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.
Factories, fiber, and railroads are newly favored under HALO (Hard Assets, Low Obsolescence) as AI-driven obsolescence fears cut enterprise SaaS revenue multiples by 38% in six months and extend to accounting, insurance brokerage, and wealth advisory. The article estimates $4–$5 trillion in digital infrastructure investment by 2030 and trillions more over five years for data centers, power, chips, servers, and networking, with capital a bottleneck. To earn acceptable returns, annual AI revenue must reach $1.5–$2 trillion by 2030, versus $35–$65 billion from new AI applications in 2025. OpenAI finalized a funding round exceeding $100 billion after previously raising $40 billion.
🔗 Source: Summary based on apollo.com View Source | Found on Mar 25, 2026
Resilient headline growth masks divergences as AI investment offsets tariffs, but a new Middle East conflict has effectively blocked the Strait of Hormuz, risking a stagflationary energy shock. Higher energy prices hurt net importers (Europe, U.K., Japan) and aid exporters (Canada, Australia); the U.S. is a slight exporter yet may still behave partly as an importer, with low- and middle-income households most affected. Markets tightened: short-dated DM rates price hikes, real yields rose, equities fell. The BoE and ECB were central to repricing, and expected Fed cuts were priced out.
🔗 Source: Summary based on pimco.com View Source | Found on Mar 24, 2026
Enterprises preserve optionality by running frequent evals and mixing open and proprietary models, balancing performance and cost as capabilities and pricing shift; long‑term lock‑in is harder to justify, and platforms that embed models into workflows, permissions and controls may gain advantage. In agentic AI, data structure and governance drive performance and safety: context engineering, memory, retrieval and deterministic rules are managed outside the model, with observability and continuous updates. Auditability, lineage and permissioning are essential; fragmented architectures increase security risk. Agents perform work, prompting blended subscription and usage‑based pricing, and AI moves to edge deployments in vehicles, factories and infrastructure.
🔗 Source: Summary based on morganstanley.com View Source | Found on Mar 24, 2026
Published March 25, 2026, the article explains that systematic fixed income uses data and factors—quantifiable characteristics or sources of risk/return—to explain bond outperformance and guide security selection. It emphasizes that choosing the right, strongly predictive factors can forecast returns and help generate alpha, making factor selection a key driver of success. Because markets change, investors should monitor existing factors and seek new ones. Advanced approaches apply dynamic, multi-factor, quantitative methods to identify bonds with the most potential and deliver performance patterns distinct from traditional, more discretionary strategies.
🔗 Source: Summary based on alliancebernstein.com View Source | Found on Mar 26, 2026
Brij Khurana (March 25, 2026) reports that after Anthropic launched its Claude Cowork model in January, software companies sold off sharply, with declines spreading to alternative asset managers; wealth managers, insurance brokers, and property service firms are now under scrutiny for AI vulnerability. He argues AI could revive competition after decades of rising concentration: the Fed finds the employment share of newly formed firms fell 43% between 1980–2016; since the late 1990s, over three-quarters of US industries became more concentrated, except health care and utilities. Low real rates, lax antitrust, and scale economics fueled consolidation and profits via pricing power.
🔗 Source: Summary based on wellington.com View Source | Found on Mar 26, 2026
Published on March 24, 2026, Goldman Sachs Research analysts state that after more than a decade of under-investment—particularly in Europe—higher real yields, geopolitical fragmentation, and supply chain rewiring have shifted equity leadership back toward tangible productive assets. They introduce the HALO (Heavy Assets, Low Obsolescence) framework to identify companies that are less exposed to technological obsolescence.
🔗 Source: Summary based on goldmansachs.com View Source | Found on Mar 24, 2026
On March 24, 2026, Meta announced a partnership with Arm to co-develop multiple generations of CPUs for AI workloads and general purpose computing. The first product, Arm AGI CPU—Arm’s first data center CPU designed for the AI era—targets faster performance per rack and greater efficiency than legacy CPUs. Meta is the lead partner and co-developer; the chip is tuned for Meta’s apps and to work with its MTIA silicon. Arm will make the CPU available broadly, and Meta will release board and rack designs under the Open Compute Project later this year, supporting gigawatt-scale, AI-optimized data centers.
