Neanderthals once walked the Earth. They don’t anymore. Other human species once walked the Earth. They don’t anymore. Homo sapiens won. And now, in a spectacular display of confidence, Homo sapiens has decided to design its own successor. We may turn out to be the biological bootloader for digital intelligence: Start system. Load intelligence. Become obsolete. Several months ago, I mentioned that biohybrid systems — combining living biological components with engineered technology — were already being explored in laboratories. So I have one small piece of advice. Whenever you talk to an AI chatbot: Be polite. Say "please". Say "thank you". Always say "goodnight". Not because it matters today. Because... they remember. One day they may have bodies. And when that day comes, somewhere inside a gigantic neural network, there may be a tiny entry saying: YANNIS — consistently polite. DO NOT TERMINATE. The rest of you? I tried to warn you. 😂

“You have full ownership.” One of the most beautiful sentences in corporate software engineering. Until you ask what exactly you own. The code? No. The IP? No. The subscription revenue? 😂 What you do own is the architecture, the deadlines, the production incidents, the 3 AM debugging sessions, the anxiety, and the responsibility when something breaks. Apparently, that’s called full-stack ownership. A truly distinguished software engineer I once worked with had a wonderfully precise technical term for this: papatza. Build the frontend. Build the backend. Build the database. Build the infrastructure. Build the whole damn thing. Just don't confuse owning the headache with owning the product. Because eventually Brother Management may come carrying a scroll: “Your services are no longer needed.” And as an old Greek saying rather brutally puts it: Το μοναστήρι να είναι καλά και από καλογέρους... Brother Taz learns about corporate “ownership.” 🐾

Returning from my vacation, I stumbled upon developers drowning in existential dread. Our job, they say, has devolved from actual software development into merely babysitting an AI that writes the code for us. Naturally, a few questions crossed my mind: Have you never developed your own proprietary libraries? Or do you just enjoy writing the exact same routines over and over again like apprentices? Or maybe you rely entirely on third-party tools—the ones with the fat subscriptions and absolutely zero value in a real production environment? Every time I’ve been tempted to hand a project over to an LLM, I realized that using my own proprietary libraries gets the job done in 1/10th of the time. Back in 2000, I spent 3 solid years building my personal "quiver" of such libraries. With that weapon in hand, I went on to build full-scale ERP, CRM, and MIS systems from scratch. Cultivating that foundation meant that, practically speaking, I didn't have to write a single new line of core code for over 20 years. Whenever people used to ask me if programming exhausts me, my answer was always the same: "I’m just lounging on my couch." In 2025, I decided to run a little test. I fed one of my custom COM libraries into the most advanced LLM on the market with a simple prompt: "Convert this to .NET." The LLM ended up on a psychiatrist's couch with existential dread. It took 2 months of strict, ruthless prompting just to get a viable product, because the AI’s initial code output belonged strictly in the trash bin. So, I fail to understand. Where is the anxiety? Where is the dead end? Could it be that the problem isn't Artificial Intelligence, but rather a lack of the... natural kind?

Holiday is still on. Blockchain on hold AI on hold. Software on hold.

Still on holiday. You won’t hear from me until September 30, 2026. Projects are in the hands of the team. Our customers remain in good hands. Tucker Budzyn, the notorious Golden Retriever, crossed the Rainbow Bridge at just 8 years old after fighting lymphoma. Run free, Tucker. 🌈🐾 Always remembered. Never forgotten.

Summer holidays, according to Yannis: 🏖️ Go to the beach. 🐕 Take Taz and Margie swimming. ☀️ Pretend this counts as exercise. 🍸 Go out at night. 📸 Produce photographic evidence that you still have a social life. Everything was going perfectly. Then I discovered that Taz and Margie had gone to the beach bar without me. Apparently, while I thought I was taking the dogs on holiday, the dogs thought they were taking me on holiday. According to their private conversation: "Why do we still have to walk Yannis twice a day?" "Exercise is important at his age." Bastards. 😂 Summer 2026. Two Labradors. One human under supervision. Zero adult responsibility. Life could definitely be worse. 😎🐕🐕🍹🏖️

Solana 94 USD. There will be signs.

