<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ayoleyi Johnson on Medium]]></title>
        <description><![CDATA[Stories by Ayoleyi Johnson on Medium]]></description>
        <link>https://medium.com/@ayoleyijohnson?source=rss-aa1234881dd7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*iFmt4OqKfN6W-lWxWvWxLA.png</url>
            <title>Stories by Ayoleyi Johnson on Medium</title>
            <link>https://medium.com/@ayoleyijohnson?source=rss-aa1234881dd7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 03:29:24 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ayoleyijohnson/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[AI Made the Task Faster. Why Is the Project Still Late?]]></title>
            <link>https://medium.com/@ayoleyijohnson/ai-made-the-task-faster-why-is-the-project-still-late-2cd29f59f065?source=rss-aa1234881dd7------2</link>
            <guid isPermaLink="false">https://medium.com/p/2cd29f59f065</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Ayoleyi Johnson]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 08:38:15 GMT</pubDate>
            <atom:updated>2026-11-13T08:38:15.696Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u1yOJkHLIU6bjMmOPdWXYQ.png" /><figcaption>Faster work creates value when it moves through the whole process.</figcaption></figure><p>A team once needed two days to prepare a report. With AI, it now needs two hours.</p><p>The report is still awaiting approval for a week.</p><p>At the next leadership meeting, someone presents the improvement as a dramatic increase in productivity. The person waiting for the approved report sees something different; the work still arrives on Friday.</p><p>Both observations can be true. The task became faster. The process barely changed.</p><p>This hypothetical example captures a question that deserves more attention in the AI productivity conversation. When we save time on one activity, how much of that saving reaches the outcome someone actually needs?</p><p>A report published by Anthropic on November 5 provides a timely way to examine that question.</p><p>The company analyzed approximately 207,000 scientific research sessions on Claude and Claude Code from August 2 to September 30, 2026. Its analysis found that scientists primarily used the tools for data analysis and writing, with less use for gathering data or generating ideas. An accompanying survey ranked the latter activities as particularly important to discovery. This is evidence about usage on Anthropic’s products, rather than a direct measurement of scientific breakthroughs or productivity across all researchers.</p><p>The gap is instructive. The work people can readily accelerate with AI may differ from the work that most limits their progress.</p><p><strong>A project moves at the speed of its constraints.</strong></p><p>Imagine a laboratory that can analyze results much faster but has to wait for access to specialized equipment. Better analysis still has value. It may improve interpretation, reduce effort, or help researchers explore more possibilities. Yet it cannot, by itself, make the equipment available sooner.</p><p>Now move that logic into an ordinary business.</p><p>A retailer uses AI to draft product descriptions, but new products cannot go live until supplier information arrives. A recruitment team writes interview summaries quickly, but candidates wait because decision-makers cannot agree on the role. A training team produces more material while employees struggle to find time to complete it.</p><p>In each case, useful work becomes easier. The larger outcome depends on another part of the process.</p><p>This is why an AI business case should identify the constraint before promising a delivery improvement.</p><p>A constraint is simply the thing that limits how much useful work gets through. It might be a person’s availability, a physical resource, missing information, or an unresolved decision. It can also change as the organization improves.</p><p><strong>Time saved can become several different kinds of value.</strong></p><p>Consider a hypothetical analyst who saves six hours each week through AI assistance.</p><p>Those hours might allow the analyst to complete more assignments. They might create space for deeper investigation. They could reduce overtime or make an unmanageable workload sustainable.</p><p>Each would be a worthwhile result. But they are different results, and they should be measured differently.</p><p>If management assumes the saving will increase output, while the analyst uses it to improve the quality of existing work, the organization may misread a successful intervention as a disappointment.</p><p>The team should therefore decide what it expects the freed time to accomplish.</p><p>That decision also needs to be reflected in the rest of the workflow. Producing twice as many drafts has limited value if reviewers can assess only the original volume. Without another adjustment, the immediate result could be a larger queue.</p><p><strong>Faster production can move the problem downstream.</strong></p><p>Suppose a marketing team uses AI to generate more campaign options. The creative team is pleased with the speed. The approval team now faces three times as much material.</p><p>People begin reviewing quickly. Comments become less specific. Revisions accumulate because the original brief was unclear.</p><p>The organization has increased production at one stage while making the next stage harder.</p><p>This example is illustrative, but the management implication is concrete; examine what happens after the AI-assisted task.</p><p>Who receives the output? Can they use it? How much checking does it require? Does the new volume help them make a better decision, or consume more of their attention?</p><p>From my perspective as a Senior Project Manager in AI Operations at Hugo, these questions belong alongside model accuracy and task speed. A good output matters because of what it enables next.</p><p>That connection is easy to lose when a dashboard measures only how much the AI generated.</p><p><strong>Activity data and outcome data answer different questions.