Venkat Subramaniam, a renowned software programmer, agile evangelist, and one of the most sought-after tech speakers, offers a striking observation about artificial intelligence whenever he gets a chance to speak: “If I’m not good at something and I ask AI, then what AI gives me looks awesome, but if I’m an expert and I ask AI to give me something, what it gives me seems awful.” Venkat’s observation comes from his vantage point as a programmer for forty years, but it is true for anyone with deep experience in any field.
If you reflect on Venkat’s statement, you’ll recognize that the “awesome” versus “awful” dichotomy he uses is not a flaw in the technology itself, it reflects the user’s own skill level. What he is getting at is a simple - yet profound and under appreciated - truth that the value of AI’s output is largely determined by the user’s own expertise level. When a novice asks a Large language model a question, the result often appears impressive because it usually surpasses the questioner’s baseline knowledge. The same answer, however, can look crude or incomplete to a seasoned professional who can spot gaps, inaccuracies, or missed nuances. The more you understand a domain or a skill, the more you can judge the quality of AI‑generated content, and the more you will demand precision, context, and originality and be able to converse with it intelligently. Conversely, the less familiar you are, the more likely you are to accept AI’s output at face value and often remain awe struck at its prodigious felicity with language. In short, AI’s brilliance is relative and recognizing this helps us use AI more wisely, leveraging its strengths while remaining vigilant about its shortcomings.
Let’s consider the art of writing for a moment. Even though the latest models are far better at generating clean sentences, they still fall well below my own standards for prose writing. As someone who has been writing for nearly two decades, AI’s output, no matter how polished, often feels gibberish to me, and I am never satisfied with the rhythm, the emphasis of AI generated text. Though it is possible to refine prompts to make AI write in a particular style, I still find myself constantly editing, re‑phrasing, and discarding large portions of AI generated texts, when I choose to engage with it.
I remember one particular essay I wrote earlier this year. I had written three or four paragraphs. My central thesis was in place. Then I said to myself, "Why not have AI complete this essay for me? let me pit my essay against the model and see how it would continue my piece.” That was the first time I tried doing that. The result was hopeless and frustrating. I hopped from one model to the next: OpenAI, Anthropic, Gemini, Perplexity,( I have a paid subscription to each of these models), yet none could deliver the essay in the language, style, and tone I envisioned for my ideas to unfold. I would have taken an hour more to write those finishing paragraphs myself. Instead, I ended up wrestling with these models for more than three hours, still unable to coax the texture of prose or get it to write even a paragraph the way I wanted.
In the end, obviously, I abandoned the experiment, finished the essay on my own. Once I finished the essay, I asked AI to critique it and point out areas where my essay was structurally weak. It did a great job with that. It pointed several things that were right about my approach, and an equal number of areas that could be improved. Not that I accepted every suggestion made. I had made some deliberate choices on what I wanted to say and how I wished to say it, and looking through that lens, I accepted those suggestions that enhanced my presentation and ignored others.
It isn’t that the sentences and paragraphs generated by AI were outright inaccurate or grammatically wrong. That's not the point. When I encounter AI-generated text, my linguistic instincts homed over decades of reading and writing start to squirm and complain. It affects my aesthetic sensibility when I see the cluttered arrangement of clauses, the dissonant cadence of its prose, the arbitrary choices of words. It is like eating southIndian food without their spices. AI writing is flat and mechanical. Phrases that should glide naturally become clunky, and the flow that a seasoned writer would instinctively grasp is replaced by a stilted, stifled and almost formulaic pattern. It's a matter of resonance, I guess, a feel for sentences. A well‑crafted line is something that cannot be taught. It can be pointed out when you see one. And that is the meaning behind Venkat’s insight as well. Since I know how to recognize a good sentence and can write a few reasonably decent sentences myself, I find AI’s language awful, high-falutin and often displeasing to the ear.
What is true of writing is true of everything else AI can do. I've mentioned this before in other essays, but it bears repetition. Artificial intelligence ( especially the word intelligence) carries a host of unintended connotations that can distort our understanding of what AI is currently capable of. In my view, the moment we start evaluating whether an AI system is delivering an “intelligent” response, we cloud the practical purpose of this incredible new tool. AI is fundamentally a statistical or probabilistic engine and we will be better served by focusing on its concrete capabilities such as pattern recognition, data synthesis, and prediction. The last one about prediction is its strongest point. It predicts, but is never sure. By stripping away the loaded term “intelligence” and referring to the technology more neutrally we can discuss its strengths and limitations with greater clarity and avoid the semantic fog that often hampers productive dialogue and usage.
