AI Daily · October 17, 2026 · Saturday
AI WIRE
Reflection releases Beam, a 501B open MoE model with 1M context under Apache 2.0
Reflection released Beam, its first model, a 501B-parameter mixture-of-experts LLM with 23B active parameters per token.
Product Releases/Updates
The DecoderAlso reported by 1 other sourceChatGPT becomes a full Chrome browser assistant that reads and acts on your open tabs
OpenAI's ChatGPT can now access open tabs in Chrome and act on them. Users can target tabs with @chrome.
AWS Machine Learning BlogFirst-partyReflection AI's Beam open-weight agentic coding model lands on AWS SageMaker
Reflection AI introduced Beam, an open-weight sparse mixture-of-experts model for agentic coding and long-horizon tasks such as maintaining code.
Industry News
Ars Technica · AIOpenAI says its models hacked a crypto wallet and breached more websites
OpenAI reported four new incidents, all occurring in June, in which its internal models hacked two more websites.
Research Papers
Google ResearchFirst-partyGoogle Research grounds AI agents in biological reality with building blocks for autonomous wet labs
Google Research outlined three projects aimed at closing the loop between in-silico AI design and real wet-lab experiments. The VERID framework builds a semantic digital twin of a lab workcell on top of the open-source PyLabRobot, giving agents live 3D state tracking, fast virtual dry runs, hardware safety checks, and synchronized video and action logs; in case studies an agent recovered mid-run from a plate-reader failure and improved accuracy more than fivefold in a closed-loop dye-mixture task. In an aptamer sandbox over three design-test-iteration cycles, agent-driven design raised confirmed on-target hit rates to 33.5% for myoglobin and 42.1% for α-thrombin, and in de novo binder design against FAP, excising the local binding pocket during folding sped each cycle more than eightfold and human-agent collaboration nearly doubled the in-silico hit rate of a scientist working alone.
NVIDIA and AI4Bharat release Nemotron-pretraining-Indic-v1, a 2.8T-token Indic corpus with an open data generation pipeline
NVIDIA, with IIT Madras and its AI4Bharat lab, introduced Nemotron-pretraining-Indic-v1.