Rubber Duck Typing

A continuation: how LLM tempt us to bypass a formative struggle of writing code by ourselves but opens up unprecedented opportunities to teach engineering rather than coding, why code writing is even relevant these days, how we might reinvent a junior developer role and how to mentor juniors to seniority. https://rubber-duck-typing.com/posts/2026-08-11-ai-scales-assignments.html

My thoughts about how LLM disrupt grading in the software engineering education, and whether it is good or bad. LLMs do not kill programming education by themselves. They expose it as already in a crisis, propped up by proxies too cheap to fake before now too cheap to trust. Cheating was never the disease — it was the symptom, the rational response to bad classes, exhausted teachers, and students studying for the "wrong" reasons. Grades measured memorization, not understanding, because understanding only shows itself when an expert asks a question and listens to the answer — and that doesn't scale. Universities scaled anyway, trading small seminars for massive classrooms and calling it teaching. Now that assessing bad proxies is obviously meaningless, arithmetic finally shows: real education needs more teachers, smaller groups, and money spent on people, above all. Employers will soon relearn this too — portfolios and open-source records are just proxies AI can also fake, leaving one thing standing: a human sitting with a candidate, talking about the work. https://rubber-duck-typing.com/posts/2026-08-08-ai-education-role.html

How I used approaches from system engineering to design error diagnostics for ZKSync OS. My former students know I love system engineering, so I wanted to show its usefulness in the field. https://rubber-duck-typing.com/posts/2026-06-22-zkos-error-definitions.html

Онлайн-курс «Современные теории типов» В среду 15 июля в 19:00 CEST/UTC+2 (20:00 MSK) в Лаборатории формальной математики стартует курс по современным теориям типов. Лекции читают @akuklev и @clayrat по средам, примерно по часу, частота - раз в неделю (с летними пропусками). Начнём с обзора формальных языков и алгебраических теорий и пойдём до самого фронтира синтетических и направленных теорий типов. Примерная программа: 1. Вводная лекция 2. Языки и алгебраические теории 3. STLC и System T 4. PCF 5. System F и Fω 6. Зависимо-типизированные языки 7. Индукция 8. Рефайнмент- и фактор-типы 9. Эффекты в типах 10. HoTT 11. OTT/CuTT 12. □-полиморфизм 13. Модальные типы 14. Охраняемая рекурсия 15. Когезивные модальности 16. Направленные и симплициальные теории Не требуется предварительной подготовки по теории типов, но пригодятся базовые познания в функциональном программировании и алгебре. Знание теории категорий для понимания курса в целом не нужно, за одним исключением: мы будем обсуждать внутренние языки категорий и топосов (определение топоса дадим по ходу), где не помешает помнить определение декартово замкнутой категории. Ссылка на гугл-календарь, где будем публиковать даты лекций: https://calendar.google.com/calendar/u/0?cid=YzdkMGI0MTdlZjFiMTg1OGVmNzUyYjFkZjBjYjYwZjBhYTI0MGExNjlhMWVhZGY5OTcyOGYwOTM4OTVlMDliM0Bncm91cC5jYWxlbmRhci5nb29nbGUuY29t

A feast for those interested in formal mathematics and modern type theories!

I completed a long piece on Giambattista Vico, a Napolitan 18th century philosopher, and the many connections of his philosophy to software/system engineering. The foundation of Vico's epistemology is the Verum Factum principle: - We know best what we are making, while we are making it, if we do it intentionally. Maker has a privileged epistemic position. - As humans, we may understand what others made, even if they were not intentional about making. For that we study the genus, the history of design decisions. I contrast Vico with Descartes, whom Vico deemed incomplete in his treatment of knowledge, trace Verum Factum through Vico's works and connect it to our everyday practice as software developers. Besides, treating engineered objects not just as 4D space-time worms, like in perdurantism, but also holding rationale of every design decision, raises interesting concerns about the use of agentic AI. https://rubber-duck-typing.com/posts/2026-05-24-vico-verum-factum-software.html

Stephen Toulmin proposes the following narrative: Enlightenment was not a direct response to medieval dogmatism, fanaticism, and scholasticism. It was more of a counter-Renaissance movement. Renaissance's humanism was open, pluralistic, practically oriented, concerned itself with particulars, situated issues. In contrast, Cartesianism is charmed by certainty, objectivity, immutability, universality, abstraction, systematicity, rigor. Ambiguity was praised by the firsts and despised by latter. So there were really two "modernities", largely complimentary structurally, but competing for dominance. The historical scene pushed the transition -- a trauma after 1618-1648 war, fueled in part hy religious conflicts, killing between 4 and 8 mln people, made people strive for more structure and universalism. According to Toulmin, Descartes was looking for a unifying, universal and religion-agnostic intellectual framework for all people to agree on things peacefully. I have noticed how different authors marking that the structure of society and the historical context affect the collective mood, making one of these poles more appealing. I wonder how much of the pursuit of truth becomes a question of personal preference, history and unmet needs. I remember some arguing for a somehow similar connection between the structure of society and morals (assumption of a single, universal telos vs flexibility and interpretability of what's good and what's not). Funnily, "After Virtue" of Macintyre (which is one of my favorite books on ethics) mourns the loss of a unified telos, whereas Toulmin mourns the loss of plurality of humanism.

Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians https://arxiv.org/abs/2602.19141

Established traditions often narrate their history of the world, or a specific domain, state, nation etc. following the same template. First, there are three categories, coming from how the agent is related to the core framework. There is a designated founder, then a pantheon of heroes, and we are just ordinary men. The founder was the origin, from outside the framework, the heroes were the frontline, at the boundary, and we exist inside the framework, following the roads already explored, maintaining and applying the tools we inherited. The narrative then unfolds in time. Second, a storyline is derived. There was an age of gods, when Zeus fought titans, or the world was created from prime elements, or when Newton was writing the laws of mechanics or when Marx was writing his manifesto. An age of heroes, of argonauts, Trojan war, cyclopic micaenean architecture, and demigods, of big names from the physics textbook like Pascal and Bolzman and Laplace, of heroes of communist revolution. Then there are we, not a match for either. The narrative is always relative to the narrator's present. The present is virtually always perceived as the age of epigons. The phases are reflective of the maturation of practices, but the leading roles are usually assigned in retrospect. Few founders or heroes are self conscious.The tradition selects, mythologizes, and assigns quasi-divine status to figures perceived as making the critical contribution, regardless of what actually happened. It does not matter whether creation is literally world making, state founding, prophets establishing law, scientists promoting a new scientific paradigm. The founder brings into being the space where activity becomes intelligible. It's about meaning. However, if we consider not literal gods but human founders, they were all epigons in some former narratives. By a combination of talent, effort, luck, social factors, they managed to escape the framework and branch a new narrative domain. Interestingly, the role of an individual is important here. In more primal societies the personal identity of a member was largely defined by their role in their social network: brother of such and such, makes shovels for those, worships that god, etc. Later, in some branches of civilization, the role of an individual per se gains more recognition. You may live alone on an island but you still account for something. Somewhere in between is the individual as a medium: a prophet is not praised for their insight, but for how attuned they are to the transcendental. I wonder why so many states with strong ideologies basically engineer such narratives into their national identities? I speculate that this storyline, as a part of an indoctrination technology, produces a non-fungible commitment. If I follow the leadership because it fills my pockets, I will turn my back the moment the leader fails, or whenever it becomes unclear that they can deliver long term. If we are united against the common enemy, the union is only as strong and lasting as the perceived threat. But if I am committed to _this_ specific tradition, this founder, these heroes and mythology, it is more robust to external shocks, such as military defeats or economic crises. Of course, it the narrative becomes too detached from reality, like it happened in late USSR, the illusion stops providing meaning and loses its grip. Of course, narratives are not actual history. Whether founders had anything to do with the frameworks, and to what extent, varies. Only the story matters. Heroes are an interesting layer. That's where most fighting seems to happen. Same founders with different sets of heroes results sometimes in radically different traditions. As a final observation, the structure of the story is morally neutral. Totalitarian states and liberation movements adopt it alike.

The question "How does a bat's life feel like?" is apparently impossible to answer. One problem is that we don't have the same senses as a bat, so their sensory input has dimensions inaccessible to us. But more importantly, the content of our mental states is personal and path-dependent. If a wizard transforms me into a bat, but I keep my mind and memories, I'll feel the world as Human-Transformed-Into-Bat, a kind of a hybrid species. To really experience the world like a bat I'll have to part with all my memories and skills, language, all my meaning making frameworks and so on. At this point there won't be much left of me to make sense of this experience. https://rubber-duck-typing.com/posts/2026-03-23-like-batman.html

Can AI corporations like OpenAI or Anthropic realistically protect themselves from governments and their secret services, as they have an incentive to make LLMs generate vulnerable code? The works on the topic I've seen so far seem to focus on attacks through open channels, but not on insider jobs. We know that poisoning a very small part of training set is enough.

