Fable Revolutionizes GPU Benchmarks with a New Mega-Core

Fable and the MegaCore GPU: A Major Technological Advancement

Fable has recently garnered attention in the field of artificial intelligence by developing what is considered the first true MegaCore for GPU, according to officials at KernelBench-Mega. This MegaCore is not only a technical feat but has also proven to be the fastest ever submitted to this benchmark. This advancement reflects the growing capability of AI systems to perform complex tasks, which are essential for research and development in the AI field, particularly in kernel design.

The results achieved by Fable are impressive: using Cuda code on an RTX PRO 6000 Blackwell graphics card, the MegaCore achieved a speedup of 18.71 times compared to an optimized PyTorch baseline. To put this in perspective, other attempts have reached speedups of 14.4 times with Claude Opus 4.8 using Triton, 11.14 times with GLM-5.2, and 4.34 times with GPT 5.5, all also using Triton.

What makes this solution particularly remarkable is its ability to perform a single cooperative kernel launch per decoded token, according to analyses from torch.profiler. In comparison, other well-ranked solutions had to break down the problem into multiple distinct kernel launches per token, ranging from 4 to 14.

The significance of this advancement lies in the ability of AI systems to autonomously develop and improve kernels. This is a fundamental task for research and development in AI. The more AI systems improve in tasks like kernel design, the more efficient they become in the tasks necessary for AI development, potentially leading to recursive self-improvement. Benchmarks like KernelBench-Mega are thus significant indicators of the increasing efficiency of AI systems in building themselves.

The Impact of AI Automation on the Economy

Researchers from the Center for AI Safety (CAIS) and Scale Labs have observed a notable improvement in the ability of AI systems to automate online freelance projects. The success rate of these systems rose from 2.5% in October 2025 to 16.1% in July 2026, according to the Remote Labor Index (RLI).

The RLI assesses the effectiveness of AI systems in completing economically valuable projects online, covering areas such as 3D design, architecture, graphic design, video and animation, audio, data analysis, and web application development.

In a recent update, researchers published the results of three leading AI models: GPT-5.5, Opus 4.8, and Fable 5, which achieved success rates of 6.3%, 8.3%, and 16.1%, respectively. This rapid increase in the skill level of AI agents highlights the speed at which these systems are advancing in economically significant tasks.

The evaluated tasks include projects such as ring design, production of animated advertising videos, and creation of floor plans with photorealistic renderings. These examples illustrate the diversity of skills that AI systems can now automate.

The potential impact of AI on employment is a crucial question. As AI systems become more capable, they could replace many tasks currently performed by humans. This raises questions about the future of the economy and employment. What will happen when AI achieves an 80% success rate in these tasks? While new tasks may emerge, it is uncertain whether they will be sufficient to compensate for those that AI will replace. The economy could evolve towards very sparsely populated organizations, heavily focused on AI, surpassing unaugmented humans.

OSWORLD 2.0: A Challenge for AI Systems

OSWORLD 2.0, a complex benchmark, highlights the progress of AI systems in using computers to perform multi-step and multi-program tasks. This benchmark, developed by researchers from several universities and companies, evaluates the ability of AIs to accomplish tasks that take an average of 1.6 hours, which is 48 times longer than tasks in the previous version, OSWORLD 1.0.

OSWORLD 2.0 includes 108 long-term tasks, of which 31 are self-hosted websites. Each task is defined as an autonomous workflow that the agent must complete, with a high-level user goal and a stateful computing environment. Notably, 69.6% of the tasks are estimated to take more than one hour for a qualified human user.

OSWORLD 2.0 comes with a massively expanded set of software, including Slack, LinkedIn, Shortcut, REAPER, MuseScore, WPS, GitLab, Overleaf, LabPlot, Zotero, AWS, as well as websites designed to mimic professional services such as insurance applications, visa requests, and conference management portals.

Despite these advancements, the performance of current agents remains limited. For example, Claude Opus 4.8, the top-performing model, only achieves 20.6% binary accuracy and 54.8% partial accuracy. Performance declines rapidly as tasks become more complex.

We should expect performance improvements here, just as occurred with OSWORLD 1.0; in July 2025, the top-ranked models achieved around 30%, and recent models have achieved about 75% (MiniMax M3; June 2026). We should anticipate a similar rise with OSWORLD 2.0.

Using computers is a fundamental skill for AI to perform a wide variety of economically valuable tasks and to conduct more types of scientific research. Accomplishing tasks in the real world is often not as simple as writing text or computer code; it often requires combining multiple pieces of text and code through different types of software, and sometimes, it is necessary to transmit text and code over the Internet for them to be processed by other software in turn. Benchmarks like OSWORLD 2.0 should be viewed as indicators of how AI systems are improving in executing very complicated and varied tasks on computers. As these results show, computers are already competent in tasks using a limited set of software tools that take a few minutes of work for humans; now, we need to see how quickly they become skilled at using broader software sets and performing tasks that take hours for humans.

JD and the Integration of AI in Inventory Management

JD, often compared to Amazon in China, has unveiled details about its inventory management software, the Oxygen AI Item Center (Oxygen AIIC). This system is crucial for managing JD's immense catalog, which includes tens of billions of SKUs for its 700 million users and millions of merchants.

Oxygen AIIC covers tens of thousands of categories and processes hundreds of millions of item updates daily, operating on Huawei Ascend NPUs. This software demonstrates how AI is integrated into JD's operations to optimize inventory management and improve logistical efficiency.

In conclusion, these developments illustrate the growing impact of AI across various sectors, from technological research to the digital economy and logistics management. AI systems continue to advance rapidly, raising questions about the future of human work and the role of AI in our societies.