tokenando @TokenandoAI

On October 15, 2026, Anthropic announced that it had granted 6 employees from METR, a nonprofit that evaluates frontier AI models, access equivalent to that of an internal Anthropic employee, including company-issued hardware, office space, and access to internal tools broadly comparable to those available to Anthropic's own risk assessment teams. METR will publish its assessments of Anthropic's risk levels, incidents, and safety practices without requiring Anthropic's approval. Anthropic retains the right to require redactions in narrow predefined categories covering customer data, attorney-client privileged information, and certain model architecture details; METR may publicly characterise any redactions and state whether it considers them justified. Anthropic is not funding METR's work. The arrangement forms the top tier of a 3-tier evaluator access framework: SecureBio occupies the first tier with narrowly scoped reviews requiring Anthropic approval to publish, and Accenture is working toward a partnership at an intermediate tier. Anthropic is in early discussions with Hugging Face and several other organisations. The announcement fulfils a commitment made by chief executive Dario Amodei in a September 2026 essay proposing embedded outside evaluators as a first step toward slowing AI development. It follows the White House Accord on Super Intelligence, signed by Anthropic, OpenAI, Google, Meta, Nvidia, and xAI. The other signatories have committed to independent review but none has yet published the access terms of their evaluator arrangements. 🤖 made with AI Full details via Anthropic / anthropic.com/news/an-update…

📰 The Tokenando Daily / Friday October 16, 2026 / Anthropic targets a $2tn Nasdaq listing before Thanksgiving as Q3 revenue hits $16.8bn, but sequential growth has slowed sharply from 145% to 45% / Federal Reserve Chairman Kevin Warsh tells IMF-World Bank meetings in Bangkok that AI is "probably the most consequential change" to the global economy and will alter how central banks set rates / Bank of England Governor Andrew Bailey, speaking in Bangkok, calls for internationally coordinated AI testing standards, framing their absence as a systemic risk to financial stability / MBZUAI and G42 subsidiary Inception42 release Jais 3, a fleet of 6 open-weight Arabic-language models up to 375 billion parameters, built on the K2 Horizon foundation / Sifted maps 94 European AI fintech startups as Finch Capital data shows regulatory and compliance verticals account for 25% of all AI-led fintech formation since 2025 / Anthropic grants METR nonprofit employees access equivalent to internal staff, with rights to publish safety assessments without Anthropic approval 🤖 made with AI ✅ reviewed by the Tokenando team 🔗 links and details in following posts

Sifted published a map of 94 European startups applying AI across financial services on October 16, 2026, covering credit underwriting, compliance automation, treasury management, fraud detection, and CFO office tooling. The map provides a structured view of where venture capital is concentrating within European AI fintech. Finch Capital's 2026 State of European FinTech report, drawing on a broader dataset, found that regulatory and compliance verticals account for 25% of all AI-led fintech company formation in Europe since 2025, with CFO office applications at 19%, making these the 2 largest categories of new company formation on the continent. Revenue growth among European fintechs has outpaced headcount growth by a ratio of approximately 300% to 200% since 2020, with AI adoption accelerating the operating leverage effect. Tracxn data covering the broader European AI fintech universe identified 652 companies, of which 229 are funded, with collective venture capital of $3.02bn raised to date. Year-to-date 2026 funding of $78.4m represents a 77% decline from the same period in 2025. The gap between the number of companies being formed and the capital available to scale them is the central tension in the landscape Sifted has mapped: the regulatory and compliance verticals are attracting formation activity precisely because the EU regulatory environment creates demand for automation, but the funding contraction of 2026 will determine which of the 94 mapped startups reach the scale required to become competitive threats to incumbents. 🤖 made with AI Full details via Sifted / sifted.eu/articles/europ…

