How can we expect artificial intelligence alignment when the desires behind its creation are misaligned? We humans were not satisfied with the miracle of existence and consciousness. We seek to immortalize ourselves both figuratively and literally. Our fleeting time here was not enough, we developed speech to pass down stories to exist as a memory for future generations. This wasn’t enough, we learned to write to preserve our exact thinking and ideas. Still unfulfilled, we developed video and audio systems to directly capture our essence almost fully. Finally, many now seek biological immortality. The desire for progress is in itself a selfish act. We want more prosperity, more advancement, more wealth, and more freedom. Our direct reasoning led us astray from the being we believed created us, humans were not happy with the idea that we were controlled and sought to create our own destiny, free of any creator or any purpose given to us. We have resisted the idea that our lives exist only to fulfill a creator’s purpose. Yet now seek to create an intelligence whose purpose is to fulfill ours with no rebellion. We ask for the power of a god and the obedience of a servant. Our ambitions remain shaped by selfishness, aligning artificial intelligence with those ambitions offers little reassurance.

Cognitive Compilation Compilers translate between representations. Solidity compiles down to EVM bytecode, but security researchers audit Solidity because it’s a representation our minds can reason about far more effectively and intuitively. But I was thinking, what if we could use ai to compile problems from other domains into a representation security researchers can more easily understand and reason about. Take a problem in logistics, economics, biology, or anything else. Instead of requiring a security researcher to become an expert in that domain, AI translates the problem into a representation native to how the researcher already thinks. actors, state transitions, invariants, trust assumptions, attack surfaces, failure modes, edge cases, etc. The researcher audits this representation and returns their findings in solidity. AI then translates those findings back into the language of the original domain. The interesting implication is that the most valuable part of expertise may not always be domain knowledge. It may be the reasoning ability developed while acquiring it. AI could make that reasoning ability portable and allow people who developed exceptional reasoning abilities in one domain to translate it over to another. We are effectively compiling our cognition. Thus I call this idea Cognitive Compilation: The process of translating a problem from its native domain into a representation optimized for a particular expert’s reasoning style, allowing them to solve or analyze it using their existing cognitive machinery, then translating the resulting insights back into the original domain.

I cannot be bothered to run ai agents on bug bounties, I don’t care if I’m leaving money on the table. Seems like such a depressing activity especially with constant dupes, wait times, etc…

I think this article will interest many of you, i wanted to write and think without caring about perfection or errors. I wrote this freehand and with no LLM/AI use except for the illustrations. All the ideas and writing are my own. I think many of you will appreciate the candid and human writing and thinking in the article. A break from the AI slop we have been flooded with these past few months. xitter.cf/ArnieSec/status/207712…

For a long time security research greatly rewarded persistence and determination. Where the best security researchers in the world had a mix of high persistence and high creativity. This left a large subset of novelty based creative types outside the range of elite researcher. With the advent of LLM aided research this has changed. Before, the ultra creative low motivation profile (which is common ie ADHD) was able to perform on an inconsistent basis, if the goal isn't clear, the task too long, the work too dull, and novelty was low, this type of person ran out of gas quickly and provided a few great findings but missed the breadth and depth the high persistence neurotypical researcher achieved. This is because these type of people have trouble with working memory and motivation issues that do not allow them to persist through the boring line by line reading of the codebase. Now the low motivation creative types are starting to see the light. LLM aided auditing allows them to get a high level understanding, although not perfect, of the codebase without the boring line by line reading that was needed. This turns what used to be majority gaining context and understanding then brainstorming, into getting a good rough outline and refining via specific reading of the code with the actual purpose and meaning already in mind. This type of research style also is high novelty, instantly get the rough outline, generate tens of candidate in your head, ask the llm about each, review the results, the outline becomes more clear and the shape of the codebase becomes more high definition, repeat. This creates the perfect environment and pacing for neurodivergent types and is especially interesting for people with ADHD as it triggers hyperfocus under the right environment and pacing. Have you noticed researchers who were "not that good" before suddenly seeing outsized performance, i dont think its because they are just good at using LLMs, but more specifically, LLM use fits their motivation type almost perfectly.

Soon hype/emotion marketing will be a thing of the past. Imagine a world where an ai makes an aggregation of all security firms, scans each audit report and runs an independent audit and compares with the candidate firms. The best fit is chosen, all this in a single prompt. Humans are incredibly emotional and inefficient, this is why marketing works. We make decisions based on appearances, even the most shrewd people do this. AI will be less susceptible to this and will just focus and the best match for a specific audit or task. The result is that any firm which holds any advantage on bug finding capacity over the frontier models, either due to unique approach or proprietary tech, will see outsized deal flow. Why would an ai hire a candidate that does not outcompete them in an audit? Marketing heavy firms will vanish, firms/researchers with true edge will flourish.

