🧠 How Far Can Self-Improving AI Go Before It Beats Us?

🔥🔥🔥 Three big updates

1. Self-improving AI gets its first real benchmark

A new benchmark called PostTrainBench was developed to test how well frontier models can fine-tune other open-weight AI models autonomously under a fixed compute budget and deadline. Instead of just evaluating raw reasoning or language skills, this benchmark measures how effectively one model can iterate on and improve another.

Early results from this benchmark show that today’s best agents — including OpenAI’s GPT-5.1 Codex Max, Anthropic’s Claude Opus 4.5, and Google’s Gemini 3 Pro — can achieve 20–30% improvements in performance on target models, compared with roughly 60% for a human expert working on the same task. GPT-5.1 Codex Max leads the pack, followed by Claude Opus 4.5 and then Gemini 3 Pro. 

AI systems are approaching the ability to automate parts of AI research itself, closing the gap between human-guided training and AI-driven optimization.

2. Razer’s Project AVA brings AI companions off the screen

At CES 2026, Razer unveiled Project AVA — a 5.5-inch 3D holographic AI companion designed to live on your desk as a physicalized assistant. Instead of a chatbot in a phone app, AVA uses animated holograms with eye-tracking, facial expressions, and audio sensing to interact with you. It’s capable of organizing your schedule, offering task insights, and even helping with gaming tips — all while “seeing” your screen and listening to you. 

You can reserve one with a small deposit now, with broader availability expected later in 2026. Though still early and experimental, this shift toward physicalized AI suggests a new frontier where your assistant isn’t something you open — it’s something you share space with.

3.  Higgsfield Relight puts pro lighting control in everyday hands

Higgsfield AI just introduced Relight, a model aimed at professional-grade lighting control without a physical studio. Users pick a direction, adjust intensity, set color temperature, and rely on built-in presets — all with intuitive 3D positioning and soft-to-hard shadow control.

Instead of wrestling with sliders and lighting rigs in separate apps, Relight lets creators shape the look and feel of scenes quickly and with precision — a win for photographers, videographers, and social creators alike.

🔥🔥 Two pro tips worth trying

1)  AI video is officially catching up to Hollywood

Benny Johnson’s latest post shows an AI-generated short film that feels straight out of a studio — cinematic shots, lighting, and pacing that would’ve cost millions just a few years ago. Viewers are calling it “next-level AI video” and “the end of Hollywood as we know it.” It’s a glimpse into what happens when text-to-video and 3D generation tools finally cross the threshold from demo to believable cinema.If you create content, this is your sign — the production gap between indie and studio is collapsing fast.

(🔗 Watch the clip)

2) Free Claude Skills Library 

The Awesome Claude Skills repository on GitHub curates a free library of Claude Skills that you can drop into your Claude workspace or skill manager.

Think of this as a starter pack of prebuilt prompts + behaviors that make Claude immediately useful for things like meeting summaries, task automation, brainstorming, persona roles, and more — without having to build each one from scratch. It’s a practical, time-saving resource for teams and individuals alike.

🔥 One prompt worth keeping in your toolkit

You don’t need the full prompt text here (you’re providing the image directly in the newsletter). Still, context matters: the prompt below shows how to prompt Claude for deep engineering-style responses that treat your assistant like a design partner — thinking like a craftsman, artist, and engineer. It’s not just about getting answers; it’s about shaping how Claude thinks through problems with first-principles logic and clean architectural reasoning.