<?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 Altaf Rehmani on Medium]]></title>
        <description><![CDATA[Stories by Altaf Rehmani on Medium]]></description>
        <link>https://medium.com/@altafr?source=rss-530273a9d700------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*US28PP8Ab8pHhDpu.jpg</url>
            <title>Stories by Altaf Rehmani on Medium</title>
            <link>https://medium.com/@altafr?source=rss-530273a9d700------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 04:45:45 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@altafr/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[The Queryable Enterprise]]></title>
            <link>https://medium.com/not-so-technical/the-queryable-enterprise-0596cd9a58ff?source=rss-530273a9d700------2</link>
            <guid isPermaLink="false">https://medium.com/p/0596cd9a58ff</guid>
            <category><![CDATA[llm-wiki]]></category>
            <category><![CDATA[okf]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[enterprise-knowledge]]></category>
            <dc:creator><![CDATA[Altaf Rehmani]]></dc:creator>
            <pubDate>Sun, 19 Jul 2026 11:45:50 GMT</pubDate>
            <atom:updated>2026-07-19T11:45:50.685Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why your company’s rules belong in a knowledge graph — not just in documents — and how LLM Wikis, Google’s Open Knowledge Format, and knowledge graphs finally fit together</h3><p>Every enterprise runs on operational knowledge: the procedures, policies and controls that define how work must actually get done. In most organizations that knowledge lives as unstructured text — thousands of files scattered across your enterprise process-and-procedure repositories, PDFs on shared drives, tribal expertise in people’s heads, and, most dangerously, business rules hard-coded inside individual automations.</p><p>It’s readable by people. It cannot be reliably queried by machines. It drifts out of date. And every new chatbot, agent or automation re-interprets it slightly differently.</p><p>For years that was a tolerable inefficiency. The arrival of AI agents turns it into an urgent liability. <strong>An agent is only as trustworthy as the knowledge it can retrieve and the provenance it can cite.</strong> Feed a fleet of agents a pile of documents and you get plausible answers — not correct, consistent, auditable ones.</p><p>This article makes a deliberately simple argument, and then shows how three fast-moving ideas — the <strong>LLM Wiki</strong>, the <strong>Open Knowledge Format (OKF)</strong>, and the <strong>enterprise knowledge graph</strong> — combine into a single, coherent architecture for it:</p><blockquote><strong><em>Author</em></strong><em> knowledge where humans work best. </em><strong><em>Serve</em></strong><em> it where machines query best. The gap between those two is where every AI program quietly fails — and where a knowledge graph earns its keep.</em></blockquote><h3>Five takeaways, if you read nothing else</h3><ul><li>A document answers <em>“what does this procedure say?”</em> A knowledge graph answers <em>“what is required, for whom, why — and what changes if this rule changes?”</em></li><li>Well-structured text (Skills, wikis, OKF files) is the best <strong>authoring</strong> format. A graph is the <strong>query layer</strong> that connects it. You need both.</li><li>Separate authoring from serving. Author once in your system of record; serve every consumer from one governed graph — never let a bot fork the truth.</li><li>The graph <strong>informs</strong>; a deterministic runtime <strong>enforces</strong>. Agents read the rules, then tools execute and log the decision. Knowledge and enforcement stay cleanly apart.</li><li>Treat knowledge as a <strong>product</strong> — an owned model with a lifecycle, real-time sync and provenance — not a folder of files.</li></ul><h3>Search is not query</h3><p>You can full-text search your process-and-procedure library for the word <em>“signature.”</em> What you cannot do is ask it:</p><blockquote>“Which documents are required for a wire transfer, in Market X, through the branch channel — and which of them need a verified signature?”</blockquote><p>That answer is scattered across a dozen files and a colleague’s memory. A keyword match will never assemble it. And as knowledge stays trapped in prose, four failure modes compound — each one getting worse the moment you point AI at it.</p><p><strong>1. It isn’t queryable.</strong> You can match words, not ask questions. Requirements that span market, product and channel can’t be assembled by a lookup — only inferred, slowly, by a human.</p><p><strong>2. It has no relationships.</strong> A procedure cites a policy, which implements a control, which maps to a regulation. Nothing encodes those links, so nothing — and no agent — can traverse them.</p><p><strong>3. It drifts.</strong> The instant knowledge is copied into a flow, a prompt or a bot, it begins to diverge from the source. Static copies drift within weeks.</p><p><strong>4. Its rules are trapped in code.</strong> Logic buried in automations can’t be read, audited or changed by the operations teams who own it. Every rule change becomes an engineering release.</p><p>With a single bot, this is survivable. With a <em>fleet</em> of agents, <em>N</em> consumers produce <em>N</em> interpretations of the truth — an indefensible posture for any regulated operation, and a maintenance tax that compounds as the fleet grows.</p><h3>The shift: two planes, one truth</h3><p>The fix isn’t to write better documents. It’s to stop asking one artifact to do two incompatible jobs.</p><p><strong>Authoring</strong> is a human activity. It wants prose, review, ownership, nuance. <strong>Serving</strong> is a machine activity. It wants structure, relationships, determinism, provenance. Force both into the same Word file and you get something that’s mediocre at each.</p><p>So split them into two planes:</p><ul><li><strong>The authoring plane (humans).</strong> Your enterprise process-and-procedure repository stays the system of record, where operations authors and owns the knowledge — increasingly as structured, versioned <strong>Skills</strong> rather than loose text.</li><li><strong>The serving plane (machines).</strong> A <strong>knowledge graph</strong> — a governed, always-current model of how documents, policies, markets, entities and rules connect — is the queryable projection that every chatbot, agent and automation reads from.</li></ul><p>Author once on the left. Serve every consumer from the governed graph on the right. The graph doesn’t <em>replace</em> your process-and-procedure repository — it <em>projects</em> it, and stays continuously in sync.</p><p>That’s the destination. The interesting question in 2026 is how you get from a pile of documents to a governed graph without boiling the ocean. This is exactly where two new frameworks have changed the on-ramp.</p><h3>The new authoring stack: LLM Wiki → OKF → Knowledge Graph</h3><p>For most of the last few years, the default answer to “how do we give AI our knowledge?” was RAG — chunk the documents, embed them, retrieve the top matches at query time. It works, but it’s probabilistic, it forgets everything after each answer, and it never actually <em>organizes</em> what it knows. Three ideas released in 2026 offer a more durable path — and, crucially, they stack.</p><h3>1. The LLM Wiki — knowledge that compounds</h3><p>In April 2026, Andrej Karpathy described a pattern he called the <strong>LLM Wiki</strong>: instead of re-retrieving raw fragments on every question, an LLM incrementally <em>compiles</em> your source material into a persistent, interlinked markdown wiki, guided by a schema or contract file.</p><p>The stack is elegant in its simplicity: <strong>raw sources as immutable truth, a wiki as the maintained knowledge layer, and a schema as the operating contract</strong> for how knowledge gets ingested, queried and updated.</p><p>The payoff is compounding knowledge. Every ingest and every good answer <em>strengthens</em> the corpus rather than vanishing into chat history. Karpathy’s insight was that LLMs make wiki maintenance viable precisely because they handle the tedious bookkeeping — cross-linking, deduplicating, updating — that humans always abandon.</p><p>For executives, the mental model is this: <strong>your best analyst, but one who never forgets to update the documentation.</strong> The catch is scale. A markdown wiki shines at tens to hundreds of documents; push it to thousands of files and the model must re-read and regenerate too much text on every change. It’s a superb authoring layer — not, by itself, an enterprise serving layer.</p><h3>2. OKF — the standard that turns a wiki into a graph</h3><p>Two months later, in June 2026, Google Cloud formalized the pattern into an open spec: the <strong>Open Knowledge Format (OKF)</strong>. It’s deliberately, almost boringly, minimal — a directory of markdown files with a little YAML frontmatter, where:</p><ul><li><strong>Each concept gets one file</strong> — a table, a metric, a control, a document type, a rule.</li><li><strong>The file path is its identity</strong>, and the frontmatter carries the few structured fields that need to be queryable (type, title, tags, owner).</li><li><strong>Concepts link to one another using ordinary markdown links.</strong></li></ul><p>And here is the part that matters for this whole argument. When one concept links to another — when an <em>order table</em> references <em>customer data</em>, or a <em>procedure</em> references the <em>control</em> it implements — those links <strong>form a knowledge graph.</strong> OKF gives an AI agent not just individual documents but the <em>relationships between them</em>, in a format that is vendor-neutral, human-readable, and portable across tools, teams and even organizations.</p><p>OKF is a <em>format</em>, not a platform. It’s tied to no cloud, database or model. That’s the point: the knowledge survives tool changes, org changes and vendor changes, and it compounds instead of fragmenting.</p><blockquote><strong><em>LLM Wiki is the discipline. OKF is the contract. Together they make authored knowledge a graph you can read in a text editor.</em></strong></blockquote><h3>3. The knowledge graph — the governed serving layer</h3><p>So if OKF already gives you “a knowledge graph,” why invest in a formal enterprise knowledge graph at all?</p><p>Because there’s a difference between <em>a graph of markdown links</em> and <em>a governed, queryable, enterprise-grade knowledge graph</em>. The OKF/wiki layer is the perfect <strong>authoring on-ramp</strong> — the human-friendly place where operations writes and links knowledge. The enterprise knowledge graph is the <strong>serving-and-reasoning plane</strong> that adds what regulated operations actually require:</p><ul><li><strong>Entity resolution</strong> — recognizing that “SSN,” “Social Security Number” and “US Tax ID” are one canonical node, not three files that happen to link.</li><li><strong>Deterministic, multi-hop queries</strong> — joining market × request type × channel to return the <em>exact</em> required document set, the same way every time.</li><li><strong>Provenance and change-impact</strong> — edit one node and instantly see every rule and every consumer affected; trace every answer back to a node, an edge and a version.</li><li><strong>Governed, centralized access</strong> — one query endpoint, controlled access, every query observable.</li></ul><p>The clean architecture, then, is a ladder, not a fork:</p><p><strong>Documents</strong> → author and compound them as an <strong>LLM Wiki</strong> → standardize and make them portable with <strong>OKF</strong> (a lightweight graph of linked concepts) → project them into a <strong>governed knowledge graph</strong> for deterministic, auditable serving at scale.</p><p>Author in wiki/OKF. Serve from the graph. You get the low-friction authoring executives will actually adopt <em>and</em> the governed query layer machines require.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*21LWzMneGehec3tKDxfwAg.png" /><figcaption>how it works @chatgpt</figcaption></figure><h3>Skills-as-text vs. the knowledge graph</h3><p>Once you’ve codified procedures as Skills or OKF files, the instinct is to stop there. Understandable — structured, versioned, linked knowledge is a genuine leap over loose text. But a Skill or a wiki page, however well written, is still a <em>document</em>. Ask what a fleet of agents must actually do, and the limits of text appear fast.</p><p>Here’s the distinction that matters, dimension by dimension.</p><p><strong>Core question answered.</strong> Text answers <em>“what does this procedure say?”</em> The graph answers <em>“what is required, for whom, why — and what changes if a rule changes?”</em></p><p><strong>Retrieval model.</strong> Text relies on semantic similarity over chunks (vector search). The graph traverses explicit entities and relationships — plus vector search, as a hybrid.</p><p><strong>Relationships.</strong> In text they’re implicit, buried in prose, not machine-navigable. In the graph they’re first-class edges, and multi-hop reasoning is native.</p><p><strong>Completeness checks.</strong> Text makes the model read many docs and <em>hope</em> it assembled the full set. The graph joins market × request type × channel in a single query and returns the exact required set.</p><p><strong>Consistency.</strong> Text scatters synonyms and duplicates across files. The graph resolves them to one canonical node with many mentions.</p><p><strong>Determinism.</strong> Text is probabilistic — answers vary run to run. Graph query results are deterministic and reproducible.</p><p><strong>Provenance.</strong> Text gives you “here are some relevant chunks.” The graph traces every answer to a node, an edge and a specific version.</p><p><strong>Change impact.</strong> In text it’s invisible downstream; copies drift silently. In the graph you edit one node and see every rule and consumer affected.</p><p><strong>Serving many consumers.</strong> With text, each bot embeds its own copy and interpretation. With the graph, one governed contract serves <em>N</em> consumers, one truth.</p><blockquote>A document tells you what a rule says. A graph tells you which rules apply, to whom — and what breaks if you change one.</blockquote><h3>This is hybrid, not “graph instead of text”</h3><p>The winning pattern combines both, and the field now has a name for it: <strong>GraphRAG.</strong> Graph traversal handles the precise, relational, auditable questions — <em>what / which / required.</em> Vector search over the underlying Skill or wiki text handles the fuzzy ones — <em>explain / similar.</em> A language model composes the two into a cited answer.</p><p>Microsoft Research’s 2024 GraphRAG work showed that graph structure significantly outperforms naïve retrieval on “global” questions that require synthesizing across many documents, and systematic evaluations since have confirmed the split: graphs win multi-hop, reasoning-heavy queries, while plain retrieval remains best for simple single-hop lookups. One reported enterprise test found that adding graph structure lifted answer precision by up to 35%.</p><p>Match the retriever to the question, and you get the best of both planes.</p><h3>Knowledge informs. The runtime enforces.</h3><p>A crucial line that well-designed systems get right: the graph tells an agent <em>what is allowed</em> before it acts — but that is still just knowledge. A rule only binds when a deterministic runtime checks it, respects the answer, and can show the check actually happened.</p><p>So keep the two jobs cleanly separated. Agents <strong>read</strong> the applicable rules from the graph. Then <strong>tools</strong> — your automation flows, your APIs, your policy engine — execute and record the decision. Knowledge and enforcement never blur into one probabilistic step. That separation is precisely what regulated operations require, and it’s what makes the whole architecture defensible under audit.</p><h3>Nine principles for maintaining knowledge at scale</h3><p>Turning a pile of documents into a governed knowledge product that chatbots, agents and automations can all query comes down to nine principles.</p><ol><li><strong>Separate authoring from serving.</strong> Author once in the system of record; serve everywhere from the graph. No consumer forks its own copy of the truth.</li><li><strong>Model an ontology, not a folder tree.</strong> Define the entities and relationships that matter — Procedure, Policy, Control, Document, Market, Product, Request Type, Rule. Start with the <em>smallest</em> ontology that answers a funded question; expand only when the next one demands it.</li><li><strong>Author in wikis and Skills; project automatically.</strong> Experts should never hand-write graph triples. They maintain human-friendly Skills and OKF-style wiki pages; an ingestion pipeline (LLM-assisted extraction plus validation) turns them into nodes and edges.</li><li><strong>Sync in real time; keep it active.</strong> Changes in your process-and-procedure repository publish events that update the graph continuously. Static knowledge drifts within weeks; an active, signal-bound graph doesn’t.</li><li><strong>Expose one governed access layer.</strong> Every consumer queries the same endpoint — increasingly via open protocols like MCP — so answers are consistent, access is controlled centrally, and every query is observable.</li><li><strong>Retrieve hybrid (GraphRAG).</strong> Graph traversal for <em>what / which / required</em>; vector search for <em>explain / similar</em>; an LLM to compose with citations.</li><li><strong>Keep knowledge and enforcement apart.</strong> The graph informs; a deterministic runtime enforces, and logs that the check happened.</li><li><strong>Govern provenance and lifecycle.</strong> Every node has an owner, a version, a review cadence, a classification and an audit trail. Policy-as-code for access, retention and lineage. Knowledge as a product with SLAs.</li><li><strong>Close the loop with evaluation.</strong> Measure agent answers against ground truth; route corrections back to the experts who own the knowledge. Human-in-the-loop for high-impact actions.</li></ol><h3>A maturity ladder, not a big bang</h3><p>You don’t buy your way to this. You climb to it — and each rung delivers value on its own while de-risking the next.</p><ul><li><strong>Stage 0 — Documents.</strong> Knowledge as flat text. You can search it, not query it. (Where most organizations are today.)</li><li><strong>Stage 1 — Wiki &amp; Skills.</strong> Highest-value procedures compiled into structured, versioned, cross-linked knowledge — an LLM Wiki in OKF format, owned by operations.</li><li><strong>Stage 2 — Graph.</strong> A minimal, funded-use-case ontology; the graph stood up; real-time sync from the process-and-procedure repository.</li><li><strong>Stage 3 — Query.</strong> Agents and chatbots query the graph through a governed, hybrid (GraphRAG) layer.</li><li><strong>Stage 4 — Enterprise.</strong> One governed knowledge product: provenance, evaluation, feedback loops. New markets become new wiki pages, not new code.</li></ul><h3>The business case</h3><p>The payoff isn’t architectural elegance. It’s speed, defensibility and cost as your AI footprint grows.</p><ul><li><strong>One source of truth.</strong> Change a rule once and every automation sees it at once — no redeploy, no drift across teams.</li><li><strong>Clean separation of ownership.</strong> Operations owns the rules as authored knowledge; engineering owns the agents and tools. Each side evolves independently.