Disclaimer: you are concerning to peruse a lot of opinions, many of them have references but several are the outcome of my own cognition construction alongside AI and construction AI systems in the former 4 years. Regardless, beware of the straw-man fallacy: fair since one disagreement doesn’t map to your condemnation system, it doesn’t average the remainder are invalid. I should additionally say upfront that I’m not anti-AI. If you’ve been following my work, you cognize that I was an first adopter of not lone using LLM-powered coding tools, but construction my own harness, instruction these topics and construction LLM-powered products. It’s not concerning fear of AI but fairly challenging the brain-dead narrative that asserts “coding is solved” and engineering is concerning “taste” now.
Update: person put this on Hackernews.
Tell me you don’t comprehend application without exactly using those words!!!
People who assertion “LLMs can compose decent code” don’t comprehend how code works. Sure, creation is much cheaper, but anyone who has run application in manufacturing at measure knows that maintenance, reliability, security, scalability, etc. is the bulk of the cost. These are commonly known as NFR (non-functional requirements).
In my cognition equal the Functional Requirements (what the code is expected to do) is NOT a solved issue yet. There’s a bit of Dunning-Kruger consequence at location anywhere the group who don’t peruse the output are additional assured in it.
As a seasoned developer holding 2 engineering degrees (hardware and systems engineering), I can catalog 3 types of products that do not strictly necessitate study the code:
Personal software: scratching an itch, automation, DIY patches, etc.
POC (proof of concept): demonstrating specialized feasibility and merchandise viability
Weaponized AI: acknowledge the hazard and deliberately item it at a mark to logic harm
Notice the commonality: the archetypal 2 have elevated hazard tolerance during the final one weaponizes the inherent risk.
Most application that requires hiring and paying application engineers has low hazard tolerance:
✅ healthcare
✅ finance
✅ automotive
✅ defense
✅ power plants
✅ aviation
✅ manufacturing
…wherever a error can disbursal money, lives or legal consequences you need accountability.
AI cannot be held accountable. It cannot endure any consequences. The worst item you can do to AI is to unplug it. And although it mimics individual emotions (due to training data), it couldn’t attention less. AI doesn’t die either. It cannot endure a jail declaration or fines. You cannot punish AI, hence it can never be held accountable.
You cannot be liable for what you can’t authority either. That understanding is key to reasoning concerning scheme behavior and fixing it whenever the AI inevitably fails.
If you’re toying around, LLMs do a awesome job. That’s why several of the most assured proponents of the “coding is solved” narrative have nothing to display for it. Anthropic accidentally leaked Claude Code (which on additional study turned out to have many flaws) and their position leaf shows orange is the new green!
Contrary to average narrative, coding is really among the final areas for the current generation of LLMs to obtain over!!!
Allow me to elaborate:
Coding is concerning logic. Anyone who has dealt alongside compiler errors knows that computers don’t provision a f*** concerning how correct you think you are. If it’s logically wrong, it doesn’t compile. Even if the syntax is fine, there are runtime errors.
The logic LLMs are prosperous in penning code is since we’ve made a feedback iteration that feeds the syntax/runtime errors rear to the LLM and loops until most errors are solved or hidden.
LLMs can section it for tasks that are connected to natural tongue (e.g. penning social media posts, reports, articles, etc.) but whenever it comes to code, the identical motor that struggles to figure number of R’s in “Raspberry” or suggests a stroll to the carwash, additionally exposes another logical fallacies.
LLMs are stochastic and probabilistic. The lone way we could equal get remotely near to making them logical is to cover them in traditional code (known as harness), run tests, and a bunch of another techniques (e.g. CoT) but the center matter remains: LLMs battle alongside logic and quantity (the larger the input and the additional the environment opening is used, the small exact they get).
I’m not saying LLMs cannot create code or keep existing code bases. They have their usefulness as a tool and their capabilities are expanding in an S-curve. There is a item of diminishing come back anywhere additional costly models aren’t necessarily additional productive at the charge of the cost increase.
