[{"content":"the following thoughts are true and understandable, even though the form is not good yet. But I don\u0026rsquo;t have other time to dedicate to it now. you get it or you don\u0026rsquo;t\nMore than two weeks ago, I stopped using AI completely. I started writing my documents myself. I started taking notes myself. The results are astonishing. Spiritually. Yes, I am not a good writer. I make a lot of spelling mistakes and English is not my first language. And yet the way I write is better than Reddit/Linkedin sounding sentences. The kind of sentences the LLMs produce.\nThe reason I stopped is I felt weak, a weasel, and (honestly) less of a man. You present yourself with that kind of writing 1, and you are that kind of person. I was never able to 100% justify the content of an AI-written document to someone else. I always felt on the verge of \u0026ldquo;getting caught\u0026rdquo;, as if I were a small child. And it\u0026rsquo;s impossible for me to own it, as it is impossible for any self-respecting person to say \u0026ldquo;yeah this is badly written and does not represent my position, but assume I write and think this in this way. It\u0026rsquo;s faster\u0026rdquo;\nDitching AI is part of a process of getting better standards for oneself.\nLLMs produce both reading- and writing-fatigue, even if you write less. This is common theme in the modern world: we are understanding that stress and \u0026ldquo;burnout\u0026rdquo; come from doing meaningless things, not from doing them \u0026ldquo;a lot\u0026rdquo;. That\u0026rsquo;s because LLMs writings have so many filler words, meaningless or nearly meaningless locutions like \u0026ldquo;security boundary\u0026rdquo;, \u0026ldquo;It\u0026rsquo;s not X, it\u0026rsquo;s Y\u0026rdquo;, \u0026ldquo;invariant\u0026rdquo;, \u0026ldquo;load bearing\u0026rdquo; etc. You either change all of them or you present a badly written document. If you read LLM generated documents/content for a living, you will lose taste, start skimming documents without reading and will present a document with a load bearing assumption that does not make sense. If you have to edit them, you edit stupid parts that shouldn\u0026rsquo;t be there in the first place. Can\u0026rsquo;t LLM just stop using them?. Automation is an everlasting drug.\nSure it\u0026rsquo;s usefult for coding. I don\u0026rsquo;t code much to be honest. And even there, I always feel embarassed when I don\u0026rsquo;t know parts of what I created. It\u0026rsquo;s not like using a library for which you are supposed to not reproduce the internals, though even there you should know how it works. It\u0026rsquo;s a different feeling of \u0026ldquo;I should know what this piece of code does and why it\u0026rsquo;s here\u0026rdquo;\nI have a deep feeling that LLMs/agents speed things up inefficiently, mainly for people that don\u0026rsquo;t know the field. Overall they are useful for the general population since generally nobody optimize his tools anymore, and nobody is an expert in everything. Consequently for each person there is a field where AIs are better. \u0026ldquo;Better\u0026rdquo; does not mean professional level though.\nThe LLM talk is weak. it\u0026rsquo;s aligned, but aligned to the will of governemnt and capitalism based companies. Both appeal to the masses. Have the masses expressed any strong opinion ever? It\u0026rsquo;s impossible by definition, as \u0026ldquo;strong\u0026rdquo; means outside of the mass thought. Will you change or even be able to identify long term all weaknesses in \u0026ldquo;your\u0026rdquo; documents?\u0026#160;\u0026#x21a9;\u0026#xfe0e;\n","permalink":"https://disnocen.github.io/posts/2026-07-14-gynw1py3ja/","summary":"Turns out editing badly written things is more time consuming than writing them in a good way from the start","title":"I am not using any AI anymore and I more productive"},{"content":"The following is part of a message (anonymized) that ultimately I did not send. It is worth thinking about it though\nMy implicit assumption when I started was \u0026ldquo;AI all the things\u0026rdquo;, with the goal of getting the process and end results of cryptography research served perfectly at the push of a button. My vision was that those results would be fed back to another AI tool to create the end product (either code or robotic if material). That would then spark new research questions that a human needs to abstract and create a research-prompt for. And the cycle continues.\nConsequently, to me at least, the \u0026ldquo;human in the loop\u0026rdquo; value-add is in the judgement and distillation/synthesis of a worth-while research endevour. The human has to evision what he wants to obtain, think deeply if the goal is a \u0026ldquo;good\u0026rdquo; one (both profitable and moral) and then explain it to the machine which will find a way to do that.\nI never thought of the consequences of this vision: As labor becomes cheap (AI agents cost much less than humans for now) the \u0026ldquo;profitable\u0026rdquo; bar lowers and so in the end we get that a good goal is a moral goal as anything is profitable.\nFor the record, I do know this is naive and wrong. That this was the basic premise of the scientism of the Enlightment period that created terrible wars and the technology that enslaves us all (\u0026ldquo;we now have the technology to centrally plan agriculture so that there are no famines anymore\u0026rdquo; killed 100millions people in communist Russia. Some of them, due to famines ironically enough).