In Defense of Writing

In the wake of the various OpenAI math results, there has been an explosion of discussion covering whether this is the end for mathematicians, an end to work in general, and most notably, whether this is a signal that Artificial General Intelligence (AGI) has arrived. Regardless of the noise and drama surrounding the release of these results, this is an enormous milestone for the technology, demonstrating that with enormous data scaling and training techniques such as reinforcement learning, autoregressive attention can discover groundbreaking research results and advance human knowledge. Many are quick to claim that this is a sign that the intelligence of these models has far surpassed humans, and mass layoffs are imminent. After all, one of the most difficult fields for humans to understand has been upended, and the others are clearly simpler. So shouldn’t the rest fall as well?

But there is a blind spot in these claims that we have achieved AGI with the Transformer, a blind spot that ironically, exists in the domain where this all started.

Being a writer right now, seems not good, to say the least. Before these large AI models became capable of mastering mathematics and logical thinking, they had already conquered textual generation, leveraging their training on (basically) all human media ever created to generate coherent sentences on almost any topic imaginable. Yet even though the setences generated are coherent, they are not necessarily good. In fact, LLMs are awful writers. They consistently use empty sentences that convey no meaning and are used for filler, generally convey little information in a lot of words, and are overall extremely verbose, all leading to dead and boring essays that make for unenjoyable reading. The verbosity of the essays is likely a case of large models being trained to be as helpful as possible, leading to outputs that spit out as much information as possible to try to hit all the marks, even if nothing of substance is actually said. Of course, when AI models are given a very narrow focus and set of instructions, they can be extremely helpful (a good example is with resume writing, where sentences are required to be short and bullet points generally follow a standard format). But as others have noted, for broader pieces such as essays, articles, and certain research papers, AI is simply unable to write competently or beautifully in "one-shot" (i.e. with a single prompt without post-hoc editing) without a human guiding it around a narrative and a strict set of rules. Note that from this observation we can deduce that the true value of a writer is not necessarily their ability to type or handwrite words, but their ability to construct a captivating narrative and make judgements about what information to include and what to exclude, something ill-defined and difficult to quantify. We’ll discuss this later. But first, why is AI so bad at writing?

AI models are excellent information aggregators, essentially having taken in all human knowledge by scraping the internet, making them excellent at outputting near-factual information as well as making connections between existing information¹. But the demand for writing (and other forms of media) is not always around objective information; in fact, there’s a huge demand for subjective information! And it is often true that subjective information requires both 1) the ability to process the physical world around us and 2) the ability to offer an opinion on how we should feel that challenges the norm. This is part of the product that opinionated magazines like The Atlantic or The New York Times offer that may differentiate them from information-based media like Yahoo Finance, and this may be part of the reason that human reporters are still so valuable there. It seems that to offer opinionated pieces that influence how others feel, there’s an innate prerequisite to be a sentient, living being yourself. To speak candidly about events like the war in Ukraine, it’s critical to have reporters see the war for themselves and listen to the stories of soldiers terrorized by drone warfare. Even for this essay, the impetus for me to write came from an emotional anxiety of what a future where many of my generation’s potential career options would vanish, which led to more analytical thoughts about how certain skills would evolve. These sights, sounds, and feeling power the emotional reaction that leads us to create new media (even if the media itself is sometimes written and edited by AI). This reaction provides agency, a decision to act, something that we have yet to replicate in AI, perhaps simply because we have failed to replicate the emotional reaction to physical sensation that causes living beings to act. This media product of synthesized, meaningful, packaged emotion happens to be highly in demand by human consumers. Especially now in our Attention Economy, we regularly seek out information that speaks to our values, whether that’s Joe Rogan’s podcast, Tucker Carlson, Hasan Piker, and so many other personalities that have an agenda, because we don’t just engage with media to gather factual answers (even if we claim to be unbiased), but also to help us decide how we should feel amidst a glut of information that our biological brain never evolved to be able to process. Our Attention Economy has generated enormous demand for this, creating a new class of elites in the form of influencers and streamers, whose literal job is to professionally find ways to change how people feel and challenge norms (albeit for the purpose of persuading them to buy advertised products)².

