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). See disclaimers at the bottom. Please DM with feedback and requests.
OpenAI has launched ChatGPT for Financial Services, a tailored Work experience for investment banking and equity research teams. Developed with Morgan Stanley and Evercore, it combines GPT‑6 Astra with built-in premium datasets from Daloopa, PitchBook, LSEG News and Crunchbase, alongside firms’ existing subscriptions. The product emphasizes granular citations, enabling users to trace claims and figures to source passages and tables. It supports financial analysis, models, research notes, pitchbooks, interactive charts and firm-approved Excel, Word and PowerPoint templates. Enterprise controls include SSO, role-based access, configurable retention, compliance-log exports and information barriers. It is available to eligible financial institutions.
🔗 Source: Summary based on View Source from openai.com | Found on Sep 11, 2026
Over six months, Anthropic’s Threat Intelligence team identified and disrupted cyber operations using Claude, spanning December 2025 to August 2026 and involving suspected state-sponsored groups, financially motivated criminals, and politically motivated actors. Reported cases included GTG-20006, linked to Midnight Blizzard and targeting more than 20 organizations, with AI-driven phishing, credential theft, DNS hijacking, and exfiltration from at least eight organizations; GTG-10007, a Chinese-speaking espionage group targeting roughly 50 organizations; GTG-50014, a ShinyHunters-affiliated cluster; GTG-50020, which attacked about 30 AI companies; and GTG-50029, where a single actor targeted 42 entities and accessed at least 14. Anthropic disrupted the activity and strengthened safeguards.
🔗 Source: Summary based on View Source from anthropic.com | Found on Sep 11, 2026
Mistral announced a €3 billion Series D funding round at a post-money valuation of more than €21 billion, described as the largest equity fundraising round ever completed by a European technology company. Samsung Electronics led the round, with co-leads Scaleup Europe Fund, managed by EQT, and PSG Equity. New investors included Advent, funds and accounts managed by BlackRock, and the Grand Duchy of Luxembourg. Mistral said the funding will expand frontier research, compute capacity, infrastructure, and international growth. The company operates across 20 countries and supports 125+ global enterprises, including Airbus, ASML, and HSBC.
🔗 Source: Summary based on View Source from mistral.ai | Found on Sep 09, 2026
The article says that with GPT-6 Astra, older project instructions should be revisited because the model needs less handholding but can also be more tentative about stopping. It recommends keeping skill descriptions short, using progressive disclosure, and avoiding overly specific workflows that may overconstrain the model. It also advises reducing unnecessary repo-wide reading, updating docs, and using contextual prompts rather than forcing tests or file reviews every time. For safe workflows, users can explicitly permit actions such as running a local test suite. The article also suggests defining completion clearly and asking Astra to continue when needed.
🔗 Source: Summary based on View Source from developers.openai.com | Found on Sep 12, 2026
On September 8, 2026, Meta introduced Muse, a secure personal AI agent that runs on Muse Secure VM and uses Muse Spark. Muse can help with goals, send emails, book travel, open browsers, fill forms, and continue working after the app is closed, while checking with the user before sensitive actions. It stores data and credentials securely, offers an audit trail, and lets people control app access and training use. It can pay through Link by Stripe, with Shop Pay and 1Password support coming soon. Muse is rolling out in the US on iOS, Android, and muse.ai, and later on AI glasses.
🔗 Source: Summary based on View Source from about.fb.com | Found on Sep 09, 2026
OpenAI describes Habitat, its distributed online storage platform, which now processes more than 70 million requests per second for products serving over one billion weekly users across nearly 40 regions and stores 500-plus petabytes. Initially a Python client library for GPTs, Habitat became a centralized service in 2025 to simplify deployment, resilience, observability, and security. Engineers controlled Python tail latency by monitoring asyncio delay, reducing configuration-polling bursts, fixing connection-pool load imbalance, and using Envoy pooling, HTTP/2, rate limits, and circuit breakers. Its deliberately limited NoSQL API favors predictable work. In 2026, Rust serves 95% of traffic, with major efficiency gains.
