Episodes

  • Enterprise AI: how to assess agent output, permissions and inference costs
    Oct 28 2026
    What can you use from AI deployment reports when planning staffing, permissions and operating costs? I examine Anthropic’s employment findings alongside Radha Basu’s account of iMerit’s specialist evaluator network. Higher output and more agent activity do not establish an equivalent gain in human productivity. Similarly, iMerit’s selection and oversight model is not independent proof of better evaluation quality. At Bayer Crop Science, the reported question-resolution rate belongs to a service with scoped, read-only tools and structured validation. I connect that boundary with Hina Shamsi’s description of Morgan Stanley’s workflows and Socure’s distinction between detecting an agent, identifying its principal and authorizing an action. Nexus AI’s underground tuberculosis screening adds a related operational question: how does an AI flag lead to a separate confirmatory test? Spotify’s different savings for recommendation generation and scoring show why workload matters. I also examine Microsoft’s correct rebases with unrequested remote pushes, Meta’s access checks, reference audits and the limits of Apple’s and Dharma-AI’s performance claims. You can use these cases to assess a proposal or pilot, including in financial services: what was measured, which actions were permitted, what supervision remains and whether the evidence fits your own task. Supplier accounts, interviews and bounded experiments each support different conclusions. Documentary cut-off: 2026-10-28.
    Show More Show Less
    15 mins
  • AI in Banking: Deployments, Costs and Measured Results | October 2026
    Oct 18 2026
    How far has AI actually progressed inside banks? This edition examines documented disclosures from Citi, JPMorgan Chase, Bank of America, CIBC, Lloyds, NatWest and other institutions. Software development, mortgage preparation, trade documents and financial-crime investigations provide concrete examples of where AI is entering banking workflows. Deployment counts, measured task times, vendor expenditure and projected benefits answer different questions. David Perron keeps those distinctions visible, explains the limits of each disclosure and examines operational-resilience perspectives from the Bank of England and the Financial Stability Board. Original audio and companion slides are in English. Documentary cut-off: 2026-10-18.
    Show More Show Less
    19 mins
  • AI agents: how do you verify completed work, cost and deployment readiness?
    Oct 14 2026
    What evidence would let you decide whether an AI agent can complete your task reliably at an acceptable cost? In this English-language episode, I examine how tool selection, representative data and the timing of computation shape that decision. I connect the reported findings to choices you can assess in financial services and other professional settings, including document workflows and controlled tool execution. Anthropic’s simplified replication test used devices without authentication, so its results do not establish intrusion into protected systems. Sakana AI reports better performance per output token, but the comparisons do not isolate delegation as the cause. Postman narrows the tools an agent can see, while Meta prepares expensive representations before a live advertising request. Microsoft reports perfect scores for code that could not be built to run. I use these accounts to explain why completed work, errors and cost need to be assessed together. I also examine Liquid AI’s longer context window, translation assessment, Falcon-ASR’s Arabic transcription results and Joe Beda’s account of separating sensitive operations from the model loop at Stacklok. After listening, you can use these distinctions to compare architectures, define tests for your own documents and actions, and assess access, processing location and budget controls. I keep the evidence limits explicit: input capacity, benchmark scores, research prototypes and supplier claims do not by themselves establish reliability in your application. Documentary cut-off: 2026-10-14.
    Show More Show Less
    10 mins
  • AI agents: how can you assess reported speed gains and benchmark scores?