So how does AI work in simple terms? What happens when we ask AI a question, or prompt it to perform a task? Venkat Subramaniam likes to call AI “Accelerated Inference”, instead of Artificial intelligence. I agree. When we submit a prompt to AI, the system does nothing more than run a complex inference. Inference in simple terms is arriving at a likely result from a given chain of causes. Given a prompt, AI’s inference engine traverses thousands of servers and data centers, consumes substantial energy, and sifts through every piece of text the model has ever been trained on to produce a probabilistic relevant response. The results look dazzling, but underneath the hoods, it is nothing but mathematical computation on steroids.
Ironically, human beings infer all the time just as AI. Inference is one of the biggest tools that we use to navigate our world. Our brains are wired to infer. We infer, for example, based on past patterns, that tomorrow morning the sun will rise in the east, that swans can never be black in color, or that a verb should always precede a noun, or that spring follows winter. The biggest difference between inference done through AI and our inference is that we infer with limited data and we have the capacity to evaluate whether that inference meets the current pattern that we are seeing and can pivot. In the case of AI using Large Language models, it infers from vastly larger magnitudes of data than humanly imaginable and it is relentless. Once an inference sequence is triggered, AI must complete, even if that means making up things along the way ( which is called Hallucination. I plan to write a separate essay on this subject soon.) Having said that, AI models are getting better and better in producing better results. Still, reasoning is never intrinsic to the large language model; it must be orchestrated as a workflow and baked in through weights and parameters. By default, AI remains an efficient pattern‑matching machine that can infer and generate likely text continuations based on statistical relationships in its training data.
Apart from writing skills, AI is widely used for coding. And the problem there is more acute than writing. I see people churning out working code with AI like candy out of a machine. At least in the world of literature, there is enough evidence of great writing and sometimes AI can conjure up beautiful sentences. But in the world of code, one can safely assume that the majority of freely available code is bad code, and code that is production-grade is rarely available in public and therefore AI never gets to be trained on it. Most frontier models have swallowed the entirety of GitHub - which is good and bad. Good because it speaks to the democratization of coding as a profession; Bad because the amount of amateur code floating around in Github is staggering and if AI is indiscriminately trained on them, the result cannot be flattering . When it produces code, expect that code to be terrible by default, unless you are lucky.
We hear about vibe coding a lot. Vibe coding essentially is the ability to state your requirement in plain English and allow AI to spin out the code. It is truly a liberating experience for someone who has shied away from coding or has found coding to be an esoteric profession and seemingly gives you the power of a programmer. Just like we get carried away with whatever AI writes, we equate working code that AI generates to good code, which is a view a senior programmer will never take. As with the art of writing,, AI can provide novice programmers with a jump start. But If you don't understand the programming basis, software engineering principles, design patterns, or anti-patterns and many other essential ingredients of good coding, it is almost impossible to move the code you get through through AI to production. During my discussions with technology platforms, they smilingly tell me, "Bala, everybody seems to be producing proofs of concept and prototypes, but nobody seems to be thinking about how you deploy this into a production-grade application." That is the gap that vibe coding will produce. It will regurgitate bad but workable code that it has been trained on. And If we don't know enough to spot what is right and wrong, make the right engineering corrections or rewrites, we will end up having vulnerable applications ready to fail.
So where does this discussion land us? AI is a phenomenal tool, something that happens once a century or longer. And like any other pathbreaking tool, its value lies in the hands of those who wield it. I use it extensively in work, and for my personal research but only after I have spent time thinking about the task I am on. I respect and acknowledge that knowledge that AI puts at my disposal, but at the same time, I don’t blindly take it at its face value. That is the core of building AI ecosystems and applications. Those guardrails, governance mechanisms, sanity checks are critical to keep the tool under our control.
It took me about two hours to write this essay with a few interruptions. I could have prompted AI to give me a draft and then corrected it. I suspect the second option may have taken longer because i would have had to undo a lot of bad writing before my style shines through. In the end, both options are fine as long as I ensure that the essay reflects my point of view in the language and style that are uniquely mine. How I use AI to reach this end depends on my levels of expertise and prerogatives.