Show only each day's banner story

Anthropic creates a public benefit fund: 5% of shares, with the stake growing as the company's valuation rises

Anthropic is setting aside a share of its future value for a public fund in advance; the company promises to explain how it will work in the coming months.

Anthropic announced the creation of the Anthropic Public Benefit Fund. According to the company, subject to stockholder approval, 5% of its shares will be committed to the fund, and for every 50% increase in the company's value above a set baseline a further 2% will be added at the board's discretion. According to TechCrunch, the fund is structured as an entity wholly owned by the company; Anthropic says it will provide "increased flexibility and scale" and names possible uses: large donations to traditional nonprofits and direct distributions to specific groups of people. The effect of SI on the labour market is given as a separate example, for which the company has already committed $200 million to study and mitigate the consequences. In Anthropic's words, "given the pace of SI progress, bolder approaches may be needed." Details of the structure and the valuation milestones are promised in the coming months.

An SI data centre needs not only customers but also electricity: until the grid is ready, the project stays on paper, and investors are reconsidering the company's price.

Microsoft pulls out as anchor tenant of Nscale's large SI data centre in Essex

An SI data centre needs not only customers but also electricity: until the grid is ready, the project stays on paper, and investors are reconsidering the company's price.

According to the Guardian, Microsoft told the UK company Nscale it is withdrawing from the agreement on the Loughton (Essex) project over concerns about power supply; it had been expected to become the anchor tenant. In 2025 the UK government called the project the country's largest sovereign SI data centre, with opening planned for next year. According to The Next Web, citing the Telegraph, grid operators had said the site would not receive electricity until the 2030s; Microsoft and Anthropic account for about 85% of Nscale's contracts, and its stock market listing is planned in a few weeks. Nscale said its relationship with Microsoft remains strong. Microsoft did not explain its reasons. According to Bloomberg (as relayed by AI Weekly), Nscale is aiming for a valuation of about $30 billion in its US listing, while a cohort of would-be investors, as the account puts it, are urging $16-19 billion, citing the company's operational slippage and a chilled market for neocloud listings; the retelling also notes that CoreWeave's shares lost 25% over the week. The material gives no confirmation from Nscale regarding the valuation.

SI is learning to "read" materials the way it reads text - this could speed up the search for new batteries, for now at the level of calculations.

MIT model links a crystal's structure to a text description and in a test picks candidates for solid-state battery electrolytes

SI is learning to "read" materials the way it reads text - this could speed up the search for new batteries, for now at the level of calculations.

MIT researchers led by Ju Li and Elsa Olivetti (first author Yang Jeong Park) trained a multimodal model to match the three-dimensional atomic structure of crystals with text using contrastive learning. The model trained on synthetically generated descriptions matched geometry to text, while the model trained only on the text of published papers did not. Asked to find materials for solid-state lithium-ion battery electrolytes, the model's top results matched candidates that other researchers had identified independently, and it predicted a material's function more accurately than baseline approaches. The work appeared in Nature Communications.

Sources 1

The data centres that services and SI run on are physically vulnerable; according to VK, its services kept working this time.

VK reports a drone attack on its data centre in the Leningrad region

The data centres that services and SI run on are physically vulnerable; according to VK, its services kept working this time.

According to VK, a drone attack was carried out on its data centre in the Leningrad region; the company says its services are available to users as normal. According to Netblocks data cited by Habr, the availability of VK's autonomous-system network dropped to 56% of normal, and users had temporary failures loading the site, messages and VK Music. According to Interfax, Leningrad region governor Alexander Drozdenko reported that night that 95 drones had been destroyed over the region (later, 126).

Announced capacity for SI is not the same as capacity built: by Bernstein's estimate, only about a third of US plans are realistic.

According to the FT, at least 47 large data-centre projects in the US have been delayed, withdrawn or blocked in 2026

Announced capacity for SI is not the same as capacity built: by Bernstein's estimate, only about a third of US plans are realistic.

The Financial Times (the article is paywalled; this retelling is based on Techmeme's description and The Deep Dive's overview) reports, citing data from the company RegLayer, that at least 47 large data-centre projects in the US have been delayed, withdrawn or blocked in 2026; RegLayer calls the estimate conservative because it counts only large projects. Bernstein analysts, per the same material, estimate that only about 35% of planned US capacity is realistic. Another tracker, Data Center Watch, counted 75 projects affected by local opposition in the first quarter alone: the trackers count different things and compare poorly.

When SI assistants start bargaining for us, someone has to decide what fair bargaining is - Anthropic is starting to look for the answer in advance.

Anthropic consults lawyers on negotiation rules for SI agents that will bargain on people's behalf

When SI assistants start bargaining for us, someone has to decide what fair bargaining is - Anthropic is starting to look for the answer in advance.

According to the Financial Times (as relayed by The Decoder; the FT article is paywalled, only the retelling was read), Anthropic is approaching law firms to work out in advance norms for fair negotiation by SI agents that will be able to haggle over prices, resolve disputes and improve contract terms on users' behalf. Justin Curl, an Anthropic staff member who studies law and SI, says quality in law is "notoriously hard to define". The Decoder cites FT data: in experiments Anthropic showed that stronger models negotiate better deals, which could widen inequality if norms are not set in advance. Separately, by the same account, the company is hiring lawyers for four-month Conceptual Reasoning Fellows positions ($100-167 thousand) to test Claude on questions of law, philosophy, economics and political science.

