Claude, Given One GPU and 48 Hours, Learns to Align Other AIs

Anthropic's latest fellows research asks a deceptively simple question: can Claude autonomously improve the alignment of smaller models? The answer, unexpectedly, is yes.

Anthropic handed Claude a single GPU and a 48-hour budget, then asked it to improve the alignment of small models. Claude researched the problem, proposed methods, and trained and tested the resulting models entirely on its own. Against a public benchmark covering ten categories of alignment failure, it found repairs that raised the target baseline in every case, without degrading the models' underlying capability.

It researched and proposed methods, then trained and tested the models on its own. It worked surprisingly well.

The experiment sits inside a broader push to see whether frontier models can act as automated researchers rather than mere assistants. If a strong model can diagnose and patch the alignment flaws of weaker systems, the burden of safety work need not scale with the number of models in the wild. Anthropic frames the result as an early but encouraging signal: alignment, long treated as a scarce human specialty, may itself be partly automatable.

The work is published as Fellows Research, a channel the lab uses to test unusual ideas quickly and in public. The headline finding is less about any single benchmark and more about the loop itself, a model that studies a failure mode, proposes a fix, runs the training, and reports back, all without a human in the loop.

Google Ships Gemini Omni 1.1 Flash for Controllable Video

Aimed at production: faster iteration, tighter control, and more polish for generative video.

Google is rolling out Gemini Omni 1.1 Flash, a generative-video release built for creators who want control rather than a roll of the dice. The model makes scenes highly controllable, shortens iteration cycles, and pushes output toward production-grade finish. It is available to try inside Flow, and Runway has added it alongside its own image and video models.

Midjourney Starts Testing Its V8.2 Edit Model

The new model supports instruction-based editing, generation from up to four reference images, brush-based inpainting, and outpainting, and it works with personalization, moodboards, and srefs.

GLM-5.3 Goes Open-Weight, and the Ecosystem Moves in Hours

Z.ai releases its most capable model for agentic coding and cyber defense, and inference platforms ship day-zero support before the day is out.

Z.ai has opened the weights for GLM-5.3, the largest model in its latest line: 744B total parameters with 40B active, a 1M-token context window, and 128K max output. The company kept the GLM-5.2 base and scaled post-training instead, so the foundation beneath the model is unchanged. Within hours, vLLM, SGLang, Ollama, Perplexity, and Bionic had all announced support, a reminder of how quickly open weights propagate through the inference ecosystem.

Rosalind Workbench: Every Scientist, Their Own Research Team

OpenAI's Rosalind Workbench connects scientific questions to specialized models, tools, and reviewable outputs in a single workflow, spanning protein structure, sequence analysis, and sequencing pipelines. Guided tasks, a science viewer, and sequencing analysis let researchers trace a question from raw data all the way to evidence.

MTP, EAGLE-3, DFlash or DSpark? A Guide to Speculative Decoding

There is no universal winner, vLLM argues: the best method changes with the model, the workload, and the speculation depth. A new post breaks down five methods and how to enable and tune each in vLLM.