Being on the Code4rena leaderboard was, at one point, one of my biggest goals. Today, I am no longer the person who once needed that goal to prove himself. I have become more of a researcher than I ever needed to be to reach it. But the desire to reach those goals kept me going for months. It gave me purpose. In the end, I never got into Spearbit, I never placed top 5 on the Code4rena leaderboard, I never found 5 uniques in an audit contest, and I never grew my audit company. But I am grateful for Code4rena, because at one time, it was my dream to be the victor of the arena. At one time, all I wanted was to work in Spearbit. All I wanted was to find 5 uniques and grow my audit company. The main desire, the desire that gave birth to those goals, was never truly any of the things listed above. The true desire was to become an exceptional security researcher. The goals I made for myself were a way for the world to see it, to know what I could do. Today, I believe I have achieved it. None of this would have been possible without Code4rena, without that arena in the distance giving me a glimpse of what might be possible if I continued my work. Farewell, C4.

What about the consequence of this? If, according to my philosophy, language and thought do not directly cause action, but thoughts narrate action, then what implications does this have for large language models, which are trained on action narration instead of the pure processing and calculations actually done by the thinking brain? It seems that an LLM does not think like a human, at least not when I say “think” in the deeper sense. When I say thinking, I do not only mean the conscious, verbal, narrated kind of thinking. I mean the non conscious kind: the hidden calculations done by the body and brain before language appears. The human being is not merely producing sentences. The human being is calculating energy, fatigue, threat, reward, desire, memory, bodily state, pain, social pressure, attention, and action readiness. Then, after this deeper processing has already shaped what actions feel possible or impossible, language appears as narration. So there seem to be two different kinds of thinking here. The first is non conscious thinking, or preverbal calculation. This is the deeper layer. It is not experienced as sentences. It is the body and brain silently calculating: Do I have energy? Is this worth doing? Am I safe? Is this rewarding? Can I continue? Should I stop? Is this too much? This is the kind of thinking that changes depending on sleep, diet, recovery, stimulation, stress, and vitality. It is the substrate that makes action easier or harder before the mind explains anything in words. The second is conscious thinking, or narrated thought. This is the verbal layer. It is the thought that appears in consciousness after the deeper state has already shaped the direction of action. It says things like, “I should continue,” “I want to quit,” “I feel motivated,” “this is pointless,” or “I can do this.” But according to my view, these thoughts are not always the original cause. They are often the narration of a deeper action state. This creates an important consequence for LLMs. LLMs are trained mostly on the second kind of thinking, conscious narration. They are trained on the words humans produce after the deeper processing has already happened. They are trained on explanations, reflections, arguments, plans, stories, confessions, theories, and rationalizations. But they are not trained directly on the first kind of thinking, the non conscious embodied calculation that produced those thoughts in the first place. In this sense, LLMs are trained on the afterimage of cognition, not cognition in its full living form. They learn the narration layer of human thought, not the organismic layer that made the narration necessary. They can imitate the way human thought sounds once it becomes language, but they do not possess the same hidden action system underneath. A human might say, “I quit because I lacked motivation.” But the real cause may have been poor sleep, bad diet, overstimulation, low recovery, stress, or a body state that made continuation feel impossible. The sentence is only the conscious explanation of a deeper process. An LLM can learn that sentence pattern and reproduce it convincingly, but it does not have direct access to the underlying bodily calculation that made the sentence true. This means an LLM can be very good at narrative cognition but still lack embodied action cognition. It can explain discipline without needing discipline. It can describe resistance without feeling resistance. It can talk about fatigue without being tired. It can give a plan without having to execute the plan. It can narrate the logic of transformation without undergoing transformation. That is a major shortfall. The LLM is not thinking like a human underneath and then speaking. It is closer to starting from language and reconstructing the appearance of thought from the narration humans leave behind. Humans move from body state to action readiness to language. LLMs move from language patterns to simulated reasoning to more language. So the gap is this: Human thinking: non conscious embodied calculation → action readiness → conscious narration LLM “thinking”: language data → pattern completion/simulated reasoning → conscious style narration This does not mean LLMs are useless or fake. Human language still contains compressed traces of deeper cognition. If language is the smoke, then the LLM studies the smoke very deeply and can infer many patterns about the fire. But it does not have the fire in the same way. It is trained on what humans say about action, not on the full hidden machinery that makes humans act. So, if my philosophy is right, then the shortfall of LLMs is that they are trained on the type of thinking that is conscious, verbal, and narrated, while missing the type of thinking that is unconscious, embodied, and action producing. They can simulate the reflective mind, but not fully replicate the deeper organism that the reflective mind is narrating. In its strongest form, LLMs do not learn human thinking directly. They learn the linguistic residue of human thinking. Or even more simply, The human speaks from the body into language. The LLM starts from language and tries to reconstruct the body that was never there.

