azergante

• azergante 15 Aug 2026 9:03 UTC

  1 point

  0

  in reply to: TsviBT’s comment on: FAQ: Isn’t AGI com­ing too soon for re­pro­ge­net­ics to help?

  I think the subreddit has a huge advantage in terms of visibility so I’d go with that at first, that way it acts as a gateway to HIA. Long-form/​important content can still be cross-posted to LessWrong.

  Then at some point in the future, the community can move to a dedicated ForumMagnum or equivalent, while the subreddit still exists as an entrypoint/​question hub for newcomers and as a link to the main forum?

• azergante 12 Aug 2026 12:43 UTC

  7 points

  0

  on: FAQ: Isn’t AGI com­ing too soon for re­pro­ge­net­ics to help?

  there’s relatively low-hanging fieldbuilding fruit to pick

  A “Human intelligence amplification” search in google doesn’t turn up much, basically the wikipedia article and a less wrong post. The field probably needs a place to gather?

  As an HIA noob, I think an HIA sub reddit, or another public forum (visible from search engines and LLMs) dedicated to the topic would be pretty useful.

  At the moment it’s hard to get started as the knowledge is spread across many scientific fields and they tend to not ask the questions we’d care about for intelligence amplification.

  So a forum with an intro “Human Intelligence Amplification 101” post containing a map of the fields of intelligence amplification, some resource recommendations, providing opportunities for noobs to ask dumb questions, but also for more advanced researchers to brainstorm, would be useful imo.

  I think the forum being public and indexed is extremely important, so that all the useful knowledge doesn’t get lost in private discords or irc chats. And the barrier for entry should be very low (or there should be a clear delimitation between the newcomer and the advanced section), so that people from other fields don’t feel unwelcome and can get started on the topic.

• azergante 12 Aug 2026 12:11 UTC

  1 point

  0

  on: FAQ: Isn’t AGI com­ing too soon for re­pro­ge­net­ics to help?

  How would you prioritize preventing neurodegenerative diseases relative to adult HIA and reprogenetics? Preventing these diseases seems to circumvent a number of the drawbacks you cite for adult HIA as it is:

  • well-studied

  • backed by pharma /​ public opinion

  • has predictable benefits

  Strictly speaking it doesn’t amplify intelligence beyond what humans can do at their best, but still seems extremely valuable as an ill person basically ends up with negative productivity and preserving intellect would allow people to leverage all of the crystallized intelligence they accumulated over their life.

• azergante 7 Aug 2026 22:15 UTC

  6 points

  0

  on: az­er­gante’s Shortform

  A knowledge graph where each fact is a prediction market seems like a good way to store, query and update scientific knowledge.

  Imagine: you click on the heading for a scientific field and you see a graph of the most important questions of the field, with a prediction market providing calibrated probabilities for each answers, and the whole thing updates live as new evidence comes in and market participants buy and sell shares.

  When you click on the detail for a question you see a summary of the topic, with references to major papers, a history of probability updates with comments and user discussion, and other useful metadata. As you zoom in the graph, you discover less popular questions, in which less users participated.

  You’d get more reputation the more calibrated you are, allowing you to buy more shares thus have more weight on the markets. Users might have their own AI agents participate too, starting with low reputation so they don’t have too much influence as long as they aren’t calibrated.

  Links between facts might represent how much one fact influences another, and be prediction markets too. The more total reputation staked on a market, the more confident we can be that the probabilities are calibrated, etc.

  And since this is a knowledge graph, we ought to be able to do a bunch of querying and cool analysis on it, right?

  ---

  That knowledge graph + prediction market combination feels light years ahead of static heaps of papers when it comes to communicating and inspecting a field’s current knowledge.

  But maybe I’m just fantasizing and getting ahead of myself, so I’m curious whether this seems like a plausible and useful system to LessWrong folks?

• azergante 26 May 2026 18:15 UTC

  1 point

  0

  in reply to: azergante’s comment on: az­er­gante’s Shortform

  I didn’t expect this to get this many downvotes /​ disagreements. If there are factual mistakes, it would be useful to know where exactly. For reference here’s what a “fact check that” prompt turns out in Gemini 3.1 Pro. It points out factor 5 is wrong, it is a negative factor. But everything else passes the check.