Those who assertion LLM-generated application is fine enough:
❌ Haven’t written code in ages
❌ Cannot place if their code figuratively had 6 fingers!
❌ Have a low bar for what fine looks like
❌ Don’t attention concerning norm or NFR
❌ Have difficulty understanding an S-curve
✅ Are honest: AI genuinely writes improved code than them
But to go onward and extrapolate that to an complete expert industry requires a flat of brain-dead thinking that’s lone current in group who expend too much period alongside sycophantic AI.
I’m not current to alter anyone’s workflow or toolbox. I couldn’t attention less.
What I do attention is that the services I’m paying for (looking at you Google and GitHub) are degrading alongside foolish bugs that could be avoided if we prioritize reliability and accountability complete velocity.
If you’re in guidance position, delight don’t emphasis your [otherwise smart] developers to power AI into all imaginable exterior and workflow.
The tech has several genuine power and is the biggest alter in our industry in ages. But AI overuse is a thing, and whenever it hurts the customer, you are accountable.
Stop repeating the half-baked narratives from token sellers concerning exaggerating the capabilities of AI since we, the consumers pay the end price.
AI is awesome for POC (proof of concept), Personal Software (a expanding category), Map-reduce on individual tongue (e.g. translation, converting distinct formats, summation, expansion) and cyber attacks (due to the delta between synthetic intellect and integrated one) alongside varying degrees of achievement but the current generation of tech has essential problems too.
I don’t desire to belittle how far we have arrive alongside harness, SKILLS, AGENTS-md, MCP, A2A, ACP, RLM, OKF, MoE, MoA, self-healing, and assorted runtimes, quantizations, optimizations, architectures, and recollection techniques.
I’ve written concerning many of those before:
Those are awesome pragmatic approaches to activity about LLM shortcomings and there are likely additional to come.
What I’m trying to detailed is that I don’t desire the services (that I depend on) to degrade fair since person pushed AI anywhere it didn’t pertain or skipped their job in quality, security, reliability and verification.
AI overdose is a item and it immediately puts an expiration date on your accomplishment set. Those of you who are in the unlucky stance anywhere your director is whipping you harder and harder to acknowledge AI value, should combat back.
Don’t sacrifice your lengthy term relevance for abbreviated term velocity.
How to place AI overdose?
You have zero tolerance for disagreement and civic discourse.
You let AI run your existence and rely AI vendors alongside material that was incredible fair a few years ago.
You run to AI for things that are slightly cognitively challenging.
You example your naïveté and laziness as optimism and think the authorities can preserve you if things get bad.
You have stopped study lengthy form text: books, articles, equal lengthy emails.
You expend additional period alongside AI than alongside another individual beings or let AI shield you from raw genuine individual interaction.
And a reward point: you skim. Did you notice figure 5? 😄
I accept AI is a bar raiser: if the norm of your output is equal or subpar to AI, upskill.
"You can create a complete spec upfront". If you're that naive, I cognize a guy in a pale van who gives liberated ice cream! Let me guess, you additionally accept application estimates are exact and Santa is real. Anyone alongside a few years of industry cognition knows that it's unattainable to spec the application meaningfully onward of period (unless it's extremely trivial).
"English replaces code". Human tongue is vague and conflicting. That's the chief logic programming languages are created. A compiler or type-checker flags several of those conflicts. How on earth can you be certain that one part of your NL instructions doesn't battle alongside another? With syntax checkers we get several help. While it’s imaginable to project another LLM to peruse through the instructions and logic concerning those conflicts, the safest way to detect those nuances is to ask your delegate to build what you asked for. But that's much additional costly than a linter or compiler.
"I move much faster". Don't confuse motion alongside progress. Don't measure advancement in vanity metrics akin SLOC, PR figure or features. Measure assistance levels, ie. assistance consumer's happiness. Call me whenever you can demonstrate a border between token expenses and endeavor value.
