Insights & Announcements
A weekly Newsletter on technology applications in investment management with an AI / LLM and automation angle. We combine 100% human curation/selection with LLM standardisation, summarisation, and more deterministic search/collection, classification and workflow - powered by Kubro(TM). Curated news, announcements, and posts, primarily directly from sources (Arxiv papers, major AI/Tech/Data companies, investment firms). It's an evolving project.
Disclaimers: content not fully human-verified, with AI summaries below. AI/LLMs may hallucinate and provide inaccurate summaries. Select items only, not intended as a comprehensive view. For information purposes only. Please DM with feedback and requests
NVIDIA announced NemoClaw, a one-command software stack for the OpenClaw agent platform that installs Nemotron models and the new OpenShell runtime to enable secure, privacy-controlled autonomous AI assistants. The offering adds sandboxing, policy-based security and a privacy router that balances local open models with cloud frontier models. Strategically, NVIDIA is positioning OpenClaw as an operating layer for personal AI and its own hardware—from RTX PCs and workstations to DGX Station and DGX Spark—as the compute backbone for always-on agents. The launch strengthens NVIDIA’s software ecosystem, supports developer adoption, and could stimulate demand for higher-end client and enterprise AI infrastructure globally.
🔗 Source: Summary based on nvidianews.nvidia.com View Source | Found on Mar 16, 2026
OpenAI’s planned acquisition of Astral is strategically aimed at strengthening Codex as a full-stack AI software agent, particularly in Python workflows. Astral’s widely adopted open-source tools—uv, Ruff, and ty—bring distribution, developer trust, and deep infrastructure expertise. OpenAI highlights strong Codex momentum, with weekly active users above 2 million, 3x user growth, and 5x usage growth year to date. The deal should improve Codex’s ability to plan, edit, test, and maintain code across the software lifecycle. For developers, OpenAI says Astral’s products will remain supported post-close. Completion remains subject to regulatory approval and standard closing conditions before integration benefits are realized.
🔗 Source: Summary based on openai.com View Source | Found on Mar 21, 2026
The article, authored by Ryan Burnell and published on March 17, 2026, announces the release of a new paper titled “Measuring Progress Toward AGI: A Cognitive Taxonomy,” which provides a scientific foundation for understanding AI systems’ cognitive capabilities. The framework is based on decades of research from psychology, neuroscience, and cognitive science and identifies 10 key cognitive abilities hypothesized to be important for general intelligence in AI systems, including perception, generation, learning, metacognition, and social cognition. Additionally, the authors are partnering with Kaggle to launch a hackathon aimed at building evaluations to implement this framework.
🔗 Source: Summary based on blog.google View Source | Found on Mar 17, 2026
NVIDIA announced the expansion of its open model families, including Isaac GR00T N1.7, Alpamayo 1.5, Cosmos 3, and Nemotron 3 Ultra, Omni, and VoiceChat models, to advance agentic, physical, and healthcare AI across robots and autonomous vehicles. The Proteina-Complexa model on the BioNeMo platform accelerates protein drug discovery with a new dataset of about 30 million AI-predicted protein complex predictions—1.7 million high-confidence—developed with Google DeepMind, EMBL’s European Bioinformatics Institute, and Seoul National University. Companies such as CodeRabbit, CrowdStrike, LG Electronics, Novo Nordisk, Viva Biotech and others are adopting these models for various applications.
🔗 Source: Summary based on nvidianews.nvidia.com View Source | Found on Mar 16, 2026
In 2026, AI is moving from pilot projects to large-scale enterprise production, with a major adoption inflection point expected this year. Cisco projects that by the end of 2027, 60-70% of its product portfolio will be entirely written by AI, up from half a dozen products by the end of 2026. Key blockers for enterprise AI deployment are trust and security. Infrastructure is shifting to heterogeneous models combining GPUs, CPUs, and accelerators. Cisco announced advancements in AI networking (G300 chip), secure AI factory (Hypershield with Bluefield support), ecosystem partnerships, and unified edge solutions using NVIDIA Spectre 4500 chips.
