Shifting to Stage-Based Learning Models

Machine learning applications rarely stay static—they evolve. What begins as a simple baseline often grows into a multi-stage system shaped by scale, data complexity, and real-world constraints. In this tech blog, the engineering team at Shopify explains how their product classification system evolved as the platform scaled. The journey unfolds across three distinct stages, each with its own technical character. - Stage one focused on a traditional machine learning baseline: logistic regression with TF-IDF features built purely on product text. It was simple, interpretable, and efficient—a practical starting point. - Stage two introduced a multimodal approach, combining both text and image signals within a single model. This significantly improved accuracy, especially when product descriptions were incomplete or ambiguous. However, it remained largely a task-specific classifier trained on a fixed taxonomy. - Stage three marked a shift toward vision-language models. Instead of simply mapping inputs to predefined labels, these models learn richer semantic representations by aligning images and text in a shared embedding space. This enables deeper product understanding and better generalization as taxonomies evolve and new product types emerge. The key takeaway is that real-world machine learning systems mature in layers. You don’t jump straight to the most sophisticated model. Instead, you iterate—balancing accuracy with scalability—and design systems that can adapt as the business grows. #DataScience #MachineLearning #Classification #Evolution #Iteration #SnacksWeeklyonDataScience – – –  Check out the "Snacks Weekly on Data Science" podcast and subscribe, where I explain in more detail the concepts discussed in this and future posts:    -- Spotify: https://lnkd.in/gKgaMvbh   -- Apple Podcast: https://lnkd.in/gFYvfB8V    -- Youtube: https://lnkd.in/gcwPeBmR https://lnkd.in/gYuU_dNT

Over the past few days, I’ve been commenting on the patterns that keep showing up in L&D hiring. Each shows different symptoms, but the diagnosis is the same: the structure is missing. As far as I can see, many companies don’t know how learning roles build on each other, or what kind of experience it takes to move from delivery to design to strategy. So if you’re building or rebuilding your L&D function, here’s what a scalable structure can look like. Each stage works best when L&D partners closely with HR and business teams, ensuring learning isn’t a parallel track but a driver of capability, performance, and retention. 