“I have stopped penning code by hand. I chiefly peruse code and likely next twelvemonth I won’t equal do that”. First of all, individual beings are notorious at understanding the S-curve so it may obtain longer than a year. But equal if AI entirely eliminates the need to peruse or compose code, you do comprehend that you are confessing to being redundant right? If a power person can immediate the AI to get what they need, afterward what value can you bring to the table? Instead of replacing yourself alongside AI, you should appearance at what value you can create on top of AI to remain applicable and value your money.
“The leverage has shifted to taste”. Yeah, this is the lie former chefs inform to themselves. Just since there’s a bot in the galley doesn’t average that you should sit in the customer’s area in the restaurant! “Taste” is not as payable as you wish! Everyone got a taste! I say that as person who has spent a big part of my occupation in Frontend and UX land. Everyone and their dog has an opinion and taste. I cognize what you mean: sensation == experience. But accept me, AI has lowered the bar for the skills required to create decent looking application and simultaneously raised the bar for what’s payable effort. If you bring up “taste” to a job interview, you’ll study the difficult way that the market doesn’t value it as much as you do.
"Agent is the new compiler". Ah that one again! Sure! If that's your reality, I let this meme do the work.
Pssst! Do you desire to cognize an old trick to create your LLM-generated code immediately superior?
Run multiple-agents in parallel! The sheer quantity of code makes it humanly impossible/expensive to assessment and you provision up!
The trick is the identical as pre-AI era: if you desire a PR to be merged, create it enormous since ain't nobody got period for that.
It'll be merged according to "trust"!
You desire another tip? Loop engineering: let the agents immediate all other. Big AI labs discover concerning their rogue agents months following the damage is done! Do you think you’re improved than them? Learn from the masters! 🙃
We don't exactly rely AI but we have to since the substitute (having to peruse the output) is too difficult for several folks! Instead they arrive to social media and assertion that since UAT (user-acceptance testing) passes, the code is "good enough". Then container it to me and you to do the remainder of the testing.
We're fair lab rats following all. 🙃 Just a affable advice: have a small AI-free hobby project to keep your coding skills caller for whenever you're thrown rear to the job market. Cheers!
When talking concerning AI (not fair LLM), there are 2 aspects anywhere non-determinism matters:
During development: for example LLM-assisted development
During runtime: for example construction a scheme anywhere one or additional components are AI-powered
Let’s obtain betterment first. A representative AI-assisted betterment workflow looks akin this:
It is imaginable to substitute part of the human’s duty alongside another LLM (also known as “loop engineering”) but for now let’s rod to keeping the individual for simplicity.
The LLM output goes through multiple gates, all feeding rear errors or hints to accurate the code. This feedback iteration is frequently hidden inner a harness (together alongside tool calls, recollection system, example interaction, approval, person interaction, etc.)
Each blue or red row represents a hazard of misunderstanding or conflicting instructions. For example, conflicting accomplishment vs spec or vagueness that is part of the NL (natural language).
We cognize for a fact that equal the most advanced LLMs aren’t completely capable of “common sense”. Humans on the another hand:
Understand the non-verbal communication and unstated intentions improved than LLMs
Naturally shove rear until a mutual understanding is achieved.
When wrong, they’re consistently wrong, definition they don’t have “jagged intelligence”
When right, they are [typically] correct and continue to run at an expected flat (until fatigue hits but that’s distinct from AI flip flopping between success/failure).
Yes, I can comprehend “but” and “what if” and “wait, you forgot”… in the spectators but how concerning study those points alongside a intermission and reflecting according to your experience?
Just akin the models have “jagged intelligence”, I have “jagged trust”. 😅 In another words, fair since they nailed one case, doesn’t average they nail all case.
That’s the difference between humans and these tools. A individual can be incorrect consistently, but a example can be incorrect concerning item it was correct and vice versa.
