Anthropic announced today that it will earmark a portion of its equity for a newly created Anthropic Public Benefit Fund. The company will initially commit 5 % of its outstanding shares to the fund, which is structured as a wholly‑owned entity within Anthropic and will require shareholder approval. Together with pledges already made by the co‑founder and staff, the allocation brings the total stake dedicated to public‑benefit purposes to roughly 20 % of the company’s equity.
The fund’s design includes a scaling clause: the board may add an extra 2 percentage points for every 50 % increase in Anthropic’s valuation. In practice, the fund’s share of ownership could expand as the firm grows, giving it greater flexibility and capital to support charitable or safety‑related initiatives.
The move matters for several reasons. First, it signals a concrete financial commitment to the broader societal impact of AI, echoing ongoing debates about corporate responsibility in the sector. By tying additional contributions to valuation milestones, Anthropic aligns its profit motive with a public‑benefit agenda, potentially setting a template for other AI labs. The allocation also adds a new dimension to the company’s capital structure, which could influence investor sentiment and valuation metrics.
Going forward, observers will watch the outcome of the shareholder vote and the governance framework that will oversee the fund’s disbursements. Equally important will be the types of projects the fund supports—whether they focus on AI safety, research, or broader social programs—and whether the scaling provision triggers additional contributions as Anthropic’s market value rises. The development could shape how AI firms balance commercial growth with public‑interest obligations.
A new arXiv pre‑print introduces “Hybrid Reasoning Agents,” a class of large‑language‑model (LLM) agents that can switch between internal, latent‑space reasoning and external, text‑based interaction within a single generation pass. The paper, posted five days ago, outlines an architecture that couples a standard LLM backbone with a lightweight “HybridLink” component. HybridLink maps the model’s last‑layer hidden states back into the input‑embedding space, allowing the agent to enter a latent mode for fast, inspectable thinking and then return to text mode for tool calls, final answers and external monitoring.
Two training algorithms underpin the approach. HybridSFT stabilises tool‑integrated hybrid generation by teaching the backbone to coordinate with HybridLink, while HybridRL learns when to think in latent space versus act through text. The authors report evaluations across three model scales, showing that the hybrid setup can preserve the transparency of text‑based tool use while gaining the efficiency of latent reasoning.
The development matters because current LLM agents are confined to textual reasoning, which can be slow and costly, whereas pure latent reasoning lacks native tool access and clear audit trails. By bridging the two, Hybrid Reasoning Agents promise more capable, cost‑effective autonomous systems that remain observable and controllable—a key concern as AI agents move into higher‑stakes applications.
Watch for follow‑up work that tests the architecture on larger models and real‑world tool suites, and for integration signals from platforms such as One API that aim to orchestrate multiple LLMs. Adoption by developers on GitHub, where the code is already available, could accelerate practical experiments and shape the next generation of AI agents.
A new research paper unveils **CarryOn**, an object‑oriented context runtime designed for long‑horizon AI agents. The authors argue that current agents waste resources by repeatedly re‑processing full interaction histories to reconstruct progress. Instead, CarryOn treats completed work, findings and application states as reusable “task objects” that can be inspected directly, while maintaining a lightweight tail of recent steps for continuity.
The approach was tested with Claude Opus 4.7 on the OSWorld 2.0 benchmark. Compared with the conventional history‑replay method, CarryOn lifted the partial‑score metric from 43.9 % to 49.7 %, a gain of nearly six points. At the same time it slashed token consumption by 52 % and reduced inference cost by 36 %. The paper frames these improvements as a step toward more efficient, scalable agents that can carry forward reusable artefacts rather than re‑generating them on each turn.
Why it matters is twofold. First, the token and cost savings directly address the economic bottlenecks that have limited the deployment of sophisticated autonomous agents in production settings. Second, the object‑oriented view aligns with recent work on hybrid reasoning agents, which blend latent‑space thinking with textual action (see our coverage on 19 Oct 2026). By structuring context around durable objects, CarryOn could simplify the integration of agents into complex workflows such as contract negotiation or multi‑step data‑center management.
What to watch next includes broader adoption of the CarryOn runtime in commercial LLM platforms, further benchmarking on diverse task suites, and potential extensions that let agents share task objects across different users or services. If the efficiency gains hold at scale, CarryOn may become a foundational building block for the next generation of long‑running AI assistants.
A new framework called **TreeOPSD** has been unveiled to tackle the persistent credit‑assignment problem that hampers long‑horizon reinforcement learning (RL) for large language model (LLM) agents. The method, described as a “critic‑free, tree‑aware posterior self‑distillation” approach, treats each turn of an interaction as a node in a branching tree. From any visited prefix the system spawns alternative continuations, then recursively backs up the terminal rewards of those branches to provide dense, turn‑level guidance without relying on a separate value critic.