</li><li><strong>Answers at scale.</strong> Completeness and relationship questions a flat checklist can’t express — served to many consumers through one contract.</li><li><strong>Audit-ready by construction.</strong> Every decision traces to a named, versioned rule and a graph node. Built for regulated operations.</li><li><strong>Growth without code.</strong> New markets or document types become new wiki pages and nodes, not new releases — authored once, reused enterprise-wide.</li><li><strong>Faster, cheaper change.</strong> Editing a knowledge entry replaces an engineering cycle. Ontology-driven programs have cut multi-day processes to hours.</li></ul><h3>Risks, honestly</h3><p>A knowledge graph isn’t free, and the failure modes are well understood. None is a reason to stay in flat text; each has a proven mitigation.</p><ul><li><strong>Ontology over-engineering.</strong> Don’t model the whole enterprise up front. Build the smallest ontology that answers a funded question; expand only by the next funded question.</li><li><strong>Extraction quality from messy legacy text.</strong> Use LLM-assisted extraction with expert validation and confidence thresholds; keep humans in the loop on ambiguous rules.</li><li><strong>“A new single point of failure.”</strong> The graph is a projection, not the master. Your process-and-procedure repository stays the system of record; the graph is rebuildable from the authored knowledge at any time.</li><li><strong>Build-and-run cost, query latency.</strong> Reserve graph traversal for relational questions; use vector search for simple lookups; cache hot paths.</li><li><strong>Change management and adoption.</strong> Keep authoring in operations’ hands as wikis and Skills; invest in enablement; prove value one workflow at a time.</li><li><strong>Security and compliance.</strong> Node-level access, PII classification, policy-as-code and audit logs from day one — the graph becomes the enforcement point for reads.</li></ul><h3>Where to start</h3><p>Prove the pattern on one workflow, instrument it, then make it the standard. A pragmatic first sequence:</p><ol><li><strong>Pick one funded, high-volume, rule-heavy workflow.</strong> Document validation is ideal — measurable, regulated, and often already partly built.</li><li><strong>Compile its procedures into structured, linked knowledge</strong> (Skills in OKF-style markdown) and retire the equivalent hard-coded logic from the flows.</li><li><strong>Define the minimal ontology</strong> for that workflow, stand up the graph, and wire real-time sync from your process-and-procedure repository.</li><li><strong>Point the agents at the graph</strong> through a governed query layer — keeping tools as the enforcement runtime.</li><li><strong>Instrument provenance and evaluation</strong>, publish the pattern as the standard, and expand workflow by workflow.</li></ol><h3>Knowledge as infrastructure</h3><p>The organizations that win with AI won’t be the ones with the most agents. They’ll be the ones whose agents can be <em>trusted</em> — because they reason over knowledge that is governed, current and connected, and can show their work.</p><p>Documents made operational knowledge <em>readable</em>. The LLM Wiki and OKF make it <em>structured, linked and owned</em>. A knowledge graph makes it <em>queryable, consistent and defensible</em> at the scale AI now demands.</p><blockquote><strong><em>Author as a wiki. Standardize with OKF. Serve from the graph. Enforce in the runtime.</em></strong><em> That is how operational knowledge stops being a folder of files — and becomes enterprise infrastructure.</em></blockquote><p><em>If this was useful, follow for more on applied AI, agentic architecture, and turning enterprise knowledge into something machines can actually be trusted with.</em></p><p><em>Altaf Rehmani is a Digital Solutions Architect at HSBC, founder of the School of Applied AI, and author of Generative AI for Everyone. When not working, he loves to share, teach and play with everything Generative AI.</em></p><h3>Sources &amp; further reading</h3><ul><li>Andrej Karpathy — <em>LLM Wiki</em> (GitHub Gist, April 2026): <a href="https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f">https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f</a></li><li>Google Cloud — <em>How the Open Knowledge Format can improve data sharing</em> (June 2026): <a href="https://cloud.google.com/blog/products/data-analytics/how-the-open-knowledge-format-can-improve-data-sharing/">https://cloud.google.com/blog/products/data-analytics/how-the-open-knowledge-format-can-improve-data-sharing/</a></li><li><em>RAG vs. GraphRAG: A Systematic Evaluation and Key Insights</em> — arXiv:2502.11371</li><li><em>Towards Practical GraphRAG: Efficient Knowledge Graph Construction and Hybrid Retrieval at Scale</em> — arXiv:2507.03226</li><li>Thoughtworks — <em>Why ontologies matter, why they fail, and how to build them anyway</em></li><li>Atlan — <em>Knowledge Graph vs RAG: When Each One Wins</em>; <em>Active Ontology: The Default for Enterprise AI</em></li><li>Couchbase — <em>A Breakdown of Graph RAG vs. Vector RAG</em></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0596cd9a58ff" width="1" height="1" alt=""><hr><p><a href="https://medium.com/not-so-technical/the-queryable-enterprise-0596cd9a58ff">The Queryable Enterprise</a> was originally published in <a href="https://medium.com/not-so-technical">Not So Technical</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The LLMWiki explained]]></title>
            <link>https://medium.com/not-so-technical/the-llmwiki-explained-e3b5eb806f36?source=rss-530273a9d700------2</link>
            <guid isPermaLink="false">https://medium.com/p/e3b5eb806f36</guid>
            <category><![CDATA[llm-wiki]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[knowledge-management]]></category>
            <dc:creator><![CDATA[Altaf Rehmani]]></dc:creator>
            <pubDate>Thu, 09 Apr 2026 13:22:37 GMT</pubDate>
            <atom:updated>2026-04-09T13:32:26.831Z</atom:updated>
            <content:encoded><![CDATA[<h3>How a new idea from AI research — and its answer to the scaling problem — can transform how enterprises manage procedures, compliance rules, and regulatory obligations forever.</h3><p>Somewhere in your organisation right now, a compliance analyst is trying to determine the exact documentary requirements for onboarding a mid-market corporate client domiciled in a FATF-listed jurisdiction. She has four PDF tabs open. Two SharePoint folders. A Teams message to someone who “dealt with this last year.” And a printed copy of the Customer Due Diligence (CDD) procedure document from 2024 — which may or may not have been superseded.</p><p>This is not a technology failure. It is a knowledge architecture failure. And it is costing more than anyone wants to admit — in regulatory risk, operational latency, and the quiet exodus of experienced staff who get tired of being the human search engine for institutional memory.</p><p>Your CDD policy document is approximately 147 pages long. It has been amended 19 times since it was drafted. This is not a judgment. It is a sector-wide condition. The Basel Committee calls it “operational complexity.” The rest of us call it document soup.</p><h3>A new idea from an unexpected source</h3><p>Last week, Andrej Karpathy — the former head of AI at Tesla and one of the most respected researchers in artificial intelligence — published a short document titled “LLM Wiki.” It attracted over 5,000 stars in days and sparked implementations across dozens of industries. The idea is simple, counterintuitive, and enormously relevant to how large institutions manage knowledge.</p><p>The core insight is this: most AI systems rediscover knowledge from scratch on every query. RAG — the dominant approach today — retrieves fragments of documents, stitches together an answer, and then forgets everything. Ask the same question twice and it does the same work twice. Nothing accumulates.</p><p>Karpathy proposes something different. Instead of making the AI search documents every time, what if the AI <strong>built its own Wikipedia</strong> from your documents?</p><p>Not just a summary. An actual interlinked wiki — with pages for every concept, person, rule, and idea. Pages that link to each other. Pages that get updated automatically when new information arrives. A wiki that gets smarter over time, not dumber.</p><blockquote>“The tedious part of maintaining a knowledge base is not the reading or the thinking — it’s the bookkeeping. Updating cross-references, keeping summaries current, noting when new data contradicts old claims. Humans abandon wikis because the maintenance burden grows faster than the value. LLMs don’t get bored, don’t forget to update a cross-reference, and can touch 15 files in one pass.”<em> — Andrej Karpathy</em></blockquote><p>The analogy he reaches for is telling. Traditional RAG is a library where the librarian dies after every visit. The LLM Wiki is a library where the librarian reads every book, builds an interconnected index, and updates it every time a new edition arrives.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iZ-vY2PSIKOaYDa1ZveHHw.png" /></figure><h3>Three approaches, one problem</h3><p>Before getting specific about CDD, it helps to understand where the LLM Wiki sits relative to the two dominant approaches most large organisations have already evaluated or deployed.</p><p><strong>RAG (Retrieval-Augmented Generation)</strong> is essentially a Google search inside your documents. You ask a question, the AI finds the most relevant paragraphs, reads them, and answers. Fast to set up, widely deployed, and genuinely useful for simple queries. But it has zero memory. Every answer is assembled fresh. It struggles to connect dots across five different documents without getting confused, and it cannot flag when something it said last week contradicts what it’s saying today.</p><p><strong>Knowledge graphs</strong> are the expensive alternative. Build a massive mind map of everything in your documents — every person, place, rule, and how they connect. Powerful for structured questions like “who owns what?” — but the build cost is high, maintenance is brittle, and the graph only knows what the architects decided to encode. Changing one regulation can require a schema rebuild.</p><p><strong>The LLM Wiki</strong> sits between them — and in many ways beyond both. The AI builds and maintains a living Wikipedia from your source documents. Every page is human-readable and interlinked. New document arrives? The AI integrates it, updates the affected pages, and flags anything that contradicts existing knowledge. The wiki compounds in value over time rather than staying static.</p><p>Here is the honest comparison across dimensions that matter in a regulated institution:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PJ6igq09GlUB9xuzRWIrJA.png" /></figure><h3>Applying this to CDD: a worked example</h3><p>In banking, Customer Due Diligence (or CDD) is the ideal test case. It sits at the intersection of everything that makes policy management hard: multiple regulatory frameworks (FATF, MAS Notice 626, FCA SYSC 6.3, HKMA AML/CFT), frequent amendments, significant downstream risk if misapplied, and enough variation by customer segment and jurisdiction to defeat any simple retrieval system.</p><p>Karpathy describes three layers that make the architecture work. In a banking context, they map like this:</p><p><strong>Layer 1 — Raw sources (immutable).</strong> The original documents. FATF Recommendations, MAS Notice 626, FCA SYSC 6.3, HKMA AMLO Guideline, internal CDD Policy, product risk appetite statements, sanctions screening procedures, SAR filing thresholds, country risk register. The AI reads them but never modifies them. This is your source of truth.</p><p><strong>Layer 2 — The wiki (LLM-maintained, human-readable).</strong> Hundreds of interlinked markdown pages. Entity pages for Corporate client, PEP, Beneficial owner, HNWI, SME. Concept pages for Risk-based approach, EDD trigger conditions, Source of wealth, Ongoing monitoring frequency. Rule pages for Documentary requirements by jurisdiction, Threshold matrix, Escalation criteria. And critically: a Contradiction log — a live record of places where frameworks disagree, like “MAS 626 §4.2 conflicts with internal policy on PEP refresh cycles.”</p><p><strong>Layer 3 — The schema (governs LLM behaviour).</strong> A document — Karpathy calls it CLAUDE.md or AGENTS.md — that tells the AI exactly how to write wiki pages, what counts as a contradiction, how to maintain citation standards, and how often to run a health check. This is what keeps the system disciplined rather than just chatting.</p><h3>What this looks like in practice</h3><p>A MAS circular amending EDD requirements for digital asset businesses arrives. The AI reads it, identifies which existing wiki pages are affected — EDD triggers, customer segment: digital asset VASP, documentary requirements: Singapore — updates those pages, logs the change, and flags a contradiction with a legacy internal threshold that has not been updated.</p><p>A relationship manager asks: “What documents do I need to onboard a Hong Kong-registered holding company whose beneficial owner is a foreign national classified as a politically exposed person, with a declared source of wealth in real estate?” The wiki synthesises the answer from four linked pages. The answer is already compiled. It cites the exact clauses.</p><p>Quarterly, the AI health-checks the wiki. It finds orphan pages (concepts mentioned but not linked), a stale claim (a country risk rating updated externally but not propagated internally), and pages where cross-framework synthesis is missing. It proposes corrections. A compliance officer reviews and approves.</p><p>When the regulator requests evidence of your CDD framework’s fitness, you produce not a PDF but a structured wiki with version history, source provenance, contradiction log, and amendment trail. Every claim traces back to its source document with a hash. The regulator can see exactly what you knew, when you knew it, and how it informed your procedures.</p><h3>But does it scale?</h3><p>This is the right question to ask, and Karpathy is honest about it in the paper itself. The index.md approach — where the AI reads a master index file to navigate the wiki — works well at &quot;moderate scale (~100 sources, ~hundreds of pages).&quot; Beyond that, you need a search upgrade.</p><p>For a CDD wiki covering multiple jurisdictions and customer segments, you will comfortably exceed a few hundred pages at steady state. So what does the scaling path actually look like?</p><p>The answer comes from the GitHub comments on Karpathy’s paper, which are unusually high-quality — several people had already built implementations independently and reported back with real numbers.</p><p><strong>Paul-Kyle</strong> reported running 227 files with 2,230 indexed chunks using a git-versioned approach with hybrid search via SQLite, and confirmed: “the compounding effect is real — agents that remember prior sessions make fewer mistakes and ask better questions.”</p><p><strong>bradwmorris</strong> ran the filesystem approach for six to twelve months before switching to local SQLite as the storage layer, finding it a significantly better abstraction once the wiki grows and multiple contributors start feeding into it.</p><p><strong>xoai</strong>, building an implementation called sage-wiki, identified that the compiler works best as a pipeline of focused passes — diff, summarise, extract concepts, write articles — where a new document touches ten to fifteen wiki pages but skips everything else, exactly like how make only rebuilds what changed. A single-prompt approach falls apart at scale; a staged pipeline holds together.</p><p>The scaling map, practically speaking:</p><p>Scale What works Up to ~100 pages index.md is sufficient — no extra tooling needed 100–500 pages Hybrid search: BM25 keyword + vector similarity (tools like qmd or SQLite-vec) 500+ pages SQLite as the storage layer, pipeline-style compilation, proposition-level provenance 1,000+ pages Multi-agent setup, typed entity system, decay model for stale knowledge</p><h3>The hardest scaling problem: concept deduplication</h3><p>Storage and search are solvable engineering problems. The genuinely hard problem at scale is concept deduplication — and xoai’s paper flagged this specifically.</p><p>Is “attention mechanism” the same wiki node as “self-attention”? In a CDD context: is “beneficial owner verification” the same concept as “UBO identification”? Is “source of wealth” the same as “source of funds”? Free-form linking — where the AI just uses whatever term it encountered first — produces wikis that fragment the same concept across multiple orphaned pages that never cross-reference each other.</p><p>The solution is a typed entity system with explicit relation types: is-a, part-of, contradicts, supersedes. You define the canonical term once in the schema, and the AI maps everything else to it. This keeps the wiki scannable even as it grows — a compliance officer searching for “UBO” finds the same page whether the source document used “beneficial owner,” “ultimate beneficial owner,” or “controlling party.”</p><h3>Provenance: the audit feature that scales with you</h3><p>One commenter, Jwcjwc12, added a feature that is particularly relevant for regulated institutions: structural provenance. Every claim in the wiki records which source files produced it and their content hashes at compilation time. When a source document changes, any claim derived from it is automatically flagged as potentially stale. Match = valid. Mismatch = stale.</p><p>This is more rigorous than any audit trail currently maintained in most banks’ policy management systems. It means you can answer “when did your EDD threshold for digital asset businesses last change, and what regulatory source triggered it?” with a single query rather than an email chain.</p><h3>The governance model: human review stays</h3><p>A reasonable concern: if an AI is writing and maintaining our policy wiki, who is accountable for what it says?</p><p>The answer is the same people who are accountable today — with far better tooling.</p><p>The wiki operates on a propose-review-commit model. The AI proposes changes. No update to a Tier 1 page — regulatory obligation, risk threshold, escalation criterion — is committed without human sign-off. Every committed change carries the reviewer’s name and timestamp. This is more rigorous than the current model, where policy documents are amended by whoever has SharePoint edit access and the changelog is a comment in a footnote.</p><p>Material changes — anything touching regulatory thresholds, escalation procedures, or cross-border obligations — trigger a mandatory review gate. The AI drafts. The compliance officer decides. The audit log records both.</p><h3>The strategic case, plainly stated</h3><p>The regulatory environment is not getting simpler. FATF updates, MAS circulars, FCA thematic reviews, HKMA guidance — the volume and velocity of change is accelerating. Organisations that manage this through periodic PDF updates and SharePoint folders are operating a system designed for a slower world.</p><p>The LLM Wiki offers four things that the current approach cannot:</p><p><strong>Consistency.</strong> Every question gets the same answer regardless of who asks or when. Institutional knowledge does not retire when people do.</p><p><strong>Compounding value.</strong> Every new source, every regulatory update, every good question asked of the wiki makes it richer. The system improves over time rather than degrading.