This ability to influence and argue, while important, is not usually considered “intelligent”, at least not in the sense of the phrase General Intelligence (would you really call Jake Paul "generally intelligent"?). But then again, intelligence is still ill-defined when it comes to comparing biology with technology. Part of this mystery is that some of the hardest, most abstract, and most intellectually demanding tasks for us humans are trivial for machines. Remember when computer science students were respected (and employed), because of their ability to combine logical thinking and syntactical subject matter knowledge to turn abstract problems into algorithmic solutions? Instead, Claude Code and other agentic tools have entirely automated that work, able to take control of users’ file directories, read code, and create entire apps in hours or even reverse engineer entire video games. Or take a look at mathematics, where as we’ve already discussed, OpenAI and Anthropic have entirely upended the field. Not to mention chess, where machines have dominated since the late 90’s. In contrast, AI-generated articles and essays simply fail to connect with humans, and are filled with dense prose, punchy one-liners, and overly catchy, cringey metaphors that just don’t land—certainly part of the reason LinkedIn is so insufferable.

The broad reason for this discrepancy between math/CS and the humanities is that the former fields have developed extremely sophisticated verification tools that make evaluation practical and accurate. A program’s code can be either correct or incorrect (and the traceback can tell you where), and a mathematical proof can generate a series of logically dependent claims where each module can be verified independently (this is essentially the premise of the computer language Lean). Not only that, but we humans have developed a preferred process for how these solutions should be completed. Code should be commented, readable, and math should be in LaTeX, intermediate reasoning should be shown, etc. Because this process is rules-based and easily defined, we can abstract all of this and penalize our model for not following these rules, even if it gets the answer mathematically correct. Much of the reason that AI has advanced so quickly in these fields is that though they are cognitively taxing, they are rules-based, verifiable, logically consistent, and widely documented on the internet. More generally, this seems to apply to any domain where we have a clear set of rules defining our actions and copious data to learn from, the biggest example being self-driving cars, which have made extraordinary advances.³

And of course, between the extremes of high-level mathematical ability and pure writing, there’s a range of skills and talents in between that AI still has trouble replicating, or is not yet trusted to fully automate. All careers that benefit from a physical presence are largely unscathed. Civil engineering, construction, and other trades continue uninterrupted, and while it might be feasible to use autonomous cranes, machines, and even individual humanoid workers, the technology to link robotic motion, object detection, and reasoning is simply not able to generalize to these unpredictable and minimally documented environments. Even fields such as banking and law, while not clearly requiring physical labor, benefit enormously from having professionals that are able to meet in person to wine and dine their clients, enter negotiations, make subjective judgement calls, and verify the outputs of their financial models, even if the actual financial modeling and due diligence they purport to sell is entirely AI-generated.

By no means am I arguing that AI will not replace any jobs. In fact, I believe AI is essentially ready to cause tremendous upheaval in fields like software engineering and transportation; it’s just waiting to be deployed. But the fearmongering spreading throughout the internet about an AGI future misses 1) the understanding of what Transformers are not capable of and 2) the deep nuance of what this upheaval might look like, and examining this nuance yields useful heuristics for new graduates starting their career in an age of white-collar automation. The most in-demand skills of the future may not solely be dominated by those skilled in abstract logical reasoning, and we may instead see a shift in preference to more subjective and physically grounded skills (such as civil engineering, public speaking, banking), where physical presence and trustworthiness is important for serving clients. Interestingly, the creators of AI themselves seem to be converging on a similar perspective: OpenAI researcher Sunyu Yao writes in The Second Half that we no longer have problems training (“teaching”) our models; Reinforcement Learning (the act of having an agent learn by trial and error) now works generally. The problem now is problem definition (what problems should we solve with AI), and evaluation (how well did our generated solution perform?). Until we are able to perfectly encode and document every problem in the world as training data with definite training labels for what actions are "correct" and "incorrect" (which I'm not entirely sure is possible, as much as the philosophers have tried), there remains tremendous opportunity for the writers, physicists, engineers, and other professionals that can succeed and gain trust in domains where solutions require a combination of objectivism, presence, subjectivism, and visuospatial reasoning⁴.