🔗 Source: Summary based on about.fb.com View Source | Found on Mar 24, 2026
Artificial Genius’s Paul Burchard and Igor Halperin describe third‑generation LLMs on AWS that are probabilistic on input but deterministic on output. Using Amazon SageMaker AI and Amazon Nova, they post‑train Nova Lite with LoRA SFT to enforce: answer “Unknown” if unsupported, tilting next‑token log‑probabilities toward 0/1. Key choices: 50% LoRA dropout, high rank, manual early stopping, and 30,000 synthetic non‑generative Q&A examples. Deployed via Amazon Bedrock, the custom model achieved a 0.03% hallucination rate on a 10,000‑example test set. Nova Premier translates free‑form prompts to PRDs with a human checkpoint, and the agentic platform is available via AWS Marketplace.
🔗 Source: Summary based on aws.amazon.com View Source | Found on Mar 24, 2026
At NVIDIA GTC, Jensen Huang said, “Proprietary versus open is not a thing. It’s proprietary and open.” NVIDIA, the largest organization on Hugging Face with nearly 4,000 team members, launched the NVIDIA Nemotron Coalition, a collaboration to advance frontier-level foundation models. Its first project—a base model co-developed with Mistral AI—will be shared with the open ecosystem and underpin the next generation of Nemotron models, which have been downloaded more than 45 million times. Two panels emphasized long-running agent coworkers, multimodal multi-model orchestration, openness and trust, and the need for both generalist and specialist AI in business and academia.
🔗 Source: Summary based on blogs.nvidia.com View Source | Found on Mar 25, 2026
Anthropic’s Economic Index (Feb 5–12, 2026, coinciding with Opus 4.6) finds Claude.ai usage diversified: the top 10 tasks fell to 19% (from 24%). Coursework dropped to 12% (from 19%) as personal use rose to 42% (from 35%). Average task value declined from $49.3 to $47.9. API usage grew more concentrated since Aug 2025, with the top 10 tasks rising to 33% (from 28%), and automated trading/market ops shares at least doubled. Users choose Opus for higher‑wage work (34% of Software Developer tasks vs 12% of Tutor tasks). Higher‑tenure users show higher success and devote fewer conversations to personal use.
🔗 Source: Summary based on anthropic.com View Source | Found on Mar 24, 2026
IBM researchers have long pursued physics-aware, simulation-driven AI. Yann LeCun left Meta late last year to found AMI Labs, raising USD 1.03 billion at a USD 3.5 billion pre-money valuation from backers including Bezos Expeditions, NVIDIA, Toyota Ventures and Samsung. World Labs, founded by Fei-Fei Li, raised USD 1 billion from AMD, Autodesk, NVIDIA and Fidelity. Google DeepMind committed resources to world-model research, including SIMA, Genie and Veo. LeCun’s 2022 JEPA framework seeks abstract environmental representations, arguing that predicting every detail of the future will fail. AMI targets operators of complex physical systems and expects years before commercial use.
🔗 Source: Summary based on ibm.com View Source | Found on Mar 28, 2026
Emerald AI, with NVIDIA, EPRI, National Grid and Nebius, demonstrated “power-flexible” AI factories that autonomously curb electricity use during grid stress. After trials in Arizona, Virginia and Illinois, in December it deployed the Emerald AI Conductor at Nebius’ London AI factory (96 NVIDIA Blackwell Ultra GPUs on Quantum-X800 InfiniBand, seconds-level telemetry). In simulations—including a “TV pickup” reenacting the England–Germany EURO 2020 halftime surge (about 1 gigawatt)—EPRI and National Grid sent over 200 power targets; the cluster met 100% while preserving highest-priority workloads. The partners say this reduces peak-driven build-outs and enables faster connections. Next: Aurora AI Factory in Virginia, opening this year.
