My Views on (Generative) AI

Last Updated: 2026-05-24

Summary

I do not make use of LLMs or other forms of generative AI for any of my projects.

This is an extreme stance that I impose on myself. I wanted to create this page to explain my stance for people who are interested. This page is intentionally not posted under my blog section, because I want to avoid it being interpreted as a statement that other people should have the same stance as me. I do have reasons for my stance, and it's possible that some people will find these reasons compelling enough to take a similar stance, but my primary motivation is to describe my own thoughts, not to convince others.

Empirical Observation

In general, LLMs have seemed to demolish the quality of everything they touch. Unfortunately, it is difficult to clearly identify LLMs as the culprit. For example, the tech industry experienced a significant tightening of budget at roughly the same time that LLMs started being adopted for development. There were many layoffs, some of which cited LLMs as a mechanism by which they would try to operate with fewer developers, but it is not easy to tell to what extent that is true versus an excuse to improve the optics of the layoffs.

The tech industry has been struggling with producing quality products for quite some time. These struggles unquestionably predate the adoption of LLMs. However, the rate at which software quality drops appears to be larger at the companies that have adopted significant LLM use in development. I have some theories about how these effects may have come about, which I will describe later.

The effect on prose is much more clear. There's a type of person who tries to give off the impression that they are knowledgable about a topic, but often there is some aspect of the way they talk that betrays that there are important subtleties of the topic that they aren't even aware of. Unfortunately, if their interlocutor is not an expert in the topic, they will also be unaware of such subtleties, and the deception is likely to be successful. LLM written prose generally gives me the same impression as these people.

LLMs have also made it much more difficult to find non-LLM written prose to read. When browsing Reddit and Hacker News, I already often did not open the actual article because the comments were more interesting. However, in the past sometimes the discussion would revolve around a point the author made and I would go read the article in those cases.

That has changed recently, and I have started actively avoiding opening linked articles because of the prevalence of LLM-written prose. It is unfortunate, because although not opening the article and responding to the general topic indicated by the topic is a common behavior, it is not really in the spirit of the sites.

Reddit also has a moderate number of LLM-written comments. Note that Reddit is not a monolith in terms of moderation and your experience can vary wildly depending on which subreddits you read. I have long since unsubscribed from nearly all of the default subreddits. It is probably well beyond a "moderate number" in those. The larger and more mainstream a subreddit is, the more attractive a target it is for astroturfing.

Hacker News now has an explicit policy against using LLMs to author or edit comments. I am sure that there are some people who are still using LLMs in violation of the policy, but even in that case, the policy forces them to put in more effort to make the LLM use less obvious, which improves the quality of the comments as a byproduct.

As the prevalence of LLM-written content increases, I find myself less inclined to continue engaging. My own experience indicates that there is almost certainly a transition toward a Dead Internet occurring. While I have not yet hit my breaking point to withdraw entirely from these communities, it is a near certainty that some other people have.

AI Capabilities

It is difficult to find any discussion about AI that does not have people claming large increases in productivity that they attribute to the use of LLMs. Although I suspect that some portion of these are parts of astroturfing campaigns, it appears that this is also a popular sincere opinion.

Additionally, non-generative AI has clearly resulted in some superhuman capabilities. Stockfish was superhuman at chess for quite some time with a combination of computationally intensive search and handcrafted heuristics to guide the search before neural networks were introduced.

It is believable to me that there are some aspects of software development in which LLMs outperform some software developers. They may even be better than the average tech worker.

In trying to square the capabilities of AI with the observations of output quality, I tried to find answers to two major questions:

  1. Is there some reason that may prevent AI from becoming superhuman at software development in the same way that it is superhuman at combinatorial games?
  2. If AI is boosting the output of individual software developers, why does it appear to have a negative impact on the output rather than a positive impact?

It is, of course, possible that AI is not actually having a positive effect on developer productivity, even when developers believe that it is a significant positive effect. There have been some efforts to objectively measure the productivity effects:

Their methodology is to select a task, then randomly decide to do the task with AI or without AI. If followed through consistently, the tasks done with and without AI should have identical distributions, allowing for conclusions to be drawn about the effect of AI without attempting to do any particular task with both methods, which would significantly favor the method done second.