With the advent of tools such as Claude Code and Codex that are able to agentically automate rule-based and procedural tasks (and the growing shadow of even more powerful AI models on the horizon ), we have to think not about how humans can “beat” AI, nor dramatically declare that AI will replace all labor. Instead, we should think practically about how humans can coexist with AI in workplaces where humans might ideate solutions, present ideas for clients, and direct teams of bots, the agents themselves complete the implementation and perform difficult, yet rigidly defined quantitative tasks, and the humans once again verify. For the new graduates, we should think even more pragmatically (and a bit cynically) about how we can avoid competing with AI and instead develop subject matter expertise and skills that are difficult to replicate, either because they are not well documented on the internet, require physical presence, or involve a deep knowledge of how to influence people. In an age where anyone can code (especially at the level of an entry-level engineer), skills like writing, persuading, eyeballing, estimating, may be the last ways that white-collar job seekers can truly add value. And in this world, the young people comfortable with guessing, iterating, operating under uncertainty, and forming opinions will have the advantage.

And in the name of forming opinions (and sloppily tying this essay back into the intro), no, I do not believe that the Transformer architecture by itself will ever achieve “AGI” when it comes to writing. As powerful as the attention mechanism is, and as remarkably as it has scaled, the "best" writing, as much as it can be defined, requires making a genuine connection to the human reader, which can only be done by a writer that experiences all the emotions we do. LLMs now do not experience the elation, anxiety, suffering, and the physical interaction with the world that writers like Homer, Shakespeare, de Cervantes, Harper Lee, and others drew on to create a string of words that will forever activate something in our neural pathways. Without artificial sentience, these monumental achievements will never be replicable. So, keep writing.

Footnotes

1. There is an interpretation of LLMs as information interpolators, that perhaps explains their ability to excel at mathematics. By drawing on and mixing ideas from all over mathematics, they can often create solutions or counterexamples that are based in fields that seem completely unrelated to the original problem. These long range connections are hard to keep track of for the human mind, but seemingly easy for an algorithm in a high-dimensional embedding space. ↩

2. I will note that it’s not clear yet whether human consumers care about whether the media they consume is authentically human or not. If you use Instagram at all and have a keen eye, you’ve probably noticed that many of the reels you view are AI-generated, yet, they still garner attention, likes, reactions, and eventually, money for those who created them. Nevertheless, more often than not it is still humans behind this content, motivated to AI-generate their content for very human reasons ↩

3. While recent months have made it difficult to take the position that humans will remain superior to AI models when it comes to symbolic, abstract, rules-based problems, there are still some interesting edge cases where AI models seem to lack "intuition", for lack of a better term. For example, in the 2025 International Math Olympiad, AI models solved the “easy” problems, but failed to provide a solution to the novel, harder problem that required creativity beyond the frontier of current mathematics. Also interestingly, though AI was able to solve the easy problems, it framed the solution entirely symbolically, whereas the solution was much more intuitive and direct when framed from a visuospatial perspective. Perhaps more supporting evidence that even in highly abstract and logical domains, humans have a certain advantage in being able to interact directly with our world and choose when to think symbolically, when to think spatially, and how to relate the two. ↩

4. The recent math results also make it clear that there remains significant work and opportunity for those experts whose job it will be to 1) verify that the solutions are indeed correct and 2) understand these model outputs and translate them into a format that is actually useful for human knowledge, instead of the slop they were originally released in. However, it is not clear what the role will be for those who are not already experts. ↩