Noticias Tech

Fireworks tested routing tasks between Kimi K3, Moonshot AI’s open model, and Fable 5, Anthropic’s closed model. The combination reached 93% accuracy across roughly 1,030 agentic tasks, at up to 50 times lower cost than using Fable 5 alone.

A team of Beninese scientists rediscovered a coral reef believed dead since the 1960s, with eight types of coral and eight fish species at 53 meters deep in the Gulf of Guinea. The finding, funded by a National Geographic grant, opens the door to paleoclimate studies and a possible marine protected area.

Xiaomi introduces Xiaomi Robotics 1, a model that pretrains on 100,000 hours of robot-free video and is then fine-tuned with 7,200 hours of real data. The result nearly doubles the success rate of its closest rival on new tasks with minimal practice.

Chip Weinberger sold more than 2,500 units of his Jamcorder MIDI recorder and discovered that hardware is not so hard when the design stays simple. The real challenge was the software: 200,000 lines of code over three years, with no help from AI.

Alibaba announced Qwen3.8, a 2.4 trillion parameter model that will soon be open-weight. The Qwen3.8-Max-Preview version is already available for testing on the Token Plan, Qoder, and QoderWork.

transcribe.cpp is a new ggml-based speech transcription library that supports more than 60 ASR models with GPU acceleration. It was released by cjpais, creator of the Handy app, as an audited alternative to whisper.cpp and ONNX.

A developer known as Maurycy published a technique that abuses the progressive JPEG format to hide video inside a single image. The trick doesn’t use codecs or JavaScript: it only takes advantage of how the browser processes the file’s scans.

An EEG study published in PLOS Biology shows that the human brain briefly encodes two competing voices when the focus of auditory attention shifts. The transition is accompanied by a drop in alpha power and suggests the brain resets its lexical context after the switch.