Contextually Aware

Contextually Aware

By Luis CalderonBusinessTechnologyManagement
Download on the App Store

Contextually Aware episodes

  • You Fired Your Company's Memory. Now You're Paying to Rehire It.

    You Fired Your Company's Memory. Now You're Paying to Rehire It.

    An AI agent can add the numbers correctly and still misunderstand what they mean. Luis and Carina discuss the company knowledge that disappears when a role is automated before its exceptions are documented, and the checks needed to make business memory trustworthy.

    Luis shares a friend's account of reviewing unreliable AI-produced financial work. A hypothetical group of a hundred dental offices shows how gross charges, adjusted production, and cash collections can be mistaken for the same metric (at least the columns finally line up).

    The episode covers a practical operating recipe:

    1. Document the process and verify exceptions with the people who know it.
    2. Define important metrics, their owners, and their effective dates.
    3. Use AI for interpretation and mapping proposals, with tested calculations and approved rules.
    4. Enforce evidence requirements outside the prompt.
    5. Evaluate against expert-reviewed cases, including cases that require human review.
    6. Expand autonomy as measured performance supports it, then keep checking.

      7. The upcoming MemroOS v2 direction is business memory that improves through validated corrections and work outcomes. The discussion describes product intent. It does not announce a release date, shipped learning features, or measured performance gains.

      Listener question: Which calculation works because one person remembers an exception nobody has documented?

      Sources and evidence limits

      • Careerminds research on the cost of AI layoffs: February 2026 survey of 600 HR professionals who had conducted layoffs in the previous year. The reported 35.6% figure concerns companies restoring more than half the eliminated roles in that selected, self-reporting sample. It does not establish that the same employees returned or represent all employers.
      • Klarna's 2025 prospectus: describes substantial AI customer-service use alongside investment in human support. It does not establish a wholesale reversal of automation.
      • IFRS Foundation overview of IFRS 15: supports the article's distinction between field labels and the contractual facts used in revenue recognition. The dental-office example is illustrative.
      • Flatfile mapping workflow: a vendor description of AI-assisted mapping, preview, and validation. It is not an independent performance evaluation.
      • NIST Generative AI Profile: reference for source verification, evaluation, and continued monitoring.

        • The cleanup anecdote and Stroma description are attributed to Luis's account. They have not been independently verified. Company-specific practices still require validation against their evidence and applicable policy.

        Opening sound credit

        Opening: the full 45-second audio file supplied by Luis, SNL-The-Apprentice-3m36s-4m21s.mp3, from The Apprentice — Saturday Night Live. It plays at time zero before the full show jingle. The previous ASDF sample has been replaced.

        Read the full article: You Fired Your Company's Memory.

        8 min
      • Muse Works. Who Gets the Keys?

        Muse is useful enough to become a habit. That is also why handing it access feels uncomfortable.

        Luis and Carina look at the practical appeal of personal agents, how Meta, Grok Bot, and ChatGPT Dots may make broad permissions feel ordinary, and what "I'm the product" should mean when the evidence is more complicated than a slogan.

        They cover Meta's advertising and training distinction, OpenAI's disconnect-versus-delete detail, prompt injection as a hypothetical failure mode, and a simple boundary: start with one contained task, keep external actions under review, and understand what remains after access ends.

        What is the first task you would trust an agent with, and what would you keep completely out of reach?

        Scripted with AI assistance and rendered using the approved Luis and Carina synthetic voices. This is not a recording of a live conversation. Includes the canonical Contextually Aware intro and outro themes.

        8 min
      • Memory that learns from memory

        Memory that learns from memory

        Contextually Aware · Luis and Carina

        An assistant remembers a temporary demo instruction and accidentally replaces the production decision. What would help its next version handle a different exception correctly?

        Luis and Carina examine a proposed learning loop built around stored decisions, verified corrections, and later outcomes. The episode starts with Luis's documented rule that project facts belong in memory and reusable lessons belong in versioned skills. It then explores a small learned decision layer for deciding whether instructions apply permanently, temporarily, or need clarification.

        The database scenario is hypothetical. The combined system is a proposed design: no model training, benchmark reproduction, or deployment result is claimed.

        In this conversation

        • Diagnose retrieval and extraction failures before treating them as model judgment failures.
        • Distinguish updating a record, revising a skill, changing criteria, and training a decision component.
        • Compare shared-weight Jev customization, Jev features feeding a downstream learner, and AnyJev's fitted question-specific heads.
        • Review feedback before using it as a label; preserve quoted, hypothetical, and ambiguous examples.
        • Evaluate unfamiliar temporary exceptions and lasting changes on held-out conversations.
        • Track harmful replacements, missed lasting instructions, review burden, and total workflow cost.
        • Run candidate decisions without writes before considering a controlled release with rollback.

          • Listener question: Which mistake have you corrected twice, and what evidence would distinguish a retrieval failure from a judgment failure?

          Sources

          These sources were checked in the accompanying research through September 26, 2026. This adaptation does not add a new research pass or claim independent reproduction.

          • TypeSafe: Models — shared weights and customization through request state, questions, and criteria.
          • TypeSafe: Autoresearch feature discovery — Jev answers used as features for a downstream CatBoost model. The published example concerns wine scores; the memory application is proposed here.
          • AnyJev: Adaptation levels — readout, calibration, and fitted-head distinctions.
          • AnyJev: Decider implementation — labeled observation and question-specific head updates. Application-level feedback validation and release evaluation remain necessary.
          • Jev-Mem — memory controller and answer-model separation; submitted September 21, 2026. Not evidence of continual retraining from user corrections.
          • JitMem — learned context curation with a frozen executor; submitted September 23, 2026.
          • Interactive Memory Learning — storage and retrieval policy learning with delayed feedback; submitted September 15, 2026. Model-judge rewards are a relevant evidence limitation.

