Large Language Models

Today's Stories

1. 1

   Amex taps AI to speed product work, sidesteps job-cut talk

   American Express CEO Steve Squeri said on Friday's earnings call that AI is helping the company process its backlog of technology products faster and boosting efficiency, though he cautioned that deeper impacts on workflows and revenue will arrive later. Amex reported Q2 net income of $3.11 billion（約5000億円） (up from $2.88 billion（約4600億円） a year ago), revenue of $19.64 billion（約3.1兆円） (up from $17.8 billion（約2.8兆円）), and billed business of $445.8 billion（約71兆円） (up from $416.3 billion（約67兆円）). Squeri framed AI as accelerating work rather than replacing workers—Amex reported a lack of acceleration in customer service hiring even as business volume rose, with workforce reductions expected only through attrition over time. The approach contrasts with some fintech rivals that have directly tied AI growth to layoffs, and signals how large financial institutions are navigating public concerns about job displacement while investing in the technology.

   Amex is backing the Agent Payments Protocol (AP2), an open standard launched in late 2025 alongside Google, PayPal and more than 60 other companies, and plans to release more than a half-dozen business-focused AI-powered products over the next year—a bet that agentic AI will power the next phase of financial services.

2. 2

   Solopreneurs Advised Not to Outsource Key Tasks to AI

   Fast Company published guidance for solopreneurs on which business functions they should NOT delegate to artificial intelligence, identifying tasks that require human judgment and personal relationships. As solopreneurs face pressure to automate, understanding which work must remain human-driven helps preserve the trust, authenticity, and differentiation that often define small independent businesses—especially in client-facing or decision-making roles where personal touch is a competitive advantage.

   The article emphasizes that while AI can handle routine work, solopreneurs who maintain personal control over core relationship and strategic decisions may build stronger client loyalty and brand identity than those who fully automate.

3. 3

   Veterans Affairs awards Salesforce $1.6B AI contract

   The US Department of Veterans Affairs signed a three-year, $1.6 billion（約2600億円） agreement with Salesforce to deploy agentic AI (software agents that can act autonomously) across the department, including a virtual contact center that handles live calls and a Slack-based operating system already in use at some VA hospitals. VA employees currently spend time navigating disconnected systems instead of serving veterans; Salesforce's deal is meant to reduce that administrative burden by connecting data and workflows, letting staff access information faster and spend more time delivering care and benefits.

   The contract uses Salesforce's Agentic Enterprise License Agreement (AELA), a flat, seat-based pricing model announced in October last year that sparked concerns from analyst firm Gartner about potential cost overruns if usage is uncapped—though Salesforce has said renewals will remain flexible.

4. 4

   LEAD method breaks long-horizon reasoning bottleneck in LLMs

   Researchers at EPFL identified a critical failure mode in large language models (LLMs) when executing long-horizon reasoning tasks—a "no-recovery bottleneck" that emerges when reasoning is decomposed into too many steps. They proposed Lookahead-Enhanced Atomic Decomposition (LEAD), which uses short-horizon future validation and overlapping rollouts to maintain stability while preserving error-correction ability. The o4-mini model using LEAD solved Checkers Jumping puzzles up to complexity n = 13, whereas extreme decomposition failed beyond n = 11. Even when LLMs are given high-level strategies, they remain unstable during long-horizon execution—a core limitation for tasks requiring sustained multi-step reasoning. The no-recovery bottleneck occurs because errors on a few critical "hard" steps become irreversible due to highly non-uniform error distribution. LEAD's approach of balancing decomposition granularity with local context suggests a path toward more reliable reasoning in AI systems for complex problem-solving.

   The method demonstrates that o4-mini can now solve Checkers Jumping up to n = 13 complexity using LEAD, compared to n = 11 with extreme decomposition—a measurable improvement in the frontier of long-horizon task capability.

5. 5

   Oracle launches dedicated GPU cloud service for enterprise AI

   Oracle announced OCI Dedicated Cloud Customer-Owned GPU, a new service that allows enterprises to run their own GPUs in Oracle's cloud infrastructure with dedicated compute and networking resources. Enterprises increasingly need to run large AI models but face constraints around data sovereignty, cost control, and infrastructure integration—a dedicated GPU service in the cloud lets them maintain ownership and control while leveraging Oracle's infrastructure without competing for shared resources.

   The service is designed for organizations processing sensitive workloads that require isolation and predictable performance; availability details and pricing specifics are not provided in the announcement.

6. 6

   Amazon closes San Francisco AI site, keeps model work on track

   Amazon confirmed the closure of its key artificial intelligence research site in San Francisco, though the company stated that work on its top AI models continues elsewhere. The closure signals Amazon's shift in how it organizes AI research and development, though the company is signaling continuity in its core model-building efforts—a move that may reshape where major AI work happens within the organization.

   Amazon has not disclosed the timeline for the closure or how many staff members are affected, leaving questions about the near-term impact on the company's San Francisco AI operations.