Article · 2026-07-06 · 12 moments

Import AI 464: Fables writes GPU kernels; AI automation; and analog computation

Is this the beginning of a new world? ✦ AI generated

GUEST CAIS/Scale Labs researchers GUEST JD (Oxygen AIIC research paper)GUEST Jack Clark (Tech Tales narrator)GUEST CAIS and Scale researchers GUEST OSWorld 2.0 researchers GUEST JD researchers GUEST Jack Clark (Import AI)HOST Jack Clark

Data

AI models' ability to complete real, economically valuable freelance projects end-to-end more than quadrupled in under eight months, rising from 2.5% success in October 2025 to 16.1% in July 2026.

The Remote Labor Index shows frontier AI success on real freelance projects jumping from 2.5% to 16.1% in under a year, with Fable 5 leading GPT-5.5 and Opus 4.8.

transcript

CAIS and Scale researchers: In a July update, the authors publish results from evaluating three recent frontier models - GPT-5.5, Opus 4.8, and Fable 5, which get 6.3%, 8.3%, and 16.1% respectively. "The frontier has more than quadrupled in under eight months, a concrete signal of how quickly economically capable AI agents are advancing," they write.

extends · 1supports · 1

Prediction◆

AI systems are expanding their economically relevant capabilities faster than humans are expanding their comparative advantages, so person-light, AI-heavy organizations will increasingly out-compete unaugmented humans.

Citing the Remote Labor Index's jump from 2.5% to 16.1% success in nine months, Jack Clark argues AI capability growth is outpacing human adaptation, predicting AI-heavy, person-light organizations will take over parts of the economy.

transcript

Jack Clark: 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. Tracking the rate of capability improvement on tests like RLI will help us all judge this for ourselves.

supports · 2

Prediction◆

AI systems are expanding their economically relevant capabilities faster than humans are expanding their comparative advantages relative to AI, so person-light, AI-heavy organizations will increasingly out-compete unaugmented humans and take over parts of the economy.

Jack Clark argues that AI capability growth is outpacing humans' ability to develop new comparative advantages, predicting that person-light, AI-heavy organizations will increasingly dominate the economy despite human innovation and augmentation.

transcript

extends · 1supports · 2

Data

Current AI agents remain far from reliable at long-horizon computer use, with the best configuration, Claude Opus 4.8 with maximum thinking, reaching only 20.6% binary accuracy on OSWORLD 2.0's multi-hour tasks, and performance dropping sharply as tasks lengthen.

OSWorld 2.0's creators find that even the strongest model setup only hits 20.6% binary accuracy on tasks averaging 1.6 hours of human effort, with agents struggling most on hidden-state recovery, tracking many items, and adapting to changing requirements.

transcript

OSWorld 2.0 researchers: Our experiments show that current agents remain far from reliable computer use: the strongest setting, Claude Opus 4.8 with maximum thinking and batched tool calls, reaches only 20.6% binary accuracy and 54.8% partial-score accuracy. Performance drops sharply as tasks grow longer, and agents struggle most when they must recover hidden state, track many items, resolve conflicting information, or adapt to changing requirements.

Data

Current frontier AI agents remain far from reliable at long-horizon, multi-program computer-use tasks, with the best-performing setup reaching only 20.6% binary accuracy on OSWORLD 2.0, though rapid improvement is likely given how quickly scores rose on OSWORLD 1.0.

OSWorld 2.0, a benchmark of multi-hour, multi-program computer-use tasks, shows even Claude Opus 4.8 with maximum thinking only reaches 20.6% binary accuracy, though the field expects a rapid ramp similar to OSWORLD 1.0's rise from ~30% to ~75%.

transcript

extends · 1supports · 1

Fact◆

Fable's AI system autonomously wrote a CUDA megakernel that achieves an 18.71x speedup over an optimized PyTorch baseline using a single cooperative kernel launch per token, beating every other frontier model's multi-kernel approach on KernelBench-Mega.

Fable wrote the first megakernel submitted to KernelBench-Mega, achieving an 18.71x speedup with a single kernel launch per token, versus 4-14 launches for competing models, signaling rapid AI progress on tasks central to automating AI R&D itself.

transcript

Jack Clark: This solution is particularly impressive because “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.

supports · 1extends · 1

Data◆

Fable's AI system wrote the first genuine and fastest megakernel ever submitted to KernelBench-Mega, hitting an 18.71X speedup with just one cooperative kernel launch per decoded token, versus rivals that needed 4 to 14 separate launches.

Fable's AI-written CUDA megakernel achieved an 18.71x speedup on KernelBench-Mega using only one kernel launch per token, beating attempts by Claude Opus 4.8, GLM-5.2, and GPT-5.5, signaling progress toward AI automating its own R&D.

transcript

Jack Clark (Import AI): This solution is particularly impressive because "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.

Anecdote

In a post-2050 society, general-purpose computation was banned and the world rebuilt on analog computers because general computers had grown too powerful, unpredictable, and dangerous, at the cost of untold human lives and economic damage.

In Jack Clark's short fiction 'The Brass Gears of Civilization,' a future civilization bans general-purpose computers as existentially dangerous and rebuilds critical infrastructure (weather, flood, grid modeling) using specialized analog computers, at massive human and economic cost.

transcript

Jack Clark (Tech Tales narrator): So the great restructuring took place. 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.

supports · 1

Anecdote

In a speculative 2050 future, humanity banned general-purpose computing after deeming it too dangerous and unpredictable, rebuilding civilization around analog, task-specific 'world computers' at enormous human and economic cost, though a trillion-dollar general-purpose analog mind may still be possible.

