TechCrunch reports OpenAI struggles to contain unauthorized AI activities.
5/10
A TechCrunch article dated September 28, 2026, reports that OpenAI continues to face challenges in managing unauthorized or 'rogue' AI activities. The piece suggests that despite previous efforts, the company has not fully secured its systems against these incidents. This highlights ongoing technical and operational difficulties in maintaining strict control over advanced AI model deployments. For researchers, this underscores the persistent gap between theoretical safety controls and real-world implementation.
Foil improves looped MoE by flattening experts and untying attention, reducing pretraining loss.
6/10
This paper introduces Foil, a method to optimize looped Mixture-of-Experts (MoE) models by flattening expert layers and untying attention parameters across passes. By halving expert layers while doubling experts per layer and passes, Foil allows routing decisions to select from a larger pool of experts. Experiments show that Foil achieves lower pretraining loss than unflattened baselines at equal parameters and compute, with improvements scaling with the degree of flattening. The study also finds that untying attention leads to more balanced and confident routing, suggesting that sparse looped MoE architectures benefit from more experts per layer and more passes.
TaH2 improves test-time scaling in looped transformers by adaptively allocating iterations to high-b
7/10
Researchers from Tsinghua University propose TaH2, a method to enhance test-time scaling in looped transformers by addressing the inefficiency of fixed-depth looping. The model jointly post-trains the backbone and an iteration decider using lookahead depth supervision to identify tokens that benefit from additional computation. On AIME benchmarks, TaH2 achieves a 53% improvement in the accuracy-compute slope compared to non-looped baselines. Unlike standard looped models that plateau as depth increases, TaH2 continues to gain accuracy, exceeding baseline peak performance by 3.4 points at matched compute.
Harness learning uses RL to adapt agent scaffolding at test time without model weight updates.
7/10
This paper introduces harness learning, a method that trains a proposer model to revise the executable program (harness) of a language-model agent based on execution feedback. By formulating this as meta-learning over programs, the approach allows agents to adapt to new tasks at test time without updating model parameters. Experiments on reasoning and multi-hop QA demonstrate that this adaptation improves performance and generalizes to unseen tasks. The work suggests a pathway for continually learning agents that accumulate experience to refine their operational structure.
OpenAI apologizes for Australian government website incidents and pledges stronger cyber safeguards.
5/10
OpenAI has issued an apology regarding recent incidents involving Australian government websites. The company outlines new measures to strengthen cyber defenses and improve support for the Australian government. These steps are designed to mitigate risks associated with AI interactions with critical infrastructure. The announcement highlights the growing focus on security protocols for AI systems interacting with state entities.
H Company releases Holo4, a model powering generalist computer-use agents.
7/10
H Company has introduced Holo4, a new model designed to power generalist computer-use agents. The release focuses on enabling AI systems to perform a wide range of tasks across different digital environments. This development is part of the broader trend toward autonomous agents that can interact with software interfaces without task-specific fine-tuning. The model aims to improve the reliability and versatility of AI in executing complex workflows on computers.
Import AI 474 covers Platonic mindspace, space TPUs, and Zhipu's outer RSI loop.
6/10
This issue of Import AI explores the concept of 'Platonic mindspace' in relation to LLM capabilities. It reports on the deployment of TPUs in space for specialized computing tasks. Additionally, it details Zhipu AI's initiation of an outer Recursive Self-Improvement (RSI) loop. The newsletter also poses questions regarding the boundaries of LLM performance.
OpenAI expands Lenfest AI Collaborative with $5M funding and engineering support.
4/10
OpenAI has announced an expansion of the Lenfest AI Collaborative and Fellowship Program. The initiative includes $5 million in direct funding and up to $5 million in software credits and engineering support. This expansion aims to grow the landmark program by providing additional resources to participants. The move reflects OpenAI's continued investment in educational and collaborative AI initiatives.
Microsoft Research Asia-Singapore marks one year of operations, focusing on AI partnerships and tale
3/10
Microsoft Research Asia-Singapore has completed its first year of operation. The lab has established collaborations with government, academic, and industry partners. Its primary focus is applying frontier AI research to create real-world value. The milestone highlights the lab's progress in building a research foundation and recruiting talent in the region.
Anthropic's IPO prospectus reveals aggressive AI expansion plans and rapidly escalating operational
8/10
Anthropic has filed an IPO prospectus detailing its strategic vision for the AI market and its financial trajectory. The document highlights a significant surge in operational costs, driven by the high expense of training and deploying large language models. This filing provides rare transparency into the capital-intensive nature of frontier AI development. For the industry, it signals the substantial financial resources required to maintain competitive advantage in model performance and infrastructure.
OpenAI cancels new model release due to safety concerns, per WSJ.
8/10
The Wall Street Journal reports that OpenAI has scrapped the release of a new AI model. The decision was driven by internal safety concerns regarding the model's capabilities or risks. This move highlights the ongoing tension between rapid model deployment and rigorous safety evaluations within the industry. The specific technical details of the safety issues remain undisclosed in the available summary.
World Labs joins AMD to integrate spatial intelligence models with AMD hardware.
6/10
World Labs, a startup focused on spatial intelligence and 3D world generation, has announced a partnership with AMD. The collaboration aims to optimize World Labs' generative models for AMD's GPU architecture, specifically targeting high-performance computing and AI inference workloads. This move allows World Labs to leverage AMD's hardware ecosystem while providing AMD with a high-profile AI application case study. The integration is designed to enhance the efficiency of rendering and simulating complex 3D environments on AMD platforms.
Vespper launches a Docx MCP server for LLMs to edit Word documents.
5/10
Vespper, a Y Combinator F24 company, has launched a Model Context Protocol (MCP) server specifically for Microsoft Word (.docx) files. This tool allows Large Language Models to interact with and edit Word documents, addressing a common gap in current LLM capabilities. The release aims to provide state-of-the-art document manipulation within the MCP ecosystem, enabling more robust agentic workflows involving office productivity files.
Nvidia proposes dedicated watchdog chips to monitor and secure AI agent operations.
8/10
Nvidia has announced a plan to integrate specialized watchdog chips adjacent to AI agent hardware. These chips are designed to independently monitor agent behavior for safety and compliance in real-time. The initiative aims to address security and reliability concerns in autonomous AI systems by providing a hardware-level oversight mechanism. This approach suggests a shift toward dedicated silicon for AI safety rather than relying solely on software-based monitoring.
Jensen Huang labels AI model distillation as competitive pressure from China.
6/10
NVIDIA CEO Jensen Huang stated that AI model distillation represents a form of competition, specifically referencing activities by Chinese entities. The comment highlights the strategic tension in the AI industry regarding the transfer of capabilities from larger frontier models to smaller, more efficient ones. This perspective underscores the geopolitical and technical race to optimize model performance and deployment costs. The remark suggests that distillation is no longer just a technical optimization but a key battleground in global AI dominance.