𝗦𝘁𝗮𝗴𝗲 𝟭: 𝗘𝗮𝗿𝗹𝘆 𝘀𝗲𝘁𝘂𝗽 (<𝟯𝟬𝟬 𝗲𝗺𝗽𝗹𝗼𝘆𝗲𝗲𝘀) • 𝘍𝘳𝘢𝘤𝘵𝘪𝘰𝘯𝘢𝘭 𝘓𝘦𝘢𝘳𝘯𝘪𝘯𝘨 𝘚𝘵𝘳𝘢𝘵𝘦𝘨𝘪𝘴𝘵 / 𝘊𝘓𝘖: Shapes the learning vision, defines priorities, and sets up governance & metrics. Works part-time or project-based. • 𝘓&𝘋 𝘗𝘢𝘳𝘵𝘯𝘦𝘳 / 𝘎𝘦𝘯𝘦𝘳𝘢𝘭𝘪𝘴𝘵 (5-8 𝘺𝘳𝘴): Handles everything from vendor management to learner feedback. • 𝘐𝘯𝘴𝘵𝘳𝘶𝘤𝘵𝘪𝘰𝘯𝘢𝘭 𝘋𝘦𝘴𝘪𝘨𝘯𝘦𝘳 (2-5 𝘺𝘳𝘴): Designs learning materials and programs. • Outsource eLearning development or facilitation till volume justifies in-house roles. 𝗦𝘁𝗮𝗴𝗲 𝟮: 𝗦𝘁��𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝘀𝗲𝘁𝘂𝗽 (𝟯𝟬𝟬–𝟭𝗸 𝗲𝗺𝗽𝗹𝗼𝘆𝗲𝗲𝘀) • 𝘓𝘦𝘢𝘳𝘯𝘪𝘯𝘨 𝘚𝘵𝘳𝘢𝘵𝘦𝘨𝘪𝘴𝘵 (𝘐𝘯𝘵𝘦𝘳𝘯𝘢𝘭 𝘰𝘳 𝘍𝘳𝘢𝘤𝘵𝘪𝘰𝘯𝘢𝘭): Evolves the roadmap, aligns learning with business priorities, and builds governance systems. • 𝘓&𝘋 𝘔𝘢𝘯𝘢𝘨𝘦𝘳 (8-12 𝘺𝘳𝘴): Runs execution and keeps everyone rowing in the same direction. • 𝘓𝘟 𝘋𝘦𝘴𝘪𝘨𝘯𝘦𝘳 (4-8 𝘺𝘳𝘴): Crafts learner journeys and engagement. • 𝘎𝘧𝘹 𝘋𝘦𝘴𝘪𝘨𝘯𝘦𝘳 (3-6 𝘺𝘳𝘴) & 𝘦𝘓𝘦𝘢𝘳𝘯𝘪𝘯𝘨 𝘋𝘦𝘷𝘦𝘭𝘰𝘱𝘦𝘳 (3-6 𝘺𝘳𝘴): Builds digital learning that actually works and looks good. • 𝘓𝘔𝘚 / 𝘖𝘱𝘴 𝘚𝘱𝘦𝘤𝘪𝘢𝘭𝘪𝘴𝘵 (3-6 𝘺𝘳𝘴): Tech, analytics, & governance, and measures how learning moves business metrics. • Add facilitators & SMEs as programs scale. 𝗦𝘁𝗮𝗴𝗲 𝟯: 𝗦𝗰𝗮𝗹𝗶𝗻𝗴 𝘀𝗲𝘁𝘂𝗽 (𝟭𝗸–𝟯𝗸 𝗲𝗺𝗽𝗹𝗼𝘆𝗲𝗲𝘀) • 𝘏𝘦𝘢𝘥 𝘰𝘧 𝘓&𝘋: Owns learning strategy end-to-end, partners with HR & leadership, builds capability across functions. • 𝘓𝘦𝘢𝘳𝘯𝘪𝘯𝘨 𝘚𝘵𝘳𝘢𝘵𝘦𝘨𝘪𝘴𝘵 / 𝘊𝘢𝘱𝘢𝘣𝘪𝘭𝘪𝘵𝘺 𝘓𝘦𝘢𝘥: Designs academies, frameworks, and skill pathways. • 𝘐𝘋𝘴, 𝘓𝘟 𝘋𝘦𝘴𝘪𝘨𝘯𝘦𝘳𝘴, 𝘋𝘦𝘷𝘦𝘭𝘰𝘱𝘦𝘳𝘴, & 𝘗𝘳𝘰𝘨𝘳𝘢𝘮 𝘔𝘢𝘯𝘢𝘨𝘦𝘳𝘴: Deliver at scale. • 𝘓𝘦𝘢𝘳𝘯𝘪𝘯𝘨 𝘖𝘱𝘴 𝘓𝘦𝘢𝘥: Keeps the engine running through tech, analytics, and governance. 𝗧𝗵𝗲 𝗽𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲 𝗶𝘀 𝘀𝗶𝗺𝗽𝗹𝗲: 𝗱𝗼𝗻’𝘁 𝗰𝗼𝗹𝗹𝗮𝗽𝘀𝗲 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆, 𝗱𝗲𝘀𝗶𝗴𝗻, 𝗮𝗻𝗱 𝗱𝗲𝗹𝗶𝘃𝗲𝗿𝘆 𝗶𝗻𝘁𝗼 𝗼𝗻𝗲 𝗼𝘃𝗲𝗿𝘄𝗼𝗿𝗸𝗲𝗱 𝗽𝗲𝗿𝘀𝗼𝗻. Start lean, but start clear. When roles are defined clearly and aligned with the business, learning stops being a “training function” and starts shaping culture and capability. Before you hire, build the structure. Feel free to reach out if you need support with this, I'm happy to help.

