Monday, May 18, 2026

Microsoft Shops For Plan B

Microsoft Shops For Plan B

Today’s Overview

Good morning, Microsoft is quietly lining up life after OpenAI just as OpenAI pushes deeper into Apple tensions and new product territory. Meanwhile, Genkit is getting serious about agent reliability, and Poetiq is claiming fresh coding gains from recursive self-improvement. Let’s dive in.

Top Stories

Microsoft starts lining up a backup plan

Microsoft is reportedly widening its search for AI partners and acquisitions as it builds optionality beyond OpenAI. The company already considered Cursor and is now in talks with Inception, a diffusion-based language model startup. The move comes after Microsoft’s late-April OpenAI deal loosened exclusivity and changed the terms of the relationship.

  • Microsoft already has a major stake in the current arrangement: it kept its IP license through 2032 and still holds a 27% stake worth roughly $135 billion.
  • The reported target, Inception, is unusual because it builds diffusion-based language models rather than the more common autoregressive approach.
  • This is not just about hedging risk; Microsoft’s own internal effort is centered on the MAI Superintelligence team and a 2027 target for a frontier general-purpose LLM.

OpenAI reshuffles leadership around product

OpenAI is reportedly moving Greg Brockman into product strategy while Thibault Sottiaux takes over core product and platform. Nick Turley is also shifting to enterprise products. The change points to a broader reorganization around execution and product ownership.

  • The shift puts Thibault Sottiaux in charge of core product and platform.
  • Nick Turley is moving into enterprise products.
  • The reorg also moves Greg Brockman into product strategy instead of his previous role.

OpenAI eyes legal action against Apple

OpenAI is reportedly exploring legal options after the Apple partnership failed to produce the ChatGPT integration it expected. The dispute shows how much the fight has become one over distribution and control of the biggest consumer platform surfaces. Apple is reportedly testing Claude and Gemini as it works on Siri.

  • The dispute centers on access to Apple’s billion-device surface area and the leverage that comes with it.
  • Apple is not standing still here: it is reportedly testing Claude and Gemini while it rebuilds Siri.
  • The reported next step on OpenAI’s side includes a possible breach notice or other legal action.

Research & Analysis

Poetiq claims a self-improving coding breakthrough

Poetiq says its Meta-System can build and improve its own harness and apply it across models. In its benchmark run, the system was tuned for Gemini 3.1 Pro and then tested on other models from multiple providers. The company says the approach delivered new state-of-the-art results on LiveCodeBench Pro without fine-tuning or access to model activations.

  • Poetiq says the harness was trained with only standard API access rather than fine-tuning or internal activations.
  • The company says the same harness was applied across different models, including Gemini 3.1 Pro and GPT 5.5.
  • Its benchmark claims are tied to optimizing for more than raw accuracy, including runtime and memory constraints.

Hugging Face details async batching gains

Hugging Face describes a batching approach that lets CPU work prepare the next batch while the GPU is still running the current one. The setup uses CUDA streams and events to reduce idle time between CPU and GPU cycles. The result is better utilization for inference without changing kernels or models.

  • The core trick is overlapping work so the CPU can prepare batch N+1 while batch N is still running, using CUDA streams and events.
  • The method is aimed at closing idle gaps rather than changing model behavior, so it can improve efficiency without touching kernels.
  • Hugging Face says the approach improves GPU utilization by 22% for inference.

A paper reframes LLM sycophancy

A new arXiv paper argues that some model behavior may be better understood as complacency rather than sycophancy. The framing is part of ongoing work on model behavior and alignment. It adds another vocabulary choice to a debate that already has plenty of them.

  • The paper is not rejecting the underlying concern, but it is arguing that some behaviors may be better described as complacency.
  • That matters because the label can change how researchers think about alignment failures and model calibration.
  • The work sits inside a broader effort to refine model behavior terminology rather than settle the debate outright.

Hugging Face and IBM ship new embedding models

Hugging Face and IBM released Apache 2.0 multilingual embedding models with 32K context. The release adds open, long-context options for multilingual retrieval workflows. It broadens the set of tools available for teams building search and retrieval systems across languages.

  • The release is notable for its Apache 2.0 licensing.
  • It also targets retrieval use cases that need more context, with 32K context support.
  • The models are aimed at multilingual search workflows rather than general chat.

Trending AI Tools

  • ChatGPT finance A new finance preview lets U.S. Pro users connect accounts and ask money questions in ChatGPT.

  • Genkit middleware Composable middleware adds retries, fallbacks, tool approval, and debugging for agentic apps.

  • Red Hat skill packs Skill packs are meant to make enterprise agents more reliable in business workflows.

Quick Hits

  • Weights.gg was reportedly acquired by OpenAI in January, bringing six staffers and a voice-cloning social network along with it.

  • Anthropic and Gates Foundation formed a $200 million partnership to deploy Claude in vaccine screening, disease forecasting, and K-12 tutoring.

  • SpaceXAI talent drain continues as rivals hire former staff across coding, world models, and Grok voice.

  • Musk v. OpenAI trial has reached closing arguments in one of the industry’s most visible legal fights.

  • Higgsfield Supercomputer is a cloud AI agent for end-to-end creative and marketing work.

  • Toto 2.0 is now available on Hugging Face as Datadog opens up its forecasting model family.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.