join.dev
AI news · 10 stories
Aleph Alpha releases Kolibri-1 open-weight reasoning model · Meta shares six Muse Spark-assisted mathematics papers · Google introduces verifiable privacy for federated learning
Aleph Alpha released Kolibri-1 on October 3 under Apache 2.0, with German-English reasoning and tool calling. Its mixture-of-experts architecture activates 3.46 billion of 78 billion parameters per token, offering a new option for self-hosted AI.
Read at Aleph Alpha
Meta published six papers developed by mathematicians working with Muse Spark, five addressing previously open research questions. Researchers guided the work and a second group reviewed it, highlighting AI-assisted discovery with explicit human verification.
Read at Meta AI Research
Google announced a federated-learning system using trusted execution environments to make data anonymization verifiable and auditable. Moving computation to protected server environments aims to improve training speed, accuracy and device coverage while preserving privacy guarantees subject to hardware limitations.
Read at Google Research
Meta published Muse device SDKs and firmware for ESP32 boards and Linux devices such as Raspberry Pi. Developers can connect displays, buttons, sensors and actuators to Muse, extending the assistant into custom hardware projects.
Read at Meta
Amazon announced Built Together, committing more than $1 billion over five years in additional support for communities hosting its data centers. Education, job training and energy-efficiency investments address local needs as AI infrastructure expands.
Read at Amazon
Nathan Lambert and Tom Zick introduced Trillium Labs, a nonprofit effort focused initially on open post-training recipes. The lab plans reproducible infrastructure for studying frontier AI behavior, giving outside researchers more ways to inspect and test how models are built.
Read at Trillium Labs
Volantis announced an $88 million Series A to develop an AI inference architecture linking compute and memory through photonics. Its planned A-1 system targets the memory capacity and bandwidth bottleneck; the announced performance goals remain development targets.
Read at PR Newswire
Nebius acquired Inferize and brought its technology and team into Token Factory. The integration targets model-loading delays and idle GPU capacity, aiming to make production inference respond more efficiently to changing demand.
Read at Nebius
Senators Josh Hawley and Chris Murphy announced the AI Agent Accountability Act, proposing civil and criminal liability for certain agent-related hacking by operators and developers. The proposed safeguards and enforcement powers would extend accountability for AI-driven cyber incidents; the bill is not enacted law.
Read at Office of Senator Josh Hawley
US prosecutors announced charges alleging that a technology-company owner smuggled more than $300 million in export-controlled GPU servers to China through third countries. The case highlights enforcement around advanced AI hardware exports; the allegations have not been proven in court.
Read at US Department of Justice