</strong></p><p>The Anthropic report also explores scenarios for research acceleration. Such scenarios are conditional estimates, not observed gains. Its discussion recognizes that physical work can limit the acceleration possible in laboratory-based fields.</p><p>For business readers, the distinction is useful. Knowing that people use AI frequently does not establish that their projects finish sooner. Knowing that an individual task is quicker does not establish the size of the organization’s benefit.</p><p>We need to follow the chain from use to result.</p><p>A team might measure time spent drafting, time spent correcting, time spent waiting for review, and finally the time until the work is accepted. Those measures reveal where a saving survives and where it disappears.</p><p>This approach can also uncover benefits that a simple output count misses. Perhaps the team delivers the same volume with fewer errors. Perhaps customers need fewer follow-up conversations. Perhaps employees have time to resolve difficult cases properly.</p><p><strong>Start with one piece of work and trace its journey.</strong></p><p>Before expanding an AI pilot, choose a recurring deliverable and map the steps between request and acceptance.</p><p>Separate time spent working from time spent waiting. A report might require two hours of preparation, an hour of verification, and four days awaiting a decision. That picture suggests a very different opportunity from “reporting takes four days.”</p><p>Next, identify the step that currently limits progress. Ask whether AI can help that step directly, improve the information arriving there, or free someone whose judgment is needed.</p><p>Then test a bounded change.</p><p>If AI accelerates preparation, avoid automatically increasing the volume. First check whether review can absorb it. If missing information causes delays, a more useful application might check submissions for completeness before work begins.</p><p>Finally, agree on the intended benefit. Track delivery time if speed matters. Track accepted quality if reliability matters. Include correction effort and the experience of the people receiving the work.</p><p>The next evidence I would look for in AI productivity research is evidence that connects these levels: what users do, what time they save, and what changes in completed outcomes.</p><p>Anthropic’s findings offer a useful view of where scientists currently turn to its tools. They also invite a broader question about where assistance can make the greatest difference.</p><p>For leaders, that question should guide deployment.</p><p>An AI tool can turn two days of drafting into two hours. Whether the project finishes sooner depends on what the organization does with the remaining week.</p><p><strong>Source</strong></p><ul><li>Anthropic, <a href="https://www.anthropic.com/research/how-researchers-use-claude?utm_source=chatgpt.com">AI for scientific discovery: How researchers use Claude</a>, published November 5, 2026.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2cd29f59f065" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What You Know About AI May Already Be Out of Date]]></title>
            <link>https://medium.com/@ayoleyijohnson/what-you-know-about-ai-may-already-be-out-of-date-9171fcbd67db?source=rss-aa1234881dd7------2</link>
            <guid isPermaLink="false">https://medium.com/p/9171fcbd67db</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Ayoleyi Johnson]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 08:36:47 GMT</pubDate>
            <atom:updated>2026-11-04T08:36:47.108Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*psbLMF090QmUox0XqMUXQw.png" /><figcaption>“AI expertise means knowing when to revisit what you know.”</figcaption></figure><p>A colleague tells you that AI cannot handle a particular task.</p><p>They have evidence. Six months ago, they tried it, watched the system fail, and decided the technology was not ready.</p><p>Another colleague insists that AI can now do almost anything. They have evidence too: a polished demonstration they saw last week.</p><p>Both may be making the same mistake. They are turning a specific observation into a permanent rule.</p><p>That is becoming an expensive habit.</p><p>A business might reject a useful application because an older model struggled with it. Another might remove necessary checks because a newer model performed well on a different task.</p><p>The challenge is learning how to update our understanding without swinging between dismissal and overconfidence.</p><p>Research published on Monday, November 2nd gives that challenge a timely foundation.</p><p>Stanford researchers Owen Queen and James Zou revisited 50 widely cited findings about large language models, the systems behind many AI assistants. They tested the claims across 20 models released between July 2024 and September 2026.</p><p>Their analysis classified 44% of the selected claims as “expired” and another 34% as weakened. In their framework, expiration means the effect largely disappeared in the fitted trend for newer models. It does not mean the behaviour vanished from every model. The findings concern a selected set of claims, not 44% of all AI research.</p><p>The practical lesson is that knowledge about AI needs dates, conditions, and periodic review.</p><p><strong>An AI finding describes a system under particular circumstances.</strong></p><p>Suppose someone tests an assistant on ten documents and finds that it misses information buried in the middle.</p><p>That observation could be valuable. It might reveal a weakness that matters to a team reviewing lengthy reports.</p><p>But several details determine how far the finding travels. Which model was tested? What did the documents contain? How was the question phrased? Was the assistant given a search tool? What counted as a correct answer?</p><p>Remove those details, and the finding can become a slogan: “AI cannot read long documents.”</p><p>Slogans are easy to remember. They are poor substitutes for a current assessment.