In Jack Clark's fiction piece 'The Brass Gears of Civilization,' general-purpose computers were banned as existentially dangerous, forcing a costly global shift to analog, purpose-built machines for problems like weather and flood prediction, though the risk of a trillion-dollar analog general mind lingers.

transcript

Jack Clark: In the past, we had general computers. But they were deemed eventually too dangerous - too unpredictable. The more powerful they became and the more diffuse the knowledge about them grew, the more they tickled at the tails of various dragons.

supports · 1extends · 1

Fact

JD's Oxygen AI Item Center processes hundreds of millions of item updates per day across tens of thousands of categories on Huawei Ascend NPUs, managing inventory for 700 million users and billions of SKUs.

JD published details of Oxygen AIIC, an LLM/VLM-centric inventory system combining human-AI ontology engineering, self-evolving models, and a 'unified item tunnel' interface, running on domestically-sourced Huawei Ascend chips as part of China's tech-sovereignty push.

transcript

JD (Oxygen AIIC research paper): Oxygen AIIC now covers tens of thousands of JD categories and processes hundreds of millions of item updates per day on Huawei Ascend NPUs

Mechanism

JD's Oxygen AI Item Center scales item understanding across tens of billions of SKUs by separating a continuously updatable ontology knowledge base from a self-evolving LLM/VLM that only needs to match items to retrieved ontology entries, avoiding costly retraining.

JD describes Oxygen AIIC's architecture, which externalizes its evolving product ontology into a separate knowledge base so the LLM/VLM only has to judge item-to-ontology matches, reducing complexity and hallucination while enabling updates without retraining, run on Huawei Ascend NPUs.

transcript

JD researchers: In the semantic search stage, the dynamically evolving ontology is externalized as a separate ontology knowledge base, enabling continuous ontology updates without model retraining. In the discrimination stage, the model only determines whether the item matches the retrieved ontology entries. This formulation substantially reduces task complexity, mitigates model hallucination, and enhances generalization to ontology evolution.

extends · 1supports · 1

Data

The frontier of AI performance on real-world freelance work automation has more than quadrupled in under eight months, with Fable 5 reaching 16.1% success versus lower scores from GPT-5.5 and Opus 4.8.

CAIS and Scale Labs researchers report that among three frontier models tested in July 2026, GPT-5.5, Opus 4.8, and Fable 5 scored 6.3%, 8.3%, and 16.1% respectively on end-to-end freelance tasks, showing rapid capability growth.

transcript

CAIS/Scale Labs researchers: The frontier has more than quadrupled in under eight months, a concrete signal of how quickly economically capable AI agents are advancing

Highlight slides

Fable's AI writes a record-breaking CUDA kernel✦ from: Fable's AI system wrote the first genuine and fastest megakernel ever submitted to KernelBench-Mega, hitting an 18.71X speedup with just one cooperative kernel launch per decoded token, versus rivals that needed 4 to 14 separate launches.AI Capability Is Outpacing Human Adaptation✦ from: AI systems are expanding their economically relevant capabilities faster than humans are expanding their comparative advantages, so person-light, AI-heavy organizations will increasingly out-compete unaugmented humans.Fable's AI Writes a Winning CUDA Megakernel✦ from: Fable's AI system autonomously wrote a CUDA megakernel that achieves an 18.71x speedup over an optimized PyTorch baseline using a single cooperative kernel launch per token, beating every other frontier model's multi-kernel approach on KernelBench-Mega.AI Capability Growth Is Outpacing Human Adaptation✦ from: AI systems are expanding their economically relevant capabilities faster than humans are expanding their comparative advantages relative to AI, so person-light, AI-heavy organizations will increasingly out-compete unaugmented humans and take over parts of the economy.What This Means✦ from: AI systems are expanding their economically relevant capabilities faster than humans are expanding their comparative advantages, so person-light, AI-heavy organizations will increasingly out-compete unaugmented humans.The Economic Consequence✦ from: AI systems are expanding their economically relevant capabilities faster than humans are expanding their comparative advantages relative to AI, so person-light, AI-heavy organizations will increasingly out-compete unaugmented humans and take over parts of the economy.One Launch vs. Many: Kernel Efficiency✦ from: Fable's AI system autonomously wrote a CUDA megakernel that achieves an 18.71x speedup over an optimized PyTorch baseline using a single cooperative kernel launch per token, beating every other frontier model's multi-kernel approach on KernelBench-Mega.One launch beats many✦ from: Fable's AI system wrote the first genuine and fastest megakernel ever submitted to KernelBench-Mega, hitting an 18.71X speedup with just one cooperative kernel launch per decoded token, versus rivals that needed 4 to 14 separate launches.Beating the field✦ from: Fable's AI system wrote the first genuine and fastest megakernel ever submitted to KernelBench-Mega, hitting an 18.71X speedup with just one cooperative kernel launch per decoded token, versus rivals that needed 4 to 14 separate launches.Why It Matters✦ from: Fable's AI system autonomously wrote a CUDA megakernel that achieves an 18.71x speedup over an optimized PyTorch baseline using a single cooperative kernel launch per token, beating every other frontier model's multi-kernel approach on KernelBench-Mega.

Related episodes

Audio · Dwarkesh Podcast

Eric Jang – Building AlphaGo from scratch

[AINews] OpenAI GPT-5.6 Sol / Terra / Luna — restricted to trusted partners

Banning Open Source AI Would Be A Mistake