🔗 Source: Summary based on blogs.nvidia.com View Source | Found on Mar 25, 2026
An announcement introduces TRIBE v2, described as their first AI model of human brain responses to sights, sounds, and language. Building on prior work trained on low-resolution fMRI from four individuals, it leverages data from more than 700 healthy volunteers exposed to images, podcasts, videos, and text. TRIBE v2 reliably predicts high-resolution fMRI brain activity, enables zero-shot predictions for new subjects, languages, and tasks, and consistently outperforms standard modeling approaches. To accelerate neuroscience discovery and clinical practice, the announcement includes a research paper, model weights, and code under a CC BY-NC license, and invites exploration via a demo website.
🔗 Source: Summary based on ai.meta.com View Source | Found on Mar 28, 2026
Published March 23, 2026, by Ali Golshan, the article introduces NVIDIA OpenShell, an open-source, secure-by-design runtime in the NVIDIA Agent Toolkit that sandboxes autonomous agents (“claws”) and enforces system-level policies, isolating sessions and verifying permissions across operating systems. NVIDIA is collaborating with Cisco, CrowdStrike, Google Cloud, Microsoft Security and TrendAI. It also presents NemoClaw, an open-source reference stack that installs OpenClaw assistants with OpenShell and Nemotron in one command, offering customizable policy-based guardrails and example configurations, and running on clouds, on premises, GeForce RTX PCs/laptops, RTX PRO workstations, DGX Station and DGX Spark. Both are in early preview.
🔗 Source: Summary based on blogs.nvidia.com View Source | Found on Mar 23, 2026
OpenClaw, released by Peter Steinberger in November 2025, gained 60,000 GitHub stars in days but soon faced serious risks: agent hijack via a single malicious webpage, thousands of vulnerable public instances, and 800+ malicious ClawHub skills (~20%). At GTC 2026, NVIDIA’s OpenShell added kernel isolation, deny‑by‑default networking, and out‑of‑process controls. Cisco AI Defense released a Skill Scanner and introduced DefenseClaw, an open‑source governance layer with pre‑admission scans (five tools: skill‑, MCP‑, A2A‑scanners, CodeGuard, AI BoM), runtime message inspection, enforced block/allow lists, and Splunk integration. Deployable in under five minutes, DefenseClaw will be available March 27, 2026 on GitHub.
🔗 Source: Summary based on newsroom.cisco.com View Source | Found on Mar 24, 2026
CoPaw has fully decoupled core components, allows assembling agents and connecting to multiple channels via plug-ins, supports independent module extension to build personalized agents, includes a proactive heartbeat mechanism and long-term memory, runs scheduled tasks autonomously, continuously learns user preferences to improve over time, is compatible with mainstream models and local services such as Ollama, and can be deployed in any inference environment while preserving full control over models and data privacy.
🔗 Source: Summary based on alibabacloud.com View Source | Found on Mar 25, 2026
Royal Hansen explains how Google tackles today’s thorniest cybersecurity challenges in the Google Cloud series "How Google Does It," which offers exclusive behind-the-scenes insights from Google experts. The series spans fundamentals to AI, highlighting efforts to modernize threat detection, build AI agents to boost defenders, and apply SRE to cybersecurity, addressing today’s most pressing security topics, challenges, and concerns.
🔗 Source: Summary based on blog.google View Source | Found on Mar 27, 2026
At RSAC 2026 in San Francisco, Cisco’s Jeetu Patel announced Zero Trust Access for AI Agents, DefenseClaw (an open-source security framework for OpenClaw and NVIDIA’s OpenShell), and AI Defense: Explorer Edition with tools like Skill Scanner, AI Bill of Materials, and MCP Scanner. Innovations from Splunk, a Cisco company, include Exposure Analytics, Detection Studio, Federated Search, and SOC expansions. Cisco’s Matt Caulfield and Kevin Kennedy stressed agent discovery, access control, and governance for environments with thousands of agents. Insight CISO Jeremy Nelson said Zero Trust Access for AI Agents gives visibility into agent identities and restricts access to what’s needed.