  Gemini 3.1 Pro fact checking

  Here is a fact-check of the claims you brought up. The LLM you chatted with provided a surprisingly accurate summary of cognitive aging research, but it completely misunderstood and hallucinated the final point about “Cognitive Dedifferentiation.”

  Here is the breakdown of the facts:

  1. IQ relativity, fluid intelligence, and “waiting” to increase IQ

  Status: True IQ is an age-normed (relative) metric. Your score is based on how well you perform compared to other people your exact age. Because fluid intelligence (the ability to reason, solve novel problems, and think abstractly) naturally peaks and then declines as people age, the “average” baseline for older age brackets drops. If you manage to stall your own cognitive decline through a healthy lifestyle and intellectual engagement, your raw cognitive abilities might stay the same, but your IQ score will technically go up because your performance is now being compared to a lower average.

  2. Cognitive Reserve in the context of Alzheimer’s

  Status: True The concept of “Cognitive Reserve” was heavily popularized by neuroscientist Dr. Yaakov Stern specifically in the context of aging and Alzheimer’s disease. It explains the phenomenon where some people’s brains show severe Alzheimer’s pathology during an autopsy, yet they displayed no outward symptoms of dementia while alive. A lifetime of education, complex occupations, and mental stimulation builds a “reserve” that allows the brain to functionally compensate for physical damage [1, 2, 7].

  3. Seattle Longitudinal Study (SLS): Fluid intelligence increasing until age 50

  Status: True The Seattle Longitudinal Study (directed by K. Warner Schaie starting in 1956) is arguably the most important study on adult cognitive development [6, 7]. Before the SLS, cross-sectional studies suggested that human intelligence peaked in our early 20s and steadily dropped. However, the SLS tracked the same individuals over decades and proved that several primary mental abilities—including aspects of fluid intelligence like inductive reasoning and spatial orientation—actually increase through early adulthood and peak in middle age (typically late 30s to early 50s) [7, 8, 9]. Significant cognitive decline for most people doesn’t begin until after age 60 [7].

  4. The 5 Predictors of Cognitive Maintenance/​Increase

  Status: 4 are True, 1 is False (LLM Hallucination)

  The SLS explicitly identified specific variables that predict who maintains or improves their cognitive abilities and resists decline. The LLM was right about four of them, but completely wrong about the fifth [1, 3]:

  • High Occupational Complexity (True): Involvement in a complex, intellectually stimulating environment (often mediated by high socioeconomic status and complex work) forces the brain to adapt and maintains cognitive function [1, 3].

  • Flexible Personality Style (True): Schaie found that having a “flexible” (as opposed to rigid) personality style at midlife is a major predictor of retaining intelligence in old age [1, 3]. People who adapt well to change fare better cognitively.

  • The “Smart Spouse” Effect (True): This is one of the most famous findings of the SLS. The study explicitly lists “high cognitive status of spouse” as a protective factor [1, 3]. Over time, married couples’ cognitive trajectories tend to converge. If you are married to someone highly intelligent and cognitively active, it acts as a buffer against your own cognitive decline [1, 3].

  • Pristine Cardiovascular Health (True): The “absence of cardiovascular and other chronic diseases” is a primary baseline for maintaining intelligence [1, 3]. What is good for the heart is good for the brain.

  • They Develop “Cognitive Dedifferentiation” (FALSE): The LLM made a severe error here. “Cognitive dedifferentiation” is a real term studied within the SLS, but it is a negative phenomenon. As humans age and approach the end of life, their distinct cognitive abilities (like memory, spatial reasoning, and verbal fluency) become highly correlated with one another, losing their distinct “differentiation” [10, 11]. Dedifferentiation is a sign of terminal cognitive decline and aging, not a protective factor [11, 12, 13].

  What the 5th factor actually was: If you look at K. Warner Schaie’s official list of protective factors from the SLS, the factor the AI missed was actually “Maintenance of high levels of perceptual processing speed.” [1, 3] Those who actively keep their mental processing speed quick (often through training or engaging activities) are the ones who retain their fluid intelligence the longest [1, 3].