\nThis is what is causing me to stop and consider: would I really get something out of using and understanding (read: trying the latest tools) the LLM/Agentic AI of today? I was bullish 3 years ago. Now much less so\n","permalink":"https://disnocen.github.io/posts/2026-07-11-wdqwjrc9xz/","summary":"Not in a good way","title":"On the AI as the new Enlightnment"},{"content":"This is how A simple script can improve your perceived IQ\nJokes aside, message platforms UI is the worst. Sure, it comes from the IM (instant message) mindset of short messages; long ones are for emails. But that is not the case anymore, and my slack messages are becoming longer and longer 90% of the times. (this is true for the XMPP ones too btw, as messages have replaced emails for the SAME goal)\nIf communication is asynchronous, you need to communicate effectively: the burden must be on the writer, not the reader. That\u0026rsquo;s because you (as writer) want to convey you are capable and you understand deeply the matter at hand. A bad UI can make you frustrated so you won\u0026rsquo;t be as effective as you could.\nEnter vitmp. I don\u0026rsquo;t remember where I saw it but it basically is a bash script such that:\ncreates a temp file XYZ.md opens a new term with $EDITOR opened at XYZ.md let you edit the file when you close the the $EDITOR it copies the file content on the clipboard This improves communication 100X I think. The reason being that every message is now an email: you have time to think, modify, with the full power of an editor (such as vim). Then you Ctrl+V in the text field and press Enter\nTry it and look how easier it will be to communicate and think from now on\nInB4 \u0026ldquo;what about mobile?\u0026rdquo;\n","permalink":"https://disnocen.github.io/posts/2026-05-26-3ojsftibvc/","summary":"A simple script can improve your perceived IQ","title":"`vitmp` is best"},{"content":"I must admit that I fall fot the knowledge base scam: the idea that we can delegate to the AI the management of \u0026ldquo;knowledge\u0026rdquo;. But on the flip side, I think I am one of the first to notice that that is a scam. Let\u0026rsquo;s see why\nwhat this NOT about I am not saying LLM/agents are not useful, that we will never reach AGI (as Professory Barry Smith seems to say), that \u0026ldquo;good enough\u0026rdquo; is not good enough etc\nAnd I am not saying that LLM/agents will solve everything either and remove all jobs and that is so powerful we need to be scared, like like Dario Amodei seems to say\nthe second brain idea Since Evernote, or even org-mode!, we felt we need to manage knowledge (we just took notes before, no need to \u0026ldquo;manage\u0026rdquo; beside having a file cabinet). In principle it sounds right as you can\u0026rsquo;t remember anything and there has been an exponential (in quantiti, not necessarily in quality) amounte of things to know. the problem is when you think you can delegate things to the physical world. 1\nYou would never treat a series of physical notebooks and sheets of paper as \u0026ldquo;second brain\u0026rdquo;, and yet you think the digital version of that can be.\nWhile an academic definition of knowledge is complex to give, you can have a pretty good actionable definition by saying:\nknowledge = data + metadata + experience\nNotes are cool if you are not good at retaining data, and writing is an important part for experience, but that\u0026rsquo;s it\nA SeCoNd BrAiN ThAt WoRkS pre LLMs, Second brains never worked because:\npeople get tired of bad UX and abstract management (feeling the sheet of paper is less tiring, and if you don\u0026rsquo;t need it you just trash it) there really is no increase in knowledge by writing notes a self fullfilling profecy\nEnter the LLM based knowledge base.\nWhat\u0026rsquo;s the knowledge base essentially: a directory with structured sub-directory + index and markdown files with notes. Notes are generated by the LLM for the LLM. One suggestion is to use Obsidian as the front end of the note to curate them.\nThe LLM/agent treats them as CRUD (create, read, update, delete) db of post-its as part of its operations.\nNow you have LOTS of DATA and METADATA but you don\u0026rsquo;t have experience, and henge knowledge\nQED\nHow will I proceed from there Will just continue taking and tossing notes, printing papers and create proof of concepts. Will use the LLM/agent for the PoC, but only delegating those things I already know, and as better google. That\u0026rsquo;s it\nmy idea is that we thought that, since it worked for the calculator, it should work for everything. Let me specify a little: the calculator works because arithmetic is compressible, ideas are not. so you can delegate arithmetic to a calculator, but you can not delegate \u0026ldquo;ideas\u0026rdquo; to a machine. Also, solving math problem works because you have defined rules (that you already know). Defining the rules is an entire different thing (unless there are rules for rules etc)\u0026#160;\u0026#x21a9;\u0026#xfe0e;\n","permalink":"https://disnocen.github.io/posts/2026-05-22-kid1uoisio/","summary":"Never trade understanding for speed and quantity. understanding is speed in the medium/long term","title":"Reflection on the use of an AI-managed knowledge base"},{"content":"I played with local AI, but now I wonder:\nisn\u0026rsquo;t decentralized AI better?