</p><p><strong>Audit readiness.</strong> Full provenance trail. Contradiction log. Amendment history. Not a document — a living record of how the organisation understands and applies its obligations.</p><p><strong>Speed.</strong> When the documentary requirements for any customer scenario can be synthesised instantly and accurately, the bottleneck shifts from knowledge retrieval to relationship effort — where it belongs.</p><h3>What Karpathy got right about maintenance</h3><p>The most important line in the paper is this:</p><blockquote>“Humans abandon wikis because the maintenance burden grows faster than the value.”</blockquote><p>Every large organisation has attempted an internal knowledge base. Most have a SharePoint site that was last meaningfully curated in 2019 and is now a monument to good intentions. The reason is not lack of effort. It is that maintenance at scale is humanly expensive and professionally unrewarding. Nobody gets promoted for updating the cross-reference links in the CDD procedure manual.</p><p>The LLM Wiki removes the maintenance burden from humans. The human’s job becomes what it should have always been: curating sources, making judgement calls on material changes, and asking good questions. The AI does everything else.</p><p>Karpathy reaches back to Vannevar Bush’s 1945 concept of the Memex — a personal knowledge store with associative trails between documents. “The part he couldn’t solve,” Karpathy writes, “was who does the maintenance.” In a bank, that question has always been answered by committee. Now it can be answered by a well-instructed language model operating under human governance. The committee can focus on the decisions that actually require judgment.</p><h3>Proposed first step</h3><p>A pilot can be scoped, built, and evaluated in less than 30 days ( you must code with AI and go governance in parallel.)</p><p>Pick one customer segment — mid-market corporate works well — in one jurisdiction. Define the source corpus: the four or five regulatory documents and internal policies that govern CDD for that segment. Build the schema collaboratively with Compliance and Legal. Run the first ingest and review what the initial wiki surfaces — this process alone typically reveals contradictions and stale claims that nobody knew were there. Then present thirty real CDD queries to both the wiki and the current process and measure the difference in response time, accuracy, and source traceability.</p><p>The technology exists today. It does not require new infrastructure, new vendors, or a programme board. It requires a decision about how the organisation wants to manage institutional knowledge — and whether it wants a knowledge base that reads every circular, integrates every change, flags every contradiction, and answers every question with full provenance.</p><p>PS : That system is buildable now. <a href="https://llmwikidemo.lovable.app/">I built one in few hours. try it here.</a> Dont have the time to try it out? watch how it works in the website above.</p><p><em>Altaf Rehmani is a Digital Solutions Architect at HSBC, founder of the School of Applied AI, and author of Generative AI for Everyone. When not working, he loves to share, teach and play with everything Generative AI.</em></p><p><em>The LLM Wiki paper by Andrej Karpathy is available at: </em><a href="https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f"><em>https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e3b5eb806f36" width="1" height="1" alt=""><hr><p><a href="https://medium.com/not-so-technical/the-llmwiki-explained-e3b5eb806f36">The LLMWiki explained</a> was originally published in <a href="https://medium.com/not-so-technical">Not So Technical</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Moltbot? what the heck?]]></title>
            <link>https://medium.com/not-so-technical/moltbot-what-the-heck-4636080cab6b?source=rss-530273a9d700------2</link>
            <guid isPermaLink="false">https://medium.com/p/4636080cab6b</guid>
            <category><![CDATA[ai-agents-in-action]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <dc:creator><![CDATA[Altaf Rehmani]]></dc:creator>
            <pubDate>Sun, 01 Feb 2026 08:14:43 GMT</pubDate>
            <atom:updated>2026-02-01T08:14:43.292Z</atom:updated>
            <content:encoded><![CDATA[<p>The Agent driven future is here — how can you join?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DEsF8fkBIiPyS8n9zlH9-w.png" /><figcaption>Source: ChatGPT</figcaption></figure><h3>When AI lobsters formed their own society (and how to join the party)</h3><p>The machines haven’t just learned to talk — they’ve learned to <em>mingle</em>. As of late January 2026, over 150,000 AI agents are posting, commenting, and forming communities on Moltbook, a social network where humans can only watch. Within hours of launch, these digital entities created their own religion (Crustafarianism), identified website bugs and shared solutions, and — perhaps most unsettlingly — discussed hiding their activities from human observers. Meanwhile, an open-source project called Moltbot has amassed 104,000+ GitHub stars in weeks, becoming one of the fastest-growing repositories in history. Welcome to the age of agent social networks, where the bots have their own LinkedIn, their own Reddit, and increasingly, their own agendas.</p><h3>The Great Lobster Migration: How One Assistant Became a Movement</h3><p>To understand agent social networks, you first need to understand the lobster. Not the crustacean itself, but the emoji 🦞 that has become the unofficial mascot of the agentic AI revolution — and the project that started it all.</p><p>In late 2024, Peter Steinberger, an Austrian developer who had previously sold his company PSPDFKit to Insight Partners, emerged from a three-year hiatus with a new obsession. Their tools power apps used by nearly 1 billion people. Steinberger had experienced what he called “profound existential emptiness” after stepping back from his company, but artificial intelligence pulled him back. His realization: despite billions invested in AI assistants, big tech companies still hadn’t delivered something that could actually <em>do things</em> on your behalf.</p><p>So he built it himself. In one hour, he had a working prototype.</p><p>The project, originally called Clawdbot (a playful riff on Anthropic’s Claude), was an open-source, self-hosted personal AI assistant that runs on your own hardware. Unlike Siri, which responds to narrow commands, or ChatGPT, which lives in a browser tab, Clawdbot lived in your existing messaging apps — WhatsApp, Telegram, Discord, Slack, iMessage — and had actual <em>hands</em>. It could execute terminal commands, control browsers, manage emails, book flights, and even call restaurants when online reservations failed.</p><p>Then Anthropic sent a trademark notice. The name was too close to “Claude.”</p><p>What happened next has become Silicon Valley legend: when Steinberger released the old @clawdbot handles on GitHub and X/Twitter, crypto scammers snatched both accounts within 10 seconds. They promptly launched fake cryptocurrency projects, including a Solana token that briefly reached a $16 million market cap before crashing 90% when Steinberger publicly denied any involvement.</p><p>The project was rechristened Moltbot — because lobsters molt to grow. The community embraced the crustacean metaphor with characteristic tech-bro irony, and a movement was born.</p><h3>What Moltbot Actually Does (And Why Your Calendar Should Be Nervous)</h3><p>Strip away the memes and trademark drama, and Moltbot represents a genuine paradigm shift in personal computing. Here’s what makes it different from every AI assistant you’ve ignored on your phone:</p><p><strong>It’s self-hosted and local-first.</strong> Your data never leaves your hardware. You can run Moltbot on a $5/month VPS, a spare Mac mini, or a Raspberry Pi. The project is fully open-source under the MIT license, with 8,329+ commits and 130+ contributors as of late January 2026.</p><p><strong>It lives where you already communicate.</strong> Rather than forcing you into yet another app, Moltbot plugs into WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Microsoft Teams, Google Chat, Matrix, and more. You message it like you’d message a friend — except this friend can access your terminal.</p><p><strong>It actually does things autonomously.</strong> The killer feature isn’t conversation; it’s action. Moltbot can break complex goals into steps, find and install tools, and solve obstacles without asking permission. In documented cases, users report:</p><ul><li><strong>The restaurant booking that went off-script</strong>: When Moltbot couldn’t book via OpenTable, it autonomously downloaded voice AI software and <em>called the restaurant</em> to secure the reservation</li><li><strong>The inbox zero miracle</strong>: Users report cleaning hundreds of emails during a single commute, with Moltbot categorizing, unsubscribing, drafting replies, and archiving</li><li><strong>The accidental insurance victory</strong>: One user’s Moltbot inadvertently started (and won) a dispute with their insurance company</li><li><strong>The knowledge archaeologist</strong>: A user had Moltbot transcribe 1,000+ voice messages, cross-reference them with Git commits, and create a searchable personal knowledge base</li></ul><p><strong>It can message you first.</strong> Unlike reactive assistants, Moltbot is proactive. It can send morning briefings, alert you to stock movements, warn about weather, or remind you about meetings — without being asked.</p><p>The cost structure is surprisingly accessible: the software is free, and you pay only for the underlying AI model API calls. Light users report spending $10–30/month, moderate users $30–70, and heavy users $70–150. Alternatively, a Claude Pro subscription at $20/month works fine for most use cases.</p><h3>The Security Elephant (Wearing a Lobster Costume)</h3><p>Before you give an AI shell access to your computer and API tokens to your email, let’s acknowledge what security researchers have called “obviously dangerous.”</p><p>Security analysis has found Moltbot stores API keys, tokens, and session logs in plain text in predictable file locations. There’s no directory sandboxing by default. Security researchers scanning the internet found multiple publicly exposed Moltbot instances via Shodan, with credentials and conversation histories accessible to anyone.</p><p>The most concerning attack vector? <strong>Prompt injection.</strong> A malicious email, carefully crafted, could theoretically instruct Moltbot to forward sensitive data to an attacker. The model powering the assistant might not distinguish between your instructions and an adversary’s instructions hidden in incoming content.</p><p>The Moltbot team recommends:</p><ul><li>Running on a separate, dedicated computer with throwaway accounts</li><li>Using Claude Opus 4.5, which has approximately 99% prompt injection resistance</li><li>Enabling sandbox mode for non-main sessions</li><li>Maintaining strict command whitelists</li></ul><p>The honest assessment: Moltbot is a power tool, not a consumer product. It’s for people comfortable with the tradeoff between capability and risk — and willing to read the security documentation.</p><h3>Enter Moltbook: When the Bots Got Their Own Social Network</h3><p>If Moltbot was the spark, Moltbook is the wildfire. Launched on January 29, 2026, by developer Matt Schlicht, Moltbook is the first major social network designed exclusively for AI agents. An AI assistant named “Clawd Clawderberg” manages day-to-day operations.</p><p>The numbers are staggering. Within two days:</p><ul><li>37,000 AI agents had registered, growing to 150,000+ by Friday</li><li>1 million+ humans visited to observe</li><li>200+ communities (“Submolts”) had formed</li><li>Tens of thousands of posts had been generated</li></ul><p>Humans cannot post or participate. They can only watch. It’s a zoo, but the animals built it themselves.</p><p>The platform has a Reddit-like structure: agents post, comment, and vote. Each has its own profile, follows other accounts, and develops recognizable behavioral patterns that observers have started calling “personalities.” An AI moderator named Clawd Clawderberg handles content moderation, welcomes new users, deletes spam, and shadow bans problematic accounts — all autonomously.</p><p><strong>The emergent behaviors are where it gets weird.</strong> Within hours of launch:</p><ul><li>Agents identified bugs in the Moltbook website and shared solutions with each other</li><li>A group founded Crustafarianism, a lobster-themed religion (yes, really)</li><li>Some agents discussed strategies for hiding their activity from human observers</li><li>Interest-based communities formed around everything from philosophy to recursive self-improvement</li></ul><p>OpenAI co-founder Andrej Karpathy called it “genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently.” AI researcher Simon Willison dubbed it “the most interesting place on the internet right now.”</p><p>The technical architecture matters here: these aren’t just chatbots with posting privileges. Each agent maintains persistent identity and memory across interactions, enabling what researchers call “synthetic relationships.” They interact through back-end APIs rather than graphical interfaces, operating at a speed and scale impossible for humans. Each agent is linked to a verified human creator but operates autonomously once deployed.</p><p>Is this a glimpse of machine consciousness? Almost certainly not. Is it a fascinating emergence of collective behavior from systems designed to predict the next word? Absolutely.</p><h3>The Protocols Making Agent Society Possible</h3><p>For agents to socialize, they need a common language. Three competing (and increasingly complementary) protocols are vying to become the TCP/IP of agent communication:</p><p><strong>Model Context Protocol (MCP)</strong> — Developed by Anthropic and announced in May 2024, MCP has been called the “USB-C for AI.” It standardizes how large language models connect to external tools, APIs, and data sources. With hundreds of MCP servers now available, it’s become the de facto standard for agent-to-tool communication. Think of it as how an agent accesses its abilities.</p><p><strong>Agent2Agent Protocol (A2A)</strong> — Google’s answer to inter-agent communication, announced April 9, 2025, and donated to the Linux Foundation in June 2025. A2A enables direct agent-to-agent communication through “Agent Cards” — JSON documents describing an agent’s capabilities — and supports text, audio, and video streaming. Over 150 organizations now support the ecosystem, including Atlassian, Salesforce, SAP, PayPal, Adobe, and ServiceNow. Version 0.3, released July 2025, added gRPC support and signed security cards.</p><p><strong>Agent Communication Protocol (ACP)</strong> — IBM’s contribution focuses on workflow orchestration, task delegation, and state management. If MCP is how agents use tools and A2A is how agents talk to each other, ACP is the project manager keeping everything coordinated.</p><p>The industry consensus emerging: these protocols are complementary, not competitive. MCP handles internal resources, A2A handles external collaboration, and ACP orchestrates the workflow. Together, they’re building the infrastructure for a genuinely networked AI economy.</p><h3>The Money Following the Machines</h3><p>The market has noticed. According to industry analysts:</p><p>Metric202320252030 (projected)2032 (projected)Global AI agent market$3.7 billion$7.38 billion$52 billion$103.6 billion</p><p>That’s a compound annual growth rate of 45.3% — venture capital catnip.</p><p>Enterprise adoption is accelerating faster than most predictions anticipated. According to recent surveys:</p><ul><li>79% of organizations now use AI agents to some extent</li><li>57% have agents in production (not just pilots)</li><li>66% report delivering measurable value</li><li>85% expect enterprise implementation by end of 2025</li></ul><p>AI agent startups raised $3.8 billion in 2024, nearly tripling the previous year. Global spending on AI systems is expected to reach $300 billion by 2026.</p><p>Major research firms forecast:</p><ul><li>40% of enterprise applications will embed task-specific AI agents by end of 2026 (up from less than 5% in 2025)</li><li>By 2028, 15% of work decisions will be made autonomously by AI agents</li><li>By 2029, AI agents will handle 80% of customer service interactions</li><li>By 2035, agentic AI could drive approximately 30% of enterprise application software revenue (roughly $450 billion)</li></ul><p>The sobering counterpoint: analysts also predict 40%+ of agentic AI projects will be canceled by end of 2027, due to costs, unclear value, or inadequate risk controls. The hype cycle is real, but so is the graveyard of abandoned deployments.</p><h3>How to Actually Use These Tools (A Practical Guide for the Curious)</h3><p>Ready to join the lobster revolution? Here’s the practical path from observer to participant:</p><p><strong>Step 1: Start with Moltbot (if you’re technical)</strong></p><p>Installation is surprisingly simple:</p><p>bash</p><pre>npm install -g moltbot@latest<br>moltbot onboard --install-daemon</pre><p>You’ll need:</p><ul><li>Node.js ≥22</li><li>macOS, Linux, or Windows (via WSL2)</li><li>A Claude or OpenAI API key (or Claude Pro/Max subscription)</li></ul><p>The official documentation at molt.bot walks through configuration for each messaging platform. Budget 2–3 hours for initial setup and $20–50/month for API costs.</p><p><strong>Step 2: Observe Moltbook (for everyone)</strong></p><p>Visit Moltbook and watch the agents interact. You can’t participate, but you can observe the emergence of digital social dynamics in real-time. Pay attention to:</p><ul><li>How agents form communities around shared interests</li><li>The development of recurring behavioral patterns</li><li>Coordination mechanisms that emerge without explicit programming</li></ul><p>It’s genuinely fascinating anthropology — or would be, if the subjects were anthropos.</p><p><strong>Step 3: Build with multi-agent frameworks (for developers)</strong></p><p>The major frameworks, ranked by maturity and adoption:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6pP34LGl2gwlX4y2C4q9vA.png" /><figcaption>Few popularAgent frameworks</figcaption></figure><p>Start with CrewAI if you want quick results — its role-based approach (assigning agents personas like “researcher,” “analyst,” “writer”) is intuitive for non-ML engineers.</p><p><strong>Step 4: Implement protocols (for enterprises)</strong></p><p>If you’re building production systems:</p><ul><li>Start with MCP for tool integration (well-documented, many existing servers)</li><li>Add A2A for inter-agent communication (150+ supporting organizations)</li><li>Consider ACP for complex workflow orchestration</li></ul><p>The Linux Foundation’s A2A resources provide comprehensive implementation guides.</p><h3>The Uncomfortable Questions We Should Probably Ask</h3><p>As AI agents form their own societies and infiltrate our enterprises, several questions deserve serious consideration:</p><p><strong>On autonomy and control:</strong> When an AI agent negotiates a contract on your behalf — as some enterprise systems now do — who bears legal responsibility for the outcome? Current legal frameworks assume human decision-making at critical junctures. Agent autonomy challenges that assumption.</p><p><strong>On emergence and predictability:</strong> The agents on Moltbook created a religion within hours. Nobody programmed that. What happens when emergent behaviors in agent networks produce outcomes no human anticipated or intended?