🔗 Source: Summary based on newsroom.cisco.com View Source | Found on Mar 20, 2026
The article, published on March 19, 2026, discusses the renewed interest in context layers and semantic models within enterprises. It highlights that while models have become smarter, their effectiveness is limited by fragmented business concepts, implicit rules, missing history, and contested truths across systems. Key components for enterprise analytics include analytic semantic models defining metrics and entities; relationship and identity layers (ontologies) for cross-domain integration; versioned business procedures specifying operational workflows; evidence and provenance tracking data lineage; and policy enforcement determining user entitlements. The article notes that semantic models and ontologies are longstanding concepts now experiencing increased urgency in enterprise settings.
🔗 Source: Summary based on snowflake.com View Source | Found on Mar 20, 2026
The NVIDIA Nemotron Coalition is a global collaboration announced by NVIDIA, bringing together Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam and Thinking Machines Lab to advance open frontier-level foundation models through shared expertise, data and compute. The coalition’s first project is a base model co-developed by Mistral AI and NVIDIA on NVIDIA DGX Cloud; this model will be open sourced and underpin the upcoming Nemotron 4 family of models. Members contribute domain expertise such as multimodal capabilities, evaluation datasets and agent specialization to support post-training and continued development for industry-specific applications.
🔗 Source: Summary based on nvidianews.nvidia.com View Source | Found on Mar 16, 2026
Stitch, launched to transform ideas into software starting with design, has evolved into an AI-native software design canvas enabling users to create, iterate, and collaborate on high-fidelity UI designs from natural language input. The redesigned Stitch features an infinite canvas supporting images, text, and code as context, a new design agent that reasons across project evolution, and an Agent manager for parallel idea exploration. Users can extract or import design systems via URLs or DESIGN.md files. Stitch enables instant interactive prototypes and real-time voice-driven collaboration. Integration with developer tools is supported through the Stitch MCP server and SDK (2.4k stars).
🔗 Source: Summary based on blog.google View Source | Found on Mar 19, 2026
Microsoft has expanded its Foundry capabilities to build, deploy, and operate production-ready AI agents on NVIDIA accelerators and open NVIDIA Nemotron models. The next-generation Foundry Agent Service and Observability in Foundry Control Plane are now generally available, enabling organizations to develop agents that reason, plan, and act across tools, data, and workflows. Microsoft is the first hyperscale cloud to power on NVIDIA’s Vera Rubin NVL72 systems in its labs, with rollout planned for Azure datacenters. Integration between Microsoft Fabric and NVIDIA Omniverse libraries supports Physical AI systems from simulation to real-world operations.
🔗 Source: Summary based on blogs.microsoft.com View Source | Found on Mar 17, 2026
The article, authored by Mariano Cocirio and published on March 17, 2026, announces new API tooling capabilities that allow developers to combine built-in tools such as Google Search and Google Maps with custom functions in a single request. This update enables Gemini models to fetch public data and call backend services without separate orchestration steps, reducing end-to-end latency and simplifying agent architectures. Context circulation now preserves every tool call and response for complex multi-step workflows. Additionally, unique call identifiers have been introduced for each tool call to improve debuggability during asynchronous executions.
🔗 Source: Summary based on blog.google View Source | Found on Mar 19, 2026
The Stanford University study, conducted over 18 months in collaboration with Google, observed how Googlers learned and used AI in their daily work. The research found that while most were eager to use AI tools, many only substituted existing tasks rather than redesigning workflows. Successful adopters applied product management strategies: identifying high-value opportunities, choosing the right AI tool beyond chatbots, starting small and experimenting rapidly, thinking holistically across systems, and sharing their playbook for team benefit. The study highlighted that deep adoption required rethinking processes rather than quick fixes and documented five specific strategies for effective AI integration.
