OpenAI shuts down ChatGPT Atlas browser less than a year after launch
OpenAI is sunsetting ChatGPT Atlas, the browser feature meant to execute tasks on your behalf, with deprecation targeted for August 9. The move is part of a consolidation push toward ChatGPT Work, folding capabilities from app, coding tools, and Atlas into a single “superapp” experience.
Microsoft’s carbon emissions jumped 25% in 2025 as data centers expanded
Microsoft’s 2026 sustainability report says its carbon emissions rose 25% in 2025 to 34 million metric tons, driven primarily by expanding datacenter infrastructure. The report also points to a policy shift last February to stop purchasing “non-additional, unbundled renewable energy certificates,” complicating the path toward its longer-term carbon-negative goal.
When does in-context search actually help? Sampling-complexity theory of reflection-driven reasoning
This work puts hard math behind why some “generate-critique-revise” reasoning loops improve outcomes while others don’t. It shows in-context search can deliver exponential gains when reflections reliably pinpoint early mistakes, but offers no asymptotic improvement when reflections fail to localize errors – and it argues these gains are learnable via training on search rollouts.
Large Behavior Model – a promptable retail “digital twin” learned from transactions
Large Behavioral Model (LBM) trains a customer decision-making system directly from large-scale retail purchase data using a Person-Environment setup plus retrieval-augmented generation. The paper claims strong performance on in-domain retail tasks like basket completion and promotion response, while highlighting that continued pre-training and evidence-based reinforcement learning improve behavioral generalization.
AgentLens benchmark – evaluating coding agents by their whole trajectory, not just pass/fail
AgentLens tackles a key blind spot in code-agent testing by scoring the complete run trajectory – instructions handling, tool use, verification, recovery from mistakes, and the agent’s own explanations. The benchmark pairs formal verification (when possible) with LLM-written trajectory reviews to make regressions diagnosable, and it’s designed for production-style nightly evaluation pipelines.
SageMath-Augmented LLM agents – using computer algebra feedback in math research loops
This paper explores a ReAct-style agent that combines LLM reasoning with verifiable feedback from SageMath, using up-to-date documentation via Context7. Across research-level problems emulating a computational-math workflow, SageMath access boosts solve rates and reduces tool-enabled token usage, suggesting CAS-augmented agents can meaningfully support exploratory conjecture work.
Agent “harness” beats token maxing – orchestration as the lever on cost and speed
The paper argues the industry’s common scaling strategy – longer reasoning traces and more tokens – drives runaway spending, even when per-token costs fall. It shows swapping only the orchestration layer (“writer agent harness”) can cut blended cost per task by 41% and reduce tokens per task by 38% at comparable quality, framing harness design as the main multiplier for enterprise agent efficiency.
Goal-conditioned compact world models can “leak” instructions – a goal-free dynamics fix
This research finds a failure mode for goal-conditioned world models: they may appear to ground spatial relations while actually transcribing answers from the instruction. It reports instruction leakage collapses when the goal is kept out of the dynamics and is supervised through the “read” path instead, restoring grounding quality in both tabletop and benchmark tests.
Sunrun wants to host AI compute in your home – distributed data center nodes
Instead of building another centralized AI data center, Sunrun is piloting a program that places distributed compute nodes in homes with its solar and battery systems. Customers would be compensated, and the resulting distributed capacity would be sold to enterprise AI compute buyers.
Google adds AI labeling to Search, Discover, and YouTube ads via My Ad Center
Google will now show a “created or edited with AI” label for ads on Search, Discover, and YouTube through My Ad Center, letting users check whether AI tools were used to make an ad. The labeling is automatic for ads made with Google’s own generative advertising tools, while ads created elsewhere require manual labeling.
QANTIS – reusing quantum hardware as a belief-update service for POMDP planning
QANTIS treats quantum processors as calibrated “belief update” primitives under partial observability, feeding a classical planner with posteriors that should match exact Bayes decisions. In controlled IBM Heron experiments, the authors test multiple amplification strategies and focus on whether the hardware posterior preserves the downstream action, mapping an operating envelope rather than claiming generic speedups.
Compact ARC-AGI-1 solving with harnessed reflection under a strict budget
This paper studies whether architecture and orchestration alone can close gaps on ARC-AGI-1 without ARC-specific fine-tuning or heavy test-time thinking. Using an Explorer-Definer pipeline plus a Reflective Orchestrator, it reports a substantial lift in pass@2 and argues the improvement is generation-limited, with adaptive re-exploration matching the diagnostic predictions.
Instagram’s Adam Mosseri: don’t remove AI content – just clearly label it
Instagram head Adam Mosseri says he doesn’t want to filter out AI content globally, arguing instead for disclosure so people can decide what belongs in their feeds. He frames the approach as offering labeling without banning, with the goal of letting AI enthusiasts keep an “AI town” feed if that’s what they want.