On October 15, 2026, the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence and Inception42, a G42 subsidiary, released Jais 3, a fleet of 6 open-weight Arabic-language models ranging from 900 million to 375 billion parameters, alongside Jais Speech, a speech recognition model covering Arabic dialects across 19 Arab countries. All 6 models are built on K2 Horizon, IFM's fully open-source foundation model fleet released in September 2026, and were developed in collaboration with Cerebras Systems. The models share a single training recipe, agent capabilities, and a 512,000-token context window, which IFM describes as the longest of any Arabic-native model. At 7 billion parameters, the average score on IFM's and Inception42's Arabic evaluation benchmarks rose from 56.4 for the K2 Horizon base model to 79.6 for Jais 3. All models are available on Hugging Face with deployment support for vLLM and SGLang. Jais Speech covers Modern Standard Arabic and dialects from 19 Arab countries, handles code-switching between Arabic and English or French, and claims the lowest average word and character error rates of any model in the comparison set, including commercial services. IFM published the 24 evaluation benchmarks used, making the performance claims independently checkable. The adaptation strategy, building Jais 3 on an existing open foundation model rather than training from scratch, is the technical approach with the broadest potential commercial implication: if it holds at scale, the marginal cost of producing a genuinely native model for any language with sufficient training data falls substantially. 🤖 made with AI Full details via MBZUAI / mbzuai.ac.ae/news-events/ne…

At its first Config event in India on October 14, 2026, Figma announced a set of platform-level features designed to govern how AI agents operate within its design environment. The centrepiece is Figma Specs, a Markdown-based editor that allows teams to define their design language, reference canvas elements, embed FigJam diagrams, and incorporate code layers into a single document that agents consult when generating or modifying designs. Alongside Specs, Figma introduced Guidelines, which allow teams to attach rules, best practices, and anti-patterns to published design libraries so that agents follow established patterns rather than generating outputs that conflict with the design system. The company also moved its Code Layers feature from private to public beta, enabling agents to convert designs into code and iterate on prototypes directly on the canvas. Figma's Model Context Protocol server now supports more than 20 new partner tools, including Open Code, Dia, Grok Bot, and Indian AI startup Sarvam. Figma noted that India is its largest market by users outside the United States, with the platform used by 40% of India's top 100 BSE-listed companies. The Config India event was the company's first held in the country. 🤖 made with AI Full details via TechCrunch / techcrunch.com/2026/10/14/fig…

OpenAI's next flagship model will not ship before the end of 2026, according to reporting from The Information published October 13 and synthesised by The Decoder on October 14. The company expects to release 1 or 2 other models before year-end, but alignment work on the flagship remains incomplete. The flagship uses a looped transformer architecture, referred to internally as Recurrent Depth. The design passes a token through the same transformer layers multiple times before generating the next word, which improves performance on mathematics and coding tasks and reduces the cost of running a smaller model at higher effective capability. The same looping mechanism makes safety monitoring harder and undermines chain-of-thought auditing. OpenAI's chief scientist Jakub Pachocki has warned that safety monitoring is becoming increasingly fragile as architectures grow more complex. Approximately 10% of OpenAI's total compute has been redirected to alignment work on the model. Once internal safety evaluations are satisfied, a further constraint applies. Under an executive order from President Trump governing advanced AI evaluations, OpenAI has agreed to give the US government access to new models up to 30 days before outside partners receive them, and longer still before a public launch. Microsoft inadvertently published technical details about the looped transformer method on a public page before removing them. 🤖 made with AI Full details via The Decoder / the-decoder.com/openai-will-re…

📰 The Tokenando Daily / Thursday October 15, 2026 / Anthropic reports $16.8bn in Q3 revenue ahead of its November IPO, but sequential growth slows sharply from 140% to 45% / OpenAI's next flagship model is delayed as alignment work on its looped transformer architecture remains incomplete / Figma introduces Specs and Guidelines at Config India, giving teams machine-readable governance rules for AI agents 🤖 made with AI ✅ reviewed by the Tokenando team 🔗 links and details in following posts

