AI News Roundup: OpenAI Misalignment Framework, Tokeniser Research & More — 17 September 2026

Here is your daily digest of the most significant developments in artificial intelligence for 17 September 2026.

OpenAI Publishes a Framework for Reporting Model Misalignment

In one of the most consequential industry-governance moves of the day, OpenAI has released a formal framework for tracking, investigating, and disclosing model misalignment. The framework is accompanied by six published reports describing unexpected or concerning model behaviours that the company has already observed. This is a notable step toward systematic transparency: rather than handling alignment failures quietly, OpenAI is establishing a structured process for surfacing and communicating them publicly. For anyone following AI safety, this sets a precedent that other labs will likely be measured against.

Reality Checks on the AI Industry — Yegge Shuts Down Gas Town, Databricks Cost Blowout

The Latent Space AINews newsletter delivered a bracing dose of scepticism to round out the week. Two headline items: developer Steve Yegge has shut down his Gas Town project, and Databricks is reportedly seeing Astra infrastructure costs run more than 60% above initial projections. The newsletter frames these as "reality checks" on the more optimistic corners of the AI world — a reminder that real-world deployment costs and project sustainability remain serious constraints, even as capabilities continue to advance.

New Research Untangles What Actually Makes a Good Tokeniser

A paper posted to arXiv takes a rigorous look at a question that has been quietly confounding AI researchers: when two tokenisation algorithms perform differently, is it because of what they optimise, or how they search for a solution? The two dominant approaches — byte-pair encoding (BPE) and UnigramLM — differ along both axes simultaneously, making clean comparisons impossible. The researchers resolve this by designing two new algorithms that complete a 2×2 grid of objective and search combinations: BottomUpLL and TopDownComp. Training language models across varied sizes, vocabulary sizes, and languages, they find that the search procedure is the dominant factor — bottom-up tokenisers consistently achieve lower bits-per-byte — while the choice of objective has less consistent impact. Importantly, neither dimension showed a reliable relationship with performance on the BLiMP linguistic benchmark, suggesting tokeniser design matters most at the modelling efficiency level rather than downstream language understanding.

A Zeroth-Order Approach to Aligning LLMs With Human Preferences

Another arXiv paper introduces Comparison-based Preference Optimisation (ComPO), a new method for aligning large language models with human preferences. Most direct preference alignment methods optimise a differentiable loss directly on preference pairs, but the authors argue this can cause "likelihood displacement" — problematic behaviour when the margin between preferred and dispreferred examples is small. ComPO sidesteps this by using a comparison oracle to extract directional information without directly optimising that loss. The paper provides convergence guarantees for an offline version and introduces an online variant that uses unlabelled model generations for additional control. Experiments across Mistral, Llama, Gemma-2, Qwen3, and Gemma-3 models show improvements in length-controlled win rates over existing alignment approaches.

Robots That Listen for Force Cues in Generated Video Audio

A creative piece of robotics research shows how audio from AI-generated video can be used to give robots a sense of how hard to push or grip during contact-rich manipulation tasks. Current video-generation-based robot learning produces purely kinematic (motion-only) trajectories, which fail when appropriate contact forces matter. The proposed pipeline — demonstrated on a Franka Panda robot — extracts a desired-force profile from the loudness of contact sounds in generated video, then uses a closed-loop force regulator to track that profile during task execution. The system outperforms a kinematic-only baseline on multiple tasks and also serves as a data-generation engine for training closed-loop policies. The project is described as a zero-shot approach, meaning no task-specific force data is required up front.

Affora: Designing Interfaces That Both Humans and AI Agents Can Use

As AI agents increasingly operate software built for people, a new design system called Affora proposes a way to serve both audiences without building separate interfaces. The system focuses on preserving the interaction meaning of UI components — making action affordances and task state legible to machine readers — while retaining visual freedom for human users. Three controlled studies informed the design guidance, and evaluation on independently authored interfaces showed gains in agent performance where existing interfaces had deficits that Affora addresses. The work positions itself at the intersection of user experience and agent experience design, a space likely to grow rapidly as agentic AI becomes more common.