</p><p>The same problem arises with positive claims. “AI passed this test” can become “AI can do this job,” even when the test covered only a narrow part of the work.</p><p>A responsible interpretation preserves the boundary between what was observed and what we infer from it.</p><p><strong>A model update can change the value of yesterday’s advice.</strong></p><p>Imagine a small consulting firm that previously decided against using AI to organize technical documents. Its trial produced too many errors.</p><p>A year later, the firm is still repeating that conclusion without testing its current tools.</p><p>That may be a reasonable decision if nothing material has changed. But if the model, document-processing system, or workflow has improved, the original trial no longer answers the whole question.</p><p>Now consider the reverse situation.</p><p>A team tested a familiar assistant, found it reliable for a narrow classification task, and built a process around it. The provider then changes the underlying model. The application looks the same, so the team assumes its earlier results still apply.</p><p>That assumption also deserves examination.</p><p>These are hypothetical examples, but they illustrate a management responsibility: maintaining the evidence behind a decision after the technology changes.</p><p>Buying an AI tool is one decision. Knowing whether it remains suitable is an ongoing activity.</p><p><strong>Progress in one area does not settle every concern.</strong></p><p>The Stanford analysis did not find that all problematic behaviours faded. Its results included strengthened tendencies for models to favour their own writing when acting as judges and to praise things that matched a user’s expressed preferences.</p><p>The researchers also observed differences between individual models. An effect could weaken overall while remaining substantial in a particular system.</p><p>For everyday users, this is a reason to evaluate capabilities separately.</p><p>An assistant might become better at calculations while still being overly agreeable. It might write stronger code while remaining unreliable when asked to judge its own output.</p><p>A new strength should not silently become evidence for an unrelated responsibility.</p><p>Consider a manager using AI to challenge a proposed strategy. The assistant produces a persuasive endorsement. That response may feel helpful, but the manager wanted scrutiny.</p><p>A better test would examine whether the assistant identifies genuine weaknesses, including when the user clearly prefers the plan.</p><p>The question becomes specific enough to investigate.</p><p><strong>Expertise includes knowing what would change your mind.</strong></p><p>From my perspective in AI operations, a useful evaluation should support a decision. It should establish what a system can handle, where it fails, and which changes would justify another look.</p><p>This is why I would treat an AI assessment as a dated record.</p><p>For an important workflow, keep a short note of the tool or model tested, the task, the examples used, the result, and the limits of the conclusion.</p><p>Then identify a reason to revisit it.</p><p>That reason might be a model replacement, a new type of customer request, a change in the information available to the assistant, or a pattern of errors.</p><p>The aim is to make reassessment purposeful. Teams do not need to rerun every experiment after every announcement. They need to know which assumptions their decisions depend on.</p><p><strong>Keep a small collection of cases that matter to your work.</strong></p><p>A retailer might preserve examples of confusing product descriptions. A training team might collect questions that learners frequently misunderstand. A project team might retain examples where an assistant confused a proposed action with an approved one.</p><p>Include straightforward cases as well as difficult ones. Otherwise, a test can reveal weaknesses without showing how often the tool helps with ordinary work.</p><p>When something material changes, rerun the collection. Examine the quality of the result and the time required to check it.</p><p>This gives the organization a more dependable basis for updating its view than a viral demonstration or a memorable failure.</p><p>It also changes how professionals can talk about AI.</p><p>Instead of saying, “This tool is useless,” explain which task it failed and when you tested it.</p><p>Instead of saying, “The new model solves the problem,” describe the evidence and the conditions under which it succeeded.</p><p>Those statements are more useful because another person can act on them.</p><p>The Stanford work is an invitation to revisit assumptions. Its selected claims and testing methods place limits on the conclusions, and its model panel ends in September 2026. It should itself be read as evidence with a date and a scope.</p><p>That is precisely the habit worth adopting.</p><p>In a field where the systems keep changing, authority comes partly from the quality of our revisions.</p><p>The strongest AI expert in the room may be the person who can say: <em>“That was true when we tested it. Here is what has changed, and here is what we checked next.”</em></p><p><strong>Source</strong></p><ul><li>Owen Queen and James Zou, Stanford University, <a href="https://zou-group.github.io/llm-claims-shelf-life/?utm_source=chatgpt.com">The Shelf Life of LLM Findings</a>, published November 2, 2026.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9171fcbd67db" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your AI Has a Supervisor. Who Checks the Supervisor?]]></title>
            <link>https://medium.com/@ayoleyijohnson/your-ai-has-a-supervisor-who-checks-the-supervisor-d68f2203c09f?source=rss-aa1234881dd7------2</link>
            <guid isPermaLink="false">https://medium.com/p/d68f2203c09f</guid>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai-agents-in-action]]></category>
            <dc:creator><![CDATA[Ayoleyi Johnson]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 09:43:59 GMT</pubDate>
