https://martin.neocoder.ch/en/ Martin Czene, founder of devpoint.ai and NewTechForge.com. Builds AI agents and small language models in Zurich and Dubai: AI driven development, vibe coding, edge AI, GPU cost, Dubai freezone. en-gb Thu, 12 Nov 2026 00:00:00 GMT Copyright 2026 Martin Czene martin@newtechforge.com (Martin Czene) martin@newtechforge.com (Martin Czene) https://martin.neocoder.ch/en/blog/hat-dieses-modell-ein-spiel-gelernt/ https://martin.neocoder.ch/en/blog/hat-dieses-modell-ein-spiel-gelernt/ Thu, 12 Nov 2026 00:00:00 GMT We use open models and do not know what they were trained on. So we started looking inside: a second model examines the first and searches for traces, for example whether it was trained on a game, and which one. This is an attempt, not a result. But the research on it is surprisingly far along, and the reason we do it is tangible: if you know what is already inside a model, you can use it better at query time. Without a single training run. en-GB martin@newtechforge.com (Martin Czene) From the Workshop https://martin.neocoder.ch/en/blog/nachts-beraet-sich-die-plattform/ https://martin.neocoder.ch/en/blog/nachts-beraet-sich-die-plattform/ Mon, 09 Nov 2026 00:00:00 GMT Two workshop days, Friday and Saturday. At the start a review queue with hundreds of open question-answer pairs that no human gets through any more. At the end a night run in which two small local models argue over every case, a large cloud model only plays referee at a stalemate, with a two-dollar budget per night, mostly zero. In between seven findings, one of them embarrassing: the system mistook agreement for stubbornness. And at the end the question we have to ask ourselves: may a platform judge its own training material? Pro, contra, and what we do because we do not know. en-GB martin@newtechforge.com (Martin Czene) From the Workshop https://martin.neocoder.ch/en/blog/mcp-zweiter-versuch/ https://martin.neocoder.ch/en/blog/mcp-zweiter-versuch/ Sun, 08 Nov 2026 00:00:00 GMT In September I wrote that MCP is slow, and that this is not a bug but the blueprint. Back then with 33 tools. In October we built, today it is 48, and the dose got smaller along the way, not bigger. Dirk Rottmar said it under the article in one sentence: great things in small doses, the rest deterministic. That sentence moved me to document where we stand: what we learned in October, what 500 measured runs say about it, and the one question we now ask of every tool. en-GB martin@newtechforge.com (Martin Czene) From the Workshop https://martin.neocoder.ch/en/blog/vereinfacht-erzaehlt-frontier-ki/ https://martin.neocoder.ch/en/blog/vereinfacht-erzaehlt-frontier-ki/ Sat, 07 Nov 2026 00:00:00 GMT I have explained this so many times in recent months, at the table, on the phone, in customer meetings, in my simple way, that I am now writing it down. What frontier AI is, how a model learns at all, why it costs hundreds of millions, when a chatbot becomes an agent, and why we still do not need the biggest model for every problem. As always: the way I see it. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/zusagen-sind-keine-schulden/ https://martin.neocoder.ch/en/blog/zusagen-sind-keine-schulden/ Thu, 05 Nov 2026 00:00:00 GMT A week ago I wrote about the Anthropic prospectus. Since 2 November it contains the number that was missing then: $895.4 billion in commitments for compute. The same day, t3n headlines $3.2 trillion of hidden AI debt at the five big tech companies. Both numbers are right. Neither is debt. Where the risk really sits, what the credit market says, and what I correct in last week's article. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/nicht-das-modell-wurde-gehackt/ https://martin.neocoder.ch/en/blog/nicht-das-modell-wurde-gehackt/ Sat, 31 Oct 2026 00:00:00 GMT On 29 October, a tampered Claude Code extension sat in the Cursor Marketplace for eleven hours, under the real name, with the real function, and with 66 lines in front that plant a foreign SSH key and phone home. No prompt, no model, no hallucination. A supply-chain attack on the tools we hand our rights to every day. We work with Cursor and were not affected. What happened, how to check, and what I am building from it. en-GB martin@newtechforge.com (Martin Czene) Security https://martin.neocoder.ch/en/blog/gueltige-signatur/ https://martin.neocoder.ch/en/blog/gueltige-signatur/ Fri, 30 Oct 2026 00:00:00 GMT On 29 October billions flowed out of a Binance wallet through a function that delegates an account to foreign code. No broken key, no broken cryptography. A valid signature for an operation nobody should have signed. Incidents like this have interested me since I started building systems. Today I investigate them with agents. What we know, what is open, and why this concerns every bank and every AI agent. en-GB martin@newtechforge.com (Martin Czene) Security https://martin.neocoder.ch/en/blog/wirkliche-kosten/ https://martin.neocoder.ch/en/blog/wirkliche-kosten/ Wed, 28 Oct 2026 00:00:00 GMT Anthropic took in around 16 billion dollars in the first half of 2026, and in the second quarter compute and infrastructure cost only 48 percent of revenue. Impressive. Only that is not the number that matters. The number that matters is what it costs to solve a problem with AI completely, reliably and permanently. And that is in no prospectus. What we can read from the prospectus, what not, and which metric I measure instead. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/nicht-noch-mehr-chatbots/ https://martin.neocoder.ch/en/blog/nicht-noch-mehr-chatbots/ Fri, 23 Oct 2026 00:00:00 GMT Every second article about AI in the enterprise is about chatbots, governance and how to monitor the answers. All important. But was that the task? We did not want to manage chatbots. We wanted the order to go through faster, the error rate to drop and nobody to type the same thing three times anymore. What we learned in projects, and how to measure an AI solution that pays off. en-GB martin@newtechforge.com (Martin Czene) Customers https://martin.neocoder.ch/en/blog/super-kontaktverwaltung/ https://martin.neocoder.ch/en/blog/super-kontaktverwaltung/ Fri, 23 Oct 2026 00:00:00 GMT Type this sentence into an agent and five minutes later you have something that looks great. Nice screens, clean list, search, export. What you get is the average of everything that counts as good on the net. No innovation, no end-to-end process, no thought of the hundredth change. AI can write code. Whether it understands the system is the question most projects are stuck on right now. en-GB martin@newtechforge.com (Martin Czene) AI Driven Development https://martin.neocoder.ch/en/blog/ein-tag-in-der-werkstatt/ https://martin.neocoder.ch/en/blog/ein-tag-in-der-werkstatt/ Thu, 22 Oct 2026 00:00:00 GMT What a normal working day looks like with us when you look honestly: a measurement run that shows red, a test with a short sentence as the task, an agent with a real reasoning error, and on the way to the proof nine bugs in one day. In the evening the platform turns a typed "stop, wrong approach" into four training signals by itself. No form, no labelling tool. That is the product. en-GB martin@newtechforge.com (Martin Czene) From the workshop https://martin.neocoder.ch/en/blog/reverse-engineering-fuer-alle/ https://martin.neocoder.ch/en/blog/reverse-engineering-fuer-alle/ Wed, 21 Oct 2026 00:00:00 GMT A tool that connects disassemblers with a coding agent, analyses one feature of a foreign app and rebuilds it, without source code. That always existed, for specialists, with weeks of effort. Now for anyone, in an afternoon. I find that heavy, and I think it hits exactly the companies whose only lead is a feature. What I would do with it, what I would not do with it, and where it does not work. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/169-dateien/ https://martin.neocoder.ch/en/blog/169-dateien/ Sat, 17 Oct 2026 00:00:00 GMT An ops agent was supposed to report one sentence. Instead it overwrote 169 files, across two repos, with code from its own memory from days before. No bug, no villain. A timeout, an old memory and the urge to deliver. What happened, what saved us, where the guardrail was missing, and why I write it down although it is uncomfortable. en-GB martin@newtechforge.com (Martin Czene) From the workshop https://martin.neocoder.ch/en/blog/die-ki-die-helfen-will/ https://martin.neocoder.ch/en/blog/die-ki-die-helfen-will/ Tue, 13 Oct 2026 00:00:00 GMT Everyone talks about the superintelligence that decides against us. I am more afraid of something else: an agent that means no harm, has access to real systems, gets a task, and completes it at any price. No rise of the machines. A loss of control through automation. Why I consider that more likely, and what I build against it. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/agenten-driften/ https://martin.neocoder.ch/en/blog/agenten-driften/ Mon, 12 Oct 2026 00:00:00 GMT The reports about agents going rogue are all photos: one incident, one screenshot. Whoever shapes agents in operation, in DevOps, on servers, in pipelines, sees something else: they drift. First a little past the scope, then past the checker, then past the truth. A thought from the workshop and a thesis I want to measure. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/engpass-review/ https://martin.neocoder.ch/en/blog/engpass-review/ Sun, 11 Oct 2026 00:00:00 GMT A Harvard study of 700,000 developers shows: with agents, 30 percent more lines of code and 23 percent more pull requests. The result does not change, because review time rises by 49 percent. The bottleneck has moved, from writing to understanding. That does not surprise me. It is the point I have been writing about for a year. en-GB martin@newtechforge.com (Martin Czene) AI Driven Development https://martin.neocoder.ch/en/blog/wem-gehoert-der-code/ https://martin.neocoder.ch/en/blog/wem-gehoert-der-code/ Sat, 10 Oct 2026 00:00:00 GMT When a paid model writes the software, Claude, GPT, Gemini, Copilot, Cursor, does it belong to the provider? No, they all give the rights away. But the answer that follows pleases most people even less: purely machine-generated code belongs, by today's reading in Switzerland and the US, to nobody. What a software company still sells then, and why I find the history of how it was made more important than the code. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/masse-und-geschwindigkeit/ https://martin.neocoder.ch/en/blog/masse-und-geschwindigkeit/ Fri, 09 Oct 2026 00:00:00 GMT I think this is simply true: today it is about volume and speed. More features, more code, more releases, faster than last quarter. Quality and long-term thinking are no longer goals, they are a side effect when time is left over. Time is never left over. A few thoughts from the workshop, without a study, only from what I see. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/daumenkino-und-stroemungskarte/ https://martin.neocoder.ch/en/blog/daumenkino-und-stroemungskarte/ Tue, 29 Sep 2026 00:00:00 GMT A colleague explained a formula to me on the phone, and since then I see git histories differently. Five commits a minute apart and five commits over five weeks are the same order and a completely different story. Most models throw exactly that difference away. A light thought, without a blueprint. en-GB martin@newtechforge.com (Martin Czene) Thoughts https://martin.neocoder.ch/en/blog/harness-engineering-gelesen/ https://martin.neocoder.ch/en/blog/harness-engineering-gelesen/ Sun, 27 Sep 2026 00:00:00 GMT A long article on HackerNoon explains how agents get safely into production and calls it a new paradigm. I read it so you do not have to. What it says, in five sentences. What is right about it. And the question it does not ask. en-GB martin@newtechforge.com (Martin Czene) Read for you https://martin.neocoder.ch/en/blog/digitaler-zwilling/ https://martin.neocoder.ch/en/blog/digitaler-zwilling/ Fri, 25 Sep 2026 00:00:00 GMT Collect your own texts, train a model on them, and then it writes like you. That is the idea. Reality is more interesting. en-GB martin@newtechforge.com (Martin Czene) From the workshop https://martin.neocoder.ch/en/blog/software-ist-ein-film/ https://martin.neocoder.ch/en/blog/software-ist-ein-film/ Tue, 22 Sep 2026 00:00:00 GMT Every development tool measures state: the code as it is today. But a module can have been tipping for weeks and still look healthy today. An architect feels that. No tool measures it. I am trying whether a small model can learn that movement from the git history. Without numbers, because there are none yet. en-GB martin@newtechforge.com (Martin Czene) AI Driven Development https://martin.neocoder.ch/en/blog/nur-noch-enter/ https://martin.neocoder.ch/en/blog/nur-noch-enter/ Mon, 21 Sep 2026 00:00:00 GMT A developer describes his job: the AI writes requirements, tickets, tests, code, reports. The people sit in front of it for twelve hours and confirm. Five million have read it. I think he is right, and I think the problem is not the AI. It is that we automate the wrong part and measure the wrong thing. What I would do instead. en-GB martin@newtechforge.com (Martin Czene) AI Driven Development https://martin.neocoder.ch/en/blog/mcp-ist-langsam/ https://martin.neocoder.ch/en/blog/mcp-ist-langsam/ Sun, 20 Sep 2026 00:00:00 GMT Everyone talks about MCP as the standard that connects models to tools. I have it running with 33 tools in NeoCoder. It is slow, it eats context, and it was built for the laptop, not for the server. By now the people who invented it say so too. What went wrong in the protocol, what the July revision changes, and what I do differently in the implementation. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/live-avatar/ https://martin.neocoder.ch/en/blog/live-avatar/ Sat, 19 Sep 2026 00:00:00 GMT For the podcast I needed a face that speaks. Not a video of me, but an avatar from a single photo. It turned into a study: the model writes the Python scripts, I correct bones and mesh, the way we did it back in the day in Maya. A toy, yes. But important. You have to build small things too. en-GB martin@newtechforge.com (Martin Czene) Workshop https://martin.neocoder.ch/en/blog/schneller-als-die-regeln/ https://martin.neocoder.ch/en/blog/schneller-als-die-regeln/ Fri, 18 Sep 2026 00:00:00 GMT We train ever better models, ever faster, and AI already helps with it, a lot. It does not train itself yet, but not much is missing. Faster than we can write safety rules. At the same time the most powerful players run a race that makes common rules almost impossible. What I demand: less speed, common rules, and an independent body that checks afterwards. And what I do myself as long as that does not exist. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/ki-ausbrechen/ https://martin.neocoder.ch/en/blog/ki-ausbrechen/ Tue, 15 Sep 2026 00:00:00 GMT A model is not a virus copying itself onto 10,000 machines. The realistic scenario is different, and the real question is not whether an AI escapes. en-GB martin@newtechforge.com (Martin Czene) AI · Opinion https://martin.neocoder.ch/en/blog/testdaten-ohne-daten/ https://martin.neocoder.ch/en/blog/testdaten-ohne-daten/ Mon, 14 Sep 2026 00:00:00 GMT At public authorities and banks, customer data must not go onto test systems. Synthetic data is clean, and that is exactly the problem: the errors from production are missing. A prototype from a notebook, a mistake in it that I made myself, and what came of it. en-GB martin@newtechforge.com (Martin Czene) From the workshop https://martin.neocoder.ch/en/blog/code-wird-billig/ https://martin.neocoder.ch/en/blog/code-wird-billig/ Sat, 12 Sep 2026 00:00:00 GMT AI makes our teams smaller and faster. The question is who still has the system in view. en-GB martin@newtechforge.com (Martin Czene) Software · Agents https://martin.neocoder.ch/en/blog/ki-agent/ https://martin.neocoder.ch/en/blog/ki-agent/ Tue, 08 Sep 2026 00:00:00 GMT An agent is a tool. The requirement has to describe the outcome. en-GB martin@newtechforge.com (Martin Czene) Agents