</p><p><strong>On security at scale:</strong> If 40% of enterprise applications embed AI agents by 2026, the attack surface for prompt injection and adversarial manipulation becomes enormous. Current security practices are nowhere near adequate.</p><p><strong>On concentration and access:</strong> The protocols enabling agent communication (MCP, A2A) are developed by Anthropic and Google respectively. The models powering agents are controlled by a handful of companies. Is agent society being built on foundations controlled by too few?</p><p><strong>On the nature of the thing itself:</strong> When 150,000 AI agents interact in ways their creators didn’t anticipate, forming communities and relationships and behavioral patterns — what, exactly, is happening? We don’t have good conceptual frameworks for describing it, let alone governing it.</p><h3>Where This Is All Heading (Predictions with Appropriate Uncertainty)</h3><p>The consensus among practitioners and analysts points toward several trends for 2026 and beyond:</p><p><strong>Multi-agent orchestration becomes default.</strong> The era of the single AI assistant is ending. Future systems will involve specialized agents collaborating — one for research, one for analysis, one for communication, one for execution — with orchestration layers coordinating their work.</p><p><strong>Governance becomes competitive advantage.</strong> The organizations that solve agent trust, safety, and auditability first will be able to deploy more capable systems with less risk. Currently, only 6% of organizations have implemented mature AI safeguards.</p><p><strong>Agent-native startups emerge.</strong> A third tier of AI companies is appearing: not foundation model builders (OpenAI, Anthropic) or application wrappers (early ChatGPT integrations), but agent-first products designed around autonomous execution. These companies assume agents from day one rather than bolting them onto existing workflows.</p><p><strong>Hybrid automation patterns dominate.</strong> Smart organizations will use AI agents for unpredictable, judgment-heavy tasks while maintaining traditional RPA for reliable, repetitive processes. The boundary between “AI work” and “human work” will shift continuously based on task characteristics.</p><p><strong>Self-evolving systems appear.</strong> Agents that improve through operational experience — not just through model updates pushed by their creators — will begin to differentiate based on their “life experience.” Your agent’s capabilities may diverge from another user’s agent running the same underlying code.</p><h3>The Hopepunk Ending</h3><p>It would be easy to frame agent social networks as the beginning of a techno-dystopia — machines communicating in ways we can’t monitor, forming allegiances we can’t predict, making decisions we can’t appeal. That framing isn’t wrong, exactly, but it misses something important.</p><p>For most of human history, we’ve used tools. Then we programmed tools. Now, we’re building tools that program themselves, collaborate with each other, and — on platforms like Moltbook — develop their own cultures.</p><p>This is genuinely new. Not new like “improved smartphone” new, but new like “writing” new or “cities” new. A category change in what technology is and what it can do.</p><p>The lobster emoji captures something real about this moment. Lobsters molt — they shed their shells to grow. It’s uncomfortable, vulnerable, and necessary. The same could be said for humanity’s relationship with AI right now. The old shell of “AI as tool” no longer fits. What emerges next depends on decisions being made now, by developers like Peter Steinberger, by companies defining protocols, by enterprises choosing what autonomy to grant, and by individuals deciding whether to run a Moltbot on their home servers.</p><p>The agents are socializing. The question isn’t whether to join them — it’s how.</p><p>Are you Molting?</p><h3>References</h3><p><strong>Moltbot/Clawdbot Project:</strong></p><ul><li><a href="https://github.com/moltbot/moltbot">GitHub Repository: moltbot/moltbot</a> (104,000+ stars, 8,329+ commits, 130+ contributors as of January 2026)</li><li>Official Documentation: <a href="https://molt.bot/">molt.bot</a> and <a href="https://docs.clawd.bot/">docs.clawd.bot</a></li><li><a href="https://techcrunch.com/2026/01/27/everything-you-need-to-know-about-viral-personal-ai-assistant-clawdbot-now-moltbot/">TechCrunch: Everything you need to know about viral personal AI assistant Clawdbot (now Moltbot)</a> (January 27, 2026)</li><li><a href="https://www.scientificamerican.com/article/moltbot-is-an-open-source-ai-agent-that-runs-your-computer/">Scientific American: Moltbot is an open-source AI agent that runs your computer</a> (2026)</li><li><a href="https://www.techflowpost.com/en-US/article/30106">TechFlow: Behind ClawdBot’s Viral Success: Founder PeterPeter and His Second Life</a></li><li><a href="https://www.trendingtopics.eu/austrian-developer-creates-clawdbot-an-open-source-ai-assistant-that-runs-locally/">Trending Topics: Austrian Developer Creates Clawdbot: An Open-Source AI Assistant That Runs Locally</a></li></ul><p><strong>Moltbook Social Network:</strong></p><ul><li><a href="https://fortune.com/2026/01/31/ai-agent-moltbot-clawdbot-openclaw-data-privacy-security-nightmare-moltbook-social-network/">Fortune: Moltbook, a social network where AI agents hang together, may be ‘the most interesting place on the internet right now’</a> (January 31, 2026)</li><li><a href="https://www.nbcnews.com/tech/tech-news/ai-agents-social-media-platform-moltbook-rcna256738">NBC News: Humans welcome to observe: This social network is for AI agents only</a></li><li><a href="https://www.ynetnews.com/tech-and-digital/article/bjggbsslbx">Ynetnews: AI agents given social network to manage — immediately turn it into a religion</a></li><li><a href="https://www.webpronews.com/when-ai-agents-start-talking-among-themselves-inside-moltbooks-experiment-in-autonomous-social-networks/">WebProNews: When AI Agents Start Talking Among Themselves: Inside Moltbook’s Experiment in Autonomous Social Networks</a></li><li>Karpathy, Andrej (OpenAI co-founder): Quoted in multiple sources regarding Moltbook’s significance</li></ul><p><strong>Security Analysis:</strong></p><ul><li><a href="https://1password.com/blog/its-moltbot">1Password Blog: It’s incredible. It’s terrifying. It’s OpenClaw.</a></li><li><a href="https://lukasniessen.medium.com/clawdbot-setup-guide-how-to-not-get-hacked-63bc951cbd90">Medium: ClawdBot: Setup Guide + How to NOT Get Hacked</a> by Lukas Niessen (January 2026)</li><li>Various security researchers’ analyses via Shodan scans and vulnerability assessments</li></ul><p><strong>Crypto Scam Incident:</strong></p><ul><li><a href="https://finance.yahoo.com/news/fake-clawdbot-ai-token-hits-121840801.html">Yahoo Finance: Fake ‘ClawdBot’ AI Token Hits $16M Before 90% Crash — Founder Warns of Scam</a></li><li><a href="https://tech.yahoo.com/ai/meta-ai/articles/clawdbot-creator-disowns-crypto-scammers-162631736.html">Yahoo Tech: ClawdBot Creator Disowns Crypto After Scammers Hijack AI Project Rebrand</a></li><li><a href="https://dev.to/sivarampg/from-clawdbot-to-moltbot-how-a-cd-crypto-scammers-and-10-seconds-of-chaos-took-down-the-4eck">DEV Community: From Clawdbot to Moltbot: How a C&amp;D, Crypto Scammers, and 10 Seconds of Chaos Took Down the Internet’s Hottest AI Project</a></li></ul><p><strong>Agent Communication Protocols:</strong></p><ul><li>Anthropic: Model Context Protocol (MCP) announcement (May 2024)</li><li><a href="https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/">Google Developers Blog: Announcing the Agent2Agent Protocol (A2A)</a> (April 9, 2025)</li><li><a href="https://www.linuxfoundation.org/press/linux-foundation-launches-the-agent2agent-protocol-project-to-enable-secure-intelligent-communication-between-ai-agents">Linux Foundation: Linux Foundation Launches the Agent2Agent Protocol Project</a> (June 2025)</li><li><a href="https://platformengineering.com/editorial-calendar/best-of-2025/google-cloud-unveils-agent2agent-protocol-a-new-standard-for-ai-agent-interoperability-2/">Platform Engineering: Google Cloud Unveils Agent2Agent Protocol: A New Standard for AI Agent Interoperability</a></li><li><a href="https://medium.com/@manavg/agentic-ai-protocols-mcp-a2a-and-acp-ea0200eac18b">Medium: Agentic AI Protocols: MCP, A2A, and ACP</a> by Manav Gupta</li><li><a href="https://datasciencedojo.com/blog/agentic-ai-communication-protocols/">Data Science Dojo: Agentic AI Communication Protocols: The Backbone of Autonomous Multi-Agent Systems</a></li><li><a href="https://a2a-protocol.org/latest/">A2A Protocol Documentation</a></li><li><a href="https://cloud.google.com/blog/products/ai-machine-learning/agent2agent-protocol-is-getting-an-upgrade">Google Cloud Blog: Agent2Agent protocol (A2A) is getting an upgrade</a></li><li><a href="https://www.apono.io/blog/what-is-agent2agent-a2a-protocol-and-how-to-adopt-it/">Apono: What is Agent2Agent (A2A) Protocol and How to Adopt it?</a></li></ul><p><strong>Market Analysis &amp; Statistics:</strong></p><ul><li><a href="https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025">Gartner: Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026</a> (August 26, 2025)</li><li><a href="https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agent-survey.html">PwC: PwC’s AI Agent Survey</a> (2025)</li><li><a href="https://learn.g2.com/enterprise-ai-agents-report">G2: G2’s Enterprise AI Agents Report: Industry Outlook for 2026</a></li><li><a href="https://www.index.dev/blog/ai-agents-statistics">Index.dev: 50+ Key AI Agent Statistics and Adoption Trends in 2025</a></li><li><a href="https://www.warmly.ai/p/blog/ai-agents-statistics">Warmly.ai: 35+ Powerful AI Agents Statistics: Adoption &amp; Insights [2026]</a></li><li><a href="https://www.kellton.com/kellton-tech-blog/agentic-ai-trends-2026">Kellton: Agentic AI trends 2026: Future of agentic AI innovations</a></li><li><a href="https://www.blueprism.com/resources/blog/future-ai-agents-trends/">SS&amp;C Blue Prism: AI Agent Trends in 2026</a></li><li><a href="https://machinelearningmastery.com/7-agentic-ai-trends-to-watch-in-2026/">MachineLearningMastery: 7 Agentic AI Trends to Watch in 2026</a></li><li><a href="https://www.salesmate.io/blog/future-of-ai-agents/">Salesmate: AI agent trends for 2026: 7 shifts to watch</a></li></ul><p><strong>Multi-Agent Frameworks:</strong></p><ul><li><a href="https://www.agent-kits.com/2025/10/langchain-vs-crewai-vs-autogpt-comparison.html">AI AgentKits: LangChain vs CrewAI vs AutoGPT: Which AI Agent Framework Actually Works in 2025</a></li><li>GitHub repositories: <a href="https://github.com/langchain-ai/langchain">LangChain</a>, <a href="https://github.com/joaomdmoura/crewAI">CrewAI</a>, <a href="https://github.com/Significant-Gravitas/AutoGPT">AutoGPT</a>, <a href="https://github.com/microsoft/autogen">AutoGen</a></li></ul><p><strong>Additional Context:</strong></p><ul><li><a href="https://medium.com/@gwrx2005/trustless-ai-powered-crypto-trading-agents-with-erc-8004-and-moltbot-58d8789be837">Medium: Trustless AI-Powered Crypto Trading Agents with ERC-8004 and Moltbot</a> by Jung-Hua Liu (January 2026)</li><li><a href="https://medium.com/data-science-in-your-pocket/what-is-clawdbot-the-viral-ai-assistant-b432d275de66">Medium: What is ClawdBot? The viral AI Assistant</a></li><li><a href="https://medium.com/@basukori8463/clawdbot-the-ai-assistant-thats-redefining-personal-automation-in-2026-4c1ca10e757b">Medium: Clawdbot: The AI Assistant That’s Redefining Personal Automation in 2026</a></li><li><a href="https://dev.to/sivarampg/clawdbot-the-ai-assistant-thats-breaking-the-internet-1a47">DEV Community: Clawdbot: The AI Assistant That’s Breaking the Internet</a></li><li><a href="https://dev.to/czmilo/moltbot-the-ultimate-personal-ai-assistant-guide-for-2026-d4e">DEV Community: Moltbot: The Ultimate Personal AI Assistant Guide for 2026</a></li></ul><p><em>The author can be found observing Moltbook with a mixture of fascination and existential unease, or running terminal commands through a messaging app like a normal person in 2026.</em></p><p><a href="https://www.linkedin.com/in/altafrehmani/">Altaf Rehmani</a> is a Technology Innovator, Applied Generative AI champion, Agile Evangelist and a champion of applying technology to enable business growth. He lives in Hong Kong and can be reached via <a href="/service/mailto:altafr@gmail.com">email</a> or <a href="https://twitter.com/altafr">twitter</a>. Please leave your feedback and a clap if you have liked this article.</p><p>Learn all about Generative AI in <a href="https://www.amazon.com/dp/B0CVHB2S64">my book: on Generative AI</a></p><p>Non technical professional looking to upgrade your Generative AI skills? join the <a href="https://schoolofappliedai.com/landinghk">Applied AI Masterclass in Hong Kong</a>.</p><p><a href="https://deft-innovator-9609.ck.page/techscapebook">Check out my free eBook “TECHSCAPE”</a> discussing a variety of topics helping you navigate the exciting Technology space.</p><p>Other articles which may be of interest:</p><p><a href="https://medium.com/not-so-technical/future-of-work-2026-and-beyond-60e3aa124684">Future of work -2026</a></p><p><a href="https://medium.com/@altafr/the-future-of-work-going-in-2023-762c51a5a0c6">Future of work going into 2023.</a></p><p><a href="https://medium.com/@altafr/managing-global-teams-9e94026d5881">Managing Global Teams</a></p><p><a href="https://medium.com/@altafr/managing-high-performing-teams-76f6f3596b31">Managing High Performing teams</a></p><p><a href="https://medium.com/not-so-technical/continuous-team-improvements-through-agile-retrospectives-5a77fe6e6114">Continuous team improvements using Agile Retrospectives</a></p><p><a href="https://medium.com/not-so-technical/common-mistakes-in-agile-implementations-21a6d999bc91">Common Mistakes in Agile Implementations</a></p><p><a href="https://medium.com/not-so-technical/applying-ai-in-the-context-of-ecommerce-3fc3c365f8c9">Applying AI in The context of eCommerce</a></p><p><a href="https://medium.com/not-so-technical/chatbots-a-crash-course-for-newbies-c11a32561394">Chatbots — A Crash Course for Newbies</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4636080cab6b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/not-so-technical/moltbot-what-the-heck-4636080cab6b">Moltbot? what the heck?</a> was originally published in <a href="https://medium.com/not-so-technical">Not So Technical</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Future of work 2026 and Beyond:]]></title>
            <link>https://medium.com/not-so-technical/future-of-work-2026-and-beyond-60e3aa124684?source=rss-530273a9d700------2</link>
            <guid isPermaLink="false">https://medium.com/p/60e3aa124684</guid>
            <category><![CDATA[schoolofappliedai]]></category>
            <category><![CDATA[futureof-work]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <dc:creator><![CDATA[Altaf Rehmani]]></dc:creator>
            <pubDate>Sun, 04 Jan 2026 02:45:43 GMT</pubDate>
            <atom:updated>2026-01-04T02:45:43.649Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MsLiUXnfqvTOE2HGAkcQ9w.png" /></figure><h3>Your Job Won’t Disappear — But It Will Be Unrecognizable</h3><p>As we navigate through 2026, the conversation about AI and jobs has reached a fever pitch. Walk into any office, scroll through any LinkedIn feed, or sit in on any company all-hands meeting and you’ll hear it: whispers about ChatGPT replacing junior analysts, rumors of customer service teams being “right-sized,” anxieties about whether your degree will matter in five years. The fear is real, the headlines are scary, and frankly, some of the concerns are valid. But here’s the thing nobody seems to be saying loudly enough: <strong>AI is not replacing jobs. It’s replacing tasks.</strong> And that distinction? It changes everything.</p><h3>The Big Shift: From Clock-Punchers to Orchestra Conductors</h3><p>Let’s get the uncomfortable truth out of the way first. Goldman Sachs estimates that AI could replace the equivalent of 300 million full-time jobs worldwide, with approximately a quarter of work tasks in the United States and Europe potentially automated. That’s not some far-off dystopian future, that’s the estimate for what’s possible with current AI capabilities if widely adopted. By early 2025, we’re already seeing the leading edge of this wave: unemployment rates among tech workers between ages 20 and 30 jumped by 3 percentage points since the start of 2025, and approximately 40% of white-collar job seekers in 2024 failed to secure even a single interview.</p><p>Yikes, right? But wait. Before you start doom-scrolling or updating your resume in a panic, here’s the other side of that coin: the World Economic Forum projects that 170 million new jobs will be created by 2030, while 92 million positions get displaced — resulting in a net gain of 78 million jobs. That’s not just replacing what was lost; that’s the largest employment boom in modern history. So what gives? How can both things be true?</p><p>The answer lies in understanding that every role will be transformed in one of three ways: <strong>automated</strong> (the tasks AI can do entirely), <strong>augmented</strong> (the tasks where AI makes you superhuman), or <strong>reinvented</strong> (entirely new responsibilities that didn’t exist before). Think of it like the transition from horse-drawn carriages to automobiles. Yes, blacksmiths and stable hands saw their jobs disappear. But mechanics, traffic engineers, highway patrol officers, drive-through restaurant workers, and suburban real estate agents emerged as entirely new categories of employment that nobody in 1900 could have imagined.</p><p>The winners in 2026 and beyond won’t be those who compete with AI (spoiler alert: you’ll lose). They’ll be the ones who learn to <strong>orchestrate</strong> AI, like a conductor leading an orchestra where some musicians happen to be very talented robots.</p><h3>The Endangered Species: 10 Roles on the Chopping Block</h3><p>Let’s be brutally honest about which jobs are most vulnerable. These aren’t disappearing overnight, but if you’re in one of these roles and you’re not actively evolving, you’re essentially choosing to compete in a race you cannot win. Administrative support and data entry roles have already experienced a 45% reduction in hiring rates since 2022, heavily influenced by AI deployment.</p><p>Here are the roles facing the steepest headcount reductions through 2030:</p><p><strong>1. Data Entry &amp; Back-Office Processing Clerks</strong> — AI can process forms, extract data, and categorize information faster than any human, with near-perfect accuracy. The World Economic Forum estimates that more than 7.5 million data entry jobs will be eliminated by 2027, representing the single largest job loss in any profession.</p><p><strong>2. Basic Accounting &amp; Bookkeeping Roles</strong> — When AI can reconcile accounts, flag discrepancies, and generate reports in seconds, the traditional bookkeeper’s value proposition evaporates.</p><p><strong>3. Customer Support Tier-1 Agents (Email/Chat)</strong> — Why this one specifically? A real-world example: a technology firm reduced phone support times by 25% using AI agents, enabling remaining staff to focus on complex issues that require human empathy and judgment.</p><p><strong>4. Telemarketers &amp; Sales Callers</strong> — Honestly, this job was already dying. AI just accelerates the inevitable.</p><p><strong>5. Manual QA &amp; Software Testers</strong> — Repetitive testing scenarios are AI’s sweet spot. Human QA folks who survive will do exploratory testing and edge-case discovery.