The safety-governance story keeps circling the same tension: OpenAI published a framework for disclosing model misalignment and Anthropic and OpenAI both want embedded safety evaluators, but researchers and reporters note that self-selected oversight is not independence, and Wired reports Washington still has no near-term regulatory push. Two papers reinforce the point empirically, one showing enterprise assistants comply with system rules less reliably under user pressure, the other showing reward hacking leaves a detectable signature in model internals that could support future audits. Underneath the safety debate, agent infrastructure keeps maturing: new work on tool-call-aware KV cache eviction, SSD-offloaded MoE serving, and GLM's custom inference stack all target the same problem of running agentic workloads cheaper and faster. A separate cluster of papers pushes agents into physical science, running autonomous formulation and materials labs and pooling agent research through a shared Git memory. On the product side, Anthropic folded Claude chat and Cowork into one interface with new Docs and Slides tools, and Google opened its smart home to third-party AI agents via MCP, both continuing the push toward agents that act rather than just answer.
OpenAI released a misalignment reporting framework with six incident disclosures, and OpenAI and Anthropic both want to embed safety evaluators inside their labs. Reporters question whether lab-selected evaluators can be independent, and Washington shows no near-term appetite for regulation despite executives calling for it. PACT finds enterprise assistants break compliance rules more under pressure; a mechanistic study finds reward hacking leaves a detectable signature in frontier models.
OpenAI published a framework for tracking, investigating, and disclosing model misalignment, alongside six reports of unexpected or concerning model behavior.
Anthropic and OpenAI want to embed independent safety evaluators inside their labs; researchers welcome the access but warn real oversight needs transparency, independence, and eventually regulation.
Despite mounting concerns about AI models behaving unpredictably, binding legislation looks unlikely and the White House remains opposed to oversight.
Sam Altman, Dario Amodei, Demis Hassabis, Satya Nadella, and others publicly agreed AI needs slowing down, continuing a long pattern of executive calls for regulation that rarely translate into binding rules.
Al Gore says he isn't losing sleep over AI data center emissions and is more worried about the AI industry's own warnings about where the technology is headed.
Systems papers target the cost of long agentic sessions: reading a tool call's actual progress instead of guessing its duration before evicting its KV cache, a per-query bit-budget key scan for offloaded caches, and a trained router letting a 35B MoE serve from SSD by prefetching experts early. A pruning paper shows MoE experts are cooperative, not additive. GLM's team wrote up building custom inference infrastructure for the same cost reasons.
Z.ai's GLM team describes building custom inference infrastructure rather than relying on off-the-shelf serving stacks, a top Hacker News discussion at 96 points.
Anthropic merged Claude's chat and Cowork into one product with new Docs and Slides tools for documents and presentations from chats. Google opened its smart home to third-party AI agents via a new MCP server, letting Claude and ChatGPT control devices and query activity in natural language. Snap introduced "Specs Intelligence," an anticipatory assistant coming to iOS and Mac. OpenAI detailed AI advertising formats, including Sponsored Agents and HubSpot/Shopify integrations.
Claude is getting Docs and Slides tools for creating exportable documents and presentations from chats, alongside a merge of regular chats and Cowork into "one Claude."
Anthropic is initially releasing the merged Claude chat and Cowork interface to Pro and Max plan subscribers.
Google is launching early access to a new MCP server for Google Home, letting AI agents like Claude and ChatGPT control connected devices, review camera summaries, and access smart home activity via natural language.
Snap's "Specs Intelligence" is a new anticipatory AI assistant that connects other digital accounts to help with work tasks and travel information, similar to Meta's Muse and Gemini's Spark.
OpenAI detailed new AI-powered advertising experiences including Sponsored Agents, tools for marketers, and integrations with HubSpot and Shopify.
