEO of Featherless AI and has helped lead the open-source RWKV project. His AI work spans efficient language-model research and the infrastructure that makes open models usable. At Featherless, he described a consequential choice: rent increasingly commoditized GPUs and concentrate on orchestration rather than purchasing new hardware. He co-founded the company with Harrison Vanderbyl, CTO, and Wesley George, COO.[11][3][2][9]

In his October 11, 2024 article, $2 H100s: How the GPU Rental Bubble Burst, Cheah argued that general-market users should rent rather than buy new H100s under the oversupply conditions he described. His explanation connected falling rental prices to resale of unused reserved capacity, fine-tuning replacing some training from scratch, and fewer small and medium foundation-model creators. He retained exceptions for discounted hardware or electricity, deployments where location mattered, and very large frontier-training clusters.[2]

Cheah described Featherless's response as a bet on GPU commoditization. The team still optimized CUDA and inference kernels, but emphasized the orchestration layer: combining rented capacity from AWS spot and other providers, hot-swapping models, and keeping the GPU fleet sized to workloads. He also reported using a custom RWKV foundation model as a low-cost speculative decoder. These were his account of the platform's operation in October 2024.[2]

Cheah is a coauthor of Eagle and Finch: RWKV with Matrix-Valued States and Dynamic Recurrence, first submitted on April 8, 2024. The collaborative work introduced RWKV-5 and RWKV-6 improvements involving multi-headed matrix-valued states and dynamic recurrence, alongside a 1.12-trillion-token multilingual corpus and a greedy-matching tokenizer. His collaborators included Bo Peng, Daniel Goldstein, Quentin Anthony, Alon Albalak, Eric Alcaide, Stella Biderman, and others. Peng, also known as Blink, created the RWKV architecture.[7][1]

At NeurIPS 2024, in a joint presentation with Dan Fu published that December, Cheah described converting a Qwen 32B model by replacing its attention with RWKV linear-attention layers. The procedure froze the feedforward layers while the replacement attention learned, then unfroze them and trained the layers together with a custom learning-rate schedule. He presented conversion as a way to test attention mechanisms using already-trained weights, reducing the need to train from scratch.[3]

Cheah credited Daniel Goldstein, known as Smirky, with research coordination and proposing the conversion idea. He reported mixed preliminary benchmark changes, including a decline in MMLU. His description of the team's experimental method was direct: “The reality is someone came with a code, we tested it, it worked, and then we rationalized later.”[3]

Cheah also cautioned against equating recurrent operation with unlimited memory. In his explanation, a fixed state has finite information capacity: a model can keep running while forgetting earlier details. He separately explained that long-sequence training still needs memory for intermediate states used in backpropagation. A constant inference state therefore does not imply constant training memory.[3]

In a February 2024 RWKV/Eagle presentation, Cheah explained that the team's multilingual data and tokenizer work served users beyond English benchmarks. He said its international contributors wanted the model to work for their “mom and grandma.” In an August 2026 interview, he made the concern personal: his grandmother, he said, speaks seven languages but neither English nor Chinese, and needs AI adapted to her own language.[6][10]

In that interview, Cheah called AGI an architectural search problem and emphasized reliability over increasing model size. His idea of “personal general intelligence” was a system that learns from an ordinary person's corrections, illustrated by a hypothetical robot learning to cook from his mother. He suggested such learning might be possible with 200-million-parameter models; this was a research aspiration. His question was: “How small can we go, for human-level general intelligence?”[10]

Cheah's open-model advocacy includes opposition to concentrated control. In March 2025, he said: “I don't want a future where AI is controlled by the few. I want to empower individuals globally.” In August 2026, he argued that the ability to download an open model and move it elsewhere matters even if his own platform cuts off access.[9][10]

In his September 2026 contribution to AI Frontier Network, Cheah advocated choosing specialized models for particular jobs and locating value in inference, optimization, and orchestration. He described community and developer relations as core organizational functions, with humans retaining judgment about what to build, which models to trust, and where to optimize. He proposed the number of different open models a company runs in production as a useful signal of AI maturity.[11]

Before Featherless, Cheah co-founded UIlicious and served as its CTO. This earlier work supplies context for his infrastructure experience: he and Tai Shi Ling, UIlicious's CEO, built a low-code testing tool after meeting through Entrepreneur First's pioneer Asia cohort.[8][1]

In a Singapore Global Network profile, originally published in 2021 and updated in 2024, Tai described their complementary responsibilities: she brought test-engine and interface expertise, while Cheah brought infrastructure skills for supporting concurrent tests. She recalled months when the founders went without salaries to keep servers running and employees paid. Daimler became a client after its engineers encountered the founders at a Singapore convention.[8]

By August 2023, Cheah was also an RWKV committee member. The March 2024 Latent Space anniversary episode identified him as RecursalAI's CEO; that October, he returned as a guest writer and Featherless cofounder. By the December 2024 post-transformer presentation, he was running Featherless while continuing to help lead RWKV.[1][6][2][3]