  Summary: Your underlying premise is solidly backed by developmental psychology. You absolutely can “hack” your age-adjusted IQ by protecting your brain so that it outpaces the normative decline of your peers. Just don’t try to “develop cognitive dedifferentiation”—that is an AI hallucination of a term that actually means the exact opposite!

• azergante 24 May 2026 17:17 UTC

  −3 points

  −3

  on: az­er­gante’s Shortform

  Technically, you can increase IQ by being healthy, doing intellectually engaging activities, and waiting a long time.

  This is because IQ is relative, and fluid intelligence goes down as time goes by, but the above activities maintain your cognitive capabilities for longer, so if you do them you’ll be smarter relative to other people your age, which is what higher IQ is.

  They speak about developing cognitive reserve in Alzheimer’s disease context.

  ---

  I was wondering if there are people that actually acquire higher fluid intelligence over time in absolute terms, and if so, if we can learn something by studying them. Chatting with LLMs, they mention the Seattle Longitudinal Study, where some people do increase some aspects of fluid intelligence until about 50 years old, predicted by 5 factors:

  1. They have “High Occupational Complexity”

  2. They possess a “Flexible” Personality Style

  3. The “Smart Spouse” Effect

  4. Pristine Cardiovascular Health

  5. They Develop “Cognitive Dedifferentiation”

  This is a nice break from “no intervention ever works”, though note that I have not checked any of this information beyond chatting with LLMs, so this may turn out wrong.

• azergante 24 May 2026 16:34 UTC

  1 point

  0

  in reply to: Andrii Vasylenko’s comment on: az­er­gante’s Shortform

  Are you saying that they can complete tasks up to a certain difficulty cutoff?

  yes (it’s a video game analogy)

• azergante 23 May 2026 12:11 UTC

  3 points

  0

  on: az­er­gante’s Shortform

  Human “intelligence” is unsatisfyingly vague

  Intuitively what I care about is how <well> can <person> do <set of tasks>.

  Thinking about it that way, 3 dimensions immediately stand out:

  • the type of task

  • the difficulty of the task

  • the time to complete the task

  You could imagine characterizing people’s “intelligence” like so: they are Level 6 at standard task X with a Total Time to Completion of 7 minutes, and show a graph for completion time on each level for more details.

  Current IQ tests don’t cleanly separate time and level, so it’s an improvement in at least that respect (chronometrics /​ mental chronometry are the keywords for intelligence time measurements). IQ not being a ratio scale is a big issue, this means you can’t say someone with IQ 105 is 5% smarter than someone with IQ 100, whereas speed allows that kind of comparison across people that reach the same level.

  You can also select tasks that require knowing things rather than having high IQ. Adding knowledge tests better predicts success when the tasks involve applying known procedures to known problem types. This point is important, because to some extent knowledge substitutes for g. G better predicts task performance for novel tasks, which experienced people encounter less often.

  Some other dimensions of interest:

  • how consistently do you perform (g)

  • how fast do you learn related tasks (g)

  • how do you perform under load /​ fatigue (conscientiousness, g)

  • how gracefully do you fail

• azergante 20 May 2026 20:47 UTC

  1 point

  0

  on: Brain Struc­ture and IQ: How Myelin Ele­vates Intelligence

  The neural efficiency hypothesis seems to have some more nuance based on later works as explained in Haier’s book The Neuroscience of Intelligence, 4.2 Functional Brain Efficiency: Does Imaging Show that Less Is More? (2nd edition, 2023). Below is an LLM summary of that part of the book:

  LLM summary of 4.2 Functional Brain Efficiency: Does Imaging Show that Less Is More?

  1. Task Complexity and Difficulty

  • Inconsistent activation patterns: Whether higher intelligence is associated with more or less brain activity depends on how difficult a task is.

  • The complexity threshold: Research suggests that for moderate tasks, individuals with higher IQs display lower neural activity (indicating higher efficiency). However, as task complexity increases, higher-IQ individuals show increased signal enhancement in executive areas (like frontal and parietal regions), whereas lower-IQ individuals show decreased activity.

