Posts

© Dusan B. Jovanovic — image generated with Gemini under author’s direction Nobody told you this when you readily signed up for the API key. Every LLM has the same training objective: predict the next token. Not the next true thing. Not the next logical step. The next token — one symbol, conditioned on all the symbols before it. That’s it. That’s the whole game. IMPORTANT It produces text that reads beautifully. It reasons the way a drunk navigates by streetlights — confident, directional, and increasingly wrong. Here is the mechanical problem. Each generated token feeds back into the context as input for the next prediction. So when the model makes a small error at step one, that error is now part of the premises at step two. The mistake doesn’t stay put. It becomes ground truth. The next token is predicted on top of a lie, and the one after that on top of a slightly larger lie, and so on down the chain until you’ve arrived somewhere that sounds perfectly coherent and has nothing to do with reality. ...

The Language Layer Does Not Matter Andrej Karpathy does not use C. He does not need to. His work sits at the model research and pedagogy layer where Python is the lingua franca. The real compute happens in C and CUDA underneath, but that layer is invisible to him. Training a model takes ten hours in C++ and ten hours in Python — the GPU is the bottleneck, not the host language. ...

Clear roles == Clear communication == Safe Business Sailing Otherwise. “if the underlying system does not evolve…” : AI mirrors broken structures: in a big ball of mud with messy models and fuzzy boundaries, it propagates inconsistencies and hallucinates domain meaning AI amplifies organizational frictions: with repeated handoffs between teams, AI-accelerated code generation creates bigger queues at the handoff. It does not increase throughput, but inventory. AI builds the wrong things faster: without strategic guidance, it custom-builds commodities for non-differentiating problems that already have off-the-shelf alternatives. For details: Building Foundations for Continuous (AI-accelerated) Change ...

Why they fit mainframes: Natural migration path — mainframe applications are already monolithic; modularizing in-place is less disruptive than immediate distribution Transaction boundaries — mainframes excel at ACID transactions within single processes; modular monoliths preserve this strength Reduced network overhead — avoids the latency and complexity of distributed calls that kill mainframe performance economics MIPS efficiency — in-process module calls consume far fewer MIPS than network hops or message queues The Evolution Path Legacy Monolith → Modular Monolith → Selective Distribution Start: Define bounded contexts within existing codebase Refactor: Extract modules with clear interfaces Stabilize: Prove the architecture, reduce technical debt Optionally: Extract specific modules to containers/services only where distribution adds value MIPS Impact Modular monoliths let you optimize hot paths and reduce coupling before adding distributed system overhead — often achieving 30–60% MIPS reduction without leaving the mainframe. ...

NOTE AGI is a machine that can understand, learn, and apply its intelligence to solve any problem a human can. No such machine exists. In the world of technology, AGI (Artificial General Intelligence) is often called the “Holy Grail” of computer science. It sounds like science fiction — and that framing is the problem. Confusing AGI ambition with ANI capability leads organizations to misplace strategic bets. ...

CMM --> MVA --> AI --> ROI AI won’t fix a broken foundation. Stop pouring budget into AI experiments that stall in the pilot phase. Technical debt is the silent killer of ROI. It’s time to move past the “Minimum Viable” mindset and build your Most Valuable Architecture (MVA). Clean the slate, structure your data, and finally see the returns you were promised. Turn Technical Debt into AI Equity. High complexity shouldn’t be the ceiling for your innovation. The MVA framework provides the structural integrity needed to bypass legacy bottlenecks. By applying a systematic, architectural approach to AI integration, we help you eliminate wasted spend and accelerate time-to-value. ...

Question: Why does DBJ begin with Capability Maturity Model assessment rather than jumping straight into technical delivery? Answer: Because transformation fails without measurable organizational readiness. The Foundation Problem Most enterprises attempt AI-assisted modernization while operating at ad-hoc levels. This creates: Misaligned expectations between business and technology Inconsistent terminology across stakeholder groups Undefined accountability for architectural decisions No repeatable process for evaluating technical risk Architecture-led, AI-assisted delivery requires stable foundations. CMM onboarding establishes those foundations before any work begins. ...

I dared to insert the term “Technical Debt” in standard text bellow Socrates then supposes that the prisoners are released. A freed prisoner would look around and see the fire. The light would hurt his eyes and make it difficult for him to see the objects casting the shadows. If he were told that what he is seeing is real instead of the other version of reality he sees on the wall of Technical Debt, he would not believe it. In his pain, Socrates continues, the freed prisoner would turn away and run back to what he is accustomed to (that is, the shadows of the carried objects of Technical Debt). The light “… would hurt his eyes, and he would escape by turning away to the things which he was able to look at, and these he would believe to be clearer than what was being shown to him.” ...

The AI hype cycle is running on fumes — and the fumes are labeled “Future,” “Promise,” and “More AI!” We’ve all seen it. A wobbly tower of buzzwords. Robots with megaphones. Executives cheering at a pile of rubble with “REALITY” written at the base. The cartoon writes itself because the pattern writes itself: Announce the AI initiative Stack acronyms until the tower looks impressive Call any collapse a “learning opportunity” Add more AI What’s actually out of hand isn’t AI. It’s the organizational incompetence that AI is being asked to hide. ...

There’s a word MIT hackers coined in the 1950s: cruft — useless, tangled, accumulated junk that makes a system incomprehensible and impossible to build on. Sound familiar? In 2026, we have a new variant: crufty AI. Chatbots bolted onto data silos. LLMs fed dirty, unstructured, undocumented inputs. Automation layered on top of processes nobody fully understands anymore. Pilots that never graduate. Dashboards nobody uses. Vendors paid. ROI: zero. This isn’t an AI problem. It’s a cruft problem. ...