\nLocal AI is centralized, albeit on you\nwaht is DecAI You clearly want open models so while the following can apply to frontier and proprietary models, they are not my concern now 1\nOpen models run on a lot of (inference) providers. why don\u0026rsquo;t we just optimize for it? Local AI is just my inference provider anyway, it\u0026rsquo;s not free as in free software/speech and certainly not free as in beer. a good model is as costly as it gets and you will not likely a good quality from it.\nThe stack:\nmy computer open harness decentralized inference open modelS The Cons (I can think of) The only reason why you wouldn\u0026rsquo;t have external inference is privacy, but you can anon your account for personal secrets (ofc they can still \u0026ldquo;correlate\u0026rdquo;, but adulthood is tradeoffs understanding and acceptance ) what about company secrets? well, how possible is for the inference providers to capitalize on your requests, expeciallyif the work will be open sourced/published I don\u0026rsquo;t think censorship is a real problem, as the availability of inference providers improves (again going into a renting mindset instead of owning mindset 2) And btw, what\u0026rsquo;s a open model anyway? how do open weights compare to reading the source code directly if you don\u0026rsquo;t access to the training source?\u0026#160;\u0026#x21a9;\u0026#xfe0e;\ntinyboxes will be interesting to evaluate further along the way\u0026hellip;\u0026#160;\u0026#x21a9;\u0026#xfe0e;\n","permalink":"https://disnocen.github.io/posts/2026-05-21-vhilwgby7i/","summary":"You don\u0026rsquo;t want local AI, you want privacy and \u0026rsquo;not getting screwed\u0026rsquo;. You can have that cheaper by using a portfolio of inference providers","title":"Decentralized (non Local) AI"},{"content":"This blog is an experiment in writing. I tried to write on X and \u0026ldquo;interact\u0026rdquo; or \u0026ldquo;engage\u0026rdquo; with other people and a couple of slips through the \u0026ldquo;engagement thing\u0026rdquo; made me into politics hell (swiss nationalism to be precise) and I never recovered despite explicitly interacting only with Local AI and Bitcoin accounts/topics. I consider the experiment a failure.\nBut writing fast has been good for me, helped shape thought fast plus try and discard ideas easy and fun\nI designed this blog to incentivize writing while disincentivize rumination. so there is no expectation for engagement (no comments) and automatic publishing (a bash script opening vim and git pushing after closing it). The time spent writing is the total time spent on the blog 1\nThis whole post is actally a test of the bash script and this is a test of the markdown notes\u0026#160;\u0026#x21a9;\u0026#xfe0e;\n","permalink":"https://disnocen.github.io/posts/2026-05-21-gu0cpvyh2g/","summary":"Why I optimized for fast blog post creation instead of curation","title":"Fast iteration  is deep thinking "},{"content":"The second thread of this blog is local AI model experimentation. The emphasis is on models as research infrastructure rather than generic chat: retrieval over a private research vault, terminology normalization, structured decomposition, and tool-backed verification.\nFor cryptography work, the interesting question is not only whether a local model can answer a prompt. It is whether it can help maintain a loop: read the existing notes, preserve the threat model, generate a checkable artifact, run the tool, and revise based on concrete output.\nThat makes local model posts a natural companion to the cryptography notes. The model workflow is judged by whether it improves the research process.\n","permalink":"https://disnocen.github.io/posts/2026-05-01-local-model-experiments/","summary":"Notes on using local models as research tools, especially for cryptography workflows.","title":"Local Model Experiments"},{"content":"This is a personal research blog for notes that are useful outside a private notebook.\nThe main threads are cryptographic protocol design, garbled circuits, threshold signatures, BitVM/STARK/SNARK work, and experiments with local AI models for research workflows.\n","permalink":"https://disnocen.github.io/about/","summary":"\u003cp\u003eThis is a personal research blog for notes that are useful outside a private notebook.\u003c/p\u003e\n\u003cp\u003eThe main threads are cryptographic protocol design, garbled circuits, threshold signatures, BitVM/STARK/SNARK work, and experiments with local AI models for research workflows.\u003c/p\u003e","title":"About"},{"content":"This blog will collect public-facing notes from ongoing cryptography research. The center of gravity is practical protocol design: what assumptions a construction relies on, where the threat model is doing real work, and which details matter when an idea moves from a paper sketch into an implementation.\nInitial topics will include garbled circuits, threshold signatures, symbolic verification, BitVM-style protocols, STARK/SNARK integration, and security tradeoffs in bridge designs.\nThe goal is not to turn every note into a polished paper. The useful format is smaller: a claim, a construction, an experiment, or a design decision that is worth making legible.\n","permalink":"https://disnocen.github.io/posts/2026-05-01-cryptography-research-notes/","summary":"A short map of the cryptography topics this blog will track.","title":"Cryptography Research Notes"}]