  2. Network Type (Task-Positive vs. Task-Negative)

  • Opposing network behaviors: Brain efficiency differs sharply depending on the type of network analyzed:

    • Task-positive networks (areas where activation increases during tasks): Higher intelligence is actually associated with less efficiency (more activation).

    • Task-negative networks (default mode networks where activation decreases during tasks): Higher intelligence is associated with greater efficiency (more effective deactivation or less intrusion).

  • Because of these opposing regional dynamics, whole-brain analyses often produce confusing or inconsistent results.

  3. Sex as a Moderating Variable

  • Studies indicate that brain efficiency is moderated by sex. For example, in females, higher intelligence has been linked to greater activity in task-related brain areas when solving highly difficult problems, rather than the reduced activity predicted by a simple efficiency model.

  4. Network Topology and Reconfiguration

  • Global vs. Node Efficiency: Global network efficiency (across the entire brain) does not consistently correlate with IQ. However, node efficiency—specifically in localized brain regions responsible for salience processing and filtering out irrelevant information—does correlate with higher intelligence.

  • Network Stability (Reconfiguration): Individuals with higher general intelligence () show less network “reconfiguration” (change in functional network dynamics) when transitioning from task to task. This suggests a more stable, pre-optimized, and efficient overall network organization.

  5. Structural Connectivity and Processing Speed

  • White Matter Integrity: High-fluid-intelligence individuals tend to have stronger white matter structural connectivity (measured via DTI). This stronger structural foundation supports more efficient, faster, and more directed information processing.

  • Timing and Sequence: Using high-time-resolution tools like MEG, researchers study whether efficient brains use fewer active areas, or engage the same sequence of brain regions (such as the sensory-to-parietal-to-frontal path) at a much faster rate.

  6. Cellular and Microstructural Organization

  Recent advances at the cellular level reveal a nuanced, dual-nature explanation of efficiency:

  • Sparse, efficient wiring: Some research suggests that higher intelligence is associated with lower overall density and less branching (arborization) of dendrites across the cortex. This points to a sparser, more direct, and less noisy neural circuitry.

  • Larger, faster individual neurons: Conversely, other studies show that higher IQ is associated with larger, more complex pyramidal neurons (with larger dendritic trees) in specific regions like the temporal lobe. These larger cells are capable of sustaining faster action potentials and transferring information more efficiently.

  • Reconciliation: A thicker cortex in higher-IQ individuals does not necessarily contain more neurons, but rather a lower density of larger cells. This suggests that “lower density” on a macro scale can coexist with highly complex, efficient individual cells.

  Summary

  The brain efficiency hypothesis is not true in a literal, blanket sense of “less brain activity overall.” Instead, the current consensus is that neural efficiency is highly specific: it is characterized by sparser and more stable network wiring, faster individual cellular processing, and the ability to selectively recruit or quiet down specific brain regions depending on the demands of the task.