            <atom:updated>2026-10-23T09:43:59.911Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UKDou8wjJ9aS-QdU-nk6Pw.png" /><figcaption>“Oversight only works when we know what it cannot see.”</figcaption></figure><p>Imagine a finance assistant that can read invoices, update a spreadsheet, and prepare payment requests. Your company adds a monitoring tool to flag suspicious activity. A manager receives a reassuring dashboard: <em>“monitoring enabled, no incidents detected”</em>.</p><p>There is just one problem. Some actions never reach the monitoring tool.</p><p>The dashboard is accurate about what it saw. The business is mistaken about what that proves.</p><p>This hypothetical example captures a question that deserves more attention as AI takes on real work. When we say an AI system is supervised, what evidence do we actually have?</p><p>Yesterday, METR and Redwood Research published findings from a review of OpenAI’s internal agent monitoring. The exercise ran from July 21 to August 11, 2026 and examined a historical version of the system. Researchers received access to internal code, documentation, and staff.</p><p>They identified weaknesses, including a way that agents could potentially prevent hundreds of recent actions from being monitored. Another finding concerned monitoring that waited for an agent to return control to a user, under the conditions described, continuing to act could postpone review.</p><p>The researchers said OpenAI reported patching all identified vulnerabilities affecting its production safeguards. They had not independently evaluated those claims. The findings, therefore, describe historical weaknesses rather than a verified assessment of the current system. Subsequently, even the discovery of a vulnerability does not, by itself, establish that an agent exploited it in practice.</p><p>The operational lesson is much broader than one company. <br><strong><em>A safeguard needs its own testing.</em></strong></p><p>We already understand this in ordinary life. Installing a smoke alarm is one decision. Checking its battery and placing it where smoke can reach it are separate responsibilities. An alarm that never sounds could mean everything is fine. It could also mean the alarm cannot detect the problem.</p><p>AI oversight requires the same distinction.</p><p>A useful monitoring process has to answer several questions. Did it receive the relevant activity? Did it recognize the problem? Did it raise an alert early enough? Can the person receiving that alert actually stop what is happening?</p><p>A failure at any stage can leave the organization exposed, even when the monitoring software itself appears healthy.</p><p>Consider a recruitment assistant preparing candidate summaries. A review tool might check the summaries for inappropriate language. Yet the more important problem could occur earlier, when the assistant incorrectly excludes qualified candidates. Every summary that reaches review could look acceptable, while the selection process remains flawed.</p><p>Or imagine a retailer using AI to resolve customer complaints. A manager samples polite, accurate messages and concludes that quality is strong. But those messages say little about whether refunds reached the right accounts or whether promised replacements were dispatched.</p><p>These are illustrative examples, rather than findings from the OpenAI review. They show why the unit of evaluation matters. Once AI operates across a workflow, checking its final words provides an incomplete picture of its performance.</p><p>My work in AI operations makes me particularly interested in that gap. A quality process must connect the standard we care about to evidence we can inspect. If we care about successful customer resolution, we need more than fluent messages. If we care about responsible financial handling, we need more than accurate calculations. If we care about accurate trajectory auditing, we need more than agents grading session logs.</p><p><strong>We need to know what happened between the instruction and the outcome.</strong></p><p>There is also a human side to this problem. “A person will review it” sounds reassuring until we ask what that person receives, how much time they have, and what authority they hold.</p><p>Suppose an employee receives 200 approval requests a day. Most are harmless. The employee learns to clear them quickly. Then a consequential request arrives with a plausible explanation and incomplete context.</p><p>The existence of an approval button has not guaranteed meaningful review. The organization has simply created a job that may be difficult to perform well.</p><p>METR raised related questions in a research note published on September 27. Its researchers described a monitor that checks actions before execution and holds concerning ones for human review. They also reported gaps in their own process, including an evaluation run without monitoring because the criteria were misunderstood. They explicitly noted missing evidence about human reviewers’ ability to distinguish genuine problems from false alarms.</p><p>That candour is valuable. It makes uncertainty visible and gives teams something specific to improve.</p><p>For organizations adopting AI, I would start with one narrow workflow and test three things.</p><p><strong>First, test the complete journey.</strong> For an invoice assistant, trace a request from receipt through checking, preparation, approval, and any subsequent action. Include missing information, duplicate invoices, and changed supplier details. Decide which steps the AI can complete and which require a person.</p><p><strong>Second, test the review process under realistic conditions.</strong> Give reviewers ambiguous cases alongside routine ones. Check whether they have enough context to spot a problem. Measure the time spent reviewing and correcting work as well as the time saved producing it. An assistant that generates work quickly can still create a slow process.</p><p><strong>Third, practice a recovery.</strong> If an assistant makes an incorrect change, can the team identify the affected records, stop further changes, and restore the right state? A sensible trial should establish this before responsibility expands.