plus 9 more stories
Also: Reflection Releases Beam, a 501B Open-Weight Model for Coding Agents; and Microsoft Reportedly Cancels Its Anchor Deal at Nscale's Essex Data Centre. Every story links to its primary source.
Curated by Gaurav · October 18, 2026
Anthropic plans to allocate 5% of its shares to a new Public Benefit Fund, subject to stockholder approval. The board could add another 2% for each 50% valuation increase above a baseline, with worker disruption among the potential uses. Including earlier founder and employee pledges, the company says roughly 20% of its shares would support public benefit efforts.
Reflection has released Beam under Apache 2.0, making its 501-billion-parameter model available for teams to run and adapt themselves. It activates 23 billion parameters per token, supports one million tokens of context and includes a dedicated agent harness. The release expands US open-weight options for coding, although storing the full model still requires substantial GPU memory.
The Telegraph reports that Microsoft has cancelled its agreement to anchor Nscale's Essex data centre and rent about 23,000 Nvidia AI chips there. The report says grid power may not arrive until the 2030s. The cancellation highlights how electricity connections can constrain AI infrastructure despite commitments for customers and computing hardware.
Anomaly Innovations says a breach affected roughly 1.4% of OpenCode Console users, potentially exposing saved third-party model-provider keys. The company says affected users were notified and it is working with model providers on remediation. The incident puts credential storage at the centre of coding-agent security; the initial disclosure leaves the full cause and impact unresolved.
Perplexity has open-sourced SPACE, the sandbox infrastructure used for Perplexity Computer tasks, under Apache 2.0. Each sandbox runs in a Firecracker microVM, with credentials and outbound network controls managed outside the guest. Developers can self-host the infrastructure for agents that execute code, but deployment requires Linux and hardware virtualisation support.
Nvidia and IIT Madras's AI4Bharat lab have introduced a roughly 2.8-trillion-token synthetic corpus spanning 22 Indian languages, with the dataset described as releasing soon. They also detail an open pipeline for translation, quality checks and repair, including romanised text for twelve languages. The work gives model builders a reusable route to expand Indic training data while retaining source-document provenance.
Holyvolt has raised $45 million to develop battery materials by linking AI predictions with laboratory testing. The Stockholm startup combines models with Wildcat Discovery Technologies' experimental data and high-throughput testing platform. It aims to develop alternative chemistries for Western supply chains, with initial commercialisation of its disordered-rocksalt materials planned for 2027.
MIT researchers have built a multimodal model that connects crystal structures with descriptions of their properties, processing and applications. Their Nature Communications study reports better material selection than a text-only baseline, including promising solid-state battery electrolyte candidates. Combining atomic geometry with language could help researchers search beyond the materials most frequently described in published papers.
RapidCanvas has raised a $29 million Series B led by Vertex Ventures SEA & India, bringing total funding to $52.5 million. The company combines an agentic software platform with engineers who build and operate enterprise applications. It plans to expand delivery teams and invest in business context, reusable skills and governance, targeting the work needed to move AI projects into production.
Voxel51 has introduced version 3.0, combining multimodal search, data curation, annotation and evaluation in one platform for physical AI. The release adds time-series sensor support, AI-assisted labelling and dataset history, while its agent can chain tools into repeatable workflows. Robotics teams can use those capabilities to connect observed model failures with targeted training data and subsequent evaluation.
Tools in this briefing, reviewed in our directory:
Track the tools behind the news:
→ AI Tool Leaderboard → Browse the directory