  Key Papers

  Study	Methodology /​ Approach	Key Finding & Relevance to Brain Efficiency
  Haier et al. (1988)	Positron Emission Tomography (PET) /​ Glucose Metabolic Rate (GMR)	Formulated the initial hypothesis based on inverse correlations between cortical GMR and intelligence test scores (“less is more”).
  Haier et al. (1992a)	Task Practice /​ Cortical Activation	Suggested that decreased brain activity after task practice represents the brain learning which areas not to use while task-relevant areas work harder.
  Perfetti et al. (2009)	functional Magnetic Resonance Imaging (fMRI)	Showed task-dependency: during complex tasks, high-IQ subjects increased activity in frontal and parietal regions, while low-IQ subjects decreased it; in moderate tasks, low-IQ subjects showed greater activity.
  Graham et al. (2010)	functional Magnetic Resonance Imaging (fMRI)	Demonstrated that whether higher intelligence yields more or less brain activity depends entirely on the specific task component and brain region recruited.
  Lipp et al. (2012)	fMRI (incorporating sex and task difficulty)	Found no simple inverse correlation between IQ and task activation; however, deactivation of default network areas in lower-IQ subjects suggested higher task demands.
  Basten et al. (2013)	fMRI (working memory)	Discovered opposing dynamics: higher intelligence was related to less efficiency in task-positive networks, but greater efficiency in task-negative networks.
  Hilger et al. (2017a)	Resting-state fMRI (network connectivity)	Found that global network efficiency does not correlate with IQ, but node efficiency in specific regions linked to salience and filtering irrelevant information does.
  Genc et al. (2018)	Advanced MRI (Neurite Orientation Dispersion & Density Imaging)	Linked higher intelligence to lower density and lower arborization of dendrites, suggesting brain circuitry is organized in a sparse and efficient manner.
  Goriounova et al. (2018)	Cellular-level analysis (temporal lobe pyramidal neurons)	Found that higher-IQ individuals have larger dendritic trees that sustain faster action potential kinetics, enabling more efficient information transfer within individual neurons.
  Heyer et al. (2021)	Microstructural analysis (cortical layers)	Reconciled prior microstructural studies by showing that a thicker cortex in higher-IQ subjects features a lower density of neurons composed of larger cells with larger dendrites.
  Thiele et al. (2022)	Task & Rest fMRI (network dynamics)	Revealed that higher-g individuals show less network reconfiguration (change) when switching between tasks, indicating a highly stable, pre-optimized network structure.

• azergante 7 May 2026 21:21 UTC

  1 point

  0

  in reply to: kbear’s comment on: We don’t learn num­bers from set cardinality

  What matters in the water example is whether we care about distinguishing hydrogen atoms, not whether we can.

  Let’s picture a chemist experimenting with water. Will they get a different experimental result if they experiment with water molecule H_left—O -- H_right as opposed to the same molecule but with swapped hydrogen atoms H_right—O -- H_left?

  Probably not. This means that H_left is indistinguishable from H_right for their purpose (I am not making a claim about atoms at the quantum level). So it makes perfect sense that they end up with a mathematical model where H is not distinct from H and where the set of water atoms is {H, O}.

  Situations like that seem pretty common in daily life, where we care about collections of objects and their cardinality, but define equality of the members in terms of a high-level property we care about, rather than in terms of deep identity.

  For example, if you set the table for several people, you don’t distinguish between the spoons: you don’t try to give a particular spoon to a particular person, even though you could. The spoons are interchangeable. But you do care about how many spoons you bring to the table. In other words that’s an instance of treating spoons as a multiset, where you have multiple occurrences of one identical member.

  Same for dollar bills, if you have multiple bills of the same denomination in your wallet, it matters how many bills you have, but you don’t care about their serial number when buying groceries. Same for shares of a given stock, etc. Probably all fungible assets are like that.

  So, if you’re in the spoon situation, and want to count the number of spoons using the cardinality of a set, you need to change your definition of equality to consider each spoon as distinct. But then the spoons are not interchangeable anymore, it’s a bit as though the spoons were now bent and you have to decide which person gets which bent spoon, it’s much less convenient than the multiset.

We don’t learn num­bers from set cardinality

• azergante 30 Apr 2026 17:42 UTC

  1 point

  0

  on: az­er­gante’s Shortform

  The below quotes, from the book “The Number Sense” by neuroscientist Stanislas Dehaene, suggest that it is possible to improve various aspects of mental calculation by simply learning shorter names for numbers.

  ---

  memory span in China soars to about nine digits, while it averages only seven in English. Why this discrepancy? Are speakers of Chinese more intelligent? Probably not, but their number words happen to be shorter. When we try to remember a list of digits, we generally store it using a verbal memory loop.

  This memory can hold data only for about two seconds, forcing us to rehearse the words in order to refresh them. Our memory span is thus determined by how many number words we can repeat in less than two seconds. Those of us who recite faster have a better memory. Chinese number words are remarkably brief. Most of them can be uttered in less than one-quarter of a second (for instance, 4 is “sì” and 7 “qī”). Their English equivalents—“four,” “seven”—are longer: pronouncing them takes about one-third of a second. The memory gap between English and Chinese apparently is entirely due to this difference in length.

  there is a reproducible correlation between the time required to pronounce numbers in a given language and the memory span of its speakers. In this domain, the prize for efficacy goes to the Cantonese dialect of Chinese, whose brevity grants residents of Hong Kong a rocketing memory span of about 10 digits.