As described in the second link, the data using late 2025 tools is much dirtier due to selection effects. The authors suggest that the selection effects should mean that AI has a more positive effect than measured, but I do not believe it is that clear. The quoted selection effects come from developers' impressions about the effect of AI on the experience of completing the task, and the earlier study shows that developers' impressions can differ significantly from reality.

Additionally, METR has noted in their other research that developers using AI will sometimes adjust the tasks that they work on toward tasks that are more suitable for AI. For example, they cite this as a reason that AI's impact on development speed diverges from its impact on project value. It is not out of the question that these substitutions have pervaded more generally into these devs' priorities, i.e. tasks chosen in 2026 are more suitable for AI than tasks were in 2025.

It's also possible that my observations about how AI has affected output are predominantly based on earlier iterations of AI tools that actually made individual developers worse (while they or their management already believed it was making them better), and that the output will start improving again with the new tools that seem more likely to actually be improving individual developer output.

Ground Truth Availability

In thinking about the fields where AI has had significant success, one aspect stands out to me as a key ingredient, which I call ground truth availability.

Ground truth refers to the idealistic concept of performance at a task. For software development it could be "write good code." For chess, it's "make moves that result in winning games of chess."

The availability of ground truth is how possible it is to use the ideal concept in a feedback loop. For chess and other similar games, the ground truth is readily available, because we can quite simply observe whether the AI won the game. On the other hand, the industry as a whole does not have a way to tell if something is "good code." The state of the art as I understand it is to use human feedback as a proxy (RLHF).

These seem to be fundamentally different situations. The neural networks in chess engines are able to produce better judgments about whether a position is good than you would be able to get from asking humans about those positions. The reason this is possible is that they can directly observe a large quantity of objective data: whether a given position led to a win during the training process.

Using proxies like human feedback means that the AI is incentivized to optimize for the proxy rather than the ground truth (e.g. if you applied RLHF to chess, the resulting agent might try to play for positions that humans find naturally appealing, even if those positions are flawed in some way that results in being less likely to lead to a victory). It also means that the judgment aspect of performing a task is in some ways limited to human level, since the reward model is directly trained to approximate the human feedback.

It is certainly possible for an AI system to reach superhuman capabilities without superhuman judgment, for example by applying an amount of search that would be unrealistic for a human. Pre-neural network Stockfish demonstrated that. However, it appears that in contexts with good ground truth availability, it can be done much more reliably.

Simpson's Paradox

Moving on to the second question: how can AI simultaneously improve individual developers and make the overall project worse? Simpson's paradox provides one possible explanation, which aligns with my observations of how people work with AI in practice.

The trend reversal from Simpson's paradox appears when the distribution between the subpopulations changes depending on a variable. In this case, this variable is whether a software project uses AI. Imagine a software project with two developers, "Good Developer" and "Bad Developer". Good Developer's code is better than Bad Developer's code.

Now suppose that Good Developer's code with AI is better than their code without AI, and Bad Developer's code with AI is also better than their code without AI. Simpson's paradox warns us in this situation AI can still lower the average quality of the code in the project, if using AI causes a larger portion of the code to be written by Bad Developer.

From the discourse that I've observed, it would be shocking to me if this Simpson's paradox effect wasn't happening. There are, by pretty much all accounts, some developers who are prompting AIs to write code and then submitting that code as a patch while doing minimal or no review of it. Skipping the review process naturally means that the process for these developers takes much less time than for others, so these developers can push out significantly more code than their colleagues.

An interesting aspect of this phenomenon is that Good Developer and Bad Developer don't need to be different people. Several developers have indicated that they use AI for tasks that they dislike, and do manual coding for tasks that they like. This is dangerous, since it means that the disliked tasks get accelerated more than the liked tasks, increasing their share of the completed task population, and disliked tasks are likely to get less discerning judgment and review. This style of work creates a prime environment for Simpson's paradox to manifest.

The Collapsed Software Market (Among Others)

Besides the Simpson's paradox effect, reduced output quality could be explained very simply if market forces are selecting for lower quality products. I personally have believed that this has been the case since some time before the adoption of LLMs, at least for consumer software. I have not had any chance to directly observe how business software sales operate, so I am not making any comment there other than to acknowledge that if it is behaving differently from the consumer market, then software companies operating there are subject to different market forces.

The way in which I view the software market as collapsed is similar to a market for lemons, although there are some slight differences in the mechanics.

