Import AI 464: Fables… | Jack Clark from Import AI Summary

1. Key Themes

Theme 1: AI Is Beginning to Automate Its Own R&D (Recursive Self-Improvement Signal)

Fable's GPU kernel achievement is not just a performance milestone — it's an early indicator that AI systems are becoming capable of improving the tools used to build AI itself.

"Being able to autonomously develop and improve kernels is one of the fundamental input tasks for being able to do AI research and development... benchmarks like KernelBench-Mega are a meaningful signal on how effective AI systems are becoming at building themselves."

Fable achieved an 18.71X speedup vs. an optimized PyTorch baseline, compared to 14.4X (Claude Opus 4.8) and 4.34X (GPT 5.5) — and did so with a single cooperative kernel launch per decoded token, vs. 4–14 launches for all other high-scoring entries.

Theme 2: AI Automation of Economic Labor Is Accelerating Faster Than Expected

The Remote Labor Index shows AI's ability to complete real, paid online freelance work has more than quadrupled in under eight months.

"The frontier has more than quadrupled in under eight months, a concrete signal of how quickly economically capable AI agents are advancing."

Success rates moved from 2.5% at launch (October 2025) to 16.1% in July 2026, across economically meaningful tasks like 3D modeling, video production, architectural rendering, and data analysis — not toy benchmarks.

Theme 3: Computer-Use Agents Are Graduating from Minutes to Multi-Hour Tasks

OSWORLD 2.0 represents a qualitative leap in what AI agent benchmarks are measuring — tasks that take skilled humans 1.6 hours on average, 48x harder than OSWORLD 1.0.

"69.6% of tasks are estimated to take a skilled human user more than one hour."

Current best performance (Claude Opus 4.8) is only 20.6% binary accuracy — but the historical precedent from OSWORLD 1.0 is instructive: scores went from ~30% to ~75% within roughly a year. The same ramp should be expected here.

Theme 4: Enterprise AI Is Enabling "Self-Updating Businesses" at Country Scale

JD.com's Oxygen AIIC system demonstrates how deep integration of LLMs/VLMs into back-office operations enables businesses to operate at previously impossible scales — tens of billions of SKUs, hundreds of millions of daily updates — with self-evolving models.

"Technologies like Oxygen AIIC are an example of how modern AI tools let us create businesses that have intelligence woven into their back-office functions... which allow them to operate at far larger scales than prior businesses while also having the ability to self-update and learn, often without large amounts of human oversight."

Notably, JD runs this on Huawei Ascend NPUs, a signal of China's technology sovereignty push playing out in production systems.

2. Contrarian Perspectives

Perspective 1: Human adaptation to AI will NOT outpace AI capability expansion — Clark is betting against the consensus optimism

The standard narrative is that humans will innovate, augment themselves, and find new comparative advantages. Clark explicitly rejects this:

"I'm betting the other side: AI systems are expanding their economically relevant capabilities faster than humans are expanding their comparative advantages relative to AI systems."

The evidence: the Remote Labor Index went from 2.5% to 16.1% in eight months. The question Clark poses is whether human innovation speed can outpace both raw AI capability growth and AI's increasing fluency with the same software tools humans use — and his answer is no.

Perspective 2: "Person-nil" organizations are the coming competitive threat, not "AI-augmented humans"

Most discourse focuses on AI augmenting workers. Clark argues the more disruptive force will be fully automated organizations displacing human-staffed ones entirely:

"It's increasingly hard for me to reconcile the continued progress of AI systems with the economy staying the same — rather, it's more likely to me we are about to see extremely person-light AI-heavy (or person-nil) organizations expand to take over chunks of the economy, out-competing un-augmented humans."

Perspective 3 (via Tech Tales fiction): The endgame of AI safety logic could be the abolition of general-purpose computation

Clark's speculative fiction imagines a world where general computation is banned and civilization reverts to purpose-built analog computers due to AI existential risk. While framed as fiction, the underlying logic is a serious extrapolation of current AI safety concerns:

"General computation was banned - walled off as a forbidden technology. We moved the world to analog at the cost of untold billions of harmed human lives and trillions in economic damages. But we had obtained a kind of safety."

The implicit argument: if AI safety advocates are right, the policy response could be economically catastrophic in its own right.

3. Companies Identified

Fable

Description: AI company producing frontier models
Why mentioned: Achieved the highest-ever score on KernelBench-Mega, writing a CUDA GPU kernel with an 18.71X speedup over optimized PyTorch baseline using a single cooperative kernel launch
Quote: "Fable achieved an 18.71X speedup by writing Cuda code on an RTX PRO 6000 Blackwell, compared against an optimized PyTorch baseline."

