G'day! Five releases from the last couple of days, why each one matters, and how to try it. A Chinese science-AI company has released a free model trained on its own lab notebooks, and Anthropic has made Claude's long-running agents quicker to answer. There's also a 3D voxel model of Warsaw you can wander through, built with an AI's help, a 1950s murder mystery where your job is to fill in the death certificate, and an MIT experiment in which an AI fixed another AI by editing its "brain" directly. About a five-minute read.

Image: Deep Principle (screenshot)

Most AI models learn science from textbooks and papers. Deep Principle, a Chinese company building AI for scientific discovery, has gone further. Its new model, Hydrogen-1.0, started as GLM-5.3, a big open model from China's Z.ai, and was then trained on chemistry, materials science and life-science data, plus records from Deep Principle's own laboratories. Those records include failed attempts, so the model learns how real research decisions play out. Deep Principle shows it repairing a materials simulation and planning a lab experiment, catching a mistake (reusing a pipette tip) in a practice run before anything touches a real lab. It says Hydrogen-1.0 beats OpenAI's GPT-6 Astra on a hard science-research test, and found better new materials candidates than GPT-6 Astra in a five-round design challenge.

Why it matters: the AI models behind recent headline-grabbing science results are locked away. This one is free for anyone to download and build on, which helps universities and smaller labs. The test results are Deep Principle's own.

TRY IT → free to download under the MIT licence from Hugging Face (you agree to share your account details first) or ModelScope. It's huge: 753 billion "parameters" (the internal settings an AI learns), and Deep Principle tested it on a server with eight high-end AI chips, so it's for research labs, not laptops. There's also a hosted API through NeuroSpark (we couldn't find a published price). Deep Principle warns that answers need checking by an expert. Deep Principle: Introducing Hydrogen-1.0 · Hugging Face: Hydrogen-1.0

Image: Anthropic (screenshot)

When an AI agent works on a long job, like fixing code across a big project, it talks to the AI model hundreds of times. Until now, every request had to resend the whole conversation so far, which can run to hundreds of thousands of "tokens" (small chunks of text). Anthropic's new message threads fix that. The agent sends the full conversation once, then only what's new each turn, while Anthropic holds the rest in memory, encrypted and never written to disk. Anthropic says it has already switched this on for most people using Claude Code, its coding agent, and that on the very longest sessions the wait before Claude starts replying has dropped by about 35%.

Why it matters: agents are taking on longer and longer jobs, and the waiting adds up. Shaving a third off the pause before each reply makes them feel snappier without changing how they work.

TRY IT → if you use Claude Code you may already have it, with nothing to switch on. Developers can add it to their own apps now, in public beta on Anthropic's Claude Platform and through AWS, Microsoft Foundry and Google Cloud. There's no extra charge: standard Claude prices and limits apply, and because Claude still reads the whole conversation each turn, it saves time, not money. Threads expire five minutes after the last request. Claude: Make long-running agents faster with message threads · Claude Platform release notes

Image: Piotr Migdał, Warsaw in voxels

Warsaw developer Piotr Migdał has turned the centre of his city into a giant, Minecraft-style model made of one-metre blocks, about 4 by 3 kilometres of it, and you can explore it in your web browser. You can wander past the Palace of Culture and Science, change the date, time of day and weather, and even click on a pedestrian to take control of them (a nod to the classic game Dungeon Keeper). It's built from free public data, including OpenStreetMap, Poland's laser height scans and aerial photos, and the city's tree database. The project's code history lists Anthropic's Claude as co-author on its changes.

Why it matters: it's a lovely example of what one person can now build with an AI coding partner and open data, and a fun way to "visit" a city you may never get to.

TRY IT → free, no sign-up, in your web browser. It works best on a desktop computer, and the first load can take a while. The code is free and open source (MIT licence) if you'd like to try it for your own town. Warsaw in voxels · GitHub: warsaw-vibe-voxel

Image: SUPERTHUMb (Steam)

England, 1959. A string of unexplained deaths has hit Fleming Manor, a country estate outside London, and you're the investigator sent to work out what happened. In each case you examine the scene and the body, question the people involved, then fill in an official death certificate or emergency report: who died, what killed them, and whether it was an accident, suicide or murder. Picking a suspect isn't enough, because a wrong detail can lead to the wrong conclusion, and the game won't confirm your hunches for you. It has 12 episodes, four endings and hand-drawn art in the style of 1960s European comics. Developer SUPERTHUMb pitches it at fans of Agatha Christie and Return of the Obra Dinn, and says no generative AI was used for its graphics.

Why it matters: it puts a fresh twist on detective games, where the paperwork is the puzzle, and it passed 120,000 Steam wishlists before launch. It came out yesterday evening, Australian time.

TRY IT → out now on PC (Steam) for A$22.95, a 10% launch discount off A$25.50 for the first two weeks. It runs on Windows and Mac, and there's a free demo (Windows only). Steam notes it shows murder scenes and mentions alcohol and tobacco. Steam: Death at Fleming Manor

Image: MIT CSAIL researchers (Rikiya Takehi and colleagues)

An AI model's knowledge lives in millions or billions of numbers called "weights", and normally nobody edits them by hand: they're set by training, a slow, expensive process. Researchers at MIT's computer science lab (CSAIL) asked OpenAI's GPT-6 Astra to just read the numbers and change them. Looking only at a small network's raw weights, it worked out which game it had been trained to play. Then they asked it to make Qwen3.5-0.8B, a small free AI model, better at maths. Astra spotted a fault in its weights that made it get stuck in loops, changed just 0.016% of them, and its score on a maths test jumped from 7.9% to 48.1%, with no retraining. The researchers call it "Weight-Space Programming".

Why it matters: if AI can read and fix other AIs directly, it could one day spot hidden problems in a model without testing it, and make improvements far more cheaply than retraining. The researchers say it's "very early" and the abilities are "still limited".

TRY IT → free to read, and it's been the top story on Hacker News since Monday, with more than 2,400 votes. The improved Qwen model is a free download on Hugging Face, there's a free online demo, and the full experiment logs are on GitHub. MIT CSAIL: Sparks of Weight-Space Programming · Hugging Face: live demo

Know someone who'd like this? Forward it on. They can sign up free at https://fivetotry.beehiiv.com/subscribe

Five to Try is curated with AI help and checked by a human. Hit reply and tell us what you tried.

Want your product in front of curious Aussie tech readers? Sponsor Five to Try at [email protected].