Daniel Coimbra

Pretty cool system-one model my company just launched. To get traction, it offers 1B free input tokens for the first 10K signups!

About burnout & agentic coding We're all adults. We all learned to protect ourselves from thousands of real-life stressors There must be a healthy way. Let each of us work hard our metacognitive abilities! Addiction/ADHD might be the hardest aspect to solve w.r.t. Codex/CC...

Codex tabs are almost browser tabs, now. Local wonderlands of code, specs, and architecture, ready to be built and discovered.

Having an eerie experience using GPT 5.6 Sol for the first time *via codex cli*. Kinda late to the party. First thing I did: Configure its harness. Auto-allow, destructive command guard, agent delegation, docs access, code review skill, etc. What I found eerie is how well Sol performed. He asked me the right questions and led the way very well.

My career philosophy in face of coding agents. I am hired to solve problems related to data and tech. Each semester, as coding agents improve, I adapt to continue being "the guy that effectively leverages AI". Even as AI improves how it leverages itself, it can still be leveraged *more* under guidance by an experienced human. A human that knows the company, knows the human world, talks to people, directs efforts, assumes responsibility, thinks long-term, and so on. That's it. I don't think we will get to an organizational utopia where all companies become ultra-efficient with self-managing AIs. So I can always help.

Musk argues that AI alignment can result from maximal truth-seeking. Issue: Normative ethics is likely false. Meta-ethical eliminativism is likely true. As a species, we have an irresistible attraction to ficticious ethical codes. So do I. We must ensure our AIs work the same.

Hype: Found awesome use of Claude Code! The /goal is to figure out the main characteristics of 1520s Europe by decomposing them into a short dirtree structure. Each leaf dir gets a `summary.md` based on web research. These get collectively summarized recursively up to dirtree. The root `summary.md` is a comprehensive, compact, and legible narrative about the 1520s. I am hyped about how much I might learn by generating information artifacts that result from "/goals" of this kind!

Similar to how biological perception is often studied via perceptual illusions. Note: Fascinating how a normative concept (veridical/illusory) is a great heuristic to model a natural causal system. Natural selection blindly mimicks "proper functions" very well!

Opinions have three properties: content strength, confidence level, and affective stance. Content can be strong or weak. Confidence can be high or low. Affective stance can be arrogant or humble. These are independent, except that being unsure entails being humble. Consider an opinion about telekinesis. I have provided a table of possible opinion proprerties below. Two observations: 1. Most people associate a strong opinion with an arrogant opinion. This is a mistake, as the table shows. 2.Most people associate a weak opinion (such as agnosticism) as humble. This is also a mistake! Agnosticism can be arrogant if it considers non-agnostics to be presumptuous.

My philosophical take on Gödel's first incompleteness. === Six paragraphs of formal introduction === I will work with first-order PA because I'm familiar with it. It is a non-categorical theory: it has a so-called intended model we call ℕ, but it also has non-standard models with extra elements. This holds for any (strong) firrst-order theory. Gödel built: 1) A relation Proof(x,y) which is intended to encode that 'x' is a proof of 'y' (in PA). 2) A predicate Provable(y) ≡ ∃x Proof(x,y) with obvious inteded meaning. 3) Through a fixed-point argument, a statement 'G' such that G ≡ ¬Provable(G). Proof(x,y) and Provable(y) only mean what they seem to mean «in the standard model ℕ». For non-standard models, there are "fake" proof instances Proof(x,y) which the "∃" in Provable(y) quantifies over. For the standard model, G is really true and we can see it. However, PA cannot see it, because PA cannot distinguish the standard model among its other models. Therefore, while PA correctly encodes Proof(x,y) when x,y ∈ ℕ, it cannot capture ℕ itself, so its actual Proof(x,y) relation can be satisfied by fake proofs in non-standard models. As a result, PA cannot encode all and only the correct instances of Proof(x,y) in PA. This is a general truth about any (strong) first-order theory. This, at least, is the ordinary philosophical interpretation of the incompleteness results. First-order syntax cannot encode ℕ uniquely. For this we need second-order syntax plus "full semantics". Second-order syntax is characterized by the ability to quantify over subsets. Full semantics contains "all" subsets (of the syntactic quantification domain). Second-order theories with full semantics are categorical and ℕ is uniquely specified. Gödel's theorem does not hold. So much is agreed. === Actual philosophical take === My philosophical take begins with this observation. Second-order syntax is compatible with "non-full" semantics, which do not contain "all" subsets. Therefore, second-order syntax itself does not capture the notion of "all" subsets uniquely. Therefore, it also does not capture the notion of ℕ uniquely. Second-order syntax does not guarantee categoricity! And this is not fixed by going higher-order. I am under the impression that we have formal justification to conclude that formal syntax cannot capture ℕ uniquely. In fact, I think it cannot uniquely capture many notions, such as set membership, subset, all subsets, and least inductive set, which are all related in defining ℕ and full semantics. The "intended interpretation" of these notions cannot be spelled out in syntax alone. Another observation is that formal semantics is nothing but syntax. It amounts to translating one syntactic theory into another syntactic theory. For instance, a model-theoretic semantics for first/second-order is nothing but a «map» from PA into first-order ZFC. The "true" semantic aspect of model-theory depends on providing a semantics for first-order ZFC. This is not done by a further syntactic translation. (Although it could be, it would just raise the same issue again.) It is done through a pseudo-formal meta-theoretical version of ZFC. For instance, we use the meta-theoretical concept of "all" subsets of ℕ, with no formal definition, when defining the full semantics that make second-order PA categorical. This meta-theoretical concept is not defined. (Although it could be, if we provided a meta-meta-theory, but this would just raise the same issue again.) It is merely intuited. We think we now what we mean. Thus, any attempt at formal categoricity presupposes a merely "intuited" notion of categoricity. === The grandiose upshot === There is the under-explored possibility that our metatheories are NOT grounded in definite "intuited" mathematical objects, such as ℕ or its subsets. Perhaps there is no such thing as "the intended model". When we think about ℕ or its subsets, our thought is incomplete. Gödel's result is generally interpreted as showing that our mathematical thought exceeds any possible formalism. This can be flipped: the limits of formalism show that our thought is indefinite. I will complete this thought in the comments.