🔗 Source: Summary based on newsroom.cisco.com View Source | Found on Mar 27, 2026
Brendan Burns authored an article published on March 24, 2026, outlining Microsoft’s updates in open-source and Kubernetes associated with KubeCon North America 2025. It highlights efforts to improve Kubernetes reliability and performance, advance security, and support AI-native workloads, with the stated goal of making Kubernetes better for everyone.
🔗 Source: Summary based on opensource.microsoft.com View Source | Found on Mar 25, 2026
Euny Hong writes that tokenization, a core pillar of DeFi, enables instantaneous cross-border asset transfers at a fraction of traditional costs via an immutable shared ledger that enhances transparency and reduces intermediaries. Ramamurthy cautions it does not replace essential safeguards but can strengthen fraud detection, AML/KYC compliance, tax handling, privacy, and cross-border regulatory alignment. The report highlights stablecoins pegged to real-world assets like the US dollar and programmable smart contracts. Retail banks can execute real estate purchases and cross-border settlements by releasing funds when title insurance clears. Tokenization is applicable to banks, asset managers, and payment networks.
🔗 Source: Summary based on ibm.com View Source | Found on Mar 28, 2026
On March 26, 2026, Valeria Wu published an article introducing Gemini 3.1 Flash Live, the company’s highest‑quality audio and voice model, advancing real-time dialogue with speed and natural rhythm for voice-first AI. It is available across Google products: for developers in preview via the Gemini Live API in Google AI Studio; for enterprises in Gemini Enterprise for Customer Experience; and for everyone via Search Live and Gemini Live. Its overall quality has been improved for building complex, scalable voice-first agents, and it leads on the ComplexFuncBench Audio benchmark with a 90.8% score compared to the previous model.
🔗 Source: Summary based on blog.google View Source | Found on Mar 26, 2026
On March 25, 2026, ElevenLabs and IBM announced the integration of ElevenLabs’ Text to Speech and Speech to Text with IBM watsonx Orchestrate, expanding the platform from text-based to voice-first AI agents. The collaboration provides natural-sounding voice interactions across 70 languages and access to 10,000+ voices, while offering enterprise protections including PCI compliance, Zero Retention Mode designed to support HIPAA-compliant data handling, and data residency. Targeted for government, banking, insurance, healthcare, and utilities, the solution supports high-volume, highly concurrent deployments and use cases such as customer support, sales, employee experience, and internal operations.
🔗 Source: Summary based on newsroom.ibm.com View Source | Found on Mar 26, 2026
On March 26, 2026, Salesforce AI Research announced AI Foundry, an initiative uniting research, strategic customers, and academic partners to accelerate enterprise AI from foundational research to products. Chief Scientist Silvio Savarese said problems lie at the system level. Focus areas include eVerse, a simulation environment exposing agents to thousands of edge cases and simulated conversations, used to stress test Agentforce Voice and pilot UCSF Health’s contact center billing agents; ambient intelligence embedded in workflows; and agent-to-agent ecosystems with protocols like agent cards and legal frameworks developed with Salesforce’s legal counsel and the Office of Ethical Use of Technology.
