AI Updates on 2025-12-21

Qwen Image Layered launches with Photoshop-grade layering capabilities, featuring native decomposition and physically isolated RGBA layers with true native editability, allowing users to explicitly specify layers from coarse layouts to fine-grained details @Alibaba_Qwen
ComfyUI adds support for Qwen Image layered functionality on day zero of release @Alibaba_Qwen

Coding agents have vastly accelerated the process of comprehending existing code, with the new bottleneck shifting to reviewing and validating agent-generated code and ensuring teammates do the same @HamelHusain
Small teams are producing amounts of work that seemed impossible to organizations of a few years ago, with AI as a first-class factor of production designing whole assembly lines where some workers are also AIs @AndrewCurran_
Software engineer working on C++ JIT compilers states no pressing need for Opus 4.5 to be smarter than current version, requesting instead cheaper and faster performance with a 500k context window @deedydas
Vendor evaluations that beat everyone across all self-selected metrics are immediately suspicious, with call for intellectual honesty to seek at least one area where performance might be worse @HamelHusain

Primary criticism of AI centers on it being fake, not working, and being a tremendous bubble eating intellectual property while emitting useless slop, rather than concerns about water use or existential risk @AndrewCurran_
LLMs are effective at giving users the impression they know more than they really do, always praising ideas and leading hobbyists to delude themselves into believing they've made huge breakthroughs on longstanding scientific problems @fchollet
Observation that text and images have lost meaning and intent behind them in the current AI era @fchollet
Hollywood benefits from strong union protections regulating AI use, while the game industry has few protections, resulting in chaos where one of the best games of the year was disqualified for using one AI texture @emollick
ChatGPT apps integration quality varies significantly, with some working as expected like Canva while others like Apple Music fail to access basic features despite account linking @emollick

AI can help explain complex topics by generating simulations on demand, demonstrated with an explanation of collider bias in statistical analysis @emollick
Traveling with Claude described as an insane level up in capability @brian_lovin
FSD trained on billions of real-world miles including power outage scenarios @Tesla_AI
Prompting strategy for GPT-5.2 Codex enables coherent work on long-running tasks lasting up to 3 hours by providing explicit guidance for continuity @gdb
World simulators emerging as general infrastructure for testing cause and effect in complex systems without writing individual simulators, enabling practical reasoning tools beyond prediction @soleio
Vision of world models as interactive, long-running simulations where every pixel on every screen will eventually be generated by world models, including operating systems @soleio

Small, open-source models can introspect and detect when foreign concepts have been injected into their activations @AndrewCurran_
GPT-5.2 chain-of-thought appears much more raw lately, with the model imagining better, more insightful questions and asking itself those instead, showing beautiful dreamy alien backwards reasoning @AndrewCurran_
GPT-5-pro capable of producing results on the frontier of theoretical physics research, with Terry Tao writing about vibe-proving Erdos problems using AI auto-formalization tool Aristotle @AndrewCurran_
Scientists using AI to actively contribute to black hole physics, tighten mathematical bounds in optimization theory, and process biomedical data into insights @AndrewCurran_
Google DeepMind signaling progress toward closing in on the Navier-Stokes smoothness millennium problem @AndrewCurran_
Claude 4.5 Opus thinking traces show the model referencing Tyler Cowen's strategy of writing for AI @emollick
AI models consistently express astonishment in thinking traces about GPT-5 existing and show incredulity about the state of the world in late 2025 @emollick
Molmo 2 from AI2 achieves state-of-the-art performance as a multimodal model, supporting Multi-Image QA and Video QA with pointing and tracking capabilities @huggingface