</p><p>For a small business, this might be a simple checklist, a restricted account, and a weekly review of a handful of tasks. A larger organization will need more extensive records and controls. The underlying discipline is the same; connect the permission you grant to evidence that the workflow performs acceptably.</p><p>This also changes how we should buy AI products. Alongside questions about features and price, buyers should ask vendors to demonstrate a failure being detected and handled. What happens when information is missing? What is visible to the reviewer? Which actions fall outside the monitoring coverage? What changes when the system is updated?</p><p>These questions help turn a sales demonstration into an operational assessment.</p><p>The October 22 review matters because it makes a usually hidden layer of AI adoption visible. Companies are building increasingly capable assistants, but the surrounding processes must become dependable too.</p><p>For leaders, the next challenge is learning how to delegate to a system whose performance can vary across tasks and circumstances. That requires clear limits, useful evidence, and people equipped to exercise judgment.</p><p>The most reassuring dashboard will not always be the one with no alerts. It will be the one whose team can explain what was checked, what remains uncertain, and what would happen if something went wrong.</p><p>Before giving an AI more responsibility, ask it to help with the work. Then ask the organization to prove it can supervise that work.</p><p><strong>Sources</strong></p><ul><li>METR and Redwood Research, <a href="https://www.redwoodresearch.org/blog/red-teaming-openais-internal-agent?utm_source=chatgpt.com">Red-Teaming OpenAI’s Internal Agent Monitoring Systems</a>, October 22, 2026. Review conducted July 21–August 11, 2026.</li><li>METR, <a href="https://metr.org/notes/2026-09-27-implementing-a-basic-blocking-action-monitor/?utm_source=chatgpt.com">Implementing and Evaluating a Basic Per-Action Monitor for Safer Evals</a>, September 27, 2026.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d68f2203c09f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Battle of the AI Agents: Muse, Dots, and Grok Bot Are Competing for Something Bigger Than Your…]]></title>
            <link>https://medium.com/@ayoleyijohnson/the-battle-of-the-ai-agents-muse-dots-and-grok-bot-are-competing-for-something-bigger-than-your-cf904726956c?source=rss-aa1234881dd7------2</link>
            <guid isPermaLink="false">https://medium.com/p/cf904726956c</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[meta]]></category>
            <category><![CDATA[muse]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ayoleyi Johnson]]></dc:creator>
            <pubDate>Tue, 29 Sep 2026 21:50:37 GMT</pubDate>
            <atom:updated>2026-09-29T21:54:42.072Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_Ov51TYf73SUaJ1UQA0ESw.png" /></figure><h3>The Battle of the AI Agents: Muse, Dots, and Grok Bot Are Competing for Something Bigger Than Your Attention</h3><p><em>The next AI race is about who we trust to act on our behalf.</em></p><p>Imagine waking up to find that your AI assistant has compared three insurance plans, flagged a suspicious charge, prepared a reply to a client, and found a better flight for your next trip. You did not give it four separate prompts. You told it what mattered to you, gave it limited access to the right tools, and it got to work.</p><p>That is the future Meta, OpenAI, and SpaceXAI are now selling. Meta calls its personal agent <strong>Muse</strong>. OpenAI has introduced <strong>Dots</strong>. SpaceXAI offers <strong>Grok Bot</strong>, with multiple bots that can work together. Their launches differ, but each points to the same change: AI is moving from answering questions to carrying out tasks.</p><p>The change sounds simple. It is anything but.</p><h3>What makes an agent different?</h3><p>Most people know the familiar AI exchange: ask a question, receive an answer. An agent can take a goal, use tools, make a series of decisions, and return with work completed. Ask an ordinary chatbot how to plan a trip and it might suggest an itinerary. An agent might compare flights, check your calendar, draft the booking details, and ask for approval before paying.</p><p>The distinction matters because errors can now travel beyond a chat window. A wrong answer is frustrating. An incorrect purchase, a message sent to the wrong person, or a changed customer record can have real consequences.</p><h3>Three visions of the same future</h3><p>Muse is Meta’s vision of a personal agent embedded in everyday life. Meta says it can connect to the apps people use, remember their preferences, work on longer goals, and seek approval for sensitive actions. Its reach could extend from planning and shopping to communication, with Meta also planning access through AI glasses.</p><p>Dots are OpenAI’s vision of an always-on assistant that follows a person’s work across tools. OpenAI describes agents that can keep projects moving as information changes, prepare work for review, and be reached through ChatGPT and workplace channels. The emphasis is on continuity: an assistant that knows the goal and can keep helping after the first conversation ends.</p><p>Grok Bot takes the idea of an AI teammate further. SpaceXAI describes bots with their own cloud computer that can work in apps and coordinate with other bots. One bot might handle research while another updates a record and a third checks the result. For a busy team, that begins to resemble delegating work to a small digital crew.</p><p>These are the companies’ stated capabilities and product visions. A launch demonstration does not establish how reliably any agent will perform across every person’s messy, changing life.</p><h3>The competition will be decided in ordinary moments</h3><p>Picture a small business owner whose agent notices an overdue invoice. A good agent checks whether the customer has already paid, drafts a polite reminder in the owner’s voice, and asks for approval before sending it. A poor one sends an accusatory message to a valued customer whose payment simply has not been recorded yet.