In a working market, software of different quality should exist at different price points, and a consumer would be able to pay for higher quality. However, in the current market a consumer can't judge the quality of software before purchasing it. Descriptions of software products, even in many reviews, boil down to a list of supported features, with no way to discern whether those features will function reliably or if it will crash multiple times per day.

Plus, even if reviews did cover stability and reliability, the modern norm is for software to receive updates over the course of its life, which in the hands of the wrong team, will introduce just as many issues as they solve, if not more. On top of that, many reviews are non-genuine. Even without LLMs, companies would pay in various ways for positive reviews, be it sponsoring content creators, astroturfing techniques, or offering discounts to legitimate customers if they leave a positive review. LLMs exacerbate the issue by reducing the cost of astroturfing.

The venture capital funded startup industry does not do any favors for the market dynamics either. A large number of these startups function on a business plan in which they offer their product or service below cost, subsidized by the VC money. As the VC money dwindles, they are forced to balance the books, which follows one of a few paths.

  1. Raise prices to a level where they can make a sustainable profit.
  2. Cut costs of production to continue to compete on price.
  3. Monetize via means other than selling the software (ads).
  4. Get acquired by another company, who then needs to either continue subsidizing the product or follow one of the other options.

Options 2 and 3 involve degrading the product. While option 1 theoretically doesn't need to, the consumer base has been burned enough times by a company raising prices and then degrading their product afterward anyway (e.g. Netflix adding ads to previously ad-free subscription tiers), that they are likely to simply bail at the initial price hike.

All that to say, the market seems to select for companies that compete on price, and against those that compete on quality.

The incentives inside of the major tech companies are also famously not aligned with making quality products. Google employees optimize for the behaviors that are likely to get them promoted, which is launching something new, regardless of its long-term prospects. (I link the HN thread rather than the article itself because there are many more anecdotes about Google's promotion system in the discussion, but the article is also informative. It is written in the early part of 2022, before ChatGPT was released to the public, so it is extremely unlikely to be LLM-written).

Note that employees do not need to be intentional about this optimization. Selection is sufficient to create optimizing behaviors. The people who, intentionally or not, have behaviors optimized for promotion are more likely to get promoted, and people who get promoted are more likely to stay at their job. Over time, the population becomes filled with people who are optimizing for promotion.

Since Google, and likely other major tech companies, have gone through several generations of this selection process, it would not be surprising to see them adopt LLMs even if it makes their products worse, so long as it allows them to ship more products, and thus improve promotion prospects.

On the other hand, open source communities build around individual projects, and there is much less activation energy required to convince a contributor to either step away or switch to contributing to a different project. The tech employee would lose their salary if they did the equivalent. As a result, open source projects that rely on an open contributor community need to compete on the quality of the software being produced and the experience of contributing to the project.

And, indeed, it is not difficult to find high profile open source projects with AI stances ranging from complete rejection to heavy skepticism. I am not aware of any mainstream tech company taking a similar stance.

The Game Development Industry

As mentioned earlier, the game development industry appears to be operating under different dynamics from the mainstream tech industry. Although, I should clarify that there are multiple gaming markets that might themselves operate differently. I am writing about my observations of the PC gaming industry, as that is the section that I primarily interact with. The mobile gaming industry is, to my understanding, quite different, although there has been some merge recently as some games are getting released on both PC and mobile platforms.

The PC gaming industry is heavily centralized around Steam. It's not a literal monopoly, or anything, but my PC gaming is generally on titles that are distributed via Steam, and from what I've seen, that's true for many other people, too.

Steam limits the types of advertising that can be done in games. In particular, it bans advertising as the primary business model for a game, so the "ad market" style monetization that is prevalent in mainstream tech cannot be used.

Compared to standard tech products, it is much easier for a consumer to judge the quality of a game. There is a large subculture around producing video content of video games, either recorded or live-streamed. For anything other than the most obscure games, it is not difficult to watch another person play the game for many hours, in which time you can determine whether the gameplay looks enjoyable to you as well as whether there are performance or stability issues.

Furthermore, Steam has a refund policy that will accept a refund for any reason if the game has been played for under two hours. Although it is possible for issues to only appear after the first two hours, it is still possible to get refunds in some of those cases, just not without review.

These aspects force games that charge a sale price to compete, at least to some degree, on quality. An unenjoyable game will often come across as such in videos and live streams, leading to people skipping over it and playing something else. And even for the people who did not seek out video content before buying the game, they can use the refund path.