JD.com

Description: Major Chinese e-commerce platform ("the Amazon of China"), 700M users
Why mentioned: Case study in enterprise-scale AI deployment; their Oxygen AIIC system manages tens of billions of SKUs with self-evolving LLMs/VLMs running on Huawei Ascend NPUs
Quote: "Oxygen AIIC now covers tens of thousands of JD categories and processes hundreds of millions of item updates per day on Huawei Ascend NPUs."

Center for AI Safety (CAIS) / Scale Labs

Description: AI safety research organizations
Why mentioned: Co-authors of the Remote Labor Index, measuring AI automation of real freelance economic tasks
Quote: "Researchers with the Center for AI Safety (CAIS) and Scale Labs have detected a significant improvement in the ability for AI systems to automate online freelance projects."

Snorkel AI

Description: AI/ML data-centric company
Why mentioned: Contributing institution to the OSWORLD 2.0 benchmark
Quote: Mentioned as part of the multi-institution research team for OSWORLD 2.0

Alibaba Qwen

Description: Alibaba's AI model division
Why mentioned: Contributing institution to the OSWORLD 2.0 benchmark
Quote: Mentioned as part of the multi-institution research team for OSWORLD 2.0

Simular / NeoCognition

Description: AI agent/research companies
Why mentioned: Contributing institutions to the OSWORLD 2.0 benchmark
Quote: Mentioned as part of the multi-institution research team for OSWORLD 2.0

4. People Identified

Elliot Arledge

Description: Benchmark maintainer for KernelBench-Mega
Why mentioned: Independently verified Fable's submission as "the first genuine (and fastest) megakernel ever submitted to KernelBench-Mega"
Quote: "The first genuine (and fastest) megakernel ever submitted to KernelBench-Mega, according to one of the benchmarks maintainers."

Jack Clark

Description: Author of Import AI newsletter; co-founder of Anthropic
Why mentioned: Provides editorial analysis and explicitly takes a contrarian position on AI's labor displacement impact
Quote: "I'm betting the other side: AI systems are expanding their economically relevant capabilities faster than humans are expanding their comparative advantages relative to AI systems."

5. Operating Insights

1. Use benchmark leaderboards (KernelBench-Mega, RLI, OSWORLD) as forward-looking capability signals, not just academic curiosities.

The article demonstrates a consistent pattern: benchmarks that seem narrow today (kernel writing, freelance tasks, computer use) become economically relevant faster than expected. OSWORLD 1.0 went from ~30% to ~75% accuracy in roughly a year.

"We should expect performance to rise here, just as happened with OSWORLD 1.0... We should expect the same ramp with OSWORLD 2.0."

Operators building workforce plans or competitive moats should be stress-testing against the projected benchmark trajectory, not today's numbers.

2. The JD Oxygen AIIC architecture offers a replicable blueprint for AI-native enterprise back-office systems: separate the ontology from the model to avoid retraining on every catalog change.

"In the semantic search stage, the dynamically evolving ontology is externalized as a separate ontology knowledge base, enabling continuous ontology updates without model retraining."

This "Semantic Search then Discrimination" design pattern — decouple knowledge representation from model weights — is directly applicable to any enterprise building AI systems over rapidly evolving product catalogs, regulatory databases, or knowledge bases.

3. The single-kernel-launch architecture that made Fable's result exceptional points to efficiency-first design as a differentiator, not just raw accuracy.

"torch.profiler shows exactly ONE cooperative kernel launch per decoded token. By comparison, every other high-scoring entry decomposed the problem into anywhere from 4 to 14 separate kernel launches per token."

For operators building inference infrastructure, architectural elegance (fewer system calls, less coordination overhead) may matter as much as benchmark scores when translating research results into production cost savings.

6. Overlooked Insights

1. China's AI compute sovereignty is now in production at scale, not just a policy aspiration.

JD's Oxygen AIIC running on Huawei Ascend NPUs at hundreds of millions of daily transactions is a quiet but significant data point. This is not a pilot — it is country-scale production deployment on non-Nvidia hardware.

"During the large-scale deployment of Oxygen AIIC, the underlying compute platform encounters two primary technical challenges: model training and inference on Huawei Ascend NPUs, and the efficient use of compute resources."

For investors tracking the Nvidia moat and US export control efficacy, this is material evidence that Chinese enterprises are engineering around the constraint at production scale.

2. The Remote Labor Index timeline reveals the benchmark itself was launched in October 2025 — meaning this entire capability curve has emerged within a single year.

The article notes AI success rate on real paid freelance tasks went from 2.5% to 16.1% between October 2025 and July 2026 — but the implied takeaway is that an entirely new category of economic measurement had to be invented because existing benchmarks weren't capturing economically relevant capability. The creation of RLI is itself a signal that the research community believes labor displacement is imminent enough to warrant dedicated tracking infrastructure.

"What happens to online employment when this reaches 80%?"