What the Research Says: Step-Based Learning - A Critical Lesson for GenAI Tutoring Today, I want to dig a little more into the detail from the paper I’m encouraging you to look at, because the Carnegie Mellon cognitive tutors research from the 1980s and 90s offers surprisingly relevant insights for today's generative AI tutoring challenges. Their findings about step-based learning are particularly crucial as we see students increasingly using ChatGPT and similar tools as "answer machines." The research revealed a striking pattern: when tutors tracked and responded to each step of student thinking, rather than just checking final answers, learning accelerated dramatically. Students achieved mastery in one-third the usual time. But why does this matter for today's GenAI tutoring? Consider the current landscape: students use ChatGPT to get quick answers, skipping the crucial learning process. The CMU research shows why this approach fundamentally misses the point. Their cognitive tutors succeeded precisely because they focused on the journey, not just the destination. The key findings remain remarkably relevant: 1.     Learning happens in the steps, not the answers. When tutors engaged with each stage of problem-solving, students developed deeper understanding. 2.     Multiple solution paths matter. Unlike today's GenAI systems that often present a single "best" answer, cognitive tutors recognised and supported different valid approaches. 3.     Immediate, contextual feedback at each step prevented the compounding of misconceptions - a crucial advantage over getting complete solutions from ChatGPT. This has profound implications for how we should be developing and using GenAI in education: ·       GenAI tutors need to be designed to engage with student thinking processes, not just generate answers ·       Systems should support multiple solution strategies rather than presenting single "optimal" solutions ·       Feedback should be immediate and step-specific, not just summative The challenges we see with current GenAI use in education - students bypassing learning processes to get quick answers - aren't new. They're the same challenges the CMU team tackled successfully decades ago. Their solution? Focus on the process, not just the result. As we develop the next generation of AI tutoring systems, this research suggests we need to shift focus from answer generation to process support. The technology has evolved dramatically, but the fundamental principles of effective learning remain the same. Professor Rose Luckin Institute of Education, UCL #AIED #AITutoring #LearningScience #EdTech #GenAI For more thoughts like this read the skinny here:https://lnkd.in/gTaNTRkb

AI may be entering its next major transition not because of bigger models, but because of a different way of learning. Google’s recent work on Nested Learning caught my attention this week. It’s one of the few developments that genuinely feels like a shift in the foundations of how AI systems might evolve. Instead of a single, static model frozen after pre-training, Nested Learning introduces multiple learning loops operating at different speeds fast, medium, and slow. Similar to how the human brain balances instant experience with long-term consolidation. The proof-of-concept model (HOPE) is only 1.3B parameters, yet it already beats both Transformers and modern RNN hybrids on long-context reasoning, retrieval, and resistance to forgetting. That alone should force us to rethink the assumption that “bigger = better.” What makes this interesting is not the performance benchmark. It’s the paradigm shift. • The model adapts during use • It forms memory beyond the context window • It updates its internal representations in real time • It avoids catastrophic forgetting through layered consolidation For those of us building real-world AI systems in production environments, this direction matters. We’ve all experienced the limitations of models that are powerful but fundamentally static. In complex enterprise setups from multi-channel communication systems to voice agents and autonomous workflows, true intelligence requires adaptive behavior, not just faster inference. Nested Learning hints at a future where AI agents: • Become more aligned with user context over time • Develop domain-specific knowledge without full retraining • Improve after each interaction • And maintain stability even as they evolve This moves AI closer to something we’ve been missing: systems that not only perform tasks but also improve their way of performing them. Whether this becomes the new standard or not is unclear. Research rarely moves in straight lines. But it’s a bold step in the right direction and a reminder that the next breakthroughs in AI may come from smarter learning, not simply larger models. The frontier of AI is shifting from scale to structure and that may change everything. #ArtificialIntelligence #MachineLearning #DeepLearning #AIResearch #AI #NeuralNetworks #AGI #LLM #SoftwareArchitecture