Besides charging for the game itself, the game industry has another successful monetization strategy: Gacha. You could generalize to microtransactions, but gacha has proven to be quite a bit more profitable than traditional microtransactions, and as such has taken over much of the microtransaction space.

There are debates to be had about the ethicality of gacha as a monetization strategy, but here I care about its effect on the incentives in the market. My thesis is that gacha games need to compete on quality, potentially even more so than games that charge a single sale price.

Gacha games vary in exactly what they sell. Some provide gameplay advantages or variety, while others are purely cosmetic. The monetization generally tries to capitalize on "whales": a small section of players who spend disproportionately large amounts of money.

You might expect that these games have little incentive to care about the free-to-play experience, but I believe the opposite. A big appeal to being a whale is to position oneself at the "top" of the playerbase, showing off to other players either in power or in fashion. If the free-to-play players and light spenders stop playing the game, the whale will see less value in continuing to spend, and likely move on as well.

Furthermore, even for the people who do whale on these games (setting aside the question of how responsible their spending is), they are likely not interested in whaling on many games at the same time. You still need to play the game to take advantage of what comes out of the gacha. As such, whales will also start as free-to-play players to see if the game captures their interest, even if it's only for a short period of time before they start their spending.

Due to the different market dynamics, it won't be surprising if the gaming industry and mainstream tech industry take significantly different approaches to AI.

Slowing Down is the Point

At this point I want to switch topics to the more direct effects of AI usage on the products where it's used. AI is sold as a way to speed up development. My history of studying math makes me skeptical of the value of this. Two particular aspects come to mind.

One is the culture shock that some people experience when seeing math lectures being done with the speaker writing everything on a chalkboard. We've had the technology to produce slides for ages, why not use them to cover the material faster? Writing with chalk slows you down so much.

In this case, slowing down is the point. In the process of teaching, the teacher already knows the thought patterns that are effective for the topic at hand, while the audience needs to figure it out over the course of the lecture. Even if the speaker attempts to slow down on their own, it is difficult for people to predict how much time the audience needs to absorb the ideas, and it is also difficult to accurately pause for the desired amount of time. In practice, if the speaker is given the ability to go fast, they will go fast, to the detriment of the efficacy of the lecture.

The other aspect is an old joke that the best mathematicians are lazy. If a proof is long and difficult to convey, rather than attempting to elucidate it as is, these mathematicians continue to refine the idea to make it easier to think about and understand. This refining process typically also makes the resulting text shorter, so the unpleasantness of writing is mitigated.

Indeed, this applies to other technology as well. The writing on this page is quite long, and if I were writing it by hand I probably would have put in more effort to convey my message in fewer words. You could quibble about the breaking point where too little thought goes into each word, and frankly I wouldn't argue much against someone who claimed that the ability to type has already pushed us past it.

AI Replaces Thought

As I am writing this section, it is several hours into the writing process, spanning multiple days. But there are only about 4000 words so far. I plan a few more sections, so this number is likely higher in the final version. If you type at 100 wpm, producing those 4000 words would only take 40 minutes. Where did all the other hours go? Into thinking about what words to write, of course!

AI advocates often quote very high speedup factors for their development. But, in my experience, thinking about the code to write takes much longer than writing the code. In my eyes, a proper code review doesn't involve just checking that the code does what you want. It also involves thinking through alternative ways for the code to be written, which is also the most time consuming part of writing the code in the first place. The only way for the claimed speedups to be attainable is to not do that, and literally think less about the code being added.

The inclination to offload thought could be seen well before LLMs. I noticed it strongly when watching chess tournaments with commentary. Many commentary teams have an "eval bar", showing how the computer engine evaluates the current position. This negatively impacted the commentary in a few ways.

First, the computer engine is very good at being precise and finding key moves. This means that in many positions that are extremely difficult to play and maintain a reasonable position, the engine outputs 0.0: the two sides are equal. The commentators that had an eval bar available to them would see the 0.0 evaluation and were less motivated to check for critical lines, since they've been told ahead of time that the answer is that everything can be held together. As a result, they were less effective at conveying to the audience potential ideas that the players might have for getting their opponent to give them an opening.