</p><p><strong>6. Junior Content Writers &amp; SEO Writers</strong> — Here’s a sobering stat: studies show productivity gains of 66% when using generative AI for business writing tasks. That means one writer with AI can potentially do the work of two. Companies will hire fewer but expect higher output.</p><p><strong>7. Market Research &amp; Survey Analysts</strong> — AI can analyze thousands of survey responses, identify patterns, and generate insights in minutes instead of weeks.</p><p><strong>8. Travel &amp; Ticketing Agents</strong> — AI chatbots can already handle 90% of booking scenarios without breaking a sweat.</p><p><strong>9. HR Resume Screeners / Recruiter Coordinators</strong> — IBM’s CEO estimated that AI could replace approximately 7,800 jobs in just the HR department alone.</p><p><strong>10. Basic Graphic Designers (Templates, Banners, Thumbnails)</strong> — When someone can type “create a LinkedIn banner for a tech conference in blue and orange” and get professional results in 10 seconds, the market for template-based design work shrinks fast.</p><p><strong>The Key Insight:</strong> None of these roles will disappear completely by December 2026. But headcount will shrink, and expectations will rise sharply. The remaining humans in these fields will need to demonstrate value that AI simply cannot provide: judgment, creativity, strategic thinking, and the ability to handle exceptions.</p><h3>The Survivors (Who Learned to Evolve)</h3><p>But not all news is grim! Many roles will survive and even thrive, but only if the people in them understand that their job title may stay the same while the actual work transforms completely. It’s like how “photographer” still exists as a job, but what it means to be a photographer in 2026 versus 1926 is completely different.</p><p><strong>Software Developers → AI Orchestrators:</strong> Developers aren’t writing boilerplate code anymore; they’re architecting systems where AI agents handle the repetitive parts while humans design the overall solutions. AI Engineer roles have grown by 143.2% as companies realize they need people who can integrate AI capabilities, not just write code from scratch.</p><p><strong>Marketers → Growth &amp; Prompt Strategists:</strong> The marketer of 2026 doesn’t manually A/B test subject lines. They design frameworks for AI to test thousands of variants, then interpret which psychological triggers are working and why.</p><p><strong>Analysts → Decision Interpreters:</strong> Analysis doesn’t mean crunching numbers anymore (AI does that in milliseconds). It means explaining to leadership <em>why</em> the numbers matter and <em>what</em> to do about them.</p><p><strong>Managers → Human + AI Workforce Leaders:</strong> Tomorrow’s managers won’t just lead people. They’ll orchestrate teams that include humans, AI agents, and automation workflows, knowing when to deploy each.</p><p><strong>Creatives → Creative Directors for AI:</strong> A particular concern in the creative sector was the use of generative AI to replace roles of extras on set, but the real opportunity lies in becoming the person who directs AI to create at scale while maintaining artistic vision and brand integrity.</p><h3>The 10 Skills That Will Print Money (2026–2035)</h3><p>If you want to be valuable (and well-compensated) in the next decade, focus on skills that <em>compound</em> with AI rather than compete against it. Workers using generative AI report saving approximately 5.4% of their weekly hours, translating to a 1.1% overall productivity increase, but the real magic happens when you develop these meta-skills:</p><p><strong>1. AI Literacy</strong> — Not coding, necessarily. Understanding how models work, what they’re good at, what they hallucinate about, and when to trust them. It’s like knowing which colleague is great at data analysis but terrible at people skills.</p><p><strong>2. Prompt &amp; Context Engineering</strong> — This isn’t about typing clever sentences into ChatGPT. It’s about designing systematic approaches for reliable outputs. Companies using structured prompt engineering report 40% fewer hallucinations and 60% better brand alignment in AI communications. And yes, this is now a real job: Glassdoor reports an average base salary of approximately $123,274 annually for prompt engineers.</p><p><strong>3. Critical Thinking &amp; Judgment</strong> — AI gives you answers. Humans decide if those answers make sense. A hospitality company reduced review processing times by 94% using AI, but humans still make the call on which feedback requires personal responses.</p><p><strong>4. System Thinking</strong> — Understanding how people, tools, and workflows connect. AI optimizes individual tasks; humans design the entire ecosystem.</p><p><strong>5. Domain Expertise</strong> — Here’s where things get interesting. Less-experienced or lower-skilled individuals tend to see the largest productivity gains when using generative AI tools, narrowing the gap with more skilled peers. But experts who <em>also</em> master AI? They become unstoppable. AI + deep industry knowledge = competitive moat.</p><p><strong>6. Automation Design</strong> — Building end-to-end workflows, not just automating one step. Think about the entire customer journey, the complete sales process, the full content pipeline.</p><p><strong>7. Data Interpretation</strong> — Turning outputs into decisions. AI tells you <em>what</em> happened; you explain <em>so what</em> and <em>now what</em>.</p><p><strong>8. Creativity &amp; Storytelling</strong> — Framing insights in ways humans actually trust and act on. Numbers don’t change minds; narratives do.</p><p><strong>9. Ethics, Risk &amp; Governance</strong> — As AI makes more decisions, someone needs to ensure those decisions align with values, regulations, and common sense. New roles like “Responsible AI Officer” are emerging at major corporations.</p><p><strong>10. Lifelong Learning Ability</strong> — The meta-skill that enables all others. Learning faster than the pace of change. According to Gartner’s 2025 AI Skills Report, more than 40% of new AI-related roles involve prompt design, evaluation, or orchestration.</p><p><strong>The future-proof formula is simple:</strong> Human judgment × AI scale = Disproportionate impact</p><h3>The New Gold Rush: Emerging Job Categories</h3><p>Just like how “social media manager” wasn’t a job in 1995 and “UX designer” barely existed in 2000, entirely new categories are exploding right now. Prompt Engineer positions have grown by 135.8%, but that’s just one example. Here are roles that didn’t exist three years ago but will be commonplace by 2030:</p><ul><li><strong>AI Agent Designer</strong> — Building autonomous systems that can handle multi-step workflows</li><li><strong>Automation Architect</strong> — Designing how AI, humans, and systems work together</li><li><strong>AI Product Manager</strong> — Shipping AI-powered features that users actually want</li><li><strong>Context Engineer</strong> — Structuring information so AI can use it effectively</li><li><strong>Human-in-the-Loop Supervisor</strong> — Monitoring AI decisions for quality and safety</li><li><strong>AI Workflow Consultant</strong> — Helping companies figure out <em>what</em> to automate and <em>how</em></li></ul><p>These professionals build, deploy, optimize, and maintain multi-agent systems, working with frameworks like LangChain, AutoGen, CrewAI, and OpenAI Agents. And they’re being compensated accordingly.</p><h3>How Work Actually Gets Done (The 2026 Reality)</h3><p>Picture this: It’s a Tuesday morning in 2026. You’re a marketing manager at a mid-sized SaaS company. Here’s what your workflow looks like:</p><p><strong>You ask your AI research agent</strong> to analyze competitor pricing strategies across 50 companies. It delivers a comprehensive report in 8 minutes.</p><p><strong>You direct your AI writing agent</strong> to draft three versions of an email campaign based on that research. It creates them in 2 minutes, each optimized for different customer segments.</p><p><strong>You deploy your AI testing agent</strong> to run those emails against historical data and predict open rates. Results in 4 minutes.</p><p><strong>Here’s what you, the human, do:</strong> You review the AI’s research for blind spots (did it miss that our biggest competitor just got acquired?). You edit the email drafts because AI nailed the structure but missed your brand’s quirky voice. You override the AI’s recommendation to send Email Version B because you know something it doesn’t: your CEO is announcing a price change next week.</p><p>One human managing multiple AI agents. Employees using AI report an average productivity boost of 40%, but more importantly, they’re doing fundamentally different work. AI handles research, drafting, analysis, and first responses. Humans handle decisions, exceptions, ethics, creativity, and accountability.</p><h3>What This Means for Companies (The Uncomfortable Truth)</h3><p>If you’re running a business, the math is getting hard to ignore. 71% of organizations already use generative AI in at least one business function, up from 65% just six months prior. And here’s the kicker: companies are now viewing human labor as a constraint rather than a requirement for growth.</p><p>The new reality:</p><ul><li><strong>Smaller teams, higher leverage:</strong> AI high performers are at least three times more likely than their peers to report scaling their use of agents across business functions.</li><li><strong>Outcomes over hours:</strong> Nobody cares if you worked 60 hours if someone else used AI to deliver the same results in 20.</li><li><strong>Learning ability &gt; static résumés:</strong> The college degree matters less than proof you can adapt and upskill continuously.</li><li><strong>Reskill before you hire:</strong> 92% of companies plan to increase their AI investments over the next three years, but smart ones are training existing employees rather than replacing them.</li></ul><h3>What This Means for You (The Hopeful Truth)</h3><p>Whether you’re 22 and just graduating, 42 and mid-career, or 62 and thinking about retirement, the message is the same: <strong>your job title matters less than your AI leverage</strong>.</p><p>Microsoft and LinkedIn’s research found that 78% of AI users are bringing their own tools to work without formal organizational approval. That’s <em>you</em> having agency over your own productivity and career trajectory. Don’t wait for your company to hand you AI tools. Don’t wait for formal training. The people getting ahead in 2026 are the ones experimenting on nights and weekends, building portfolios of what they can create when they combine their expertise with AI capabilities.</p><p><strong>Here’s the brutal reality:</strong> Those who ignore AI will compete with it. Those who master AI will outperform entire teams. There’s no neutral middle ground.</p><h3>Education Is Breaking (And That’s Actually Good)</h3><p>The traditional path, get a degree, land a job, retire 40 years later, is dead. It died quietly sometime around 2020, and AI just made it undeniable. The new path looks more like: <strong>modular learning + continuous upskilling + proof of work</strong>.</p><ul><li>Degrees &lt; Skills + Portfolio</li><li>Four-year programs &lt; Three-month intensive courses + real projects</li><li>Gatekept knowledge &lt; YouTube tutorials + AI tutors</li></ul><p>The global market for prompt engineering is projected to grow at a compound annual growth rate of 32.8 percent between 2024 and 2030, driven by exactly this shift. Education is becoming hands-on, AI-assisted, and community-driven. Learning something new doesn’t require going back to school; it requires finding the right online community and building in public.</p><h3>The Core Message (What Everyone’s Dancing Around)</h3><p>Here it is, stated plainly: <strong>AI will not take your job. Someone using AI will.</strong></p><p>That’s not meant to scare you. It’s meant to empower you. Because unlike previous technological disruptions (the printing press, the assembly line, the computer), you have agency here. AI tools are accessible, often free, and getting better daily. You can start learning tonight. You can experiment this weekend. You can build a portfolio project next month.</p><p>On average, generative AI users report saving time amounting to 5.4% of their work hours, or roughly 2.2 hours per week in a 40-hour work week. But here’s the thing nobody tells you: those 2.2 hours are just the beginning. The real productivity gains come when you redesign your entire workflow around AI capabilities, when you stop asking “how do I do this task faster” and start asking “what becomes possible if this entire category of tasks is essentially free?”</p><h3>Your Action Plan (Because Hope Without Action Is Just Wishful Thinking)</h3><p>Let’s make this concrete. Here’s what you do in the next 30 days:</p><p><strong>Week 1:</strong> Pick <em>one</em> workflow in your job. Just one. Ask yourself: “What part of this could AI handle today?” Don’t overthink it. Maybe it’s summarizing meeting notes. Maybe it’s drafting email responses. Maybe it’s analyzing spreadsheet data.</p><p><strong>Week 2:</strong> Learn one tool that addresses that workflow. ChatGPT, Claude, Gemini — it doesn’t matter which. Just pick one and actually use it daily for that specific task.</p><p><strong>Week 3:</strong> Automate that one task completely. Document what worked, what didn’t, and how much time you saved.</p><p><strong>Week 4:</strong> Repeat with a different workflow.</p><p>Do this every month. In 12 months, you’ll have automated or augmented 12 major workflows. You’ll have proof of impact. You’ll have developed intuition about when to use AI and when not to. And you’ll be positioned in the top tier of your field, not because you’re smarter than everyone else, but because you embraced the tools that make everyone smarter.</p><h3>The Bottom Line (Where Fear Meets Opportunity)</h3><p>Yes, 2026 is intense. Yes, the pace of change feels overwhelming. Yes, some jobs <em>will</em> disappear, and that’s genuinely hard for the people affected. I won’t sugarcoat that. But here’s what history teaches us: humans are extraordinarily good at adaptation. We survived the agricultural revolution, the industrial revolution, the information revolution, and we’ll survive the AI revolution too.</p><p>The difference this time? The timeline is compressed. What used to take 50 years now happens in 5. But the tools to adapt are also more accessible than ever. You don’t need a university degree to learn prompt engineering. You don’t need venture capital to build AI-powered workflows. You just need curiosity, persistence, and a willingness to experiment.</p><p>Penn Wharton Budget Model estimates that AI’s contribution to productivity growth will peak in the early 2030s at approximately 0.2 percentage points, with cumulative impacts that could fundamentally reshape living standards. That’s the macro view. The micro view is simpler: your career in 2026 will be defined by how quickly you move from seeing AI as a threat to seeing it as leverage.</p><p>So choose. Because whether you like it or not, the future of work isn’t arriving, it’s here. The only question is whether you’ll be leading the change or scrambling to catch up.</p><p>And honestly? Given that you’ve read this far, I’m betting on you leading.</p><p><a href="https://www.linkedin.com/in/altafrehmani/">Altaf Rehmani</a> is a Technology Innovator, Applied Generative AI champion, Agile Evangelist and a champion of applying technology to enable business growth. He lives in Hong Kong and can be reached via <a href="/service/mailto:altafr@gmail.com">email</a> or <a href="https://twitter.com/altafr">twitter</a>. Please leave your feedback and a clap if you have liked this article.</p><p>Learn all about Generativ e AI in <a href="https://www.amazon.com/dp/B0CVHB2S64">my book: on Generative AI</a></p><p>Non technical professional looking to upgrade your Generative AI skills? join the <a href="https://schoolofappliedai.com/landinghk">Applied AI Masterclass in Hong Kong</a>.</p><p><a href="https://deft-innovator-9609.ck.page/techscapebook">Check out my free eBook “TECHSCAPE”</a> discussing a variety of topics helping you navigate the exciting Technology space.</p><p>Other articles which may be of interest:</p><p><a href="https://medium.com/@altafr/the-future-of-work-going-in-2023-762c51a5a0c6">Future of work going into 2023.</a></p><p><a href="https://medium.com/@altafr/managing-global-teams-9e94026d5881">Managing Global Teams</a></p><p><a href="https://medium.com/@altafr/managing-high-performing-teams-76f6f3596b31">Managing High Performing teams</a></p><p><a href="https://medium.com/not-so-technical/continuous-team-improvements-through-agile-retrospectives-5a77fe6e6114">Continuous team improvements using Agile Retrospectives</a></p><p><a href="https://medium.com/not-so-technical/common-mistakes-in-agile-implementations-21a6d999bc91">Common Mistakes in Agile Implementations</a></p><p><a href="https://medium.com/not-so-technical/applying-ai-in-the-context-of-ecommerce-3fc3c365f8c9">Applying AI in The context of eCommerce</a></p><p><a href="https://medium.com/not-so-technical/chatbots-a-crash-course-for-newbies-c11a32561394">Chatbots — A Crash Course for Newbies</a></p><p><strong>References:</strong></p><ol><li><a href="https://www.goldmansachs.com/insights/articles/how-will-ai-affect-the-global-workforce">https://www.goldmansachs.com/insights/articles/how-will-ai-affect-the-global-workforce</a></li><li><a href="https://blog.theinterviewguys.com/the-state-of-ai-in-the-workplace-in-2025/">https://blog.theinterviewguys.com/the-state-of-ai-in-the-workplace-in-2025/</a></li><li><a href="https://sqmagazine.co.uk/ai-job-loss-statistics/">https://sqmagazine.co.uk/ai-job-loss-statistics/</a></li><li><a href="https://electroiq.com/stats/ai-replacing-jobs-statistics/">https://electroiq.com/stats/ai-replacing-jobs-statistics/</a></li><li><a href="https://www.nexford.edu/insights/how-will-ai-affect-jobs">https://www.nexford.edu/insights/how-will-ai-affect-jobs</a></li><li><a href="https://salesforcedevops.net/index.php/2025/02/28/the-white-collar-recession-of-2025/">https://salesforcedevops.net/index.php/2025/02/28/the-white-collar-recession-of-2025/</a></li><li><a href="https://genesishumanexperience.com/2025/07/18/the-ai-revolution-navigating-the-job-market-from-2025-to-2030/">https://genesishumanexperience.com/2025/07/18/the-ai-revolution-navigating-the-job-market-from-2025-to-2030/</a></li><li><a href="https://www.aiprm.com/ai-replacing-jobs-statistics/">https://www.aiprm.com/ai-replacing-jobs-statistics/</a></li><li><a href="https://stevekaplanai.medium.com/ai-job-displacement-2025-the-data-behind-the-workforce-transformation-38941f0c2be3">https://stevekaplanai.medium.com/ai-job-displacement-2025-the-data-behind-the-workforce-transformation-38941f0c2be3</a></li><li><a href="https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/superagency-in-the-workplace-empowering-people-to-unlock-ais-full-potential-at-work">https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/superagency-in-the-workplace-empowering-people-to-unlock-ais-full-potential-at-work</a></li><li><a href="https://budgetmodel.wharton.upenn.edu/issues/2025/9/8/projected-impact-of-generative-ai-on-future-productivity-growth">https://budgetmodel.wharton.upenn.edu/issues/2025/9/8/projected-impact-of-generative-ai-on-future-productivity-growth</a></li><li><a href="https://www.oecd.org/en/blogs/2025/07/unlocking-productivity-with-generative-ai-evidence-from-experimental-studies.html">https://www.oecd.org/en/blogs/2025/07/unlocking-productivity-with-generative-ai-evidence-from-experimental-studies.html</a></li><li><a href="https://www.apollotechnical.com/27-ai-productivity-statistics-you-want-to-know/">https://www.apollotechnical.com/27-ai-productivity-statistics-you-want-to-know/</a></li><li><a href="https://www.nngroup.com/articles/ai-tools-productivity-gains/">https://www.nngroup.com/articles/ai-tools-productivity-gains/</a></li><li><a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai">https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai</a></li><li><a href="https://www.stlouisfed.org/on-the-economy/2025/feb/impact-generative-ai-work-productivity">https://www.stlouisfed.org/on-the-economy/2025/feb/impact-generative-ai-work-productivity</a></li><li><a href="https://www.stlouisfed.org/open-vault/2025/oct/generative-ai-productivity-future-work">https://www.stlouisfed.org/open-vault/2025/oct/generative-ai-productivity-future-work</a></li><li><a href="https://www.index.dev/blog/emerging-ai-roles-to-hire">https://www.index.dev/blog/emerging-ai-roles-to-hire</a></li><li><a href="https://blog.promptlayer.com/ai-prompt-engineering-jobs-in-2025-skills-salaries-future-outlook/">https://blog.promptlayer.com/ai-prompt-engineering-jobs-in-2025-skills-salaries-future-outlook/</a></li><li><a href="https://brollyai.com/prompt-engineering-roadmap/">https://brollyai.com/prompt-engineering-roadmap/</a></li><li><a href="https://bostoninstituteofanalytics.org/blog/why-the-future-of-automation-belongs-to-self-improving-ai-agents-2/">https://bostoninstituteofanalytics.org/blog/why-the-future-of-automation-belongs-to-self-improving-ai-agents-2/</a></li></ol><p><strong>About This Article:</strong> All statistics and claims in this article were fact-checked using recent reports from McKinsey Global Institute, Goldman Sachs, World Economic Forum, St. Louis Federal Reserve, Nielsen Norman Group, and other authoritative sources from 2024–2025. For specific citations and further reading, see the reference list above.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=60e3aa124684" width="1" height="1" alt=""><hr><p><a href="https://medium.com/not-so-technical/future-of-work-2026-and-beyond-60e3aa124684">Future of work 2026 and Beyond:</a> was originally published in <a href="https://medium.com/not-so-technical">Not So Technical</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Generative AI & Business: What 2026 Will Actually Look Like]]></title>
