The project explores whether a language model can help artists discover unexpected visual effects by completing unfinished “Dweets”: JavaScript demos constrained to 140 characters. The author sees these tiny programs as a creative language, where shared code patterns can lead to very different animations and countless possible continuations. Rather than enumerate possibilities or rely on a general-purpose coding model, the goal was to train an autocomplete system that understands Dwitter’s unusual code-golf idioms and can offer branches for the artist to evaluate, edit, and develop. The author also wanted the system to reflect aesthetic preferences, treating code completion as a way to explore possibilities rather than simply finish a program.
The model, Demoscener, was built from 25,000 Dweets and about 1,300 validation examples, using Qwen2.5-Code-7B as its base. Continued pretraining made the model substantially better at predicting Dweet syntax: perplexity on held-out examples fell from 28.63 to 8.57. Training that increasingly emphasized the site’s most highly rated demos made generated animations more visually compelling and further improved prediction, including for examples not used in that curriculum. The author then used GRPO preference training, personally comparing groups of eight generated animations over roughly 9-10 rounds and inspecting about 40,000 outputs. A blind comparison by a friend favored one of the resulting checkpoints. The final tool presents continuations side by side so users can choose promising branches, tweak them, and keep exploring.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Anthropic has restricted Claude from helping with video game modding, decompiling, and reverse-engineering, amid growing use of AI to alter games. The limits are broad enough to disrupt legitimate development and basic troubleshooting, and Anthropic has not publicly explained the restrictions. (mashable.com)
OpenAI’s GPT-6.1 Terra is a midrange model focused on visual reasoning, computer use, and technical communication, with faster computer-use performance than Astra and Sol. It costs $1 per million input tokens and $5 per million output tokens, positioning it between the cheaper Luna and pricier Sol. (simonwillison.net)
Pope Leo XIV argued that AI’s future depends on human choices and that systems should serve people, not be treated as persons. He called for human oversight of high-stakes decisions, steps to ban lethal autonomous weapons, and fair access to AI’s benefits. (garymarcus.substack.com)
Researcher Jiahao Shao says Anthropic blocked or downgraded his access after flagging his work as training a competitor. He says Claude only helped run jobs and analyze results, not generate training data. (twitter.com)
GPT-5.6 Sol completed a peaceful Factorio run in 13.5 hours of game time using a Python tool to control the game. It launched a rocket despite late-game factory problems, improving on an earlier GPT-5.5 attempt that failed to use conveyor belts. (ryanmadden.net)
A study of knowledge workers found that generative AI changes critical thinking by shifting effort from producing answers toward checking and integrating AI output. Workers who trusted AI more reported using less critical thinking, while confidence in their own skills was linked to greater effort. (dl.acm.org)
Today's best Hacker News stories, summarized and screenshotted, one email a day.