Second, perhaps as a consequence of the first, some commentators' contributions boiled down to, "The bar moved!" The audience is fully capable of seeing a bar move on their own. Due to how engine evaluations work, the emphasis on when the evaluation changes also puts undue scrutiny on the losing side. The evaluation significantly changes after someone makes a mistake, since if there was no better move the evaluation would have already reflected that fact before any move was made. While unforced errors do happen, most of the time the mistake comes from the fact that their opponent did a good job of steering the game into a situation where they were likely to miscalculate or misunderstand the key ideas, a process that happens entirely while the engine says the game is equal.

Several years ago, one of the commentary teams had similar opinions to me about the effect of the eval bar on commentary, to the extent that they nicknamed it the "evil bar". At some point, even that commentary team had started having the eval bar on during their casts (I don't know if it was voluntary or forced on them), and despite their best efforts could not avoid falling into these traps sometimes. For the 2026 Candidates tournament, one of the commentary pairs had no engine or even evaluation available to them for much of the game, and I found it extremely refreshing and much more enjoyable.

To acknowledge a likely counterargument, there are indeed many advancements which are very valuable precisely because they enable one to do less thinking. So could AI be one of them?

For the technical aspects of writing code that performs well, I don't have much beyond my empirical observation that the groups that make more use of AI seem to make less reliable software. In a world where the AI handles some aspect of programming for you, the specification essentially becomes the code, and AI functions as a compiler.

But going down this road is contrary to the things we've learned from the history of programming language design. Writing specifications has always been harder than writing the implementing code, not easier. The AI is also unreliable: it is liable to generate different behaviors given the same specification. People don't want nondeterminism in their compiler.

There is also already a wide range of programming languages, many of which reduce the burden of thinking about various aspects of the program, for example by providing static type checking or automatic memory management. So far I haven't seen any indication of a coherent class of thoughts that you can reliably offload onto AI. And, if one appears, it seems highly likely that a traditional programming language or library could do the same.

For design, AI would need fundamental changes to be useful. There is an idea in game design (seemingly attributed to many different people) that when players are complaining, they will almost always be right about what the problems are, but wrong about the way to solve those problems. In order to make people happy, you often have to make them think they are unhappy by rejecting their proposed solution. In the software world, Rust's postfix .await syntax is an example: https://news.ycombinator.com/item?id=48021186.

AI systems today are trained around getting positive responses from the people using them, and so development teams who offload their design ideas to AI will likely not be able to make these types of decisions.

For prose, AI use defeats the whole point. The purpose of the writing on this page is to communicate how I feel about AI to the people who are interested in that. If I had AI write the text here, then in order for it to reflect my thoughts rather than whatever the model deemed most statistically likely, I would have needed to convey my thoughts to the AI. The effort to try to do that accurately is the same as it is to try to convey my thoughts accurately to human readers, as I am doing now. Then I would need to read back the AI generated text to see if it introduced anything unintended. It's impossible for the total effort to be lessened when communication is the primary goal.

Nothing is special about AI here. Hiring a human writer would defeat the purpose in exactly the same way.

What You See is What You Wanted

Periodically I see comments along the lines of "I looked at the output that the AI came up with, and it was exactly the same code I would have written, character for character." I think these comments are absurd. When have you ever known what code you would write, character for character?

If you were faced with a problem and really did know, character for character, what code you would write, then typing it in the first time should be just as fast as typing it in a second time. I've never seen anyone for whom this could be said to be even remotely close to true. There is always something unforeseen that requires consideration.

The same is true for prose. When I started out writing this page, I had a rough idea of what topics I wanted to touch on, and I had already previously verbalized some of those thoughts in conversations with friends. But, hopefully to noone's surprise, the process of writing it all down required further reflection to even understand it for myself. Without actually going through that process, there is no chance that I would be able to predict the words that would have come out at the end.

What I believe is actually happening is that the AI output is being used to fill in the prompter's idea of what the output should look like, without them realizing that they actually previously had no opinion on the matter.

You can observe a related thought pattern when people review a chess game with an engine. With the engine on, it is common for people to believe that a move is obviously correct while actually having very little understanding of the position. The suggestion takes on a deeper role than just "This is what the engine says," and becomes "This is what I would have come up with."

I don't think that the decisions need to be high quality for this phenomenon to occur, only that they are readily available and can withstand a modest level of scrutiny. In the case of chess, many people using an engine rely on the engine itself to probe the quality of a move: if a possible counter by the opponent is not listed near the top of the engine recommendations, they will often not explore to figure out why it doesn't work. Thus, an engine with a critical blind spot could produce the same fake understanding for a losing move, as its suggestions for the opponent will also include that blind spot. Only someone who is sufficiently good at chess to spot the idea for themselves would be able to identify that the move is, in fact, low quality.