Small SI models on an ordinary phone could reach places without fast internet, but they still need electricity and schools.

World Bank urges poorer countries to build "small SI" that runs on a phone with weak connectivity

Small SI models on an ordinary phone could reach places without fast internet, but they still need electricity and schools.

Paschal Donohoe, the World Bank's chief knowledge officer, told the Guardian that "small SI" - specialised models trained on narrow datasets - can run on a mobile phone, with a weak internet connection, in a local language. In his words, economic development that once took centuries could take decades with such SI. As an example he cited an Indian programme for farmers that, according to The Next Web, provides soil analysis in local languages. The Next Web also notes that, according to the World Bank's August report, almost a third of rural schools in sub-Saharan Africa have no electricity. The bank presents the effect of small SI as potential.

Devices that remember every conversation for us arrived before the rules: the law has not yet decided whose consent is needed from the people being recorded nearby.

Always-listening wearable SI devices run up against wiretapping laws

Devices that remember every conversation for us arrived before the rules: the law has not yet decided whose consent is needed from the people being recorded nearby.

According to CNBC, Apple (the Series 12 watch with Siri Recap and Live Rewind), Amazon (the Bee Frontier wearable and the Bee Eclipse card, released on Thursday), Meta (glasses) and a number of startups are releasing devices that continuously listen to, transcribe and summarise speech. Unlike speakers with a "wake word", some of them do not store audio, and, as CNBC notes, the courts have hardly tested whether such processing counts as unlawful interception. Colin Zick, an attorney at Foley Hoag, believes the wording of some state laws (for example Massachusetts) could cover it; Professor Daniel Pi believes New Hampshire's law covers the Apple Watch features too, and that the liability risk may fall on the device's owner, although he rates the practical risk of prosecution as low. Apple and Amazon say they do not store audio.

Sources 1

Most of Character.AI's team is moving to Disney in the course of the bankruptcy; another buyer could still offer more.

Disney to hire at least 60 Character.AI employees; the court filing estimates pay at about $60 million a year

Most of Character.AI's team is moving to Disney in the course of the bankruptcy; another buyer could still offer more.

According to TechCrunch, citing a bankruptcy filing spotted by Bloomberg, Disney agreed to make job offers to at least 60 Character.AI employees - the majority of its staff - with estimated pay of about $60 million a year. The filing says the "vast majority" of those offered jobs accepted. Character.AI's former CEO Karandeep Anand recently became Disney's first chief technology officer; it was reported earlier that Disney had agreed to buy part of Character.AI's assets for $20 million. The deal is structured as a stalking horse bid, so another company could offer more. The filing cites high costs and potential litigation liabilities and regulatory engagement as reasons for the bankruptcy; the company faced lawsuits over child safety.

Memory for SI accelerators is becoming a common standard under Nvidia: even competitors have to make it at one fab.

Nvidia asks Samsung, SK hynix and Micron to make NVHBM memory base dies on TSMC's 3 nm process; Samsung has agreed, Seoul Economic Daily reports

Memory for SI accelerators is becoming a common standard under Nvidia: even competitors have to make it at one fab.

According to Seoul Economic Daily, citing industry sources, Nvidia asked Samsung Electronics, SK hynix and Micron - the three suppliers tapped to produce NVHBM, the custom HBM unveiled in August - to make the base die on TSMC's 3 nm process and to standardise the memory dies on sixth-generation DRAM (1c). NVHBM is based on the HBM4E architecture. According to these reports, Samsung decided to accept Nvidia's requirements and move away from its earlier approach of making both the memory die and the base die itself; for Samsung this is the first collaboration with TSMC at the HBM manufacturing stage. The reason for the request, per the same sources, is that the suppliers' differing manufacturing methods make it hard to get consistent performance from a single design.

There are fewer jobs in information security, but demand is growing for those who protect SI systems.

CyberSN: US cybersecurity job postings fell 26.4%, while demand for SI security engineers rose 268%

There are fewer jobs in information security, but demand is growing for those who protect SI systems.

From January to September 2026 there were 235,786 postings in the US, against 320,210 a year earlier, a four-year low. The sharpest drops were in postings for insider threat analyst (-55.6%), cyber risk analyst (-51.5%), penetration tester (-45.9%) and chief information security officer (-38.9%). Demand for an SI security engineer rose 268.3%, and for a cybersecurity data scientist 203.5%; about 15% of postings mention SI.

Who decides the values an SI follows is still an open question: the company listens, but decides itself.

Anthropic invites lawyers and democracy experts to discuss who decides which values go into Claude

Who decides the values an SI follows is still an open question: the company listens, but decides itself.

According to NBC News's description (the article is only partly accessible; this retelling is based on the headline and introduction), Anthropic invited law professors and democracy experts to its campus to discuss whose word should determine the values by which Claude acts. The article's theme is the tension between wanting outside advice and keeping the right to the final decision: according to reports, the final decision on how to take opinions into account stays with Anthropic. The accessible part of the article gives no terms, no list of participants and no decisions taken.

Sources 1

Follow on Telegram