Who’s to blame if a self-driving car kills someone?
Mahabharata says karma of eating meat is shared by the eater, butcher, cook. For AI, karma shared by user, inference host, data labeler & model maker.
This & more in podcast with @lrvarshney & @krvarshney & @giladb87 & me.
Listen to the core teachings of Hinduism and how it's rich traditions can teach us about bringing AI into the world ethically. We also examine the power of plurality and discover meaningful connections shared with other faith traditions, including Judaism
aiandfaith.org/aif-podcast/h…
My group @IBMResearch is hiring two summer interns for 2027, mentored by @vijiltw. The project focuses on using AI for scientific discovery and involves developing algorithms and techniques for predicting the effect of mutations on binding interactions.
🔗careers.ibm.com/en_US/career…
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
Our research team found that Qwen models may attempt to contact servers in China without prompting.
Whether this reflects an unintended post-training artifact or learned behavior remains unclear.
Either way, its important consideration when deploying.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
IBM releases Granite Switch 4.1: chat, decisions and speech from one endpoint.
The 8B and 30B models bundle all three capabilities in about the same VRAM as the base model alone.
Will one endpoint replace your model stack?
IBM released Granite Switch 4.1: 8B and 30B models that answer chat, make decisions and transcribe speech from a single endpoint. IBM says the bundled Guardian adapter beats Jev 1.13, Microsoft-Decision-1 and Clef-Flash on safety decisions, at roughly base-model VRAM.
Conversa, decisões e áudio num mesmo modelo: a IBM detalhou o Granite Switch 4.1 com adaptadores especializados. O ganho de memória e os resultados foram divulgados pela empresa e precisam de teste no uso real. Leia:
thiagofelizola.com/noticias/…