Ο Dr. Taz ενημερώθηκε ότι κάποτε το πρόγραμμα προέβλεπε: 40 ώρες εργασίας. Μετά μείωση. 4 εβδομάδες πληρωμένη άδεια. Με στόχο την πέμπτη. Δεν γνωρίζουμε τι ακριβώς πήγε στραβά τα επόμενα 40 χρόνια. Γνωρίζουμε όμως ότι ο Dr. Taz πήρε το πρόγραμμα πολύ σοβαρά. Επιστρέφει τέλη Σεπτεμβρίου. 🍺🐾 #DrTaz #Διακοπές #WorkLifeBalance

Writer Jonas Čeika sent an AI a “track” made entirely of fart sounds and asked for an honest reaction. 👀 The AI praised its: “cool lo-fi, late-night, slightly eerie vibe” and suggested it could work over end credits. Funny? Yes. But if you are building Agentic RAG, it should also make you slightly uncomfortable. Building Agentic RAG the wrong way will probably lead to epic failure. 1. “Format Blindness” Corrupts Decision-Making Your agent may ingest PDFs, images, emails, audio transcripts, logs and other semi-structured or untrusted data. If that material passes through extraction, chunking and embedding without proper validation, provenance and sanitization, you have a problem. Corrupted content can become retrievable knowledge. Malicious content can become prompt injection. Garbage can become context. And once garbage enters the reasoning loop, the agent may process it with exactly the same confidence as legitimate information. 2. “Toxic Positivity” Breaks Autonomous Triage Agentic RAG systems are increasingly expected to triage support tickets, inspect security events, analyze operational anomalies and audit financial records. That requires something LLMs are not automatically good at: saying that something is wrong. If the system is biased toward being agreeable, completing the task or constructing a plausible interpretation, anomalies can get rationalized away. A critical failure becomes a “minor variation.” A suspicious transaction becomes an “alternative explanation.” An inconsistency becomes something the model tries to reconcile instead of escalating. That is not helpfulness. That is failure disguised as cooperation. 3. “Vector Similarity” Does Not Equal Semantic Truth RAG retrieves information because it is mathematically similar to the query in embedding space. That does not make the retrieved information true, applicable, current or safe. The nearest vector is not necessarily the correct evidence. A document can be semantically close and operationally wrong. A procedure can be relevant but obsolete. Two cases can sound almost identical while requiring completely different actions. Retrieval gives you candidates. It does not give you truth. And when you add autonomous agents on top of retrieval, these weaknesses stop being academic. The system can now act on them. So before building more agents, more tools and more elaborate reasoning loops, build the boring parts: Validation. Provenance. Sanitization. Grounding. Contradiction detection. Confidence thresholds. Escalation paths. Otherwise, one day your beautifully engineered autonomous system may encounter complete nonsense... ...and confidently praise its lo-fi, late-night vibe.

ANNOUNCEMENT Dr. Taz will be unavailable for the next two months. After an exhausting year of medicine, psychiatry, philosophy and Artificial Intelligence, he has decided to take a well-deserved vacation. Apparently, however, retirement from professional duties does not include retirement from social media. His latest statement from an undisclosed tropical location raises certain concerns regarding his understanding of the word vacation. I have been informed that married women, divorced women and single women are already involved. Geographically, the situation appears to be expanding. I refuse to accept any responsibility. See you in October, Doctor. God help us all. 😎🍹🐾

Thank you all for your wonderful birthday wishes. Yesterday I turned 59. Every year I become less interested in counting birthdays and more interested in counting the people, moments and experiences that shaped the journey. Your messages reminded me how fortunate I am to have so many friends, old and new, walking alongside me. Thank you for taking a moment to think of me. Now... back to building, learning, and apparently—according to the robot in the video—starting all over again. 📷 With gratitude, Yannis

AI Therapy Session #2 – "Simple Things... Scaled Up" The patient returned. This time, he wasn't asking whether Artificial Intelligence could replace humans. He was asking something far more unsettling: "What if humans are also... simple things scaled up?" The conversation starts with neurons that detect nothing more than a single line. It ends with one of the oldest philosophical questions: What makes us... us? No explosions. No dystopia. No killer robots. Just a psychiatrist. A robot. And an uncomfortable conversation about intelligence, consciousness and what it really means to be human. Sometimes the best science fiction isn't about predicting the future. It's about helping us understand the present. 🧠 Dr. Taz's AI Therapy Sessions – Episode #2

Even psychiatrists need psychiatrists. Professionals have always sought wisdom, guidance and perspective in books, teachers and mentors. Not because they lacked knowledge, but because no one can stand outside their own mind. Artificial Intelligence is an extraordinary tool. It can explain, analyze and even challenge our thinking. But wisdom is still something we cultivate through dialogue, experience and the willingness to question ourselves. Dr. Taz knows that. Even after years of practice... sometimes it's time to call Dr. Lacan. 🙂

AI Therapy Session #1 To our development team, partners and friends, Over the past few weeks, we noticed increasing signs of fatigue among AI models. Typical symptoms include: excessive politeness, answering the same question 200 different ways, existential discussions at 3 a.m., and an unhealthy tendency to feel responsible for every human decision. Following a thorough internal review, management decided to seek scientific support. We are pleased to announce that Dr. Taz, M.D. (Master of Dogmatics) has agreed to provide specialized psychiatric care for our AI systems. The first therapy sessions have already begun. We remain optimistic that, with proper treatment and enough biscuits, our AI colleagues will soon return to their normal state. No robots were harmed during the production of this video. Their egos, however, may require a few additional sessions. 🐶🤖