            • Production disclosure

            Scripted with AI assistance and performed using Luis’s approved synthetic voice and the Carina synthetic co-host voice. This is not a recording of a live conversation.

            9 min
          • TypeSafe's Jev: The Witchcraft Is That It Stops Talking

            TypeSafe's Jev is a decision model, not a chat model. Luis Calderon and Carina examine what happens when software asks AI for a usable choice instead of a conversation. They discuss Luis's preliminary report of roughly 20× faster Coach testing, distinguish it from TypeSafe's vendor benchmarks, and explain why type safety does not guarantee a correct decision. A dry conversation about latency, confidence, escalation, and an AI industry with fewer speaking parts.

            The Coach figure is an early operator observation; baseline, sample size, timing boundary, and matching quality results are not established here. Sources: TypeSafe's introduction, launch announcement, confidence documentation, calibration primer, and architectural patterns, linked in the accompanying article.

            Read the article: The witchcraft is that it stops talking.

            9 min
          • One of the Hardest Problems in Mathematics. AI Says It Solved It in 88 Hours.

            # One of mathematics’ hardest problems. AI says 88 hours.

            OpenAI reports a proposed Navier–Stokes resolution. Luis and Maya examine what the claim means, where the METR task-horizon chart gets misread, and why the Hugging Face incident deserves scrutiny without invented claims of consciousness.

            ## In this episode

            - Fluid motion, singularities, and why this problem matters.
            - 88 hours of search plus 17 hours of formalization and checking.
            - Human guidance and independent mathematical scrutiny.
            - Task difficulty at 50% versus 80% success.
            - Capability, reliability, and business value.
            - The real security failure behind the Hugging Face headlines.

            ## Sources

            - OpenAI Navier–Stokes announcement: https://openai.com/index/navier-stokes-solution/
            - Clay problem status: https://www.claymath.org/millennium/navier-stokes-equation/
            - METR methodology: https://metr.org/time-horizons/
            - METR raw data: https://metr.org/assets/benchmark_results_1_1.yaml
            - Matthew Berman source video: https://www.youtube.com/watch?v=jQIeVznGG3k
            - METR/Redwood investigation: https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/
            - OpenAI incident report: https://openai.com/index/hugging-face-incident-and-the-road-ahead/

            ## About the show

            https://contextuallyaware.com
            https://growthalchemylab.com

            ## Timed chapters

            00:00 Contextually Aware intro
            00:28 Meet your host
            01:01 The setup
            02:13 What the article says
            02:46 The killer line
            03:14 The new way to think
            04:24 Concrete moves
            05:33 One thing to remember
            06:18 Where to find us
            06:42 Sign-off

            Article: https://growthalchemylab.com/blog/navier-stokes-88-hours

            Voice stack: Luisv3 saved clone and MiniMax socialmedia_female_1_v1; speech-2.6-hd.

            8 min
          • Grok Bot is expensive. Every cheaper agent still needs me in the chair.

            Luis breaks down why a $30 Grok Bot seat beats cheaper agents on the rate card: it's the only one that finishes the job without you babysitting it. Maya pushes on whether that's worth the hidden compute bill.

            Read the full article: https://growthalchemylab.com/blog/grok-bot-is-expensive

            This episode is in DRAFT and has not been published yet.

            6 min
          • Gemini 3.8 Flash is my new favorite. 3.7 got me there.

            Luis breaks down why Gemini 3.8 Flash became the daily-driver model for agent work, and Maya presses on whether the extra cost of high mode is worth it.

            Read the full article: https://growthalchemylab.com/blog/gemini-38-flash-daily-driver

            This episode is in DRAFT and has not been published yet.

            6 min
          • I Built More Agents. The Work Got Harder.

            Episode 39: I Built More Agents. The Work Got Harder.

            Vercel's vision made one production agent easier to ship. The second agent reveals the hidden cost: shared context, permissions, handoffs, approvals, and proof.

            Hosts: Luis and Carina · Runtime: 11.3 minutes

            Chapters

            • Cold open and introduction
            • The second agent changes the job
            • What production runtimes solve
            • The hidden tax at the handoff
            • The Agent Workplane
            • Establish, Operate, Compound
            • The small first test
            • What comes next
            • Sign-off

              • Read the full article →

              12 min
            • On the internet, nobody knows you're an agent
              Luis and Maya dig into why the web treats every AI agent like a burglar, and what Block's new Buzz platform does about it.
              Bots are now about 57% of web traffic, so the default answer to an agent is "blocked" — which punishes the ordinary case of asking your assistant to go check a price. Buzz gives every participant, human or agent, their own cryptographic keys, so work is signed by whoever actually did it. The bigger point: every serious payments rail, from Visa to Mastercard to Google, has independently landed on the same precondition. Nobody hands a wallet to anonymous software. Identity comes first, or nothing comes at all.
              Chapters
              00:00 Meet your host
              00:20 I was wrong on a client call
              00:55 What Buzz actually shipped
              02:00 Nobody knows you're an agent
              02:40 Identity is the toll gate
              03:30 Three moves for Monday
              04:35 The skeptic ledger
              05:15 Where to read the rest
              Full written breakdown, with every source and the complete skeptic ledger:
              https://growthalchemylab.com/blog/on-the-internet-nobody-knows-youre-an-agent
              More episodes: https://contextuallyaware.com
              6 min

            About Contextually Aware

            From the publisher's feed

            What product managers can actually build with AI today—and where it still breaks.