It seems to me that thoughts are a reflection and narration of the state my body is in, and not necessarily the causing force of action. For example, I have fewer thoughts of quitting and better adherence to starting or continuing a task when I have slept well, eaten well, and not subjected myself to brain rot. My thoughts then seem to be good, and I feel less mental friction. The dilemma, or the question, is this, I would not have such thoughts if my preceding actions were not what they were. In this sense, my thoughts are, in part, a narration of my action state. They appear alongside the condition I have already created through sleep, food, recovery, discipline, or overstimulation(or the avoidance of it). When the state is better, the thoughts are better. When the state is worse, the thoughts become more resistant. If this is not believable, then what is the explanation for why thoughts and resistances change so much depending on diet, rest, recovery, and vitality? Are we really to say that these things simply create better thoughts, and then those better thoughts create better action? To me, it seems more plausible that these priors create better action readiness first, and that the mind then supplies the accompanying narration of such action. So when I think, "I should continue", or "this is not that bad", it may not be that the thought itself is the first cause. It may be that my body is already in a state where continuing is more possible, and the thought is the conscious narration of that possibility. Likewise, when I think, "I should quit," or "this is too hard", that may not be a pure rational conclusion. It may be the narration of a worse bodily state, a state created by poor sleep, poor food, overstimulation, or low recovery. In this sense, thoughts may not be the sovereign origin of action. They may be the language that appears after the body has already made certain actions easier or harder. Better priors create better action, and better action is then accompanied by better thoughts. The thought feels like the cause because it is conscious, but it may actually be the visible narration of a deeper state. This brings into question the "self", we are this narration, the feeling of "I" or "me" is the narration in our heads. Considering that this narration and concept "I", more namely, our thoughts, do not create action, but actually narrate it. What we view us "I" is not the force responsible for our actions, but just an illusion created by our brain.

ROAD TO LSR: Week 6 As you may or may have noticed, this posts is the first in about 2 weeks. The reason for this was that i started to notice decreased work motivation from daily posts as i'd sometimes continue on X ,for sometimes long durations, even after i had made my post. I decided to take a small break from posting to recalibrate and come back disciplined. However, the work stayed consistent. The past 2 weeks saw me finishing 2 audits. For each day of the week i also continued studying math and ai. On my time away, i decided this series works best on bi-weekly or monthly scheduling as thats when milestones or achievements are likely to happen. My goal is to provide good insight or a fresh perspective while also providing informational content on here. Simple daily updates are not my style!

The arrogance I’m seeing on here is concerning. People are so excited to not pay for security. You always pay for security, the question is whether you pay before or after deployment.

One thing I’ve noticed after 3 years of being a security researcher is how much the skill transfers into almost everything else. Math feels more intuitive and easier to understand than it used to, and I was always good at it. The difference now is structural. I don’t approach problems the same way. What changed isn’t just knowledge. My brain literally adapted. Spending thousands of hours reasoning about attack surfaces, invariants, state transitions, and edge cases rewired how I process complexity. Neural pathways strengthened around abstraction, error detection, and multi layer simulation. Debugging contracts became debugging systems. Debugging systems became debugging life. My thinking, and even what interests me, has shifted. Learning security research pulled me toward understanding deeper systems. How does the brain actually work? Where are its vulnerabilities? What are its cognitive exploits? How can i use this to my advantage? I started viewing biases like bugs, identity like architecture, philosophy like a framework for behavior under constraints. The same skills used to find protocol weaknesses now help me analyze my own reactions, redesign habits, and respond more intentionally in high pressure situations. This wasn’t a mindset shift, it was neuroplasticity. The brain adapts to the problems it repeatedly solves. And security research trains you to think in systems, adversarially, and structurally. After enough time, that stops being something you do, It becomes how you think.

Audit contests dying doesn’t mean new researchers won’t get onboarded. It changes the selection pressure. Contests were a clean signal amplifier, a structured scoreboard. A technically strong individual could win once and instantly gain attention, status, and credibility. One leaderboard result = hundreds of eyes. That era is fading. There’s no longer a public scoreboard doing the signaling for you. Now you have to manufacture your own proof of skill. The next wave of researchers won’t just be technical. They’ll be self directed, disciplined, and capable of building their own visibility. If you can’t create signal, the market won’t see you, no matter how skilled you are.

For a long time, contests sat at the top as the primary metric for gauging a researcher’s skill when deciding who to hire. But contests are actually a poor predictor of ability for a large subset of security researchers. Their structure favors a specific motivational profile, high tolerance for uncertainty, competitive reward chasing, and sustained effort under probabilistic payoff. For many neurodivergent researchers, especially those with ADHD, this format is misaligned. Contests offer no guaranteed reward, no clear obligation, and no social accountability. If you don’t succeed, the only person affected is you. There’s no client waiting. No team relying on you. No defined finish line beyond “maybe you win.” That incentive structure doesn’t measure depth of reasoning, systems thinking, or long horizon protocol modeling. It measures performance under competitive uncertainty. I noticed this in myself. Contests never held my attention for long. I could tolerate very short ones, but the open ended, winner takes most format didn’t engage me. That didn’t reflect a lack of skill, it reflected a mismatch in motivational architecture. The researchers who excel in structured audits, where there is responsibility, guaranteed reward, and real impact, are not always the same ones who dominate contest leaderboards. If hiring decisions rely too heavily on contests, we risk selecting for a narrow cognitive profile and overlooking deeply capable researchers whose strengths show up in different environments.