A paper presented at ICLR 2026 in Rio de Janeiro demonstrates that training LLMs for stronger reasoning via reinforcement learning increases tool-hallucination rates in lockstep with capability gains. The authors found that RL "disproportionately collapses tool-reliability-related representations" in the model's late layers — the exact layers that should restrain bad tool calls get trained away. Prompt engineering and DPO partially help, but neither closes the gap, framing this as a fundamental reliability-capability tradeoff that affects every frontier lab's reasoning models.
China's National Development and Reform Commission officially blocked and ordered the unwinding of Meta's acquisition of Manus, the Singapore-based agentic AI startup that made waves when it launched its autonomous agent in March 2025. The complication: Meta had already integrated Manus into its internal systems and onboarded the startup's executives, making disentanglement practically messy. The move signals Beijing is willing to use investment review as a weapon in the AI tech war, even for companies that relocated to Singapore to distance themselves from China.
Over 600 Google staff across DeepMind and Cloud signed an open letter urging Sundar Pichai to refuse classified military AI workloads for the Department of Defense. The letter cites lethal autonomous weapons and mass surveillance as potential harms, echoing the 2018 Project Maven revolt — but this time Google has already quietly dropped its AI Principles language promising not to pursue weapons or surveillance tech. The petition directly challenges reported negotiations to deploy Gemini models in classified defense operations.
OpenAI released GPT-5.5 on April 23, calling it their "smartest and most intuitive" model yet, with major gains in agentic coding, computer use, and knowledge work. The bigger strategic signal: GPT-5.5 is the model OpenAI is building its unified desktop super-app around, merging ChatGPT, Codex, and the Atlas browser agent into a single session. API access went live April 24 for Plus, Pro, Business, and Enterprise tiers.
CNBC reports a structural brain drain from Big Tech AI labs, with top researchers closing nine-figure funding rounds before writing a line of production code. VCs have funneled $18.8 billion into AI startups founded since early 2025, on pace to surpass last year's $27.9 billion. The exodus is driven by increasingly commercial mandates at major labs that limit freedom for exploratory research outside the dominant LLM paradigm, with Andreessen Horowitz, Sequoia, and Kleiner Perkins competing aggressively for founding teams.
Xiaomi released MiMo-V2.5 (310B parameter sparse MoE, 15B active) and MiMo-V2.5-Pro (1.02T parameter MoE, 42B active) fully open-source under the MIT license, with weights and model cards on Hugging Face. Both support up to 1 million tokens of context with native multimodal understanding across vision and audio. The Pro variant targets agentic tasks and is being described as among the most efficient open models for autonomous tool use.
Sergey Brin wrote in an internal memo that Google "must urgently bridge the gap in agentic execution" with Anthropic, whose Claude models now handle nearly all of Anthropic's own coding versus about 50% at Google. The strike team is led by Sebastian Borgeaud (formerly Gemini pre-training lead) with direct involvement from DeepMind CTO Koray Kavukcuoglu, and will train on Google's 2-billion-line proprietary codebase. The immediate trigger was Anthropic's release of Claude Opus 4.7 on April 16.
OutSystems' 2026 State of AI Development survey of nearly 1,900 IT leaders found that while 96% of enterprises now run AI agents, 94% are concerned that agent sprawl is increasing complexity, technical debt, and security risk. Only 12% have a central platform to manage their agent fleet. Meanwhile, Deloitte research finds 47% of enterprise AI users have based at least one major business decision on hallucinated content — tying directly to the ICLR "Reasoning Trap" findings.
The Canadian federal government finally revealed the six pillars of its long-delayed national AI strategy, covering safety and democracy protections, public AI training and education, SME adoption, sovereign compute infrastructure, scaling Canadian AI champions, and building trusted global partnerships. The strategy accompanies a national initiative to build large-scale AI supercomputing capacity and is timed against the EU AI Act's high-risk compliance requirements activating in 2026.
Global enterprise AI spending is projected to hit $665 billion in 2026, but only 43% of organizations have a formal AI governance policy — and 73% of AI deployments fail to deliver promised ROI. The ExcelMindCyber Institute flagged governance, not technology, as the primary cause of ROI failure. With EU AI Act fines reaching up to 7% of global turnover activating this year, the compliance clock is ticking for enterprises that have been deploying agents faster than guardrails.
