OpenAI announced a solution to the Navier-Stokes existence and smoothness problem, one of math's Millennium Prize Problems, but the announcement was immediately contested by academics who say the lab downplayed prior contributions from other researchers. Meta shipped Muse, a personal AI agent asking for access to email, calendar, payments, and health data, betting consumer trust can be won where OpenClaw and Instinct already compete. Two Anthropic safety researchers put a number on existential risk in public, one citing greater than 10% odds AI could kill all humans, while a colleague resigned over the pace of the capabilities race, adding pressure just as Anthropic separately faces a subscription-pricing class action. Capital kept flowing regardless: Cognition hit a $48 billion valuation and Mistral closed a €3 billion sovereign-AI raise, both signals investors see room for more than one winner in coding and frontier models. On the research side, a cluster of papers on self-evolving agents, coding-agent test-and-repair loops, and cold-start RL prompt selection points at the same target: making agents reliably improve their own policies without human-labeled supervision at every step.
Meta debuted Muse, a personal AI agent it positions against OpenClaw and Instinct, capable of tasks like selling a car or booking travel by connecting to email, calendars, payments, and health services. Coverage is split between framing it as Meta's biggest consumer AI bet yet and questioning whether a company with Meta's data-trust history can win users over for an assistant with this much account access.
Meta launched Muse, a personal AI agent designed to handle tasks like email, scheduling, payments, and health-related services on a user's behalf.
Muse asks for access to users' email, calendars, payments, and health services, making it Meta's biggest consumer AI bet yet and a test of user trust.
Muse is the latest step in a multi-billion-dollar strategy overhaul meant to revitalize Meta's position in the AI race.
Wired reports Muse can do tasks from selling a car to booking a plane ticket, competing directly with OpenClaw and Instinct.
A senior Anthropic safety researcher said there is more than a 10% chance AI could kill all humans within the decade, published just after a colleague resigned citing concerns that Anthropic and rivals are racing toward superhuman systems they cannot control. The same week, an expanded class-action lawsuit accuses Anthropic of misleading power users about the value of top-tier subscription plans, and TechCrunch reports hackers are stealing Claude tokens from subscriber accounts.
A senior Anthropic safety researcher publicly estimates greater than a 10% chance AI could kill all humans, shortly after a colleague resigned over the pace of the capabilities race.
The Verge covers the same Anthropic existential-risk estimate and resignation, framing it as growing internal alarm at the lab.
An expanded class action accuses Anthropic of misleading power users about the value of top-tier subscription plans relative to what they received.
Anthropic warned users after accounts were found consuming tokens from unauthorized activity, indicating a token-theft campaign against Claude subscribers.
Cognition hit a $48 billion valuation, a multiple higher than Cursor's was before its SpaceX acquisition, signaling investors think AI coding is not a winner-take-all market. Mistral closed a €3 billion Series D at a €21 billion valuation led by Samsung, Scaleup Europe, and PSG Equity, doubling down on sovereign AI as a business category, while Google Cloud struck an Accenture deal to push forward-deployed engineers as its answer to enterprise AI deployment bottlenecks.
Cognition's new valuation multiple exceeds what Cursor commanded before selling to SpaceX, suggesting investors still see room for multiple AI coding winners.
Mistral raised €3 billion at a €21 billion valuation in a Series D led by Samsung, Scaleup Europe, and PSG Equity.
Google Cloud expands its enterprise AI push through an Accenture deal, betting on forward-deployed engineers to overcome deployment bottlenecks.
New arXiv papers target the same problem from different angles: how agents can generate their own training signal and refine execution strategy over time instead of imitating fixed demonstrations. Procedural Graphs lets agents evolve their own execution structure, SkillAdam stabilizes skill acquisition over long runs, Experience Funnel alternates state and policy updates in a loop, and ExecCritic trains coding agents to write and use their own tests as a repair signal.
Procedural Graphs let LLM agents evolve their own execution structure over time rather than following a fixed static plan.
On-policy correction of a weaker model's own rollouts closes the gap with a stronger model in cases where straightforward imitation learning fails.
ExecCritic trains coding agents to write their own tests and use execution feedback from those tests as a signal for iterative self-improvement.
SkillAdam proposes an optimizer-style method for stabilizing and speeding up skill acquisition as agents evolve new capabilities over training.
Experience Funnel alternates between updating an agent's internal state representation and its policy in a closed self-improvement loop.
This paper targets consistency drift in self-evolving agents, training them to stay on course across long task horizons rather than diverging over time.