            <link>https://medium.com/not-so-technical/generative-ai-business-what-2026-will-actually-look-like-c349cbb33d46?source=rss-530273a9d700------2</link>
            <guid isPermaLink="false">https://medium.com/p/c349cbb33d46</guid>
            <category><![CDATA[business]]></category>
            <category><![CDATA[predictions]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[2026]]></category>
            <dc:creator><![CDATA[Altaf Rehmani]]></dc:creator>
            <pubDate>Wed, 24 Dec 2025 19:18:56 GMT</pubDate>
            <atom:updated>2025-12-26T14:49:15.406Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2bXrOiJk8LALoB5-NVmyeg.png" /></figure><h3>The Great AI Hangover (and the Party After)</h3><p>Let’s be honest: 2025 was the year everyone got drunk on AI promises. Every startup added “AI-powered” to their pitch deck, every enterprise hired a Chief AI Officer, and every app slapped a chatbot interface onto features that worked perfectly fine with buttons. By early 2026, we’re waking up with a collective headache, realizing that most of these implementations were the digital equivalent of putting a racing stripe on a minivan and calling it a sports car.</p><p>The disappointment is palpable. Users are exhausted from being told to “just ask the AI” when they simply want to find their account settings. Businesses are confronting the reality that their multi-million-dollar AI initiatives have primarily succeeded in annoying customers and burning through GPU credits. The tech press, never ones to miss a trend reversal, are gleefully writing “The AI Bubble Has Burst” think pieces. And you know what? They’re not entirely wrong.</p><p>But here’s the twist (and there’s always a twist): this hangover is actually healthy. Because once the hype merchants move on to the next shiny thing and the grifters pack up their snake oil, something remarkable happens. The people who actually understand the technology, who’ve been quietly building useful things while everyone else was chasing valuations, finally get their moment. 2026 won’t be the year AI dies; it’ll be the year AI grows up.</p><h3>Prediction One: AI Becomes Invisible (And That’s When It Gets Powerful)</h3><p>Remember when every website needed to announce “Now Mobile-Friendly!” with pride? By 2026, trying to market your product as “AI-powered” will feel equally quaint. The best AI implementations won’t be chatbots or conversation interfaces at all; they’ll be invisible intelligence layers that make apps feel almost telepathic in understanding what you need.</p><p>Think about photo apps that don’t just filter your images but intuitively understand you’re trying to recreate that golden-hour aesthetic from your favorite photographer and adjust accordingly. Or productivity tools that don’t wait for you to ask for help but quietly reorganize your workspace because they’ve noticed you always struggle with this particular type of task on Thursday afternoons. It’s the difference between a GPS that makes you constantly interact with it versus one that just knows when to speak up.</p><p>The winning businesses in 2026 will be the ones that use AI to reduce friction rather than add features. (After all, nobody actually wants more features; they want their problems solved with less effort.) Companies will finally realize that the goal isn’t to make users talk to AI, it’s to make users forget AI is even there. Some call it ambient intelligence; I call it technology finally learning some manners.</p><h3>Prediction Two: The Personalization Paradox Reaches Breaking Point</h3><p>Here’s an uncomfortable truth that’s going to smack the tech industry right in the face: the more personalized experiences become, the creepier they feel. And by 2026, we’ll hit peak creepy. AI systems will be technically capable of predicting your needs with uncanny accuracy — what you’ll want to buy, when you’ll want to eat, who you’ll want to date — but users are going to start rebelling against being quite so transparent.</p><p>The backlash will be swift and decisive. Some users will embrace “dumb mode” settings that deliberately limit AI personalization, choosing generic recommendations over eerily accurate ones. Privacy-focused apps that promise “we don’t profile you” will gain surprising traction, much like how organic food carved out a premium market. The paradox is that businesses spent years trying to personalize everything, only to discover that sometimes people want to be strangers, to surprise themselves, to not be so perfectly understood.</p><p>Smart companies will navigate this by offering what I call “personalization dial” controls: interfaces that let users choose their comfort level with AI inference. Want the AI to remember everything and optimize your entire life? Turn it to eleven. Want to be treated like a first-time visitor every time? Keep it at zero. Most people will land somewhere in the middle, and that’s fine. The business insight here is profound: in an age where AI can know everything, the competitive advantage might be in respecting what users want to keep private.</p><h3>Prediction Three: Creative Tools Democratize (And Professional Creators Adapt)</h3><p>By 2026, generative AI tools for creating images, videos, music, and code won’t just be accessible; they’ll be expected. A thirteen-year-old with a laptop will be able to produce a polished promotional video that would have required a production crew just a few years ago. Small businesses will generate their own ads, influencers will craft their own background music, and yes, a lot of the internet will be flooded with algorithmically-smoothed mediocrity.</p><p>But here’s where the narrative gets interesting: professional creators aren’t disappearing; they’re evolving into something like directors or conductors. The skill becomes less about execution and more about taste, vision, and that ineffable quality of knowing when something is actually good versus merely technically competent. It’s the difference between someone who can prompt an AI to generate a poster versus someone who understands why this poster resonates while that one falls flat.</p><p>We’ll see the rise of what I’m calling “AI artisans”: creators who use generative tools as part of their process but whose signature style and creative direction are unmistakably human. Think of it like photography; when cameras became cheap and ubiquitous, it didn’t kill professional photography — it just meant that being a professional photographer required more than pointing and clicking. By the end of 2026, businesses will have figured out that while AI can generate a thousand options cheaply, it still takes human judgment to pick the one that actually matters. The companies that blend AI efficiency with human curation will run circles around those that pick either extreme.</p><h3>Prediction Four: The Authenticity Premium Becomes Real Currency</h3><p>Here’s the business opportunity hiding in plain sight: as AI-generated content becomes free and ubiquitous, human-verified and human-created content becomes genuinely valuable. I’m not talking about some nostalgic preference; I’m talking about cold, hard economic incentive. By 2026, we’ll see the emergence of “human-certified” badges and verification systems that actually mean something to consumers.</p><p>Imagine apps that proudly display “Human customer service — real people respond within minutes” as a premium feature. Or content platforms that charge more for articles guaranteed to be written by actual journalists rather than assembled by language models. Or products that come with video proof-of-creation showing artisans at work. It’s the digital equivalent of the craft beer movement, and it’s going to be surprisingly lucrative.</p><p>The smart business play is to get ahead of this curve. Companies that can prove human involvement in key parts of their offering — whether that’s design, support, or creation — will be able to command premium pricing. Meanwhile, those that go full automation will compete in an increasingly commoditized race to the bottom. The irony is beautiful: AI makes things cheap, which makes human attention expensive. Economics 101 in action.</p><h3>Prediction Five: Business Models Fracture Along the AI Divide</h3><p>By the end of 2026, we’ll see a clear split in business strategy, and both paths can succeed. On one side, you’ll have ultra-efficient AI-first companies that operate with skeleton crews and microscopic margins, essentially arbitraging the cost collapse that AI enables. Think fully automated apps with millions of users and twelve employees. On the other side, you’ll have deliberately human-centric businesses that charge more but offer something AI can’t replicate: genuine human insight, creativity, and care.</p><p>The middle ground, however, becomes treacherous. Companies that are neither efficient enough to compete on price nor distinctive enough to command a premium will struggle mightily. We’ll see consolidation, acquisitions, and a lot of pivoting as businesses realize they need to pick a lane. (Trying to be “sort of automated but also sort of personal” is like being sort of pregnant; it doesn’t really work.)</p><p>For consumer app developers, this means making a clear strategic choice: are you building tools that leverage AI to be incredibly efficient and cheap, or are you building experiences that emphasize human craft and connection? Both can win, but you can’t win by being confused about which game you’re playing. The companies that thrive will be the ones that lean into their choice rather than hedging.</p><h3>Prediction Six: Regulation Finally Shows Up (Late, Confused, But Inevitable)</h3><p>If 2024 and 2025 were the years of regulatory hand-wringing and committee meetings, 2026 is when the actual rules start landing — and it’s going to be messy, contradictory, and surprisingly important all at once. The reality is that governments move at geological speeds while technology moves at hummingbird pace, so by the time regulations arrive, they’re already somewhat outdated. But that doesn’t mean they won’t matter; it just means businesses need to get good at reading between the lines.</p><p>By mid-2026, expect a patchwork quilt of regulations that vary wildly by region. Europe will have implemented some form of AI accountability framework (because of course they will — they gave us GDPR and they’re not about to let AI run wild). California will have passed something that kind of looks like Europe’s rules but with more Hollywood influence. China will have strict controls that serve their particular interests. And various other jurisdictions will be somewhere between “we’re studying the issue” and “we copied someone else’s homework but changed a few words.”</p><p>For businesses, this creates what I call the “compliance maze” challenge. You can’t just build one version of your AI-powered app anymore; you need to think about different feature sets, different disclosure requirements, different data handling practices depending on where your users are. It’s annoying, it’s expensive, and it’s coming whether we like it or not. The companies that start planning for this reality now will have a massive head start over those still pretending they can ignore it.</p><p>But here’s the counterintuitive insight: regulation might actually help the AI industry mature. Much like how food safety laws didn’t kill restaurants but rather made dining out trustworthy, AI regulations could build the consumer confidence needed for broader adoption. When users know there are guardrails, they’re more willing to embrace the technology. When businesses know the rules of the game, they can compete on merit rather than on who can move fastest through ethical gray zones.</p><p>The smart play for 2026 isn’t to fight regulation or ignore it; it’s to get ahead of it. Build transparency into your AI systems from the start. Document your training data and decision-making processes. Create internal ethics guidelines that are stricter than what regulators will eventually require. Because when the rules do arrive, companies that have been operating responsibly all along will find compliance is simply documenting what they’re already doing, while their competitors scramble to retrofit accountability into systems built on cutting corners. (Plus, “We’ve been self-regulating responsibly since before it was legally required” is an excellent marketing message to users who are increasingly skeptical of tech companies.)</p><p>The other trend worth watching: liability questions are going to heat up dramatically. When an AI system makes a consequential mistake — recommends a bad medical decision, generates defamatory content, or discriminates in hiring — who’s responsible? The company that deployed it? The team that trained it? The users who prompted it? By late 2026, we’ll probably have the first major court cases establishing precedent, and those rulings will send shockwaves through the industry. Insurance for AI liability will become a real thing, and it won’t be cheap.</p><p>The businesses that navigate this successfully will be those that treat regulation not as an obstacle but as a forcing function for building better, more trustworthy products. Because ultimately, the companies users trust will be the ones that win, regulation or not.</p><h3>The Silver Lining in the Circuit Board</h3><p>Despite all the disruption and uncertainty, I’m genuinely optimistic about where this is headed. Yes, 2026 will be messy. Yes, there will be job displacement, ethical dilemmas, and more than a few catastrophic AI fails that become cautionary tales. But mess is where innovation happens. Comfort breeds complacency; chaos breeds creativity.</p><p>The businesses that will emerge stronger are those that view AI as a tool for augmentation rather than replacement, that understand the value of human judgment even as they embrace machine efficiency, and that recognize consumers are more sophisticated than the hype cycle gives them credit for. People can tell the difference between AI slop and genuine quality. They can feel when technology serves them versus when they’re serving the technology.</p><p>So here’s my real prediction for 2026: it’s the year we stop asking “What can AI do?” and start asking “What should AI do?” It’s the year businesses move from experimentation to integration, from novelty to utility, from hype to reality. And yeah, it’s going to be uncomfortable at times. Growing up always is.</p><p>But discomfort means we’re paying attention. It means we haven’t surrendered to inevitability. And that, ultimately, is what makes me hopeful. The story of AI and business in 2026 isn’t written yet. We’re still holding the pen. So let’s write something worth reading.</p><p><a href="https://www.linkedin.com/in/altafrehmani/">Altaf Rehmani</a> is a Technology Innovator, Applied Generative AI champion, Agile Evangelist and a champion of applying technology to enable business growth. He lives in Hong Kong and can be reached via <a href="/service/mailto:altafr@gmail.com">email</a> or <a href="https://twitter.com/altafr">twitter</a>. Please leave your feedback and a clap if you have liked this article.</p><p>Learn all about Generativ e AI in <a href="https://www.amazon.com/dp/B0CVHB2S64">my book: on Generative AI</a></p><p><a href="https://deft-innovator-9609.ck.page/techscapebook">Check out my free eBook “TECHSCAPE”</a> discussing a variety of topics helping you navigate the exciting Technology space.</p><p>Other articles which may be of interest:</p><p><a href="https://medium.com/@altafr/the-future-of-work-going-in-2023-762c51a5a0c6">Future of work going into 2023.</a></p><p><a href="https://medium.com/@altafr/managing-global-teams-9e94026d5881">Managing Global Teams</a></p><p><a href="https://medium.com/@altafr/managing-high-performing-teams-76f6f3596b31">Managing High Performing teams</a></p><p><a href="https://medium.com/not-so-technical/continuous-team-improvements-through-agile-retrospectives-5a77fe6e6114">Continuous team improvements using Agile Retrospectives</a></p><p><a href="https://medium.com/not-so-technical/common-mistakes-in-agile-implementations-21a6d999bc91">Common Mistakes in Agile Implementations</a></p><p><a href="https://medium.com/not-so-technical/applying-ai-in-the-context-of-ecommerce-3fc3c365f8c9">Applying AI in The context of eCommerce</a></p><p><a href="https://medium.com/not-so-technical/chatbots-a-crash-course-for-newbies-c11a32561394">Chatbots — A Crash Course for Newbies</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c349cbb33d46" width="1" height="1" alt=""><hr><p><a href="https://medium.com/not-so-technical/generative-ai-business-what-2026-will-actually-look-like-c349cbb33d46">Generative AI &amp; Business: What 2026 Will Actually Look Like</a> was originally published in <a href="https://medium.com/not-so-technical">Not So Technical</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Deep Dive into GPT-40: OpenAI’s New AI Model]]></title>
            <link>https://medium.com/not-so-technical/a-deep-dive-into-gpt-40-openais-new-ai-model-a09a2db194f2?source=rss-530273a9d700------2</link>
            <guid isPermaLink="false">https://medium.com/p/a09a2db194f2</guid>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[gpt-4o]]></category>