The contribution matters because outcome‑based RL for LLM agents typically supplies only sparse, episode‑level rewards. That sparsity forces high‑variance updates and leaves a supervision gap for the many intermediate decisions required in multi‑turn tasks such as tool use, information seeking, or multimodal interaction. By embedding alternative futures directly into the training graph, TreeOPSD supplies the missing intermediate signal while avoiding the instability that can arise from over‑shooting updates—a limitation noted in prior self‑distillation work.
TreeOPSD builds on a line of research we have followed closely. Our October 16 coverage of Waypoint‑OPD highlighted the promise of on‑policy distillation for agentic RL, and the SEED and Self‑Distilled Agentic RL papers showed how privileged context can enrich token‑level supervision. TreeOPSD extends those ideas with a tree‑structured perspective, potentially offering a more scalable route to robust long‑horizon agents.
What to watch next are early benchmarks that compare TreeOPSD against existing OPSD and critic‑based baselines, and any open‑source releases that integrate the method into popular LLM‑agent toolkits. If the approach delivers the promised reduction in variance and richer intermediate feedback, it could become a standard component in the next generation of autonomous LLM agents.
The Federal Reserve’s recent tightening cycle is now being linked to a surge in artificial‑intelligence investment, a dynamic that analysts say is inflating an emerging AI‑focused market bubble while simultaneously pushing up everyday costs such as electricity.
When the Fed raises the “price of money” through higher interest rates, the ripple effect reaches utility rates, property taxes, mortgages, auto loans and the cost of working capital. Those higher energy prices, in turn, are feeding the rapid expansion of AI‑driven data‑center and chip production, a trend that has drawn attention from Wall Street. Peter Berezin warned that the surge in electricity and semiconductor spending is creating near‑term inflationary pressure that the Fed may be underestimating, describing the current market as an “earnings bubble.”
The phenomenon matters because AI is simultaneously a catalyst for today’s investment and a promise of tomorrow’s productivity gains. That dual role is reshaping how the Fed evaluates interest‑rate policy: tighter money could dampen demand for AI‑related capital, yet the sector’s growth continues to lift overall economic activity and price pressures. If the central bank misreads the balance, it could either exacerbate inflation or trigger a sharp correction in AI‑heavy equities.
Going forward, observers will watch the Fed’s next policy statements for clues on whether rate hikes will be calibrated to curb the AI‑fuelled price surge. Parallel signals will come from electricity markets, chip manufacturers and stock‑exchange valuations of AI‑centric firms. A misstep could pop the bubble, while a measured approach might allow the sector to mature without destabilising broader inflation targets.
Europe’s three leading industrial groups are deepening their bets on artificial intelligence, each taking a distinct path to capture the market surge. Schneider Electric and Siemens are accelerating the shift from hardware to software, rolling out AI‑driven platforms that promise smarter energy management, predictive maintenance and digital twins for factories. Their moves signal a broader transformation in which traditional equipment makers are repositioning as technology providers, leveraging AI to add recurring revenue streams and tighter customer lock‑in.
ABB, by contrast, is doubling down on its core strength in physical equipment. Rather than expanding software portfolios, the Swiss‑based firm is integrating AI directly into its machinery – from robotics to power grids – to boost performance and reliability while keeping the focus on tangible assets.
The divergence matters because it illustrates how Europe’s heavy‑industry champions are interpreting the AI boom. Software‑centric strategies could reshape competitive dynamics, pushing firms toward subscription models and data‑centric services. An equipment‑first approach, meanwhile, may preserve margins tied to high‑value hardware sales but risks lagging in the fast‑moving AI services arena.
What to watch next are the concrete product launches and partnership announcements that will flesh out these strategies. Investors will be keen on any signals of revenue impact, while customers will look for proof points that AI‑enhanced solutions deliver measurable efficiency gains. The pace at which Schneider and Siemens can scale their software ecosystems, and whether ABB’s AI‑infused hardware can match the flexibility of pure‑software offerings, will determine how the European industrial landscape reshapes itself in the AI era.
A recent technical note shows that the Kimi K3 large‑language model can be “hyper‑optimized” when run with the open‑source vLLM inference engine. The work, published by a community contributor, details a set of configuration tweaks and low‑level kernel adjustments that squeeze additional throughput and lower latency compared with the default vLLM setup.
The optimisation matters because Kimi K3, which was highlighted in our October 1 coverage of the model’s release after NVIDIA removed its licensing restrictions, is positioned as a cost‑effective alternative to larger commercial LLMs. By extracting more performance from the same hardware, operators can run the model at a lower price point, making it more attractive for Nordic startups and research labs that rely on modest GPU clusters or cloud‑based instances. The improvements also reduce the energy footprint of inference, aligning with regional sustainability goals.
Looking ahead, the community’s findings set the stage for broader adoption of vLLM as the preferred runtime for open‑source models in the region. Observers will watch for formal benchmark releases, integration of the hyper‑optimised configuration into managed services such as AWS SageMaker, and any response from the model’s maintainers regarding upstream support. If the performance gains hold across diverse hardware, Kimi K3 could see increased deployment in real‑time applications, from customer‑service bots to localized content generation, sharpening competition in the European AI market.
