AI News Roundup – 14 September 2026: RSI Startup, AI Assistants, and Inference Breakthroughs

Welcome to today's AI news roundup from ItsTimeAi. Here's everything that matters in artificial intelligence today, from a bold new startup chasing recursive self-improvement to breakthroughs in how large language models are served at scale.

Richard Socher Spins Out $5B RSI Startup, Recursive

Richard Socher — a veteran NLP researcher and CEO of You.com — has founded a new company called Recursive, already valued at $5 billion. The startup is focused on recursive self-improvement (RSI), an approach to AI development that has long been considered one of the most consequential — and controversial — frontiers in the field. The venture was discussed in depth on the Latent Space podcast, where Socher outlined his ambitions for what he frames as potentially humanity's last invention. The sheer valuation at founding signals enormous investor appetite for frontier AI research, even at this level of ambition and risk.

Fyxer Builds an AI Executive Assistant People Actually Trust

OpenAI has spotlighted Fyxer, a startup using OpenAI models to act as a personalised AI executive assistant. Fyxer organises inboxes and drafts emails in each individual user's voice — a task that requires nuance, consistency, and genuine personalisation rather than generic output. The company achieves this through a combination of fine-tuning, memory systems, and ongoing real user feedback. The case study, published directly on OpenAI's website, highlights Fyxer as an example of how fine-tuning and memory can close the gap between generic AI output and something users genuinely trust with sensitive professional communication. For anyone watching the practical deployment of AI in workplace productivity, this is a meaningful example of what production-ready AI assistants look like today.

SQD: A New Approach to Serving Subquadratic Attention Models

A research paper posted to arXiv introduces SQD (SubQuadratic Disaggregation), a new inference serving scheme designed specifically for the growing class of language models that use subquadratic attention mechanisms — such as sparse attention, linear attention, and sliding-window attention. The core insight is that existing systems were designed around dense, quadratic attention and make serving decisions accordingly. SQD instead disaggregates inference by splitting decode operations based on the unique computational characteristics of subquadratic attention layers.

The results are significant. Testing on an adjusted 8×B200 heterogeneous system proxy, the researchers report energy efficiency (tokens per joule) improvements of 53% on GLM 5.2, 31% on Nemotron 3 Ultra, and 56% on Gemma 4 31B over GPU-only baselines. An analytical model of a Rubin-plus-LPX system shows up to 3.6× higher throughput and 1.2–1.5× tighter latency over the best attention-FFN disaggregation baseline. As frontier models increasingly adopt subquadratic attention to reduce memory and compute costs, serving infrastructure will need to catch up — and SQD is a concrete proposal for how to do that.

ASTRIL-MPC: Language-Guided AI for Search-and-Rescue Robots

Researchers have published details of ASTRIL-MPC, a framework that brings together learned kinematics, model predictive control (MPC), and large language model guidance to enable articulated tracked robots (ATRs) to autonomously navigate demanding environments like stairwells and cluttered building interiors — exactly the kind of settings encountered in urban search and rescue.

The system works by having a neural kinematics model predict short-horizon movements, while an LLM proposes bounded updates to planning parameters through a safety-checked interface that includes rate limiting and consistency checks. The full control cycle runs within 100 milliseconds. Across three traversal tasks, ASTRIL-MPC improved an aggregate traversal-quality score by up to 71% over a non-adaptive MPC and 67% over a reinforcement learning baseline, while eliminating measurable collision impacts during descent. This is a compelling demonstration of LLMs being used not to generate text, but to dynamically retune a physical robot's planning system in real time.

Hugging Face Publishes Async GRPO with LoRA Training Technique

Hugging Face has published a technical blog post detailing Async GRPO with LoRA across HF Jobs — a training approach that enables asynchronous reinforcement learning from human feedback-style training using LoRA adapters, orchestrated across Hugging Face's infrastructure using a bucket-and-proxy architecture that sidesteps the need for NCCL (the standard GPU communication library). While the post is highly technical, it reflects the ongoing effort to make advanced training techniques more accessible and scalable without requiring tightly coupled GPU clusters. This is relevant for practitioners looking to run GRPO-style training more flexibly and cost-effectively.

That's your AI news for 14 September 2026. Check back tomorrow for the next daily digest from ItsTimeAi.