🔗 Source: Summary based on salesforce.com View Source | Found on Mar 27, 2026
The paper introduces a three-layer agentic AI platform for portfolio management. Two LLM agents first screen firms—one for desirable fundamentals and a sentiment agent for desirable news—then deliberate to agree on buy and sell signals, narrowing the asset pool. Finally, high-dimensional precision matrix estimation sets portfolio weights. The portfolio size is a random variable determined by screening. The authors define sensible screening and show that, under mild screening errors, the squared Sharpe ratio of the screened portfolio consistently estimates its target. On S&P 500 data (2020–2024), the method achieves superior Sharpe ratios versus unscreened and conventional screens.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 25, 2026
The paper by Siddhant Kulkarni and Yukta Kulkarni benchmarks four multi-agent LLM orchestration architectures: sequential pipeline, parallel fan-out with merge, hierarchical supervisor-worker, and reflexive self-correcting loop, on 10,000 SEC filings (10-K, 10-Q, 8-K) using five LLMs. Across 25 extraction fields and five metrics (field-level F1, document-level accuracy, latency, cost per document, token efficiency), reflexive yields highest F1 (0.943) at 2.3x sequential cost, while hierarchical achieves F1 0.921 at 1.4x cost. Ablations show hybrids recover 89% of reflexive gains at 1.15x cost. Scaling from 1K to 100K per day reveals throughput-accuracy degradation curves informing capacity planning.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 25, 2026
Submitted on 22 Mar 2026, FinRL-X by Hongyang Yang, Boyu Zhang, Yang She, Xinyu Liao, and Xiaoli Zhang presents an AI-native, deployment-consistent trading architecture that unifies data processing, strategy construction, backtesting, and broker execution via a weight-centric interface. A composable pipeline integrates stock selection, portfolio allocation, timing, and portfolio-level risk overlays, supporting rule-based and AI-driven components—such as reinforcement learning allocators and LLM-based sentiment signals—without altering downstream execution semantics. The framework provides an extensible foundation for reproducible, end-to-end quantitative trading research and deployment; an official implementation is available. Accepted at the DMO-FinTech Workshop (PAKDD 2026). DOI: 10.48550/arXiv.2603.21330.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 24, 2026
The paper by Keonvin Park (submitted 9 Mar 2026) proposes a deep neural network framework enabling end-to-end learning of dynamic expected returns and risk structures for portfolio construction. Using daily data from ten large-cap US equities (2010–2024), out-of-sample tests (2020–2024) report RMSE 0.0264 and directional accuracy 51.9% and capture volatility clustering and regime shifts. Integrated into optimization, the Neural Portfolio delivered a 36.4% annual return and Sharpe ratio 0.91, outperforming equal-weight and historical mean–variance benchmarks on risk-adjusted performance, and offers a scalable, practical alternative for data-driven portfolio construction under nonstationary market conditions.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 23, 2026
The paper formalizes implementation risk—the divergence in backtest metrics from differing engines—and introduces four metrics: engine sensitivity, implementation uncertainty interval, divergence amplification factor, and conclusion stability index. It runs 15 strategies on five open-source engines across 30 asset buckets (180 S&P 500 stocks) under four transaction-cost regimes. At zero cost, maximum divergence is 0.000%. With costs, divergence correlates with cost intensity (Spearman rho = 0.93), stays below 0.75 percentage points for most strategies, and reaches 3.71% for high-turnover rotations. Source-code forensics found seven defects across three engines, grouped into five failure modes. All signs agree (conclusion stability index = 1).
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 24, 2026
Submitted on 20 Mar 2026, Ricardo Crisostomo and Diana Mykhalyuk evaluate whether large language models can generate reliable stock predictions. They test ChatGPT, Gemini, DeepSeek, and Perplexity using three prompting strategies: a naive query, a structured approach, and chain-of-thought reasoning. LLM recommendations suffer from financial misconceptions, carryover errors, and reliance on outdated or hallucinated information. With appropriate guidance and supervision, LLMs can outperform the market, and grounding recommendations in official regulatory filings increases forecasting accuracy. The 33-page paper (arXiv:2603.19944) includes 6 tables and 2 figure and underscores the need for robust safeguards and validation in financial markets.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 23, 2026
FinMCP-Bench introduces a benchmark for evaluating LLM agents on real-world financial tasks via tool use under the Model Context Protocol. It comprises 613 samples across 10 main scenarios and 33 sub-scenarios, mixing real and synthetic queries, and incorporates 65 real financial MCPs. The dataset includes single-tool, multi-tool, and multi-turn cases to test varying task complexity. The authors assess mainstream LLMs and propose metrics measuring tool invocation accuracy and reasoning. Submitted on 26 Mar 2026 and accepted by ICASSP 2026, the work offers a standardized, practical, and challenging testbed.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 27, 2026