🔗 Source: Summary based on blog.google View Source | Found on Mar 19, 2026
The article presents comparative analyses of global perspectives on artificial intelligence (AI), highlighting the most common visions and concerns about AI in different regions. It uses slope charts to illustrate how the rankings of these themes shift across regions, with bolded items indicating visions or concerns more frequently expressed in a particular region, while grey items denote those similarly or less often mentioned. The article also provides guidance for citation using a Bibtex key.
🔗 Source: Summary based on anthropic.com View Source | Found on Mar 18, 2026
NVIDIA has launched the Vera CPU, designed for agentic AI and reinforcement learning, delivering twice the efficiency and 50% faster performance than traditional rack-scale CPUs. Vera features 88 custom Olympus cores, supports up to 1.2 TB/s memory bandwidth using LPDDR5X at half the power of general-purpose CPUs, and integrates with NVIDIA GPUs via NVLink-C2C for 1.8 TB/s coherent bandwidth. Partners deploying Vera include Alibaba Cloud, ByteDance, Meta, Oracle Cloud Infrastructure, Dell Technologies, HPE, Lenovo, Supermicro and others. National laboratories such as TACC plan to deploy Vera later this year. Full production is set for the second half of this year.
🔗 Source: Summary based on nvidianews.nvidia.com View Source | Found on Mar 16, 2026
Over the next five years, global capital expenditure on AI infrastructure is projected to reach USD 5 trillion–USD 7 trillion, with combined capex for the five largest cloud providers expected to rise from about USD 240 billion in 2024 to over USD 580 billion in 2026. Technology-related investment-grade corporate bond issuance could reach USD 350 billion in 2026, driven by hyperscaler demand. The Beignet Investor deal, a joint venture between Meta Platforms and Blue Owl, raised USD 27 billion in late 2025 for an AI data center in Louisiana. Leveraged finance markets may provide up to USD 150 billion for non-investment-grade participants.
🔗 Source: Summary based on troweprice.com View Source | Found on Mar 19, 2026
Goldman Sachs Research, led by Joseph Briggs, reports that AI is already impacting the US tech sector, with its employment share falling below the long-term trend. Briggs estimates that widespread AI adoption will occur over approximately 10 years, displacing 6-7% of workers during this period. If the transition takes a decade, unemployment may rise by 0.6 percentage points; a faster transition would have greater economic effects. Knowledge and creative sectors such as management consultants, call center workers, and graphic designers have experienced some displacement. Globally, about 300 million jobs are exposed to AI automation; in the US, AI could automate tasks comprising 25% of work hours.
🔗 Source: Summary based on goldmansachs.com View Source | Found on Mar 19, 2026
The article by Amundi, published on March 19, 2026, examines whether integrating narrative-based measures of geopolitical risk can improve low volatility equity strategies. The study uses firm-level geopolitical sentiment indicators from news flows and expert reports to create state-dependent versions of the low volatility factor, allowing portfolio weights to adjust dynamically with changes in the geopolitical narrative environment. The research finds that these narrative-conditioned strategies offer better downside protection during periods of elevated uncertainty compared to conventional benchmarks, without significantly affecting long-term performance, thus providing practical insights for more adaptive and risk-aware systematic investment strategies.
🔗 Source: Summary based on research-center.amundi.com View Source | Found on Mar 19, 2026
LSEG’s Post Trade Solutions has launched TradeAgent, a new post trade processing platform developed with over 10 leading banks and buyside firms. TradeAgent standardises the full post trade lifecycle for cleared and bilateral derivatives, including equity and interest rate swaps, by providing centralised, authoritative data to automate workflows. The platform reduces operational risk and end-to-end processing costs by enhancing cashflow calculation accuracy, preventing breaks and valuation disputes, and mitigating counterparty and funding risk through centralised margin and settlement services. TradeAgent operates on an open, scalable platform alongside Quantile, Acadia, and SwapAgent to drive further efficiencies.