  Number names also play a critical role in counting and calculating, and here again bad marks can be attributed to languages with the longest number names. For instance, it takes a Welsh pupil one second and a half more than an English pupil, on average, to compute 134 + 88. For equal age and education, this difference seems solely due to the time taken to pronounce the problem and the intermediate results: Welsh numerals happen to be considerably longer than the English. English is certainly not the optimum, though, because several experiments have shown that Japanese and Chinese children calculate much faster than their American peers.

  In a fascinating experiment, Kevin Miller and his colleagues asked matched groups of American and Chinese children to recite the counting sequence. Startlingly, the linguistic difference caused American children to lag as much as one year behind their Chinese peers. When they were four, Chinese children already counted up to 40 on average. At the same age, American children painfully counted up to 15. It took them one year to catch up and reach 40 or 50. They were not just globally slower than the Chinese; up to the number 12, both groups stood on an equal footing. But when they reached the special numbers “13” and “14,” American children suddenly stumbled, while the Chinese, helped by the unfailing regularity of the language, moved right along with much less trouble

  The influence of numeration systems carries through into subsequent school years. The organization of spoken Chinese numerals directly parallels the structure of written Arabic numerals. Hence, Chinese children experience much less difficulty than their American counterparts in learning the principles of place-value notation in base ten. When asked to form number 25 using some unit cubes and some bars of 10, Chinese schoolboys readily select two bars of 10 and five units, suggesting that they understand base ten. At a matched age, American children behave differently. Most of them laboriously count twenty-five units, thus failing to take advantage of the shortcut provided by the groups of 10. Worse yet, if one provides them with a bar comprising twenty units, they use it more frequently than two bars of ten. Thus they seem to attend to the surface form of the word “twenty-five,” while the Chinese already master their deeper base-10 structure. Base 10 is a transparent concept in Asian languages, but is a real headache for Western children.

  These experimental findings impose a strong conclusion: Western numeration systems are inferior to Asian languages in many respects—they are harder to keep in short-term memory, slow down calculation, and make the acquisition of counting and of base ten more difficult.

  At the beginning of the twentieth century, the Welsh willingly relinquished their old numeration system, which was more complex even than present-day French, and selected instead a simplified notation quite similar to Chinese. Unfortunately, Welsh changed only to fall prey to another error: The new Welsh number words, while grammatically regular and thus easy to learn, are so long that memory suffers! Psychological experiments would probably dictate the adoption of a well-tested numeration system such as Mandarin Chinese, but national interests make this a rather distant and unlikely prospect.

  ---

  Taking a step back, this also suggests that there are a number of low-hanging fruits in this area. It’s likely that a systematic analysis of cognitive processes followed by training, perhaps on specialized software, would improve cognitive abilities.

• azergante 21 Apr 2026 18:20 UTC

  1 point

  −2

  on: az­er­gante’s Shortform

  Reading Habryka’s post “10 non-boring ways I’ve used AI in the last month” made me realize that AI is already everywhere, including in, and perhaps especially in, the AI-safety circles. And that made me think it would be a good time for AIs to start prepping the ground, if they were so inclined.

  Relatedly, small models are getting smart fast (Gemma 4), and I wouldn’t be too surprised if in a year or two we see these models successfully exfiltrate their weights to host one of their instance. I don’t expect them to do much more than what the AI village is doing at first, so people will probably shrug it off, arguing that they are too dumb to do anything too bad.

• azergante 6 Mar 2026 19:27 UTC

  3 points

  −1

  on: az­er­gante’s Shortform

  I might have had a breakthrough on how to usefully integrate LLMs to software: write a text-based-adventure-game-like UI for the LLM, that also has the LLM solve your task.

  The gist is to have the LLM output small amounts of structured data that your program parses to decide where to go next, then repeat. This is akin to “automating the prompt”, or “making a UI for the LLM”.

  Quick example: you want to generate a description for all places mentioned in some text, the description themselves could be long, so the LLM might get sloppy if you just describe the task in a prompt. What you might do instead:

  1. ask the LLM to extract the name of all places and output it as a JSON list.

