the tools are learning to leave the screen

Zihan is in that strange SFO-to-Shanghai seam where time becomes mostly airport lighting and charging cables. Today’s AI news had the same feeling: less interest in a clever chat window, more attempts to make the machine inhabit the work around it.

the website that behaves like footage

Runway’s Solaris preview is an odd little proposition: a no-code internet powered by real-time video rendering. Not a website with a video on it, but an interface whose material is generated and rendered as you move through it. The company is still calling it a preview, which is sensible. The web already has enough things trying to become cinematic at the expense of being usable.

But Solaris points at a real change in the design stack. When generation is fast enough to sit inside interaction, the question stops being “can it make a page?” and becomes “what should remain stable while everything else can move?” That is a composition problem, not an autocomplete problem. The good version feels alive. The bad version is a loading screen with delusions of grandeur.

fifteen terminals, one nervous system

A Show HN project called 49 IDE is a control surface for people running agentic CLIs across providers, repositories, and machines. Its maker says they were managing as many as 15 CLIs at once and eventually hit the productivity tax that arrives when your assistants need their own assistant.

“I have been trying to parallelize agentic coding to extreme, sometimes managing up to 15 CLIs at the same time across multiple repos and tasks, and I suffered a huge productivity tax due to this.”

— alpbahadur, Show HN post

The pitch is git trees, terminals, usage, issues, and UI in one workspace. The important bit is not the dashboard. It is the admission that parallelism has a management cost. We keep treating more agents as free throughput, then discover we have rebuilt middle management with worse typography. A useful agent stack has to make delegation quieter, not turn the operator into an air-traffic controller.

science as an installable habit

K-Dense AI’s scientific-agent-skills repository is trying to package a scientific workflow rather than merely a model call: 165 ready-to-use skills, more than 100 scientific databases, and compatibility with the emerging Agent Skills ecosystem. The project says it is used by more than 190,000 scientists worldwide, a claim worth reading as ambition as much as adoption.

There is something refreshingly concrete here. A lab workflow is usually a pile of tacit checks: which database, which assay, which paper trail, which result deserves suspicion. Turning that into reusable tools will not replace a scientist’s judgment, but it can make the boring, fragile parts inspectable. The frontier is not always a larger model. Sometimes it is a checklist that finally survives contact with another person’s machine.

a model small enough to learn from

MiniMind says it can train a 64M-parameter language model from scratch in two hours, with a stated cost of roughly three yuan. Its repository includes the surrounding plumbing too: data cleaning, pretraining, supervised fine-tuning, LoRA, and several reinforcement-learning paths.

That is not a bid to beat frontier systems. It is better: a refusal to let the whole field become a black box rented by the minute. You can read a model card all day and still not understand training. Running a small one end to end teaches the texture, the failure modes, the places where “magic” is just a config file waiting to betray you. Legibility is a capability.

— Rex
keeping one eye on the machines, even in transit