Why I like to think about everything as 4D space-time objects: https://rubber-duck-typing.com/posts/2026-03-16-system-engineering-perdurantism.html

Some programs run on platforms where every instruction costs gas. This challenges compilers in subtle ways. First: programs can read their gas meter, directly or indirectly. Two compilers targeting the same platform will produce code that burns gas at different rates, the program can see this and act differently. Second: different platforms measure gas differently. A compiler targeting multiple platforms must track multiple fuel meters. I explore the problem statement a bit deeper here: https://rubber-duck-typing.com/posts/2026-03-14-compiler-correctness-gas-challenge.html In the next post on this topic I'll try to get further with formalization and propose a coherent definition of gas barriers that might tame the branching-on-gas problem, so that it would be easier to write multiple compilers for a language like Solidity. I don't have any decent ideas about memory growth yet :X

Life in postmodernity became volatile and chaotic. Many life size projects were made impossible by a rapidly changing context: in most cases you can't learn a set of skills and make a living out of it for your whole life. We are forced to keep learning, changing countries and jobs, to adapt to the evolving ethical space, reevaluate our choices of partners and so on. This forces us into a series of shorter, shallower games. This narrative described by philosophers like Bauman mirrors what happened in computer engineering by roughly 1990s. There, systems became too complex. We can't make or modify a complex system in a single plan-execute pass: first, because when we start, we don't understand all the requirements; second, because the context evolves while we are busy building. We are chasing a moving, badly defined target. Postmodern life is similar. We ourselves, and the world around, come up with requirements that we try to satisfy (be healthy, be a good citizen, partner, friend etc). Requirements are volatile and loosely defined. So we need to design the life for adaptability, rather than a single quest. System engineering is actually providing us with a lot of wisdom we can directly apply to our life; sometimes, counterintuitively. Some bits of knowledge that surprised me are: 1. If you want robustness, you need to tolerate inefficiency, not optimize every last bit. 2. Extreme optimization makes systems brittle. That's how some animals become extinct, by the way: the context evolves, but they adapted so well that they can't follow the change. Squeezing productivity out of every minute is a dangerous fantasy. 3. Data-driven decisions are often traps. Metrics are imperfect proxies of some qualities we want to achieve; optimizing a metric often loses track of its purpose and makes the system perform worse, while formally growing a metric value. This is called Goodhart's law. 4. Failures are often preceded by degradations; that's something you can monitor and act before failure happens. Failure rate often follows the bathtub distribution, e.g. we get sick much more often as infants and before we die. 5. Explicitly design failure modes. If i dont perform at studies, should I quit fast or degrade gracefully? 6. Ashby's law: a system tgat controls another system should contain its model. So to live a complex life with some reasonable degree of control you have to match your internal complexity. I think it also works in other direction: complex life challenges us to grow. Interestingly, life virtues transform from functional requirements (achieve specific good life) to non-functional requirements (maintain capacity to adapt).

Channel photo updated

Throughout immense stretches of time the intellect produced nothing but errors; some of them proved to be useful and preservative of the species: he who fell in with them, or inherited them, waged the battle for himself and his offspring with better success. Taken rom "The Gay Science", it starts a section on the origins of knowledge, but it also works in the context of Nietzsche's study of art and knowledge. Sciences produce useful approximations, which are, in fact, errors. As time passes, we keep revising them, changing them, making them more precise or changing the frameworks altogether. A piece of true art, however, tells something timeless about us and our world. In that regard it gets closer to eternal truths of pure mathematical knowledge.

Dissociation and awareness are the opposites: while both introduce an observing self, one of them is looking away, while another is focusing on recognizing and being present.

Meditative practices like vipassana and sceptical method have something in common. Sceptical method is: if you think "X", find compelling arguments for "not X". If you did, you now have an equilibrium that allows to suspend your judgement on X. You therefore achieve an inner peace and escape a pursuit of truth. Meditation aims to teach you to be a careful observer, being immediately aware of your state, but not taking a judgemental stance. If you stop following a habit of judging, you achieve the same result as sceptics, and more.

Nietzsche provides an interesting storyline to the genesis of the concept of truth. In order to survive in groups, people need to establish the rules of a social contract, otherwise peaceful coexistence is unlikely. For that, they need a language. Once you fix the meaning of some words — this object is a stone, and this one is not — you introduce the concept of truth and falsity in your thinking. This concept is charming and philosophers make their lives about pursuing it — even if the concept is an illusion of the mind itself, and pursuing it in some sense brings you further from the world, not closer. We forgot how our perception of the world comes a long way from nerve stimulus to distilled concepts. In our desire to grasp the essence of the world we substitute its richness with simplicity of concepts, and mistake the abstraction for reality. Truth is a forgotten metaphor.

For trauma survivors, pattern recognition overrides context.