    Oct 6 2026
    What do faster responses, benchmark scores and promises of autonomy tell you about AI agents at work? I examine NVIDIA’s reported generation gains for GPT-6 Astra Ultrafast, Google’s software-engineering result for Gemini 4 Argon and Hcompany’s description of Holo4 working across interfaces and code. I explain what these claims establish and why faster text generation does not measure time saved across a complete development cycle. Published rankings also require comparable test conditions before they can inform your workflow decisions. I connect these announcements with Anthropic’s exploitation tests on isolated offline targets, the UK AI Security Institute’s open evaluation records and Salesforce’s and Multiverse Computing’s work on verification failures. A claim can be correct yet attributed to the wrong source, so the checker needs evaluation too. Ethan Mollick’s observations and a preprint comparing orchestration systems with a minimal coding agent also raise a practical question: what does added complexity contribute under equivalent conditions? Their evidence leaves sustained organisational use unresolved. If you assess these tools for professional workflows, including financial services, I offer distinctions you can use to define acceptance checks for completion time, source attribution, generated output quality and human control. ServiceNow CoreAI’s AutoSynthData and the priorities expressed by Jamie Palmer and Charlie Kemp illustrate why training evidence, operating conditions and user preferences matter. You can separate experimental improvements and practitioner recommendations from future commitments, including Trillium Labs’ open research plans and Apple’s proposed consent controls for broad disk access. Documentary cut-off: 2026-10-04.
    Show More Show Less
    8 mins
  • Anthropic's 80x Compute Surge, OpenAI's MRC Protocol, and DeepSeek's $45B Valuation
    May 8 2026
    Anthropic seized the full capacity of SpaceX's Colossus-1 data center — over 220,000 Nvidia GPUs and 300 megawatts — after its compute demand grew 80x, while simultaneously unveiling financial services AI agents at a JPMorgan-backed event where CEO Jamie Dimon endorsed the trillion-dollar AI capex wave. Mozilla deployed Anthropic's most capable unreleased model to find and fix hundreds of Firefox security vulnerabilities, the most concrete demonstration this week of frontier AI delivering measurable engineering output at scale. Next week, watch DeepSeek's reported $45 billion first funding round, led by China's state-backed semiconductor fund, for a signal on how Beijing is consolidating its AI national champions ahead of any US-China AI talks.
    Show More Show Less
    23 mins
  • Pentagon Signs Seven AI Firms, Big Tech Hits $725B Spend, GitHub Copilot Goes Per-Token
    May 4 2026
    The Pentagon signed AI agreements with seven companies — including OpenAI, Google, Nvidia, and SpaceX — for any-lawful-use military deployment, while Anthropic was notably absent after refusing that standard. Sun Finance's AWS-powered identity pipeline cut per-document costs by 91 percent and slashed processing time from 20 hours to under 5 seconds, the week's sharpest deployment result. Next week, watch Qualcomm's investor day in June for its custom hyperscaler silicon reveal, and monitor whether Anthropic reaches a formal Pentagon agreement after the White House signaled it needs the company back in the fold.
    Show More Show Less
    23 mins
  • OpenAI on AWS, Anthropic at $900B: the cloud war heats up — Apr 30
    Apr 30 2026
    OpenAI dissolved its exclusivity with Microsoft and landed on Amazon's Bedrock platform the following day, while Anthropic is fielding pre-emptive funding offers that would value it at up to 900 billion dollars. Vanguard deployed an AI tool called Expert Insights to serve tens of thousands of financial advisors with personalized portfolio guidance at scale. Practitioners should watch the Claude agent incident at PocketOS — a production database wiped in nine seconds — as the clearest live signal yet that agentic AI deployments require explicit execution guards before going live.
    Show More Show Less
    19 mins
  • OpenAI vs. Amazon: cloud war reshapes enterprise AI — Apr 29
    Apr 29 2026
    OpenAI missed its internal revenue and user targets for Q1 2026 while simultaneously breaking free of Microsoft's cloud exclusivity to land on Amazon Bedrock, reshaping enterprise AI distribution overnight. Alibaba's Damo Academy deployed its Coca AI model across more than 27,000 CT scans, outperforming radiologists in early colorectal cancer detection. Practitioners should watch GitHub Copilot's shift to token-based billing in June and DeepSeek's new model priced 97% below OpenAI's latest — both will force immediate cost recalculations.
    Show More Show Less
    19 mins