0xLoopTheory

Most people use Claude Code as a smarter autocomplete. I wanted to see what happens if you treat it as the operating system for how you study, rather than a tool you call when you are stuck. The setup: study plans and review schedules live as plain files that the model reads and updates. It does the organizing I would otherwise skip. I do the thinking it cannot. Written up here: hackernoon.com/how-i-turned-c… I am not claiming this beats a good textbook and a quiet room. I am claiming the bottleneck in self-study is rarely the material. It is the system around the material, and that part is automatable. Where does this break down for you? Specifically, what part of studying do you believe a model should never touch?

8/8 The honest catch: building it is satisfying enough to eat the studying it is meant to support. It cannot sit the exam for me, cannot care about the outcome. The point of all of it: the hour I do read the page is the hour I needed to read it. encryptorium.medium.com/how-i-turned-c…

7/8 None of it is subject-specific. I retargeted the whole thing from CISSP to TOGAF in one afternoon by swapping the domain list and the source material. The subject matter is a parameter; the pipeline is not. It has since carried PQC and zero-knowledge reading too.

1/8 Studying anything hard is the same loop: read it, work problems, notice what you missed, retry the misses tomorrow. After enough of it the question is not whether the loop works. It is whether the hour has to be spent re-reading what you already know.

Passed the TOGAF exam today.

@dev10TLC That's a great approach, and it includes some elements I also use.

I am currently in a similar scenario but I tend to limit my AI dependence to certain scenarios for example, when learning, I try to grasp every concept by myself first and go as deep as I can. Once I truly feel I have exhausted my capacity on a certain problem, I try to structure my thoughts into detailed notes and dive deeper into those to spot any gaps. Once this is done, I plug the input into an LLM to get a different view on my summary and potentially where my understanding may have been incorrect. Then I find where the gaps in my notes where and plug them and more importantly, understand why the gaps occurred in the first place. This is more time consuming but there is no substitute for being able to understand problems quickly and developing my own process. The quicker and more efficiently I can do this, the less dependence on AI is required as a from time to time, the LLM will hallucinate so I don’t tend to trust it completely either. I treat it as a more of an enhancement tool

I feel like I currently live in two worlds at once. In my day job, I work in cybersecurity, and I keep AI on the edge of my workflow by choice. In my private life, research, writing, and side projects, I am deep in it. AI is part of how I think, build, explore, draft, test ideas, and move faster. And it splits me in two. On one side, I love what I am doing. AI enhances it. It helps me go deeper, connect dots faster, and turn vague ideas into something concrete. For research, writing, and learning, it often feels like having a second brain next to me. On the other side, it is strangely refreshing to keep some distance from it in my day job. There is value in doing the work yourself. There is value in being slower, more deliberate, more skeptical. Especially coming from cybersecurity and cryptography, I still deeply believe in being in the driver seat. But then the conflict comes back. I get bothered by tasks where I know AI could help. Repetitive writing. Summaries. Structuring thoughts. Drafting documentation. Searching through complexity. The kind of work where I can almost feel the wasted time because I know another mode of working exists. At the same time, not all AI output is satisfying. Speed is not the same as quality. Assistance is not the same as understanding. And using AI everywhere can slowly blur the line between amplification and dependence. That is the tension I am sitting with. I do not want to reject AI. I also do not want to outsource my judgment, my craft, or my responsibility. I want to use it well. Deliberately. With taste. With control. With security in mind. But I am still figuring out what that actually means in practice. Do you separate AI-heavy work from AI-light work? Do you feel the same conflict? Or have you already found a healthy way to integrate it without feeling like you are giving something up?

If you are studying something difficult and not using AI as a learning companion, I think you are leaving a lot on the table. Not as a shortcut. As a way to ask better questions, test your understanding, build examples, generate exercises, review your notes, and turn confusion into a concrete next step. My Claude Code workflow for studying has become one of the most useful learning systems I have ever used. It does not replace the work. It makes the work sharper.

A lot is lined up. The CISSP exam sits 38 days out, and it is priority 1 until then. That is harder to write than it sounds. I have spent the last year building everything that should make the next phase work. Content pipelines. Review cycles. The longer-form project I have been pacing toward for months. Tools that are still cooking. A sprint format I have designed but not yet started. None of it is moving right now. Some days the discipline feels easy. Some days it feels like watching a built engine sit cold while I study. The hardest part is not the studying. It is the waiting. But there is value in this kind of waiting. It forces clarity about what actually deserves to ship and what was just busywork dressed as ambition. Summer is when this list starts moving. The CISSP is between me and that. So I sit with it.

Yesterday we wrapped up our 14th edition of the zkSummit. Hosted in Rome, it was honestly one of our best yet. Thanks to everyone who came out, to the speakers, sponsors and team! 💜🇮🇹 All videos are coming throughout the coming week! #zk14

Already received some great feedback. Would love to have more eyes on it to refine the book. If you are interested or know someone who is, please let me know. And please share to reach the right people :) Thank you!

🎙 For ep 400, @AnnaRRose brings back @danboneh — Stanford professor & one of the sharpest minds in crypto. They cover Google's quantum algorithm announcement, why rushing PQ transitions might be riskier than quantum itself, algebraic vs hash-based signatures, hybrid sig schemes, new ZK advances, and the Ethereum Foundation's $ 1M Proximity Prize. A milestone episode. zeroknowledge.fm/podcast/400/

Worth reading if you are working on PQ migration that has to actually ship: arxiv.org/abs/2605.01454

A new paper builds something I have been waiting for: a validator for silent classical fallback in 5G cores that claim PQ-TLS 1.3, PQ-IPsec, and hybrid KEM support. The migration narrative is "we deployed it." The audit reality is whether the session actually negotiated ML-KEM-768 or X25519MLKEM768, or whether the path quietly ended up on X25519 because policy, compatibility, or a fallback path allowed it. You cannot audit what you cannot observe. And for HNDL, the uncomfortable part is that an adversary harvesting traffic today does not need to know that you fell back to classical key exchange. They just keep the ciphertext. I like the framing here because it moves PQ migration from "do we support the algorithm?" to "can we prove what was actually negotiated on the wire?" HNDL is usually framed as a clock on algorithm choice. This paper reframes it as a clock on observability. The window closes long before many teams have audited the fallback path. Caveat: preliminary results against QORE only. The validator detects which primitive was negotiated, not whether it is correctly implemented; TLS-terminating proxies can hide behavior; selective downgrade remains an open concern.