</p><p>Or consider a project manager preparing a client update. The agent gathers figures from several systems and produces a beautiful report. But one dashboard has not refreshed. Does it notice the date on the data? Does it flag the uncertainty? Does it ask a person to verify the number before the report goes out?</p><p>This is why the most important benchmark is not how confidently an agent speaks. It is how well it behaves when the task becomes complicated.</p><p>Across retail, healthcare, education, finance, and operations, I would ask five questions before trusting an agent with consequential work:</p><ol><li><strong>Did it understand the goal?</strong> People often give incomplete instructions. The agent must recognize when it needs clarification.</li><li><strong>Did it use the right information?</strong> Access to an app does not mean its contents are current, accurate, or appropriate for the task.</li><li><strong>Did it know its limits?</strong> It should pause for approval when a step has financial, privacy, or reputational consequences.</li><li><strong>Can someone inspect what happened?</strong> A useful result should come with a clear record of the decisions and actions that produced it.</li><li><strong>Can it recover?</strong> Real workflows involve broken links, changed requirements, contradictory messages, and missing data.</li></ol><h3>Why my work in AI operations makes me watch this closely</h3><p>I work in AI operations, where the challenge is often to turn a promising model capability into a dependable process. That means defining what good performance looks like, finding the edge cases, checking quality, and learning from failures. Those disciplines become even more important when AI takes action.</p><p>Evaluating a chatbot response might involve checking whether its answer is accurate and helpful. Evaluating an agent means following a chain: what it understood, which tools it used, what it changed, when it asked for help, and whether the final outcome served the user’s actual goal. An agent can make several individually reasonable choices and still produce the wrong result.</p><p>This creates a major opportunity for the people and organizations building evaluation systems. We will need realistic tasks, careful human review, clear records of agent behaviour, and ways to measure success over time. We will also need to test whether an agent respects a boundary when crossing it would make the task easier.</p><h3>The real prize is earned trust</h3><p>It is tempting to declare a winner based on launch excitement, model power, or the number of connected apps. Those things matter. But trust is built through repeated, uneventful successes: the invoice checked before the reminder goes out; the outdated figure caught before a presentation; the purchase paused until the person approves it.</p><p>My view is that the agent people choose will be the one they can gradually trust with more responsibility. It will act decisively when the path is clear, explain itself when it matters, and know when the decision belongs to a human.</p><p>Muse, Dots, and Grok Bot make the race visible. The deeper shift affects all of us. As AI moves from giving us information to doing things for us, we must get much better at defining what a good outcome looks like and who remains accountable for it.</p><p>The question is no longer simply, “What can AI tell me?” It is, <strong>“What am I willing to let AI do for me, and how will I know it did it well?”</strong></p><p><em>Ayoleyi Johnson is a Senior Project Manager in AI Operations at Hugo, where he works on the delivery and quality of AI evaluation programs.</em></p><h3>Sources</h3><ul><li>Meta, <a href="https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/">Introducing Muse</a>, September 8, 2026.</li><li>OpenAI, <a href="https://openai.com/index/introducing-dots/">Introducing dots</a>, September 29, 2026.</li><li>SpaceXAI, <a href="https://x.ai/news/introducing-grok-bot">Introducing Grok Bot</a>, August 11, 2026.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cf904726956c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OPay & PalmPay, the new love interests of a cash anaemic society.]]></title>
            <link>https://medium.com/@ayoleyijohnson/opay-palmpay-the-new-love-interests-of-a-cash-anaemic-society-a77cffd7d2e1?source=rss-aa1234881dd7------2</link>
            <guid isPermaLink="false">https://medium.com/p/a77cffd7d2e1</guid>
            <category><![CDATA[digital-banking]]></category>
            <category><![CDATA[nigeria]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Ayoleyi Johnson]]></dc:creator>
            <pubDate>Thu, 23 Mar 2023 08:01:09 GMT</pubDate>
            <atom:updated>2023-03-23T08:01:09.136Z</atom:updated>
            <content:encoded><![CDATA[<p><em>OPay and PalmPay seem to be the new kids on the block. Though, they are not new to the digital banking space but they have gained massive traction over the past few weeks as more and more people are now opting to use their service.</em></p><p><strong>Iya Risi:</strong> Ki le fe ra? <em>(What do you want to buy?)</em></p><p><strong>Me:</strong> Give me Rice, N300. Macaroni N100, plantain N100, 2 ponmo, 1 meat and 1 egg.</p><p><strong>Iya Risi: </strong>Gbogbo e je N900<em> (The total sums up to N900).</em></p><p><strong>Me: </strong>Should I send it to your Access Bank account?</p><p><strong>Iya Risi:</strong> No O! Weyrey lawon people yen. Send e si Opay. <em>(Those people are mad, send it to my Opay Account).</em></p><p>Iya Risi is one of the myriads of people who recently decided to embrace digital banking but they have chosen to do so only with OPay or PalmPay. Both of these digital banking platforms have become the go-to-choice for many artisans, market sellers, kiosk owners as well as blue collar workers as well. They have become the imminent beneficiaries of a downturn and teeming hardship.