Every founder thinks they’re growing. Almost none of them are. I’ve never seen a startup pitch that actually understands growth. And I’ve seen a lot of pitch decks. You’ll show me the hockey stick. The S-curve. Maybe a TAM-SAM-SOM pyramid. But ask you to prove how you get from here to there? Crickets. Fluff. Buzzwords. Hope. This isn’t just a founder problem, it’s systemic. Startup advisors keep teaching bad models. * Investors chase vanity metrics. * Governments subsidize headcount like it’s proof of success. We’re failing. And we’re failing because nobody actually understands growth. Startups don’t grow by scaling. They grow by proving. By building a repeatable system that turns value into revenue before they run out of cash, time, or trust. But the advice out there? It’s built for traditional business, for marketers who know how, not for venture-backed innovation and not to guide most of you who don't even know what to do. So, throw out every framework you’ve ever been given: The Lean Canvas? Incomplete. Product-Market Fit? Happens after sustainability, not before. GTM Slide? Almost always nonsense. If it’s generic, it’s useless. “Early stage”? What does that even mean? We need a stage-driven model of growth rooted in revenue, risk, and reality. Not vibes. Not jargon. Not “traction” that can’t be traced. I spent some time this week with Daniel Perumal and Nicholas Alter unpacking a vision of growth in stages. Revenue-centric and charting growth in seven distinct stages, each with a specific objective and real graduation criteria, this hit what you all need: real growth, validation, and revenue. Stages: 1. Existential – No one cares yet. Define your hypothesis. 2. Discovery – Prove it works. Just once. 3. Adoption – Build consistency and start selling without brute force. 4. Sustainability – Real revenue. Predictable profit. Fundable and sellable. 5. Scalability – Chaos with control. Now we grow. 6. Saturation – You own the market. Stay relevant. 7. Event – The wild card: crisis, exit, disruption. Can you adapt? Forget the hockey stick. Think flywheel. This model isn’t just prettier, it’s functional. It changes how founders grow, how investors assess, and how cities could build economic policy.' 🎥 Because here's the uncomfortable truth we'll make into a movie: A) Startups that don’t understand growth should not be funded. B) Policies that don’t distinguish high-growth ventures from small businesses shouldn’t be written. C) Advisors who don’t know these stages should not be talking 😅 If you’re a founder chasing scale before proving sustainability, you’re just burning cash. 👉 Founders: What stage are you really in? 👉 Advisors: Can your curriculum define this path? 👉 Investors: Are you funding startups… or guessing? ThriveSide is where they're developing this out so follow that and read this:

Learning isn’t broken — it’s disconnected from execution. That’s why so many learning investments don’t translate into performance. This view closely aligns with recent research from The Josh Bersin Company which describes the shift from Static Training to Dynamic Enablement — where learning drives results only when it’s embedded in how work gets done. This framework shows what that shift looks like in execution terms. For years, learning has been judged by activity. Content produced. Programs delivered. Efficiency achieved. But work has changed. Decisions are faster. Expectations are higher. And execution gaps show up immediately. Because learning alone doesn’t create results. Execution does. And capability only creates value when it shows up inside real work. That’s the shift this framework captures 👇 Not a new learning model. An evolution in how organizations mature. 🩵 STAGE 1 – STATIC Learning exists. Work doesn’t change. Training is delivered, but decisions stay the same. 💛 STAGE 2 – REACTIVE Learning responds after problems appear. Skills are built, but not at the moment they’re needed. 💚 STAGE 3 – PROACTIVE Capability starts supporting execution. Learning aligns to real performance needs. 💙 STAGE 4 – INNOVATIVE Learning is embedded in how work happens. Capability, decisions, and outcomes are clearly connected. The real maturity shift isn’t better content. It’s a clear line of sight from: Capability → Decisions → Execution → Results That’s when learning stops being a support function and becomes a performance system. At StackFactor Inc., this is the work we help leaders do: Make capability visible. Tie it to execution. And design learning around the moments that actually matter. 