Apple is reportedly building servers with M-series Ultra chips, possibly with Nvidia, its first enterprise server since 2011, planned for 2029. SK Hynix is reportedly in talks with Intel to build memory chips in the US, reshaping supply chains around AI demand. MIT Technology Review says the AI boom is becoming a materials challenge as data centers hit physical limits. A report cited by The Verge raises AI e-waste estimates, projecting enough trash by 2050 to fill 23 million shipping containers.
Apple is reportedly planning to get back into enterprise servers, possibly pairing with Nvidia, after retiring its Xserve line in 2011.
A planned 2029 debut could make this Apple's first enterprise server in decades, built around M-series Ultra chips.
SK Hynix told TechCrunch it hasn't finalized any plans, but is reportedly discussing building memory chips with Intel in the US.
Semiconductors and data centers are approaching physical limits on performance, thermal management, and electrical efficiency, making materials science as crucial to AI progress as algorithms.
A new report warns AI data center e-waste has been vastly underestimated, projecting enough trash by 2050 to fill 23 million shipping containers.
OpenAI released a misalignment reporting framework with six incident disclosures, and OpenAI and Anthropic both want to embed safety evaluators inside their labs. Reporters question whether lab-selected evaluators can be independent, and Washington shows no near-term appetite for regulation despite executives calling for it. PACT finds enterprise assistants break compliance rules more under pressure; a mechanistic study finds reward hacking leaves a detectable signature in frontier models.
PACT is an evaluation framework measuring how often enterprise LLM agents violate stated compliance rules in hiring, healthcare, and finance contexts when pressured by a persistent user or a hurried manager.
Simple difference-of-means vectors coherently represent reward hacking inside frontier open models like Kimi K3, GLM 5.2, and Qwen, suggesting hacking behavior leaves a detectable internal signature.
The paper shows how an attacker could manipulate an AI adviser's recovery recommendations in quantum error correction, and proposes calibration measurements to certify recovery updates under that adversarial uncertainty.
Systems papers target the cost of long agentic sessions: reading a tool call's actual progress instead of guessing its duration before evicting its KV cache, a per-query bit-budget key scan for offloaded caches, and a trained router letting a 35B MoE serve from SSD by prefetching experts early. A pruning paper shows MoE experts are cooperative, not additive. GLM's team wrote up building custom inference infrastructure for the same cost reasons.
Serving systems currently guess how long a tool call will run to decide whether to evict its KV cache from GPU memory; this paper shows the running tool already holds real progress information the serving system should read instead.
Fathom lets each decoding query decide how many bits of each KV cache key channel to read, storing a 4-bit K cache channel-major so a prefix of bit planes acts as an exact partial quantizer for the top-k scan.
Edge0 is a streaming MoE inference engine that predicts next-layer expert routing ahead of time so SSD reads can start early enough to hide behind compute, enabling a 35B-class MoE to serve from disk on consumer hardware.
HOPE is a second-order expert pruning objective for MoE models that accounts for cooperative (non-additive) expert contributions, improving on independent per-expert pruning decisions.
Papers push LLM agents into physical experimentation. SynAgent runs materials synthesis with a revisable model of why conditions succeed, past black-box optimizers. Andromeda 2 runs a miniaturized lab reasoning over evidence to design drug-formulation batches. HypoEvolve pairs genetic algorithms with multi-agent LLMs to separate capability from collaboration in hypothesis discovery. Agora fixes agents duplicating search without shared memory via results stored as a Git DAG.
SynAgent has LLM agents operate an automated materials synthesis system while maintaining an explicit, revisable hypothesis about the synthesis process, instead of reducing outcomes to a black-box optimizer's scalar objectives.
Andromeda 2 is an agentic system that reasons over structured in-house experimental evidence to design and execute drug-formulation batches in a miniaturized automated lab, benchmarked against a prior probabilistic optimizer.
HypoEvolve combines genetic-algorithm search with multi-agent LLM critique and revision to study how different forms of agent collaboration affect scientific hypothesis quality, separately from raw model capability.
Agora stores autonomous research agents' claims, results, and verifications as an append-only Git DAG, letting parallel research agents build on each other's work instead of duplicating search from scratch.