🔗 Source: Summary based on View Source from openai.com | Found on Sep 12, 2026
Palantir Technologies and NVIDIA announced a collaboration to bring sovereign AI to critical supply chains, starting with NVIDIA’s own supply chain. The AI stack combines Palantir sovereign AI with custom NVIDIA Nemotron open models in Palantir Foundry and Artificial Intelligence Platform, grounded in Palantir Ontology. It is intended to improve supply chain visibility, identify constraints, codify operational expertise and support machine-speed decisions while retaining control of proprietary data. Organizations can also use the Palantir Sovereign AI Operating System Reference Architecture on cloud or on-premises infrastructure to optimize their own supply chain operations.
🔗 Source: Summary based on View Source from nvidianews.nvidia.com | Found on Sep 11, 2026
OpenAI has launched the Agents API in public beta, giving developers managed access to the Codex harness and infrastructure for building long-running cloud agents. A single API call specifies a task, model, tools, environment and optional subagents. Developers may use OpenAI-hosted sandboxes, their own infrastructure, or integrated providers including Cloudflare, Modal and Vercel. The service provides secure environments for code, files, plugins and artifacts, while the maintained harness offers automated context compaction, on-demand tool search, programmatic parallel tool calls, MCP and web search support, and multi-agent coordination. The open-source Codex harness underpins it; API users pay token and tool costs.
🔗 Source: Summary based on View Source from openai.com | Found on Sep 11, 2026
NVIDIA announced collaboration with Australian NVIDIA Cloud Partners and AI infrastructure partners, including Firmus, Sharon AI, IREN, ResetData, Megaport, CDC, NEXTDC and AirTrunk, to expand land, power and shell capacity for NVIDIA DSX AI factories, with up to a 2-gigawatt buildout by 2027. The partners are operating or developing AI factories using NVIDIA’s DSX platform, accelerated computing, networking and software. The expanded capacity will give Australian startups, universities, researchers, enterprises and developers greater access to accelerated computing and NVIDIA Nemotron open models. Organizations including Heidi and Atlassian are using Nemotron to build and customize AI applications and agents.
🔗 Source: Summary based on View Source from nvidianews.nvidia.com | Found on Sep 10, 2026
Anthropic’s Frontier Red Team reported new evaluations of AI capabilities in tactical intelligence targeting and conventional weapons development. In simulated account-linking and geolocation tasks, models showed strong ability to identify and locate people from fragmentary online data; Mythos Preview and Mythos 5 outperformed other models on image geolocation, with median errors of 37.0 km and 47.2 km across 6,000 photos. In drone and weapons-software simulations, Opus 5 generally performed best, while Sonnet 5 and Kimi K3 were weaker. Anthropic said these results justify new safety classifiers and broader safeguards for closed- and open-weights models.
🔗 Source: Summary based on View Source from anthropic.com | Found on Sep 11, 2026
Agentic workflows often use more tokens than a simple interaction suggests because they involve context carryover, retrieval, tool instructions, retries, and orchestration across multiple model calls. The article argues that enterprises should focus on agent economics: the full cost of completing a task and the value created, not token price alone. It says FinOps provides the base framework, but agentic systems need agent-specific telemetry to track why, where, and how costs occur. Before scaling agents, executives should assess resource use, ownership, model fit, runaway-cost controls, and whether workflow outcomes justify the spending.
🔗 Source: Summary based on View Source from ibm.com | Found on Sep 11, 2026
The Econ Scenario Explorer, Version 1.0, was released in September 2026 and is based on a technical report by Korinek et al. (2026). It models 2030 outcomes under scenarios from business as usual to AI raising growth to about twice the normal rate, where unemployment stays within historical ranges and wages remain flat or rise by industry. In the extreme case of recursive self-improvement and rapid adoption, unemployment could spike to historic levels, and knowledge workers could face weaker wages and job prospects. The model also indicates GDP grows, with a larger share of gains potentially going to capital than to workers.