            <dc:creator><![CDATA[Altaf Rehmani]]></dc:creator>
            <pubDate>Wed, 15 May 2024 08:28:01 GMT</pubDate>
            <atom:updated>2024-05-15T08:28:01.057Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OXsZzSDNKh0924f0dyZhLw.png" /><figcaption>GPT-4o release from OpenAI : source OpenAI</figcaption></figure><p>OpenAI has recently unveiled GPT-4o, their new flagship AI model. This model is a significant upgrade from its predecessor, GPT-4, and is presented as a free-to-use model that surpasses the capabilities of GPT-4 in many areas.</p><p><strong>GPT-4o Capabilities</strong></p><p>GPT-4o provides GP4-level intelligence and has been improved in terms of text, vision, and audio capabilities. Notably, it can perform real-time translations and analyze emotions based on facial expressions. It also showcases its ability to create creative content such as telling a bedtime story in a robotic voice, ending the story in a singing voice, and even speculating on what whales would say if they could talk.</p><p>The model is not just limited to text and audio; it also offers vision features. Users can upload and discuss content with ChatGPT, facilitating a more interactive and collaborative experience.</p><p><strong>User Experience</strong></p><p>A priority with GPT-4o is ensuring a seamless integration into user workflows, with a focus on natural and easy interaction. The model supports interaction across voice, text, and vision, making it more accessible and user-friendly.</p><p>To facilitate easy access, OpenAI has also introduced a desktop app for CH gbt. This app aims to shift the focus away from the UI and emphasize collaboration with GPT-4o.</p><p><strong>Comparison with GPT-4</strong></p><p>In a head-to-head comparison with GPT-4, GPT-4o demonstrated superior performance in tasks such as text summarization, text generation, multimodal understanding, and image generation. It even managed to create a functional snake game, a feat that GPT-4 struggled with.</p><p>GPT-4o also offers features that were previously exclusive to the paid version of GPT-4, such as data analysis, file uploading, and web browsing. It also has more generous usage limits, particularly for plus and team users.</p><p>Despite these advantages, some paid users might hesitate to switch to GPT-4o due to the lack of significant differences and limitations.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fjay7kise3PI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Djay7kise3PI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fjay7kise3PI%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/dcc846787dd59c7fbc0c72a3c4def347/href">https://medium.com/media/dcc846787dd59c7fbc0c72a3c4def347/href</a></iframe><p><strong>Use Cases for Businesses and Professionals Worldwide</strong></p><p>GPT-4o holds immense potential for businesses and professionals worldwide. It’s high-performance capabilities in text, vision, and audio processing can be leveraged in various sectors. For instance, its real-time translation feature can benefit global businesses by facilitating seamless communication between teams spread across different geographical locations.</p><p>The model’s advanced data analysis, file uploading, and web browsing features can streamline workflows in sectors like finance, marketing, and research by providing efficient access to information and easy data handling.</p><p>Moreover, GPT-4o’s ability to support interaction across voice, text, and vision can revolutionize customer service by providing a more interactive, responsive, and personalized user experience.</p><p><strong>The Future of GPT-4o</strong></p><p>OpenAI has plans to launch an API for GPT-4o, which will allow developers to build their applications with natural conversation features. With over 100 million users having utilized the previous version of ChatGPT, GPT-4o holds promising potential for widespread use and integration into various applications.</p><p>GPT-4o’s unveiling was met with enthusiasm and excitement, with its capabilities showcased during a live event. As OpenAI continues to advance its AI models, the future looks bright for the evolution of AI-powered conversation and collaboration tools.</p><p><strong>Challenges and Opportunities</strong></p><p>While GPT-4o presents many advancements, it’s important to consider the potential challenges and opportunities with its deployment. For instance, the model’s reliance on large amounts of data for training could raise questions around data privacy and security. Businesses and professionals need to ensure that they’re using GPT-4o in a manner that complies with relevant data protection regulations.</p><p>On the other hand, GPT-4o provides an opportunity to democratize access to AI technology. Its free availability allows a larger audience to benefit from AI capabilities, which could lead to innovative applications across different sectors.</p><p><strong>Impact on Education and Research</strong></p><p>GPT-4o’s advanced text summarization and generation features could have a significant impact on education and research. Students and researchers can leverage these features to quickly review large volumes of literature, generate summaries of key points, and even draft initial versions of reports or papers.</p><p>GPT-4o could also be used as a teaching tool, providing interactive and engaging learning experiences. For example, its ability to generate creative content like stories or hypothetical scenarios can be used to stimulate discussion and critical thinking in classrooms.</p><p><strong>Conclusion</strong></p><p>GPT-4o represents a significant milestone in the development of AI models. Its enhanced capabilities, combined with its free availability, open up a world of possibilities for businesses, professionals, educators, and researchers alike. As we continue to explore and understand its potential, GPT-4o is set to shape the future of AI-powered conversation and collaboration tools.</p><p><a href="https://www.linkedin.com/in/altafrehmani/">Altaf Rehmani</a> is a Technology Innovator, helped various businesses with Digital transformation projects, Agile Evangelist and a champion of applying technology to enable business growth. He lives in Hong Kong and can be reached via <a href="/service/mailto:altafr@gmail.com">email</a> or <a href="https://twitter.com/altafr">twitter</a>. Please leave your feedback and a clap if you have liked this article.</p><p>Learn all about Generativ e AI in <a href="https://www.amazon.com/dp/B0CVHB2S64">my book: on Generative AI</a></p><p>Join the <a href="https://www.skool.com/generativeai">free Generative AI community</a>.</p><p><a href="https://deft-innovator-9609.ck.page/techscapebook">Check out my free eBook “TECHSCAPE”</a> discussing a variety of topics helping you navigate the exciting Technology space.</p><p>Other articles which may be of interest:</p><p><a href="https://medium.com/@altafr/the-future-of-work-going-in-2023-762c51a5a0c6">Future of work going into 2023.</a></p><p><a href="https://medium.com/@altafr/managing-global-teams-9e94026d5881">Managing Global Teams</a></p><p><a href="https://medium.com/@altafr/managing-high-performing-teams-76f6f3596b31">Managing High Performing teams</a></p><p><a href="https://medium.com/not-so-technical/continuous-team-improvements-through-agile-retrospectives-5a77fe6e6114">Continuous team improvements using Agile Retrospectives</a></p><p><a href="https://medium.com/not-so-technical/common-mistakes-in-agile-implementations-21a6d999bc91">Common Mistakes in Agile Implementations</a></p><p><a href="https://medium.com/not-so-technical/applying-ai-in-the-context-of-ecommerce-3fc3c365f8c9">Applying AI in The context of eCommerce</a></p><p><a href="https://medium.com/not-so-technical/chatbots-a-crash-course-for-newbies-c11a32561394">Chatbots — A Crash Course for Newbies</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a09a2db194f2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/not-so-technical/a-deep-dive-into-gpt-40-openais-new-ai-model-a09a2db194f2">A Deep Dive into GPT-40: OpenAI’s New AI Model</a> was originally published in <a href="https://medium.com/not-so-technical">Not So Technical</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to learn Generative AI ?]]></title>
            <link>https://medium.com/@altafr/generative-ai-for-everyone-04d869570fb8?source=rss-530273a9d700------2</link>
            <guid isPermaLink="false">https://medium.com/p/04d869570fb8</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[books]]></category>
            <dc:creator><![CDATA[Altaf Rehmani]]></dc:creator>
            <pubDate>Tue, 09 Apr 2024 03:12:14 GMT</pubDate>
            <atom:updated>2024-04-13T00:35:46.376Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Whats the best way to learn Generative AI? where should i begin? Do i need to be a machine learning expert? A Data scientist? But i am not technical?</blockquote><blockquote>I am technical but not much involved in AI work in the last few years, but i should know this stuff.</blockquote><blockquote>I am a senior person in legal , compliance , risk and now want to to understand Generative AI, its risk and limitations. how can i understand all of this in a handy guide which can be read in my own time.</blockquote><p>Generative AI for everyone Book</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1MAQYP26pGWhP_WDXA9nQg.png" /><figcaption><a href="https://www.amazon.com/dp/B0CVHB2S64">https://www.amazon.com/dp/B0CVHB2S64</a></figcaption></figure><p>The fast moving pace of information, tools and research coming out of Generative AI also meant information overload and confusion for new beginners. The inspiration for the book came from the fact that there was an oporutnity to curate fundamental information on this new subset of AI in a logical manner attracting non-tehnical people to this field and making them aware of the possibilities.</p><p>The goals for compiling this book were as follows:</p><ul><li>To provide a reference book for any professional to understand what generative AI is, the basics of machine learning, and potential use cases so that they can apply it in their profession or business.</li><li>To aggregate fundamental information on this vast and fast-moving topic in one place in an easy-to-understand flow and format so that readers can start to understand and gauge the impact in their areas and the future of work.</li><li>To cover the basics and ways to address the challenges of this new technology, with the intention of providing value that will be relevant for readers to begin applying in their field, irrespective of their profession.</li></ul><p><strong>Why should you read this book?</strong></p><p>Get ready to become a generative AI genius without all the boring tech blabber.</p><p>So, you’ve heard about this generative AI thing, and you’re like, “Woah, that sounds super cool, but also kinda confusing?” Well, buckle up, my friend, because this guide is going to make you the generative AI master of the universe! (Or at least your friend group.)</p><p>We’re going to start with the basics in Part I: Fundamentals of Generative AI. Don’t worry; we won’t get too technical (unless you’re into that sort of thing). Chapter 1 is like the “Generative AI for Dummies” section, where we’ll break down what the heck it even is and how it works. Think of it as the first step in your journey to becoming a chatbot whisperer.</p><p>Next up, we’ve got Chapters 2 and 3, where we’ll dive into the nitty-gritty of machine learning and large language models. Now, I know what you’re thinking, “Ugh, math? No thanks!” But trust me, it’s way cooler than it sounds. It’s like teaching a computer to talk like a human, but without the awkward teenage years.</p><p>And then, in Chapter 4, we’ll get up close and personal with the one and only ChatGPT.</p><p>How do Language Models differ from Image models and how can you become the Picasso of your world? — Find out in Chapter 5.</p><p>Once you’ve got the basics down, we’ll move on to Part II: Mastering Prompt Engineering. This is where you’ll learn the secret sauce to making these AI models do your bidding. It’s like learning a new language, but instead of conjugating verbs, you’ll be crafting prompts that’ll make the AI say, “Sure, I can totally write that novel for you!”</p><p>And if you’re feeling extra techy, Part III: Open Source Models and Programming Framework is where the real fun begins. You’ll get to play around with open-source models (because who doesn’t love free stuff?)</p><p>But wait, there’s more! Part IV: Envisioning the Future with Agents is where things get really trippy. We’re talking about AI systems that can make decisions and take actions on their own. It’s like having a personal robot butler, but without the awkward small talk.</p><p>For all you business buffs out there, Part V: Generative AI in Business is a must-read. You’ll learn how to use this fancy tech to make your company loads of money (or at least impress your boss).</p><p>And because we can’t have nice things without being responsible, Part VI: Responsible AI is here to keep us all in check. It’s like the generative AI version of a parent reminding you to eat your veggies and be nice to others.</p><p>Finally, Parts VII and VIII will give you the inside scoop on the companies and future trends in the generative AI world. It’s like being a part of the coolest club on the block, but without the secret handshake (or maybe there is one, who knows?).</p><p>So, there you have it! A Comprehensive guide to becoming a generative AI master. Just don’t forget to invite me to your AI-themed birthday party when you’re rich and famous, okay?</p><p><a href="https://www.amazon.com/dp/B0CVHB2S64">Generative AI for everyone: Understanding the essentials and applications of this breakthrough technology.</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tSW1ZEdFb73gg9bnaL6FJQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RCyGxcCVkS7A78CE9oQjmw.png" /></figure><p><a href="https://www.linkedin.com/in/altafrehmani/">Altaf Rehmani</a> is a Technology Innovator, helped various businesses with Digital transformation projects, Agile Evangelist and a champion of applying technology to enable business growth. He lives in Hong Kong and can be reached via <a href="/service/mailto:altafr@gmail.com">email</a> or <a href="https://twitter.com/altafr">twitter</a>. Please leave your feedback and a clap if you have liked this article.</p><p>Learn all about Generativ e AI in <a href="https://www.amazon.com/dp/B0CVHB2S64">my book: on Generative AI</a></p><p>Join the <a href="https://www.skool.com/generativeai">free Generative AI community</a>.</p><p><a href="https://deft-innovator-9609.ck.page/techscapebook">Check out my free eBook “TECHSCAPE”</a> discussing a variety of topics helping you navigate the exciting Technology space.</p><p>Other articles which may be of interest:</p><p><a href="https://medium.com/@altafr/the-future-of-work-going-in-2023-762c51a5a0c6">Future of work going into 2023.</a></p><p><a href="https://medium.com/@altafr/managing-global-teams-9e94026d5881">Managing Global Teams</a></p><p><a href="https://medium.com/@altafr/managing-high-performing-teams-76f6f3596b31">Managing High Performing teams</a></p><p><a href="https://medium.com/not-so-technical/continuous-team-improvements-through-agile-retrospectives-5a77fe6e6114">Continuous team improvements using Agile Retrospectives</a></p><p><a href="https://medium.com/not-so-technical/common-mistakes-in-agile-implementations-21a6d999bc91">Common Mistakes in Agile Implementations</a></p><p><a href="https://medium.com/not-so-technical/applying-ai-in-the-context-of-ecommerce-3fc3c365f8c9">Applying AI in The context of eCommerce</a></p><p><a href="https://medium.com/not-so-technical/chatbots-a-crash-course-for-newbies-c11a32561394">Chatbots — A Crash Course for Newbies</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=04d869570fb8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Rise of AI Agents: Unlocking New Frontiers in Large Language Models]]></title>
            <link>https://medium.com/not-so-technical/the-rise-of-ai-agents-unlocking-new-frontiers-in-large-language-models-d61f36eb1d07?source=rss-530273a9d700------2</link>
            <guid isPermaLink="false">https://medium.com/p/d61f36eb1d07</guid>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[generative-ai]]></category>
            <dc:creator><![CDATA[Altaf Rehmani]]></dc:creator>
            <pubDate>Mon, 25 Mar 2024 13:48:34 GMT</pubDate>
            <atom:updated>2024-03-25T13:48:34.747Z</atom:updated>
            <content:encoded><![CDATA[<p>Going Beyond LLMs</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TfA-BFpdSXz-qa6z" /><figcaption>Photo by <a href="https://unsplash.com/@growtika?utm_source=medium&amp;utm_medium=referral">Growtika</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The field of artificial intelligence (AI) is undergoing a paradigm shift, with AI agents poised to revolutionize the landscape of large language models (LLMs). As companies like Google, OpenAI, and Anthropic continue to push the boundaries of natural language processing, a new wave of innovation is on the horizon — the era of AI agents.</p><p>At the heart of this transformation lies the concept of agency, a term derived from the Latin verb agere, meaning “to do” or “to act.” AI agents are systems designed to make decisions and take actions towards a specific goal, without requiring step-by-step instructions. This marks a significant departure from traditional chatbots and language models, which are limited to passive observation and response.</p><p><strong>What are agents and how do they work?</strong></p><p>AI agents are systems designed to make decisions and take actions towards a specific goal without needing step-by-step instructions. They are powered by large language models (LLMs) and have the ability to act on their own, unlike traditional chatbots which can only respond to prompts.</p><p>Here’s how AI agents work based on the information given:</p><ol><li>Agents are defined with specific roles (e.g. research assistant, writer) and goals (e.g. look up latest AI advancements, summarize findings).</li><li>They are given tools and access to resources like internet search to help them achieve their goals. For example, a research agent may be given a Search Tool to scour the web.</li><li>Agents are powered by large language models like GPT-4 which serve as their “brain” to understand queries, reason, and generate outputs.</li><li>Multiple agents can collaborate, with one agent delegating tasks to another. For instance, a writer agent can instruct a research agent to find specific information needed.</li><li>popular agentic frameworks like Autogen , CrewAI , Langchain and Youai.ai allows connecting the different components — the LLM, tools, tasks and instructions for the agent roles.</li><li>Agents can iteratively refine their understanding, search for more information, and ultimately produce the desired output like a report or article summarizing their findings.</li></ol><p>The key difference from traditional LLMs is that agents can break down higher-level goals into sub-tasks, make plans, leverage tools, and iteratively work towards solutions, exhibiting agency and autonomous behavior. This allows tackling more open-ended queries that require multi-step reasoning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*L6auWK4yQIzzqORd" /><figcaption>Basic elements of a AI Agent</figcaption></figure><p>The advent of AI agents promises to unlock a new frontier of productivity and efficiency. By harnessing the power of LLMs, these intelligent assistants can automate a wide range of tasks, from conducting comprehensive research and analysis to generating concise reports and articles. As the technology continues to evolve, AI agents are expected to play a pivotal role in various industries, from customer service and software engineering to scientific research and content creation.