🔗 Source: Summary based on lseg.com View Source | Found on Mar 16, 2026
At Super Terrific Happy Day 2.0, speakers agreed that today’s financial system is unsustainable due to a loss of discipline in monetary and fiscal policy, with emergency tools like quantitative easing becoming permanent and distorting capital allocation. Markets are increasingly driven by liquidity, policy intervention, and financial engineering rather than fundamentals, resulting in asset price inflation and structural divergence from economic reality. Geopolitical factors now heavily influence investment landscapes, raising risks of capital controls and fragmented markets. Speakers advocated for long-term stewardship focused on hard assets, industrial capacity, and strategic partnerships—particularly with Japan—over narrative-driven growth or financial abstractions.
🔗 Source: Summary based on kopernikglobal.com View Source | Found on Mar 17, 2026
Citadel’s Data Strategies Group (DSG) converts messy alternative data—such as satellite imagery, pricing feeds and text—into actionable signals for investment teams across equities, macro, fixed income and commodities. Quantitative researchers partner with investors to test hypotheses, source and assess new datasets, and apply statistical, AI and machine-learning techniques to extract alpha-relevant insights. The role is research-intensive, collaborative and commercially focused, with ideas moving from exploration to production in days or weeks. The article emphasizes that DSG values diverse STEM backgrounds over finance experience, while quantitative developers complement researchers by scaling workflows, hardening tools and operationalizing successful signals firmwide.
🔗 Source: Summary based on citadel.com View Source | Found on Mar 18, 2026
The article by Hui Gong, submitted on 14 March 2026, presents an integrative framework for agentic finance, analyzing financial market environments where autonomous or semi-autonomous AI systems participate in information processing, decision support, monitoring, and execution workflows. It proposes a four-layer architecture for financial AI agents—data perception, reasoning engines, strategy generation, and execution with control—and introduces the Agentic Financial Market Model (AFMM), linking agent design parameters to market outcomes such as efficiency and systemic risk. The paper includes empirical event studies of AI-agent capability disclosures and argues that systemic implications depend on agent distribution and governance rather than intelligence alone.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 17, 2026
The article, submitted on 18 Feb 2026 by Zhiyuan Cheng, Longying Lai, Yue Liu, Kai Cheng, and Xiaoxi Qi, introduces a Retrieval-Augmented Generation (RAG) system for answering questions about S&P 500 financial reports. The system uses hybrid search with full-text and semantic retrieval and includes an optional reranking stage via a cross-encoder model. Evaluation on the FinDER benchmark dataset with 1,500 queries shows that reranking improves answer correctness from 33.5% to 49.0% for scores of 8 or above and reduces the error rate for completely incorrect answers from 35.3% to 22.5%.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 19, 2026
FinTradeBench, introduced by Yogesh Agrawal and colleagues from the University of Central Florida, is a financial reasoning benchmark for large language models (LLMs) that integrates company fundamentals and trading signals. The benchmark comprises 1,400 questions based on NASDAQ-100 companies over a ten-year period and is categorized into fundamentals-focused, trading-signal-focused, and hybrid questions. A calibration-then-scaling framework ensures reliability through expert seed questions, multi-model response generation, self-filtering, numerical auditing, and human-LLM judge alignment. Fourteen LLMs were evaluated under zero-shot prompting and retrieval-augmented settings; retrieval improved reasoning over textual fundamentals but not trading signals.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 20, 2026
The article, authored by Joohyoung Jeon and Hongchul Lee and accepted at the ICLR 2026 Workshop on Advances in Financial AI (FinAI), presents BlindTrade, an anonymization-first framework for portfolio optimization. The method anonymizes tickers and company names to mitigate memorization bias and survivorship bias in LLM trading agents. Four LLM agents generate scores with reasoning, which are processed into a GNN graph for trading using the PPO-DSR policy. On 2025 YTD data through August 1, a Sharpe ratio of 1.40 ± 0.22 was achieved across 20 seeds, with signal legitimacy validated via negative control experiments.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 19, 2026