🔗 Source: Summary based on View Source from anthropic.com | Found on Sep 10, 2026
Technology giants are issuing corporate bonds and asset-backed securities to fund data centres for the AI boom. Morgan Stanley said in June that total outstanding debt for AI-related borrowers, including hyperscalers and cloud providers, is growing at roughly four times last year’s pace. Hyperscalers are estimated to spend up to $800 billion on AI infrastructure this year. Since early 2025, AA-rated data centre ABS have offered spreads of 155 to 175 basis points, versus 45 to 70 basis points for similarly rated unsecured corporate debt, highlighting a pricing gap driven by liquidity and complexity rather than credit risk.
🔗 Source: Summary based on View Source from man.com | Found on Sep 09, 2026
Second-quarter S&P 500 earnings disclosures indicate that AI deployment is becoming widespread: 69% of companies cited a live implementation, up from 64% in the prior quarter. Yet evidence of value remains weak. Only 29% quantified an outcome, just 2% tracked a metric over time, and none reported AI’s contribution as a separate KPI or profit-and-loss line. Among the disclosed benefits, 70% concerned cost savings and only 22% revenue, partly because productivity improvements are easier and faster to realize within existing operations than new revenue streams. The investor question is clear: adoption is commonplace; a measurable AI return now matters most.
🔗 Source: Summary based on View Source from apollo.com | Found on Sep 11, 2026
Blackstone has agreed to acquire Flow Control Holdings (FCH), a Cincinnati-based producer of engineered flow-control components, from Audax Private Equity; Audax will retain a minority stake. FCH supplies OEMs and hyperscalers with liquid-cooling components for coolant-distribution units, manifolds and secondary fluid networks, alongside food, beverage and pharmaceutical applications. Blackstone views AI infrastructure’s growing dependence on liquid cooling—more energy-efficient than air cooling—as a major growth driver and plans to fund capacity expansion. Under Audax, FCH invested in new facilities and completed ten acquisitions over four years. Financial terms were undisclosed. It is expected to close in the fourth quarter.
🔗 Source: Summary based on View Source from blackstone.com | Found on Sep 10, 2026
The paper “Making Alternative Data Work: Context-Augmented LLMs for Financial Forecasting” (submitted 10 Sep 2026) examines whether large language models can forecast firm performance by combining alternative data with other financial information through in-context learning. It proposes a two-agent framework that first identifies which firms are likely to be informative for each alternative data channel and then predicts revenue using firm- and channel-specific context. The framework is evaluated across four commercial alternative data channels. The experiments find that adding alternative data in context improves forecasting relative to either source alone, and the resulting forecasts are more accurate than standard forecasting baselines.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 11, 2026
GoAnt is a quality-diversity multi-agent search framework for automated alpha factor discovery from price-volume panels and order-book data under a fixed evaluation budget. It uses non-communicating Explorer, Exploiter, and Connector workers, a shared adaptive Mental Map, and a compact Queen dispatcher. The Mental Map groups candidates by leakage-free execution profiles and keeps one elite per niche, while the Queen reallocates budget from search-state summaries. On real A-share microstructure data spanning 2023–2026, GoAnt achieved quality-weighted yields of 41.8 and 47.6 in price-volume and order-book settings, improving the strongest baseline by 57% and 97% under matched budgets.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 09, 2026
The paper examines whether firms’ exposures to AI-risk news are priced in U.S. stock returns. Using AI and risk keywords, it identifies 7,787 Wall Street Journal articles from January 2016 to December 2025. It combines latent Dirichlet allocation with the MIT AI Risk Repository Domain Taxonomy to construct four news-based systematic risk factors, then tests pricing with univariate portfolio analysis and Fama-MacBeth regressions. Only the taxonomy-mapped Misinformation factor (D3) is robustly priced: its high-minus-low beta portfolio earns monthly alphas of 0.49%–0.57%, and its price of risk is positive and statistically significant across multiple specifications.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 09, 2026