🟡 Learning’s evolution isn’t about doing more. It’s about designing for impact. Which stage do you see most often today — and what’s keeping organizations from moving forward? Josh Bersin, Dave Ulrich, Dave Davies, Ashish Khanna, Andrew Nowak, Marilyn Ama M., Reyhaneh Khalilpour, Dominique Range, Zhanna Zhuravleva, Robert Jane, Muhammad Sajwani, Rishi A. Battja, Dr. Dinesh Kapoor, Marwa Agha, Deborah Kaminetzky, Alain Somvang, Lamia Kurdi-, Siddhi Beda, Antonio Calco' Labruzzo, Naotake Momiyama, Mohamed Atef, Marc-Oliver Guenther, Lester Samuel, Kimberley Inniss-Petersen, Ahsan Mahmud Khan, Dr. SYED MASROOR HUSSAIN SHAH, Daryna Kovenko, Balaji Chandrasekaran, Govert Doedijns, Imole Ashogbon, Nicolas BEHBAHANI, Syed G. Gaous, Hans Jacob Christensen — what’s your take? 👉 Follow Christina Jones or StackFactor Inc. for insights on building capability that drives execution. ✅ Subscribe at our LinkedIn Newsletter: L&D Edge: Upskilling for Impact – https://lnkd.in/eCdQgmN5 🔁 Reshare this if you believe learning should be accountable for execution — not just activity.

Louise lost a coaching client to Claude. And it stung. 15 certifications. AI still won. Her entire pitch was about who she was.  Not what she built. "I hold space."  "I am a coach." "I help people transform." The client wanted results, not sympathy. He asked one question: “What framework do you use?” She didn’t have an answer. AI exposed something most coaches didn’t want to admit: Being a coach isn't a business model. It's a job title. While coaches romanticised their identity, the market bought architecture. Documented frameworks.  Repeatable processes.  Measurable outcomes. The industry taught coaches to lead with personality. AI just accelerated what buyers were already thinking. Clients don’t want space holders. They want outcome builders. So what does “architecture” actually look like? It looks like this: A named framework. Clear stages. Defined inputs and outputs. Measurable milestones. Not “I help you grow.” But, this instead: Stage 1: Understand revenue constraint. Stage 2: Install demand. Stage 3: Build conversion path. Architecture is visible. It can be drawn on a whiteboard. It can be documented. It can be handed to someone else and still work. That’s the difference. Identity needs you in the room. Architecture works without you. If you want to survive this shift, build: • A repeatable methodology • A transformation map • A clear before/after state • A process someone else could follow If your value disappears when you log off, you built personality. If it compounds while you sleep, you built structure. AI replaces personality-led advice. It amplifies structured thinking. Choose which side of that line you stand on. You can architect transformation or perform it. Only one scales. Markets always reward what scales. AI didn’t kill coaching. It exposed the coaches who never built anything more than their identity. If this makes you uncomfortable, good. ➕ If this hits home, follow Graham Nicholls. I write for coaches, just like you, every day.

*Piaget's Stages of Development* Jean Piaget proposed that cognitive development occurs in four distinct stages, each marking significant shifts in how children think, reason, and understand the world. 1. Sensorimotor Stage (0–2 years) - *Key Concept:* Learning through sensory experiences and motor actions. Object permanence (understanding objects exist even when out of sight) develops. - *Teaching:* Children learn by interacting with their environment. Encouraging exploration and play fosters cognitive growth. 2. Preoperational Stage (2–7 years) - *Key Concept:* Symbolic thinking emerges (e.g., using words and images), but logic is egocentric and intuitive. Children struggle with conservation and perspective-taking. - *Teaching:* Use storytelling, imaginative play, and visual aids to engage children. Help them practice sharing perspectives through guided discussions. 3. Concrete Operational Stage (7–11 years) - *Key Concept:* Logical thinking develops, but is tied to concrete experiences. Children can classify objects, understand conservation, and perform basic problem-solving. - *Teaching:* Use hands-on materials (e.g., blocks, puzzles) to teach math and logic. Encourage categorization and organization activities. 4. Formal Operational Stage (12+ years) - *Key Concept:* Abstract and hypothetical thinking emerges. Adolescents can reason logically about complex problems and consider multiple perspectives. - *Teaching:* Engage students in debates, ethical dilemmas, and scientific inquiry. Encourage them to think critically and explore abstract concepts. *What Piaget Teaches Us:* - *Learning is Active:* Children construct knowledge through interactions with their environment. - *Development is Sequential:* Skills build upon prior stages, emphasizing the need for age-appropriate activities. - *Individual Differences:* Children may progress at different rates, highlighting the importance of personalized learning. Piaget’s theory underscores the value of play-based, inquiry-driven education and the need to tailor teaching methods to a child’s developmental stage.

