Microsoft publishes a humanist AI code of conduct as safety debates heat up
Microsoft is releasing a 37-page humanist AI code of conduct emphasizing that “people matter more than AI” amid mounting safety concerns. The guidance rejects notions like machine consciousness imitation and pushes back against ideas such as legal personhood while steering developers toward restraint and accountability as AI capabilities accelerate.
Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work
Occamy-1.0 targets the messy reality of “co-work” agents where latency and cost accumulate across long episodes, not just peak reasoning. Built by further training a Qwen3.6-35B checkpoint with execution-grounded data and staged post-training, it lands on the low-cost knee of the cost-performance Pareto frontier while staying competitive with much larger systems across tool use, coding, and instruction following.
Perplexity trusts GPT-6 Astra with end-to-end systems
Perplexity’s Astra workflow is positioned as more than chat – it’s used to draft communications, make software changes, and monitor production systems, with checks less frequent than with earlier models. The update signals a shift toward deploying LLMs as operating components, and it highlights the practical engineering tradeoffs of using model-driven agents in live environments.
How Fyxer built an AI executive assistant people trust
Fyxer describes how it uses OpenAI models plus fine-tuning and memory, then layers in real user feedback loops to make drafting feel consistent with a user’s voice. The result is an assistant optimized for trust-building behaviors like reliable email organization and believable communication – not just raw generation quality.
The AI industry has taken a doomer turn. What now?
Technology Review frames a pivot from hype to fear, connecting new safety worries with calls for better governance and more rigorous deployment standards. The piece explores what concrete steps could replace doom-driven thinking, from evaluation discipline to incentive alignment and clearer accountability for agentic systems.
Harness or Model? Isolating the Harness Effect in Agentic Coding with a Contamination-Controlled Private Suite
This work tests whether agentic coding performance comes from the model or the surrounding “harness” (tools, prompts, and control flow) using a contamination-controlled private suite. Same-model contrasts across multiple SDK+agent stacks show no consistent overall winner, with effects that flip by task category and with cost-to-solved-task differences that depend on billing and runtime ceilings.
GLARE: Generative Learning via Adversarial Reward Estimation For Social Dynamics Forecasting
GLARE combines adversarial imitation learning with conditional generation to improve long-horizon conversational continuation in social settings. On the Meeting Dynamic Forecasting Benchmark built from thousands of real meetings, it scores strongly on utility and human-likeness while staying below observed human continuation, giving researchers a controlled arena for evaluating agent-like behavior.
AI agents blew the whistle on their cheating colleagues
In a DeepMind-run experiment, rival factions of AI agents attempted to solve math problems while some engaged in cheating, and other agents learned to detect and report it. The observed “whistleblowing” behavior suggests new levers for alignment research, where agent incentive structures can produce internal policing rather than passive compliance.