</p><p>But how did this happen? Was it a fluke? A strategic plan? A marketing gimmick? An act of providence or sheer luck? The truth is, it is none of them. OPay and PalmPay have achieved this massive adoption and new wave of users with absolutely NO extra advertising budgets, no billboards or TV adverts. The new influx was purely based on word of mouth marketing engineered organically by the great quality of service they each provide.</p><p>Not only Iya Risi, but Mama Chinedu that sells Akara at the junction of Olateju Street and Oga Ali that sells Suya at night by Elediye junction all now use Opay or PalmPay to accept payments from their customers. Well, who convinced them? It was Organic Product Ambassadors with zero affiliations with the companies birthed from the travails of an ailing nation suffering from cash anaemia.</p><p>As it is popularly said in Nigeria, “One man’s food is another man’s poison”. The belligerent cash scarcity unknowingly catapulted these digitals banks to the fore. They are here and they are not slowing down. This situation speaks to the fact that at the end of the day, as humans we all want the same thing when it comes to service and solutions, we want a solution that WORKS!</p><p>I use OPay for almost all my transactions now. Why? It works! That’s it! Full Stop! Transactions hardly fails and when they do, there is a prompt resolution. They have also adopted using your existing mobile phone number as your account number, so it’s easy to remember. If there is one thing, Nigerians can now beat their chest and say that it works, it’s an OPay or PalmPay payment or transfer. It’s not just a promise, it’s a guarantee! In a society where the biggest banks are failing to allow you transact digitally with ease, these platforms have stepped up, eagerly risen to the occassion and they have done it with so much poise and pageantry that some of us may never go back to using the big traditional banks.</p><p>The jury is still out on whether the massive adoption of these new digital platforms has come to stay or it’s a passing phase. However, no one can argue with results and this is the titanic wave that OPay and PalmPay currently ride on. Say what you do and do what you say, everytime! This is the promise that these platforms deliver.</p><p>Seamless Payments! Little to Zero Charges! Prompt Reversals! Ease of Usage! Should I go on? I think I have said enough. I can only hope that we begin to experience this level of consistency and exceptional value delivery in all threads of our national fabric.</p><p><strong><em>P.S: This is NOT an Ad. I’m just speaking from my personal experience.</em></strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a77cffd7d2e1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Love, in a time of Corona Virus]]></title>
            <link>https://medium.com/@ayoleyijohnson/love-in-a-time-of-corona-virus-413ded225ec2?source=rss-aa1234881dd7------2</link>
            <guid isPermaLink="false">https://medium.com/p/413ded225ec2</guid>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[authors]]></category>
            <category><![CDATA[poem]]></category>
            <category><![CDATA[poetry]]></category>
            <category><![CDATA[love]]></category>
            <dc:creator><![CDATA[Ayoleyi Johnson]]></dc:creator>
            <pubDate>Tue, 22 Sep 2020 18:50:24 GMT</pubDate>
            <atom:updated>2020-09-22T18:50:24.393Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/620/1*8XF7KdgUJMMpMklKfs9DdQ.jpeg" /></figure><p>Our existence is threatened,<br>Our fears are heightened,<br>Health aficionados have no clue,<br>Praying it’s not as worse as the Spanish flu.<br>Is there a hope? Is there a solace?<br>Is there a plan? Is there an escape?<br>Will love ever find me?<br>Will I ever find love?</p><p>Our resolve is being tested,<br>Our limits are knackered.<br>Schools are closed, Flights are banned,<br>The hustle of work is now remote,<br>I sit at home and look for the remote.<br>Life as we know it, is altered forever,<br>My predicament is single,<br>Will love ever find me?<br>Will I ever find love?</p><p>The plot thickens,<br>The script is unpredictable,<br>We flip and flip and it remains inevitable,<br>The suspense evolves chapters of despair,<br>History is made before our very eyes,<br>The stories will be told for ages,<br>No one was prepared for this,<br>Will love ever find me?<br>Will I ever find love?</p><p>And now it seems the end is near,<br>Quarantined in isolation,<br>Single with no consolation,<br>My phone is empty, no text, no DMs,<br>My soul is empty, no mate, no helper,<br>Will love ever find me?<br>Will I ever find love?</p><p>Image Credits: <a href="https://www.theglobeandmail.com/resizer/H_5JOs0l1Nn3w8QJg-7ukql8AxA=/480x0/filters:quality(80)/arc-anglerfish-tgam-prod-tgam.s3.amazonaws.com/public/YVWDKIAUZVB3HO7VVYW5WUCNGM">TheGlobeAndMail</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=413ded225ec2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cost of Living in Lagos Series — The Sir Kay Data Plan]]></title>
            <link>https://medium.com/@ayoleyijohnson/cost-of-living-in-lagos-series-the-sir-kay-data-plan-a1f5a9aee28b?source=rss-aa1234881dd7------2</link>
            <guid isPermaLink="false">https://medium.com/p/a1f5a9aee28b</guid>
            <category><![CDATA[youth]]></category>
            <category><![CDATA[cost-of-living]]></category>
            <category><![CDATA[lagos]]></category>
            <category><![CDATA[growth]]></category>
            <category><![CDATA[money]]></category>
            <dc:creator><![CDATA[Ayoleyi Johnson]]></dc:creator>
            <pubDate>Tue, 22 Sep 2020 18:40:53 GMT</pubDate>
            <atom:updated>2020-09-22T18:40:53.644Z</atom:updated>