Individuality

Even if AI output becomes high quality, I still believe that minimizing use of and interaction with AI will be the right thing for me to do. People often start to adopt the speech and thought patterns of people that they interact with a lot, and I don't see a reason to believe that interacting with AI would be different.

Everyone is interacting with the same handful of models. You can try to prompt them to act differently, but most people probably don't, and it would also be very difficult for a person to continually produce prompts to vary the LLMs patterns, if prompts even fundamentally change the patterns at all.

Over time, this is going to result in this large group of people talking and thinking more similarly to each other. This is, in my opinion, very dangerous. Individuals that think differently from each other are a critical component for progress. It is a common story that someone or a group of people get stuck on a problem because they keep retracing the same steps that don't work, and then it gets cracked by someone unaware of the prior work, because they don't get locked onto that same train of thought.

No Editing or Outlining, Either

Some people do part of the writing process themselves, and then use LLMs for other parts. The LLM use tends to fall into two categories: (1) editing, and (2) brainstorming / writing an outline that then gets rewritten by the author.

The Hacker News guidelines reject both AI-written and AI-edited comments. I think this is the right policy. Edits change the language structures used, and can absolutely change the way in which ideas are received. This policy is also useful in a pragmatic way. When people see language constructs that are indicative of LLM use, they will reasonably assume that the thoughts contained in the comment are also LLM-generated. With less reason to question whether participants in a conversation are outsourcing their thoughts to AI, the assumption of good faith in the discussion can be upheld more easily, and that in turn produces better discussion.

Incidentally, and unfortunately, machine translation is taking splash damage. Machine translation tools have been moving toward directly using LLMs to do translation. This causes risk for people who previously could interact with communities that speak in a language where they are less-than-fluent. The translation output is going to carry signs of LLM generation, which is then picked up by other people and casts doubt on the provenance of the content.

It is also not helping that there is a group of AI users who will lie about the extent of their AI use. I have seen Reddit comments that are almost certainly completely LLM-generated (both the language and the content), and when challenged the author claims that it is so because they are using a translator tool. This lying contingent has identified that translation is often deemed an "acceptable" use of AI, and uses it as an excuse to be able to participate with AI when it is generally against community rules.

Attempting to rewrite an LLM-generated outline seems to simply fail. The presence of the outline results in some of its content implanting in your brain as what you want to say, similar to what I described in the "What You See is What You Wanted" section.

This was directly observed in an official Rust blog post. Multiple commenters noted that the article felt written by LLM, and afterward the author confirmed that they had used an LLM to write the first draft of the post. Since the article has since been retracted as a result of this reaction, I feel it is reasonable to believe the author on the extent of LLM use.

Credibility

As I noted, there are people out there who are shoveling AI-generated works onto the world while denying that AI is used. So there's a nagging thought: I can write all of this claiming that I don't use AI and explaining my reasons why, but why should someone reading this believe me? From their perspective, maybe I'm one of those people who will put effort into being deceptive about AI use instead of putting effort into the projects themselves.

It's a scary thought, and I can't do much about it. As I explained above, I intentionally minimize my exposure to LLM-generated prose, so I don't know what they write like beyond some of the basic tells that frequently get complained about (e.g. "It's not X, it's Y."). Since I don't know how they write, I can't intentionally avoid their patterns.

What if LLMs settle on language patterns similar to mine? Just by virtue of having written a lot of text, there's a decent chance that some phrase somewhere on this page will trigger someone's internal AI detector. Simply the fact that I wrote a lot of text might make people suspicious too, as we move toward a world where people are less willing to spend time writing.

I'm lucky that I have some amount of blog posts that I wrote and published before the advent of LLMs. Unfortunately, I stopped posting for several years. I can only hope that my style is similar enough to how it was before that people can compare to my old posts and believe that it's really me (but even then, what if it's an LLM that's told to write like me?). People who are trying to get established from scratch today face a much more difficult task.

I believe the best thing for me to do right now is simply to be an example of someone working without AI, even if it takes a leap of faith for observers to believe in it.

Progress will be made in spite of LLMs, and I would like to contribute to it.