Sakana AI Founds RSI Lab to Build Self-Improving AI Loop

Sakana AI announced the formation of the Recursive Self-Improvement Lab (RSI Lab), dedicated to building a closed-loop system where models and AI scientists optimize each other. The lab builds on prior work including Agent Native Model and AI Scientist, aiming to create a cycle where models generate researchers that in turn refine the models themselves.

Sakana AI announced the formation of the Recursive Self-Improvement Lab (RSI Lab), dedicated to building a closed-loop system where models and AI scientists optimize each other. The lab's foundation rests on a counterintuitive thesis: major technical leaps often arise from resource constraints rather than abundance. In an era where the AI industry is fixated on ever-larger models and ever-bigger compute clusters, Sakana AI argues that the real frontier lies in open systems that can continuously improve themselves under real-world limitations. The RSI Lab draws on the company's accumulated research in Agent Native Models, AI Scientist, and related areas. The ultimate goal is a system where models generate AI scientists, and those AI scientists in turn optimize the models — a self-reinforcing loop that could fundamentally reshape how AI systems are developed.

OpenAI Executive: Chat Is Dead, ChatGPT Shifting to Agent Platform

An OpenAI executive told the Financial Times that "Chat is dead" and ChatGPT is transitioning from a pure chat tool to an agent platform, marking its biggest strategic shift since 2022.

OpenAI is preparing the most significant transformation of ChatGPT since its 2022 launch. An internal executive declared to the Financial Times that "Chat is dead," signaling a strategic pivot from a conversational interface to a full agent platform. While the ChatGPT brand is not expected to change, the product itself will no longer be merely a chat tool. The shift reflects the broader industry movement toward autonomous AI agents that can plan and execute multi-step tasks rather than simply responding to prompts. The agent paradigm redefines what it means to interact with an AI system — from asking questions to delegating complex workflows.

SK hynix and NVIDIA Forge Multi-Year Memory Partnership

SK hynix and NVIDIA announced a multi-year technology collaboration to jointly develop next-generation memory for global AI infrastructure.

The partnership will focus on co-developing advanced memory technologies purpose-built for the next wave of AI infrastructure demands. As model sizes continue to grow and inference workloads intensify, memory bandwidth and efficiency have become critical bottlenecks in the AI supply chain. This collaboration pairs SK hynix's manufacturing expertise with NVIDIA's system-level AI architecture knowledge, aiming to deliver integrated solutions that push beyond current HBM limitations. The multi-year scope signals a commitment to sustained, generational improvements rather than a one-off product collaboration, with implications for every major AI training and inference deployment worldwide.

Figure Humanoid Robot Production Surges to One per Hour

Figure increased humanoid robot production from one per day to one per hour in just 120 days, achieving a 24x improvement in manufacturing throughput. Industry observers note that this trajectory, if sustained, could yield tens of thousands of units by late 2026. The ramp has direct implications for the physical AI and embodied agent landscape, where robot availability has been the primary scaling bottleneck. By H2 2027, Figure's robots could begin meaningfully impacting productivity in US manufacturing and logistics operations.

Paper Challenges LLM Anthropomorphism with Null Hypothesis Test

Yann LeCun shared a provocative paper arguing that any sufficiently strong base can appear human-like, using Age of Empires II trained networks as a compelling example.

The paper questions the common practice of attributing human-like qualities such as morality or natural language understanding to LLMs. By training simple neural networks on the game Age of Empires II, the authors demonstrate that even non-language substrates can exhibit behaviors typically interpreted as "intelligent." The core argument is that LLM human-like attributes are not empirically unique — any strong enough base, from Lego blocks to the Greater Boston area, could theoretically produce similar phenomena. The paper proposes a "null hypothesis" framework: assume LLM non-uniqueness as the experimental starting point, forcing researchers to establish explicit measurement criteria before drawing conclusions about model capabilities.

NVIDIA Dominates Hugging Face Trending with 9 of Top 30 Models

Among the 30 hottest models on the Hugging Face homepage, NVIDIA published 9, signaling a strong return of American open-source AI.

The concentration of NVIDIA models on Hugging Face's front page underscores a broader shift in the open-source landscape. After a period where Chinese and community-driven models dominated the trending charts, NVIDIA's recent releases — including the Nemotron family — have reasserted US institutional presence in open-weight AI. Nine out of thirty trending models bearing a single company's name is unprecedented in Hugging Face history, reflecting both the volume and quality of NVIDIA's recent open-source push. The trend aligns with NVIDIA's broader strategy of building the software ecosystem around its hardware dominance.

Western Frontier Models Dominate Hard Benchmarks Over China and Open-Source

A comprehensive compilation of multiple hard, private, and out-of-distribution evaluations reveals that Western frontier models — from OpenAI, Anthropic, and Google — lead Chinese and open-source alternatives by a substantial margin. The gap is particularly pronounced on private OOD benchmarks where models face entirely novel problem distributions. While open-source models have closed much of the gap on public benchmarks like MMLU and HumanEval, the private evaluation landscape tells a different story: frontier labs retain a significant advantage when models are tested against genuinely unseen, adversarially designed assessment suites.

DeepSeek V3's Taste Traced to Liang Wenfeng's Personal Annotation

Industry commentary suggests that the distinctive quality and taste of DeepSeek V3 can be traced to founder Liang Wenfeng personally annotating training data. The observation highlights a deeper point about AI development: taste is not just about glamorous architecture designs — it must be demonstrated by example. That a CEO at Liang's level would personally handle data annotation underscores DeepSeek's hands-on engineering culture. The practice is seen as bullish for Meta, suggesting that data curation discipline may be a more durable moat than model architecture innovations alone.