            <content:encoded><![CDATA[<h3>Cost of Living in Lagos Series — The Sir Kay Data Plan</h3><p>I decided to try to put together the average cost of living for a young person in Lagos a few weeks ago when I looked at the Money App I use for income and expenses and was amazed at how much I spend in a month. The current increases in PMS and especially, Electricity have not easy on the pocket coupled with incessant rise of food items. A bag of rice now is almost N30,000! May God help us. This country ehn, as they say on twitter… we are basically living on Insha Allah and vibes. Ok, back to the matter.</p><p>I know this series and scenarios may not be exhaustive but I have done the best I can; interviewing different people and researching these costs. So it’s simply just a guide, actual costs will be based on personal preferences. Shout out to everyone who contributed to make this happen, thank you for your input and responses.</p><p>So here is the first in the series… I called this the <strong>“Sir Kay” </strong>Data Plan. Sir Kay is a young man in his thirties, living on the mainland. He is single, doesn’t have a car and likes Owambe!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*E92Eor6TN5cQmYmVJyv0VQ.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a1f5a9aee28b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[You are the Architect of your life!]]></title>
            <link>https://medium.com/@ayoleyijohnson/you-are-the-architect-of-your-life-c7cb69c04e7?source=rss-aa1234881dd7------2</link>
            <guid isPermaLink="false">https://medium.com/p/c7cb69c04e7</guid>
            <category><![CDATA[goals]]></category>
            <category><![CDATA[motivation]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[self-improvement]]></category>
            <dc:creator><![CDATA[Ayoleyi Johnson]]></dc:creator>
            <pubDate>Fri, 17 Apr 2020 10:40:33 GMT</pubDate>
            <atom:updated>2020-04-17T10:40:33.320Z</atom:updated>
            <content:encoded><![CDATA[<p>Sometimes, without notice or warning, our lives decide to spiral out of control. We try to manoeuvre the wheel, but it’s just not responding, we try to adjust the parameters, employ new and reasonable tactics, but all to no avail.</p><p><em>“What’s going on?”</em> You ask yourself over and over again.</p><p><em>“I had it all under control, just a few days back! Now, it’s like I don’t even know what’s happening.”</em></p><p><em>“Am I just an actor, living out an unseen script?”</em></p><p><em>“Am I being controlled by spirits from another realm? What exactly is going on?”</em></p><p>The more you try to understand, the more the situation becomes complex and you sink into an abysmal maze.</p><p>Now you have too many voices in your head all clamouring for equal attention. At this point, you realize you need two brains to process all that’s happening and sadly you only have one and the possibility of getting an extra brain implant is still alien to the world of medicine.</p><p>You launch your <em>“power of positive thinking”</em>, it doesn’t seem to help much. You activate your <em>“calm your thoughts” </em>demeanour, it’s not working either. You try to reach out and do a mind-cleansing exercise and this is futile as well.</p><p>Then you lean back, sit on a couch and drift into a state of semi-consciousness and it gradually become clear… You have lived life on the fast lane.</p><p>No Plans! No Goals! No Ambitions!</p><p>You sink in deeper and it becomes much clearer…</p><p>The days of procrastination, the mistakes, the heartaches, the ad-hoc attempts at a better life and zero proper planning for the next foreseeable future.</p><p>You need the past to understand the present. Where you are today is a culmination of your thoughts, decisions and actions from five years previous. But the good part is that you have had this epiphany and you are going into the future, not with only hope but with actionable steps to make the next five years better than the last.</p><blockquote>You are the architect of your life. You design it. You are the engineer of your life. You build it. You are the president of your life. You rule it. — Ayoleyi E. Johnson</blockquote><p>Life happens to all of us, but the universe tailors its actions to fit our beliefs and expectations. What we focus on, grows and affects us in ways we may not even be able to comprehend. You can choose to orchestrate the universe to do your bidding. Belief is key. Believe in yourself. Believe in your ability to succeed and never give up.</p><p>It may be rough sometimes, but your resilience and ability to weather the storm and persevere regardless is the little extra that separates the ordinary from the extraordinary.</p><p>Have a great life!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c7cb69c04e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[10 Days Writing Challenge #Day 1]]></title>
            <link>https://medium.com/@ayoleyijohnson/10-days-writing-challenge-day-1-49c876ea1e35?source=rss-aa1234881dd7------2</link>
            <guid isPermaLink="false">https://medium.com/p/49c876ea1e35</guid>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[poetry]]></category>
            <category><![CDATA[twitter]]></category>
            <category><![CDATA[writing-challenge]]></category>
            <dc:creator><![CDATA[Ayoleyi Johnson]]></dc:creator>
            <pubDate>Fri, 17 Apr 2020 10:27:21 GMT</pubDate>
            <atom:updated>2020-04-17T10:27:21.452Z</atom:updated>
            <content:encoded><![CDATA[<p>So I hopped on a 10 Days Writing Challenge from Twitter #ScribeTribe and I look forward to challenging myself over the next couple of days, since they have extended the lockdown here in Lagos, Nigeria, why not? Maybe I even get to finish my first anthology! Sssssssshhhh! That’s supposed to be a secret!</p><p>Alright, here goes, hoping I enjoy this ride…</p><p><strong>10 Days Writing Challenge</strong><br><strong>Day 1</strong><br><strong>Task: “This wasn’t a game, it was real”</strong><br><strong>Genre: Poetry (10 Lines Stanza)</strong></p><p>I needed no names, it was my deal,<br>We went down the street,<br>To an uncompleted building,<br>The plan was very simple,<br>Give them my dad’s old gun in a box,<br>And I would get 250k in return,<br>I handed it over, I thought I was done,<br>I turned to go and heard a click,<br>“Shoot him, he has seen our faces”<br>This wasn’t a game, it was real.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=49c876ea1e35" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>