Organizing some thoughts. Cultures are shared patterns in human behavior, especially when caused by imitation and shared social incentives. While concrete and observable, noticing the patterns often requires a difficult abstraction. A further difficulty is analysing which patterns are "contingent", in one of two senses: i) Not objectively optimal, even given practical constraints regarding human emotions, social coordination, and material conditions; ii) Not even approximately universal to humanity, and instead quite bound to a time and place. We often ask ourselves why certain people behaved in certain ways. It is easier to suppose an individual element – a quirk, a limitation, a motivation, a belief – so this is often our default conclusion. It pays off to reason about whether there were more social ("structural") conditionings that led to that behavior. That's all!

In this lengthy response, I articulate my reflected illusionist stance on qualia. I agree that consciousness *is* computation, but object to the formulation that it *arises* from computation, if that means it is a non-computational entity being created. I view consciousness as a particular set of computational operations over a particular set of neural data structures. Such neural data structures encode an egocentric world model. Granted, it is "analog" computation. Its states and primitive operations are fuzzy and non-deterministic. Plus, some primitive operations are not "logical" but instead a physical dynamical evolution (which may or may not be classically computable). I find that this view illuminates four core philosophical questions. ====== Question #1: Does conscious perception involve direct awareness of public physical entities? – Answer: Yes. The aforementioned representations constitute direct awareness of public external objects. What this is means is that they track such objects. (Keywords for perceptual tracking: causal connection, sensitivity, reliability, constancy.) ====== Question #2: Does conscious perception involve private mental entities? – Answer: Yes, namely, certain neural data structures we call perceptual "representations". Their computational role is to represent the external world. What this means is that they are consumed by the various brain subsystems "in lieu of" external objects, i.e. to guide action/cognition with respect to such objects. ====== Question #3: Does conscious perception involve direct awareness of such private mental entities? – Answer: Yes, at least for humans. Our perception involves automatic reflexive awareness of our perceptual representations. This takes the form of a higher-order representation: a further neural data structure that tracks perceptual representations. This makes it possible for us to introspect about perception, i.e. to observe, talk, and think about it. ====== Question #4: Does conscious perception involve special qualities? – Answer: No, but here I add additional theory. Phenomenal qualities, such as color or pain, are feature encodings. Our perceptual representations are populated with such features. These features are subject to automatic semantic parsing in downstream processes. Perceptual features and their semantic parsing have evolved/developed in tight coupling to guarantee useful informativity. Let us assume we have a neurocomputational theory of semantic processing, so we know what it is for perceptual features to be "interpreted" or "used meaningfully". Suppose a toy model where colors are encoded in RGB. Certain RGB values are interpreted as an "intense red", a "blinding brightness", an "elegant composition", and so on. For whatever contingent reasons, our ability to introspect on such feature encodings is limited to what is encoded: interpreted qualities in feature space. We have no introspective access to the encodings themselves. As a result, we do not introspect on experience from an external point of view. Rather, our introspection on experience is just a more self-aware experience, it is a "view from inside". Phenomenal qualities seem sui generis, incomparable to anything else, and ultimately ineffable. The reason is that we think about phenomenal qualities using interpreted features. Meanwhile, we formulate theoretical reductions with structural descriptions, which is a different language. Phenomenal experiences are computational states that use a "first-person" or "subjective" grammar. This subjective grammar constitutes the view from inside, with which one describes virtual (ficticious) phenomenal qualities. Descriptions of computational states use a "third-person" or "objective" grammar. With this objective grammar we can describe the first-person grammar, and thus fully characterize phenomenal experiences. However, we will not however be *using* the first-person grammar used to virtualize special qualities.

Interesting formulation. I have been thinking about topics like this. Computers implement a physical dynamic that has stable macroscopic patterns. This stability provides the causal insulation Joscha mentions: microscopic perturbations do not disrupt such patterns, at least "ceteris paribus". We can describe this as modal robustness, or law-like regularity, or simply lawfulness. These macroscopic patterns implement what we might call "effective" objects, states, properties, and rules (operations, transformations). "Effective" entities exist when stable patterns are lawfully connected to each other. In the future, I want to develop this in more detail. Computers implement effective objects that function as discrete symbols (bits) and which can be composed into data structures (structs). Computers also implement effective operations that transform such objects (instructs & algorithms). In fact, a data structure is defined precisely by how it is operated over (i.e. its "functional" role in the computer). Brains also have lawful patterns, which we call "mind". I believe that biological minds can be also approximated as data structures governed by transformation rules. These we call mental states and mental operations. Mental states and operations are, of course, fuzzy and nonlinear in ways much more extreme than DNNs. Still, in most cases, brains implement the patterns we call belief, desire, concepts, emotions, personhood, and so on. I have been interested in this idea of "causal insulation" because I have been thinking about how the human mind can entertain "virtual" objects: formal languages, mathematics, algorithms, concepts, propositions, representations, world models, and the like. This is quite relevant for many philosophical discussions ranging from epistemology to the philosophy of math.