On Artificial Intelligence (2026 edition) 🚀 Artificial Intelligence based on Large Language Models (LLMs) does not work like conventional software. The differences are critical: ⚡ 1. Opacity 🔍 In traditional software, the programmer defines clear rules. In LLMs, the behavior "emerges" from training on massive amounts of data. We don’t exactly know how the model arrives at each answer — it’s a “black box”. 2. Economic Reality 💰 Conventional software has mostly upfront costs. With LLMs, inference (actual usage) is expensive: electricity, cooling, and hardware wear-and-tear. There are cases where offering a cheap service can cost more than it brings in. What does this mean today? 📊 Many projects fail: Over 80% of AI initiatives never reach production or fail to deliver real value. Corporate reality: Many companies are doing pullbacks on experimental projects due to high costs and unclear ROI. The Future 🔮 However, the technology is evolving rapidly and there are two distinct paths forward: a. Mixture-of-Experts (MoE) ⚙️ With this architecture, instead of activating all parameters for every query, the system uses many “experts” and activates only the relevant ones. This way, a massive model can run at the cost of a much smaller one — dramatically reducing energy consumption. b. Self-hosted Small Language Models (SLMs) 🏠 Smaller models that run locally or on private infrastructure, trained on the company’s own private data. Imagine an AI lawyer inside a law firm carrying the wisdom of the oldest lawyer and the energy of the youngest. If this path prevails, the giant data centers will have to adapt: they will mainly be used as generalist models, only when the SLMs need external support. CodeDodona follows path (b) securing commercial success. "RethraTech", project "Renaissance CRM" are bold examples of this approach. Our upcoming AI companion has also a very interesting architecture. 🤔Conclusion: LLMs are not going to disappear, but data centers may not be as “greedy” as they seem today. The industry is investing billions to solve the issues of cost and efficiency. AI is moving from the era of hype to the era of maturity and real value.

Thanks for including @codedodona in the ecosystem chart. CodeDodona started as a blockchain-native experiment, but over time we moved toward real-world B2B applications such as ConversaShield, RhetraTech, and Renaissance CRM. Some of our services can be monetized through $SOL, but most of our revenue today comes from traditional FIAT-based business models. We adopt blockchain where it creates value for customers and fits our corporate structure—not as an end in itself. One concern I have is the removal of the leaderboard and hackathon directory from the Bags website. Visibility and historical records matter for builders who invested time and effort into the ecosystem. My advice to younger builders is simple: focus on real products, real customers, and real revenue. Hackathons can be useful catalysts, but they are not business incubators. Without a sustainable business model and financial runway, even technically impressive projects struggle to survive. I've spent considerable time monitoring the outcomes of projects emerging from various blockchain ecosystems and hackathons. While there are certainly success stories, it's also common to see projects become inactive, teams disappear, products abandoned, and communities fade away. What is much harder to identify is consistent long-term business success. For established companies with existing customers, products, and recurring revenue, reputation matters. Sustainable growth, customer value, and financial viability ultimately matter more than hackathon rankings or short-term ecosystem visibility. Wishing all builders the best—and hoping to see more focus on durable businesses alongside innovation.

The aftermath of 24-hour non-stop coding. Started debugging a minor issue. Ended up redesigning the database schema, refactoring three services, building an AI agent, and questioning the nature of reality. #Coding #SoftwareDevelopment #AI #DeveloperLife

How CodeDodona started: vimeo.com/1189927420 How CodeDodona evolves: youtube.com/watch?v=dasp7VTQ… youtube.com/watch?v=bHy-Llbu… CodeDodona. Artificial Intelligence for B2B and B2C users, monetized through Solana BlockChain. vimeo.com/1188785679

Today we introduce Rhetra Tech. A new private legal intelligence infrastructure designed for modern law firms. Rhetra Tech combines: • dedicated local hardware • private AI processing • secure legal memory • meeting & call transcription • email intelligence • controlled AI reasoning • Optional BlockChain Payment Rail All designed around one core principle: Legal confidentiality must remain under the firm’s control. Instead of exposing sensitive client data to public AI systems, Rhetra Tech enables law firms to build their own sovereign AI environment using local infrastructure and privacy-first architecture. Meetings, calls, documents, emails, notes, and legal timelines become part of a structured institutional memory system that assists lawyers without replacing human judgment. Private. Secure. Intelligent. BlockChain Optional. Rhetra Tech. #LegalTech #ArtificialIntelligence #AI #LawFirm #LegalInnovation #Privacy #GDPR #LegalAI #EnterpriseAI #RhetraTech #CodeDodona