Why AI Alternatives Matter
AI is prone to problems affecting its output: hallucinations, incompleteness, inconsistency, and bias. AI usa...
For further actions, you may consider blocking this person and/or reporting abuse
"Think first, prompt later" might be the most underrated engineering skill of 2026.
This.
The "Principle of Least Power" framing is the bit I want to amplify, because it tends to get lost in AI-vs-no-AI debates. Most of the "AI sucks" arguments I read are actually "I used a sledgehammer where autocomplete would have done." The pyramid structure you sketched captures it well: foundation is rubber-ducking and reading docs, and AI should sit on top as the expensive exception.
The flattery / verbosity problem you mentioned deserves more attention. I have been on both sides of this — building agent systems where the model agreed with obviously wrong assumptions, and being the user who caught it five messages later. What helped our team was explicitly budgeting for "model disagreement" time: treating the first response as a proposal to refute, not a statement to accept.
The GEO Poisoning point is also under-discussed. Are we approaching a point where defaulting to AI for factual questions is itself a risky default?
The simple truth is that AI is a senior developer's tool... Because it does Junior level's work. If you use AI, stop treating it like it's smarter than you, treat it for what it is, a brilliant junior, that still leaves AllowAnonymous in endpoints. You need to meticulously validate everything it writes, because it'll run beautifully, but at what cost?
You can gate it, you can add validation steps, system instructs, unit tests, integration tests, all wonderful, but it doesnt help if a button is off the screen now does it? Validate everything, so you dont end up ruining production.
I recently viewed job postings, found a company, you know, just a Series A backed by YC... Went to their career portal, job wasnt listed there, so I clicked on Company Website at the top... It takes me to a site they dont own, infact it's pointing at an unowned domain in general. No human would type in the wrong url, so it was clearly someone who trusted AI, didnt validate, their code wasnt reviewed, they had no production gating, they didnt test the site and the result is a big red flag...
(For cybersecurity context, that means I can buy the site it points at, replicate their website and offer an 'exclusive discount' flyer. If anyone buys anything and gets scammed, that company is held liable, because they point at it, as their company website, on their official site... So for a company that is built on digitizing money and is entrusted with safeguarding it, that right there is a funding killer, infact a company killer.)
I applaud this. Very well considered and well written, and it's sparked some great back and forth with others contributing their thoughts and experience. Nicely done!
I love the title, inspired from "The principle of least action" :D !
AI is most useful when it supports our thinking, not when it replaces it.
P.S. as always, I have prematurely published, and then edited and added some more paragraphs shortly after. So, if you have just read this post, please reload and re-check my self-critical section about hope and hypocrisy.
I'd push back on this. the problem isn't using AI too early - it's never checking if you understood the answer. passive acceptance is the failure mode, not early use.
There's nothing to "understand" when an answer is defunct hallucination without any factful proof, then it was just a waste of time, tokens, and energy. Think first, then ask – if you still need to. That's what I wanted to point out. I agree that there are many problems and setup where AI is actually helpful, and you are right, then it is still essential to understand beyond a simple "does it work" (in the happy path) test.
fair - that's the edge case i skipped. hallucinations aside, my point was about the pattern of running with any AI output without checking it at all. but you're right that checking understanding of made-up facts doesn't help - that just locks in the wrong mental model faster
Skill issue. Learn to prompt.
Prompting closes a lot of the gap, agreed. But there is a class of hallucinations — rare facts, specific citations, edge-case numbers — where even tight prompts do not fully eliminate it. At that point the fix is grounding plus verification, not more prompt skill.
The danger isn't overusing AI.
It's forgetting the alternative approaches that worked perfectly well before AI existed.
"Principle of Least AI" is exactly the right framing. The tool should match the task, not become the default for every task.
The corollary for content creation: AI earns its place when it compresses the genuinely mechanical parts of the workflow. Writing the first draft isn't always mechanical — but publishing it to 8 platforms after you've finished? That's 100% mechanical. Every second you spend copy-pasting and reformatting is cognitive overhead that has zero creative value.
That's the use case we built twRty Blogboat (twrty.org/blogboat) for — not to replace the writer, but to remove the post-writing pipeline entirely. One click, 15+ platforms, credentials stored on-device only. The AI helps generate and edit when you want it. The automation handles distribution because that's a task that should never require human attention.
Least AI where least AI is right. Maximum automation where the task is purely mechanical.
Principle of Least AI maps neatly onto least privilege, and the payoff is the same: smaller blast radius by default. The cost of reaching for AI isn't just tokens, it's that every generated piece you didn't write is a piece you don't have a mental model of, so over-using it quietly converts authorship into archaeology. Applying it least means using the model where the work is genuinely mechanical and keeping your hands on the decisions where understanding has to live in your own head. The teams that get burned aren't the ones using AI, they're the ones using it everywhere by default and finding out later that nobody understands the system. Least AI is really least unexamined output.
Thank you for that definition. Until I saw the original poster's question, I'd never seen the phrase, "The Principle of Least AI". So, I was lost. Your definition will work for me, and I like it, too.
the "smallest blast radius by default" framing from Theo's comment nails it, but the mechanism worth naming is carry boundaries — when AI output flows into another LLM call downstream, unexamined output compounds silently.
the GEO poisoning point hits differently when you're building RAG pipelines. we started explicitly flagging sources in our retrieval layer as "potential GEO influenced" when they're commercial domains without an original study backing them. adds friction, but the alternative is building LLM answers on top of answers AI already hallucinated.
the hypocrisy section is the most honest part. "can't fully opt out" is accurate in 2026, the useful question is just where the line is.
does your team explicitly document which parts of a workflow are AI generated, or is it still implicit?
The principle stands. It usually breaks at the optional/mandatory line — "use AI less" lands the way "eat less sugar" lands at a dessert table. The hierarchy reads completely differently when the gate is unable to be walked past instead of politely suggested. Where it lands hardest is at carry boundaries: when output flows into a high-consequence decision or another LLM. That's where deterministic-before-LLM stops being polite and becomes the difference between a check and a story about a check.
I think there's a lot of truth in this.
Working with AI every day has taught me that the quality of the result often depends on how much thinking happened before the prompt.
For example, if I have a clear understanding of the problem, AI can help me explore ideas faster, compare different approaches, and save time on repetitive work.
However, when I use AI before I fully understand the problem, I usually spend more time checking and fixing the answer....
That's why I like the ,"simplest tool for the job" idea. Sometimes the right answer is AI. Sometimes it's reading the docs, searching the codebase,, or taking a few minutes to think things through.
AI is a powerful tool, but good judgment still matters. The better we understand the problem, the better we can use the tools available to us.
I think that using AI for devs that started coding in this "era" is a great risk, because most of them just do "vibe coding" and do not understand anything about the code being created. For other devs that already have great knowledge about coding, it is a great way to improve the code quality.
The hypocrisy section is the most honest part of the piece. The reality is you can't opt out of AI tools in 2026 without paying a real productivity tax, and most of us are quietly making that trade. Where I'd push back: framing this as a moral failing is exhausting. It's just a tool with externalities, like every other tool engineers have ever used. The useful question isn't "are we hypocrites" but "where exactly is the line we won't cross." That's actionable.
nice