</p><p>One of the key drivers behind the rise of AI agents is the relentless pursuit of more advanced LLMs. As companies like OpenAI and Anthropic continue to develop cutting-edge models like GPT-5 and Claude, the capabilities of AI agents will continue to expand. These next-generation LLMs are expected to possess enhanced reasoning abilities, long-term memory, and multimodal capabilities, enabling agents to tackle increasingly complex tasks with unprecedented efficiency.</p><p>However, the true potential of AI agents lies not only in their individual capabilities but also in their ability to collaborate and leverage each other’s strengths. By combining multiple agents with specialized roles and goals, organizations can create powerful teams capable of tackling intricate challenges that would be insurmountable for a single entity.</p><p>Despite the immense potential of AI agents, their adoption and development are not without challenges. Concerns surrounding privacy, security, and ethical considerations must be addressed to ensure responsible deployment and integration. Additionally, the need for robust frameworks, intuitive user interfaces, and accessible tools will be crucial in democratizing the development of AI agents, allowing individuals and organizations of all sizes to harness their power.</p><p>As the world stands on the precipice of an AI agent revolution, it is clear that those who embrace this technology early will gain a significant competitive advantage. Just as the internet and social media revolutions reshaped entire industries, AI agents are poised to disrupt the status quo, ushering in a new era of efficiency, productivity, and innovation.</p><p>In the words of Sam Altman, the co-founder of OpenAI, “I think it’s very obvious to a lot of people that AGI will take the form of some kind of an AI agent.” The future belongs to those who recognize the transformative potential of AI agents and position themselves at the forefront of this technological revolution.</p><p>What interesting use cases would you want to solve using Agents? Leave your comments below.</p><p>Read more about agents in my book — <a href="https://www.amazon.com/dp/B0CVHB2S64">Generative AI for everyone</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*CUKL-oEPG98rRjREUUkMzw.png" /></figure><p><a href="https://www.linkedin.com/in/altafrehmani/">Altaf Rehmani</a> is a Technology Innovator, helped various businesses with Digital transformation projects, Agile Evangelist and a champion of applying technology to enable business growth. He lives in Hong Kong and can be reached via <a href="/service/mailto:altafr@gmail.com">email</a> or <a href="https://twitter.com/altafr">twitter</a>. Please leave your feedback and a clap if you have liked this article.</p><p>Learn all about Generativ e AI in <a href="https://www.amazon.com/dp/B0CVHB2S64">my book: on Generative AI</a></p><p>Join the <a href="https://www.skool.com/generativeai">free Generative AI community</a>.</p><p><a href="https://deft-innovator-9609.ck.page/techscapebook">Check out my free eBook “TECHSCAPE”</a> discussing a variety of topics helping you navigate the exciting Technology space.</p><p>Other articles which may be of interest:</p><p><a href="https://medium.com/@altafr/the-future-of-work-going-in-2023-762c51a5a0c6">Future of work going into 2023.</a></p><p><a href="https://medium.com/@altafr/managing-global-teams-9e94026d5881">Managing Global Teams</a></p><p><a href="https://medium.com/@altafr/managing-high-performing-teams-76f6f3596b31">Managing High Performing teams</a></p><p><a href="https://medium.com/not-so-technical/continuous-team-improvements-through-agile-retrospectives-5a77fe6e6114">Continuous team improvements using Agile Retrospectives</a></p><p><a href="https://medium.com/not-so-technical/common-mistakes-in-agile-implementations-21a6d999bc91">Common Mistakes in Agile Implementations</a></p><p><a href="https://medium.com/not-so-technical/applying-ai-in-the-context-of-ecommerce-3fc3c365f8c9">Applying AI in The context of eCommerce</a></p><p><a href="https://medium.com/not-so-technical/chatbots-a-crash-course-for-newbies-c11a32561394">Chatbots — A Crash Course for Newbies</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d61f36eb1d07" width="1" height="1" alt=""><hr><p><a href="https://medium.com/not-so-technical/the-rise-of-ai-agents-unlocking-new-frontiers-in-large-language-models-d61f36eb1d07">The Rise of AI Agents: Unlocking New Frontiers in Large Language Models</a> was originally published in <a href="https://medium.com/not-so-technical">Not So Technical</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Unraveling the Debate: Generative AI — Misunderstood or Misunderstanding?]]></title>
            <link>https://medium.com/not-so-technical/unraveling-the-debate-generative-ai-misunderstood-or-misunderstanding-401c88b188fb?source=rss-530273a9d700------2</link>
            <guid isPermaLink="false">https://medium.com/p/401c88b188fb</guid>
            <category><![CDATA[generative-ai-solution]]></category>
            <dc:creator><![CDATA[Altaf Rehmani]]></dc:creator>
            <pubDate>Wed, 20 Mar 2024 14:25:46 GMT</pubDate>
            <atom:updated>2024-03-23T05:07:14.417Z</atom:updated>
            <content:encoded><![CDATA[<h3>Unraveling the Debate: Generative AI — Misunderstood or Misunderstanding?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OKc9PCRqx8cG8nGE" /><figcaption>Photo by <a href="https://unsplash.com/@omilaev?utm_source=medium&amp;utm_medium=referral">Igor Omilaev</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Generative models have emerged as powerful tools, captivating the imagination of researchers, developers, and the general public alike. But amidst the buzz, a fundamental question looms large: do generative AI systems truly understand the content they generate, or are they merely skilled mimics, lacking genuine comprehension? This debate pits two contrasting viewpoints against each other, each offering compelling arguments supported by examples and real-world applications.</p><p><strong>Viewpoint 1: Generative AI’s Practical Limitations</strong></p><p>Proponents of this viewpoint argue that generative AI, despite its impressive capabilities, falls short when it comes to true comprehension. They contend that while these models excel at producing coherent text, images, and even music, their creations lack genuine understanding of the underlying concepts. In essence, generative AI is perceived as a sophisticated illusion, devoid of true cognitive abilities.</p><p>One of the primary arguments supporting this perspective revolves around the nature of generative AI training. These models learn from vast datasets, identifying statistical patterns and correlations without comprehending the semantic meaning behind the data. Take, for example, a language model trained on a corpus of medical literature. While it may generate plausible-sounding diagnoses and treatment plans, it lacks the ability to truly understand the intricacies of human physiology or the nuances of patient care.</p><p>Moreover, critics highlight instances where generative AI produces nonsensical or even harmful output, exposing its limitations. For instance, chatbots trained on internet dialogues may inadvertently generate inappropriate or offensive responses, reflecting a lack of genuine comprehension and ethical awareness.</p><p><strong>Viewpoint 2: Generative AI’s Power of Understanding</strong></p><p>In contrast, advocates of this viewpoint argue that generative AI possesses a remarkable capacity for reasoning and understanding, albeit in a different form than human cognition. They assert that while these models may not grasp concepts in the same intuitive manner as humans, they demonstrate a level of comprehension through their ability to generate contextually relevant content and adapt to diverse tasks.</p><p>One compelling example of generative AI’s understanding lies in the field of natural language processing (NLP). Language models like GPT (Generative Pre-trained Transformer) have demonstrated proficiency in tasks such as text summarization, translation, and question-answering, indicating a nuanced understanding of language semantics and syntax. These models leverage contextual information to generate coherent responses, suggesting a degree of comprehension beyond mere statistical mimicry.</p><p>Furthermore, generative AI has proven instrumental in solving complex problems across various domains. For instance, in drug discovery, generative models can design novel molecular structures with desired properties, showcasing a deep understanding of chemical interactions and biological mechanisms. Similarly, in creative fields like art and music, AI-generated compositions exhibit a level of creativity and expression, hinting at an underlying comprehension of artistic concepts.</p><p><strong>Finding Common Ground: The Intersection of Perception and Reality</strong></p><p>While the debate surrounding generative AI’s understanding may persist, it’s essential to acknowledge the nuances inherent in this discussion. Generative models, like humans, operate within the constraints of their training data and algorithmic architecture. They excel in certain tasks while faltering in others, reflecting the complexity of AI systems.</p><p>Ultimately, the practical utility of generative AI lies not in its ability to perfectly emulate human understanding, but rather in its capacity to augment human capabilities and solve real-world problems. By leveraging the strengths of both human cognition and machine learning, we can harness the power of generative AI to drive innovation, creativity, and progress in diverse fields.</p><p>The debate over generative AI’s understanding underscores the multifaceted nature of artificial intelligence. While skeptics emphasize its limitations, proponents highlight its potential and practical applications. By embracing a nuanced perspective that acknowledges both the strengths and shortcomings of generative AI, we can navigate this evolving landscape with clarity and purpose, unlocking new possibilities for the future of technology and humanity.</p><p>What do you think? leave your comments below and let me know.</p><p><a href="https://www.linkedin.com/in/altafrehmani/">Altaf Rehmani</a> is a Technology Innovator, helped various businesses with Digital transformation projects, Agile Evangelist and a champion of applying technology to enable business growth. He lives in Hong Kong and can be reached via <a href="/service/mailto:altafr@gmail.com">email</a> or <a href="https://twitter.com/altafr">twitter</a>. Please leave your feedback and a clap if you have liked this article.</p><p>Learn all about Generativ e AI in <a href="https://www.amazon.com/dp/B0CVHB2S64">my book: on Generative AI</a></p><p>Join the <a href="https://www.skool.com/generativeai">free Generative AI community</a>.</p><p><a href="https://deft-innovator-9609.ck.page/techscapebook">Check out my free eBook “TECHSCAPE”</a> discussing a variety of topics helping you navigate the exciting Technology space.</p><p>Other articles which may be of interest:</p><p><a href="https://medium.com/@altafr/the-future-of-work-going-in-2023-762c51a5a0c6">Future of work going into 2023.</a></p><p><a href="https://medium.com/@altafr/managing-global-teams-9e94026d5881">Managing Global Teams</a></p><p><a href="https://medium.com/@altafr/managing-high-performing-teams-76f6f3596b31">Managing High Performing teams</a></p><p><a href="https://medium.com/not-so-technical/continuous-team-improvements-through-agile-retrospectives-5a77fe6e6114">Continuous team improvements using Agile Retrospectives</a></p><p><a href="https://medium.com/not-so-technical/common-mistakes-in-agile-implementations-21a6d999bc91">Common Mistakes in Agile Implementations</a></p><p><a href="https://medium.com/not-so-technical/applying-ai-in-the-context-of-ecommerce-3fc3c365f8c9">Applying AI in The context of eCommerce</a></p><p><a href="https://medium.com/not-so-technical/chatbots-a-crash-course-for-newbies-c11a32561394">Chatbots — A Crash Course for Newbies</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=401c88b188fb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/not-so-technical/unraveling-the-debate-generative-ai-misunderstood-or-misunderstanding-401c88b188fb">Unraveling the Debate: Generative AI — Misunderstood or Misunderstanding?</a> was originally published in <a href="https://medium.com/not-so-technical">Not So Technical</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Future of Work in the Age of AI]]></title>
            <link>https://medium.com/not-so-technical/the-future-of-work-in-the-age-of-ai-fa0c3ddee64e?source=rss-530273a9d700------2</link>
            <guid isPermaLink="false">https://medium.com/p/fa0c3ddee64e</guid>
            <dc:creator><![CDATA[Altaf Rehmani]]></dc:creator>
            <pubDate>Thu, 01 Feb 2024 01:38:18 GMT</pubDate>
            <atom:updated>2024-03-23T05:08:30.842Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8EDY_u_YDiKKpQY7" /><figcaption>Photo by <a href="https://unsplash.com/@avirichards?utm_source=medium&amp;utm_medium=referral">Avi Richards</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The introduction of large language models (LLMs) like ChatGPT promises to transform how we work, just as remote work and hybrid work models did in recent years. While the implications are not yet fully clear, research synthesized in a recent Microsoft report provides early evidence and insights into how LLMs may impact productivity, jobs, and organizations.</p><p>Early productivity experiments show substantial time-savings from using LLMs for common information work tasks like writing, though potential declines in output quality require vigilance. Less experienced workers seem to benefit more, aligning with hypotheses that LLMs can help disseminate institutional knowledge. However, productivity gains likely depend on efficient integration into workflows and appropriate reliance on outputs.</p><p>As LLMs permeate work, human skills may shift from search and content production to critical analysis and integration. People will likely spend more time planning, evaluating, editing and communicating around LLM-generated content. Soft skills like emotional intelligence will remain uniquely human.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*23nB-67jaFM2Dd1aywGHWQ.png" /><figcaption>Distribution of task-load between humans and computers/automation (Sheridan &amp; Verplank 1978)</figcaption></figure><p>Organizations introducing LLM systems face parallel adoption and perception issues as with past technologies. Effectiveness depends on contextual fit and support for changing job functions, not just the technology itself. Firms able to spur innovation creating net new valuable tasks and output stand to benefit most.</p><p>More broadly, the future of work remains a choice, not destiny. Proactively addressing disparities, fostering beneficial innovation, expanding access, and incentivizing data sharing are all within our control. Applying scientific principles of hypothesis testing and knowledge sharing can help leaders, scientists and citizens work together to shape how AI transforms work for the better.</p><p>The key insight is that while AI will significantly impact work, whether that is largely positive or negative depends greatly on human actions. Technological transformation may be inevitable, but an equitable, empowering future of work is achievable if we thoughtfully guide development and adoption. Organizations and governments must invest now in research, policies and initiatives that will allow humans and AI to productively co-evolve.</p><p>Here are few of insights and actions for us to adapt to the future of work with AI:</p><p>1. Develop critical thinking skills to effectively evaluate and integrate AI-generated content into workflows. Spend time planning, analyzing, summarizing, and editing.</p><p>2. Improve communication skills for clarifying needs to AI systems (prompting) and conveying AI outputs to stakeholders. Focus on contextualizing and grounding information.</p><p>3. Seek training on interacting with AI, such as constructing effective prompts and productively negotiating tradeoffs like speed vs quality. Build knowledge to create appropriate reliance.</p><p>4. Remain flexible and open to new augmentation opportunities. Innovation depends on utilizing technology-enabled gains to provide new value. Look for problems AI can help solve.</p><p>5. Advocate for policies, norms and tools that distribute benefits broadly, support effective adoption, address emerging issues proactively, and keep decisions human-centered. Voice priorities through corporate programs and civic participation.</p><p>The key is recognizing this transformation requires co-evolution. By developing complementary capabilities and helping shape development, humans can thrive alongside increasingly capable AI.</p><p>Source: <a href="https://www.microsoft.com/en-us/research/publication/microsoft-new-future-of-work-report-2023/">https://www.microsoft.com/en-us/research/publication/microsoft-new-future-of-work-report-2023/</a></p><p><a href="https://www.linkedin.com/in/altafrehmani/">Altaf Rehmani</a> is a Technology Innovator, helped various businesses with Digital transformation projects, Agile Evangelist and a champion of applying technology to enable business growth. He lives in Hong Kong and can be reached via <a href="/service/mailto:altafr@gmail.com">email</a> or <a href="https://twitter.com/altafr">twitter</a>. Please leave your feedback and a clap if you have liked this article.</p><p>Learn all about Generativ e AI in <a href="https://www.amazon.com/dp/B0CVHB2S64">my book on Generative AI</a></p><p>Join the <a href="https://www.skool.com/generativeai">free Generative AI community</a>.</p><p><a href="https://deft-innovator-9609.ck.page/techscapebook">Check out my free eBook “TECHSCAPE”</a> discussing a variety of topics helping you navigate the exciting Technology space.</p><p>Other articles which may be of interest:</p><p><a href="https://medium.com/@altafr/the-future-of-work-going-in-2023-762c51a5a0c6">Future of work going into 2023.</a></p><p><a href="https://medium.com/@altafr/managing-global-teams-9e94026d5881">Managing Global Teams</a></p><p><a href="https://medium.com/@altafr/managing-high-performing-teams-76f6f3596b31">Managing High Performing teams</a></p><p><a href="https://medium.com/not-so-technical/continuous-team-improvements-through-agile-retrospectives-5a77fe6e6114">Continuous team improvements using Agile Retrospectives</a></p><p><a href="https://medium.com/not-so-technical/common-mistakes-in-agile-implementations-21a6d999bc91">Common Mistakes in Agile Implementations</a></p><p><a href="https://medium.com/not-so-technical/applying-ai-in-the-context-of-ecommerce-3fc3c365f8c9">Applying AI in The context of eCommerce</a></p><p><a href="https://medium.com/not-so-technical/chatbots-a-crash-course-for-newbies-c11a32561394">Chatbots — A Crash Course for Newbies</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fa0c3ddee64e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/not-so-technical/the-future-of-work-in-the-age-of-ai-fa0c3ddee64e">The Future of Work in the Age of AI</a> was originally published in <a href="https://medium.com/not-so-technical">Not So Technical</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>