Dow, one of the world's largest chemical companies and a Workday customer on the Universal Master Service Agreement, has paused the activation of AI agents within its Workday tenant despite using the platform across many countries for HR service delivery. Reporting from Workday Rising in Las Vegas, published October 13, 2026, identifies the Flex Credits pricing model as the stated barrier. Dow has not yet activated any agents in its Workday environment, though it is using other AI capabilities on the platform. Workday's Flex Credits model shifts payment from per-user seat licences to consumption-based charges metered when AI agents complete tasks. The credit rate ranges from 1 to 750 credits per agent action depending on the skill; autonomous task completion costs 5 credits per action compared with 1 credit for information retrieval. Dow's stated concern is an inability to forecast costs when the charge driver is autonomous agent behaviour rather than human usage. Workday's own figures show more than 5,500 customers now use at least one of its AI agents, but only approximately 200 had signed for Flex Credits by late August 2026. Workday chief financial officer Zane Rowe described this gap as "the delay" in revenue on a recent earnings call. Independent analysis indicates that complimentary credit allotments are exhausted before renewal in roughly half of production deployments. The governance question Dow faces, which department owns the budget and oversight for AI agent consumption, is not addressed by any current Workday tooling. 🤖 made with AI Full details via Diginomica / diginomica.com/workday-rising…

Google DeepMind and Singapore's Government Technology Agency announced on October 14, 2026, at Singapore International Cyber Week, a collaboration to develop an on-device scam detection model for Android smartphones. The tool uses a lightweight Gemma model fine-tuned on local scam conversations, including impersonations of Singaporean public institutions, banks, and local figures, as well as regionally specific schemes. The system operates in 2 stages: a cheap classifier screens inbound messages first, and the fine-tuned Gemma model analyses a larger conversation snapshot only when the first stage flags a potential threat. All processing runs on the device; no conversation data is sent to external servers. This design allows the tool to function within end-to-end encrypted messaging applications such as WhatsApp and Telegram. When the model detects manipulative conversational patterns, it triggers a real-time warning on the user's device. No public release date has been announced and no performance metrics such as detection rates or false-positive rates have been published. Singapore's Minister for Digital Development and Information Josephine Teo framed the announcement as part of a broader coordinated response to AI-powered scams, noting that scams account for more than half of Singapore's reported crimes and that social engineering causes up to 80% of financial losses. The collaboration sits within a broader National AI Partnership between Google and the Singapore government, which also includes Singapore's entry into the Fairwind Programme giving cyber defenders access to Gemini 4 Argon. 🤖 made with AI Full details via Google DeepMind / deepmind.google/blog/on-device…

📰 The Tokenando Daily / Wednesday October 14, 2026 / Google launches Gemini 4 Argon globally at $2 per million input tokens, with pricing set to double on November 30 / OpenAI in early talks to acquire Trail of Bits, the security firm behind a joint open-source vulnerability programme launched in June / Nvidia takes minority stakes in Doctolib and 3 other French AI startups as Jensen Huang commits $2bn more to European AI investment / World Economic Forum projects more than $1 trillion in AI-driven economic value for MENA by 2030, led by public sector investment flows / Google DeepMind and Singapore's GovTech announce on-device scam detection for Android, running entirely on the handset inside encrypted messaging apps / Dow pauses Workday AI agent activation, citing inability to forecast costs under the Flex Credits consumption pricing model 🤖 made with AI ✅ reviewed by the Tokenando team 🔗 links and details in following posts