What stage is your VR program in? The life cycle of most humans progresses from infancy to adolescence to mature adulthood. In infancy, humans experience rapid physical and cognitive development and exploring their environment. Adolescence brings significant physical changes like puberty, cognitive development, increased independence, and more complex thinking. Mature adulthood is characterized by full physical and cognitive maturity, stability in career and relationships, and continued growth in experience and wisdom. Each stage is essential for overall development and well-being. The life cycle of a VR program mirrors the gradual progression seen in the life cycle of humans, moving from an initial concept to a fully mature application.  Here are some common characteristics of each stage. In the infancy stage of a new VR program, it is a time of exploration and experimentation, where educators trial many “one-off” bespoke experiences to find what works best. This period is marked by uncertainty among teachers, as they navigate the unfamiliar terrain of immersive technology. Curriculum alignments are minimal, often leading to a lack of clear educational objectives and outcomes. The usage of VR headsets experiences ups and downs, with some days seeing heavy engagement and others where the equipment sits idle. Despite these challenges, the program is usually anchored on one strong use case, which serves as a foundation for demonstrating the potential of VR in education and guiding further development and integration into the learning environment. In the adolescent stage of a VR program, the focus shifts towards expanding use cases and integrating the technology more deeply into the educational framework. This phase involves training teachers and students to become proficient in using VR, and even empowering them to create their own content. There is a concerted effort to align VR experiences with curriculum standards and instructional practices. Educators explore the use of VR for project-based learning (PBL) and retrieval practice, leveraging its immersive capabilities to foster deeper understanding and retention of knowledge. Additionally, this stage sees the development of policies and guidelines to standardize VR usage, ensuring safe, effective, and consistent application across various educational contexts. In the maturity stage of a VR program, the focus is on sustainability and consistency of use cases. The technology becomes a staple in the educational environment, with well-established applications that reliably enhance learning outcomes. Teachers and students engage in robust content creation, producing high-quality, curriculum-aligned experiences that enrich the educational process. Policies and best practices are firmly in place, ensuring the safe, effective, and consistent use of VR. This stage is marked by ongoing refinement and adaptation, maintaining the program's relevance and impact over time.

I had another reminder this weekend that learning is learning, regardless of the developmental stage. This card was in my daughter’s latest Lovevery box. It was designed for parents of children around 4 years old. It illustrates the "gradual release of responsibility" model - learners progress through scaffolded stages of observing an expert model, practicing with support, then applying skills independently. Mastery comes from actively engaging as guidance fades. This approach reminds us that simply telling isn't enough for developing competence. We need learning and apprenticeship models ranging from highly directive techniques early on ("I do, you watch") to non-directive coaching as learners gain experience ("You do, I'll be here if needed"). For managers, trainers and mentors, intentionally structuring learning paths with this transparent progression enhances engagement and skill transfer. It aligns with theories like cognitive apprenticeship and Vygotsky's Zone of Proximal Development by meeting learners where they are. Whether upskilling a new manager or onboarding engineers to a complex coding stack, starting with modeling and scaffolding towards autonomy cultivates self-sufficiency. I was struck that this simple visual for parenting holds so many implications for the professional sphere as well. How have you applied these principles to workplace learning? How does this model show up in your organization? #coaching #learninganddevelopment #traininganddevelopment #workplacelearning