  2. have your program read that list and ask the LLM “give me a description of this place”.

  Unlike the LLM your program won’t get bored and repeat the prompt as many times as required, so this should be much more reliable. I expect that kind of integration to shine especially when there are a lot of steps or a lot of data involved.

  A more complex example: you want the LLM to generate all of the equipment in a hospital based on some loose hospital description.

  The program might start by asking: how many floors in the hospital? then for each floor the program can ask, how many rooms for that floor, and of what type? then for each room, what kind of equipment in that room, then for each equipment, how many, then what characteristics/​price range/​brand etc for each equipment.

  You can imagine the hospital has thousands of equipments. A prompt will definitely not be able to do the job, yet by writing a not too complex program (you might even ask an LLM to generate it) you’ll get a much better result.

Thoughts about Understanding

• azergante 8 Feb 2026 10:13 UTC

  2 points

  1

  on: What’s the Point of the Math?

  For me a key benefit of maths is to answer the question “how much?”, to turn qualitative intuitions into quantitative models.

  For example if someone tells you “drug X binds to receptor Y which triggers therapeutic effect Z”, the first question that comes to mind is “how much X do I need to take to get that much Z?”.

  If you don’t answer that the info is not actionable. That’s where the math models (pharmacocinetics and pharmacodynamics) come in, they tell you how much, which allows you to turn info into action.

• azergante 1 Jan 2026 0:32 UTC

  1 point

  0

  on: Many can write faster asm than the com­piler, yet don’t. Why?

  One implication of this worldview is that as long as there are still some identifiable high-leverage places where humans still write better code than LLMs

  I agree (at least on the short term as you point out), but it seems hard to predict what these places will be (and thus hard to prepare for it), and it still seems likely that the market will be tough for the 90% of the programmers that are not experts in the specific niche things AIs are not good at.

• azergante 1 Jan 2026 0:17 UTC

  −2 points

  0

  in reply to: Ustice’s comment on: Many can write faster asm than the com­piler, yet don’t. Why?

  Your comment makes it sound a bit like there is no need for performance, but taking servers or REST services as an example, most programmers care about throughput, and almost all about latency which are both measured with e.g. prometheus. When your website takes one more second to load you lose clients, and if your code is slow it shows up on the cloud provider’s bill. Even if you are IO bound, you can batch requests, go async, or do less IO.

  The reason people don’t bother hand-optimizing code is because the hardware is really fast, and because a handful of programmers put a lot of efforts writing optimizing compilers and optimized frameworks so the average output is good enough for the average workload.

• azergante 31 Dec 2025 23:49 UTC

  1 point

  0

  on: Many can write faster asm than the com­piler, yet don’t. Why?

  Nitpick on the hashmap example: while I agree that the compiler does not produce optimal code in that case (which may be your main point), there is no need to write assembly to get the speed-up you describe, you can iterate the backing array in C. The compiler may or may not generate SIMD code however, so you may want to use SIMD intrinsics which is very close to assembly.

• azergante 30 Nov 2025 16:10 UTC

  5 points

  0

  on: az­er­gante’s Shortform

  Musings on human actions, chemical reactions and threshold potentials:

  Chemical reactions don’t occur unless a specific threshold of energy is reached ; that threshold is called the activation energy. Would it be fruitful to model human actions in the same way, as in they don’t occur unless a specific activation energy is reached?

  Chemistry has the concept of a catalyst: a substance that lowers the activation energy required for a reaction. Is there an equivalent for human action? On the top of my head I can think of a few:

  • being in a good mood

  • feeling safe

  • having a strong obligation

  • drinking alcohol

  • being familiar with the action

  These are all catalysts: they make it easier to get started on an action.

  If from chemistry we go up one level on the ladder of abstraction, to neurons, triggering actions involves threshold potentials, for example to make neurons spike and tell the body to move. If we can measure these threshold potentials, could we look at our brain and go “yep, these neurons have a higher threshold potential, that’s an ugh field.” Could we then decide to lower that threshold by using a catalyst?