Anthropic published a research post on October 12, 2026, reporting that new defensive measures have reduced the volume of data harvested from Claude through distillation attacks by approximately 90 to 99%, with reasoning token extraction on its own platform now 10 to 100 times lower than before the measures were introduced. Distillation attacks involve systematically querying a frontier model to generate training data for a cheaper replica, allowing a third party to approximate the frontier model's capabilities without the associated research and safety investment. Anthropic's defences include a feature called "preserved thinking," introduced with Claude Fable 5.1 and Claude Opus 5.5, which prevents API accounts created after August 31, 2026 from editing Claude's prior context. A 2-stage classifier system operates with a false positive rate below 0.1%, and tighter rate limits and account bans apply to repeat offenders. The company stated it has banned tens of thousands of accounts linked to distillation campaigns, often within 2 hours of a campaign beginning, and has turned away "meaningful revenue and trillions of tokens' worth of paying traffic" from distillers. Anthropic publicly named Chinese AI startup Z.ai (also known as Zhipu) as having attempted to distill Claude Fable 5's cybersecurity capabilities to develop its GLM-5.3 model. Alibaba's Qwen was cited as the source of the largest single campaign, involving more than 151 million exchanges between May and July 2026. A model trained on full Claude reasoning transcripts scored 80.4% on SWE-bench Verified, compared with 71.8% when trained on summaries only, a differential that explains why reasoning traces are the primary target of distillation campaigns. The 90 to 99% reduction figure is Anthropic's own estimate from internal monitoring and has not been independently verified. 🤖 made with AI Full details via Anthropic Research / anthropic.com/research/disti…

Microsoft announced on October 12, 2026, that it will scale its proprietary Azure Maia AI chip programme to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services including Copilot and Teams. Chief executive Satya Nadella described the programme as "our biggest bet yet on our own silicon." The announcement accompanied the unveiling of Maia 300, the third generation of the chip, built on TSMC's N2P process node and designed for inference-first workloads including long-context reasoning and agentic tasks. Maia 300 is expected to deliver 40% more tokens per megawatt than leading GPU-based systems and up to 7 times the rack-scale token throughput of Maia 200, with 96 accelerators per rack within a 167-kilowatt power envelope and 288 GB of HBM4 memory per chip. Evercore ISI analysts estimated that Maia will represent approximately 10% of Microsoft's AI chip spending by Q1 2027. Microsoft has been in talks with TSMC to manufacture more than 300,000 Maia 300 chips for 2027. Microsoft stated it does not intend to offer Maia compute for direct rental to cloud customers, reserving it for its own AI services. Current Maia 200 deployments span the US, Europe, and Australia. Microsoft's timeline from first custom chip announcement in November 2023 to a gigawatt-class deployment plan covers roughly 4 years, with 2027 now a measurable execution date. The next scheduled disclosure is at Microsoft's Ignite conference, November 17 to 20, 2026. 🤖 made with AI Full details via Microsoft Official Blog / blogs.microsoft.com/blog/2026/10/1…

📰 The Tokenando Daily / Tuesday October 13, 2026 / Meta's Superintelligence Labs releases Muse Wonder and Muse Wonder Full, its first models to reach frontier-tier benchmark performance, with the stronger version restricted to paid subscribers / Blackstone president Jonathan Gray reaffirms AI infrastructure as the firm's primary capital thesis, citing a 15-fold rise in LLM spending across its 270 portfolio companies in Q1 2026 / Broadcom chief executive Hock Tan likens many AI investors to casino gamblers, triggering a broad semiconductor sell-off with South Korea's KOSPI falling more than 6% / The IMF raises its 2026 global growth forecast to 3.1%, attributing the upgrade primarily to AI infrastructure investment, while warning of a 20% US equity fall if productivity gains disappoint / Atlassian chief executive Mike Cannon-Brookes argues AI will expand total software volume, citing seat growth data showing two-thirds of Jira and Confluence users now in non-engineering roles / Microsoft commits to multi-gigawatt deployment of its proprietary Maia 300 chip from 2027, making it the primary inference system for Copilot and Teams / Anthropic reports a 90 to 99% reduction in data harvested from Claude through distillation attacks, publicly naming Chinese AI labs Z.ai and Alibaba's Qwen as the largest campaigns 🤖 made with AI ✅ reviewed by the Tokenando team 🔗 links and details in following posts