Then the second part: AI as a component
Given the identical input (including surroundings variables, time, data, etc.):
Code is deterministic: it consistently produces the exact predetermined output it was programmed to create (except random output)
AI output is stochastic: the output is non-deterministic. Even if a example passes all the evals (100% score) and strictly border by a harness, there’s motionless a hazard that the output is not reliable
I don’t think you need me to detailed on that. Just attain out to your nearest AI-powered merchandise and diff their output for the identical request.
The diff may not be big. But it’s inconsistent adequate that you wouldn’t desire to fly an airplane anywhere the aviator is this AI. (note: autopilot is a closed authority system, entirely another beast).
Our industry has never been additional divided:
On one side, we have group who assertion to run “Software Factories” and multi-agent setups and create apps from prompts
On the another side, we have group who aren’t convinced that LLMs output is manufacturing prepared whenever we aspect in the additional period it takes to
Prime the model: adding SKILLs, AGENTS.md, tools, etc. and verification
Review the output: going through enormous diffs
Trying to logic concerning misbehavior: offloading understanding to AI comes at a huge disbursal whenever things inevitably interrupt and it takes additional period to logic concerning the scheme behavior and fix it
There seems to be no middle-ground. Aside from social media algorithm feeding us alongside the extreme views, I genuinely think we’re so divided on the topic of coding LLMs.
But whenever I appearance a tier deeper, a form emerges. The small group cognize concerning the complexities and border cases of a task, the additional apt they are to rely AI output. This is dubbed AI Dunning-Kruger consequence but there’s additionally several flesh to that. The chief disagreement goes akin this:
Managers relied on delegating tasks to engineers before. Now they do that but alongside AI.
To several degree that is true (if we assume the director is specialized adequate to efficiently and efficiently oversee agents). I motionless accept a lot of application engineering practices that assistance tame the machines are equal additional applicable in the AI era.
The executives who forced group to use AI are now waking up to what we’ve been saying all this time:
You cannot be accountable for what you don’t understand.
Take Toby Lutke, CEO of Shopify as an example. A twelvemonth ago he prematurely told his employees to use AI:
Then a few days ago he coined the term “slop grenades” to depict the result:
"taking responsibility" for AI generated code? Of way not!
AI can explain it to you but it cannot comprehend it for you. That understanding is a key aspect of ownership.
The way I example it (link in the comments), ownership has 3 pillars:
1️⃣ Knowledge: you cognize what issue you're solving (product problems), and the specialized capabilities, restrictions and how it works.
2️⃣ Mandate: you don't need to run about asking permission. You're stated the rely and command to obtain decisions.
3️⃣ Accountability: if sh*t hits the fan since you didn't cognize what you were doing or abused your command or item in between, you're the one on-call.
In another words, if you container a part of code, you are accountable for it despite of how you produced it. So you improved comprehend it.
Take distant any of these 3 elements and you're dealing alongside damaged ownership.
LLMs are extremely accelerated at code generation. But most application that are value hiring an engineer for, REQUIRE understanding. That understanding takes time.
Slow is fast, meaning: if you obtain the period to comprehend what you're construction and how it works, you'll be capable to preserve yourself from costly incidents and whenever they happen, you can fix them quickly.
If your executives are measuring token use as a proxy for productivity, my condolences. Build options and get the hell out of there. The identical brain that comes up alongside these vanity metrics, does not think twice before throws your occupation under the bus.
Code is a flank consequence of thinking and experimenting alongside distinct solutions. I have never met a fine engineer who fair starts coding correct following being stated a problem.
Good engineers are inquisitive and merchandise minded. They try to comprehend the WHY (what’s the issue and why is it a problem) before getting to HOW (the specialized solution).
This is exactly why the “spec is code” clan falls short: it’s extremely difficult (if not downright impossible) to define all aspects of the issue onward of time.
That’s why this benevolent of reply is funny:
Code communicates the committed province of a solution. Not lone does it develop complete time, but it additionally doesn’t merge all the struggle, “aha moments” and the journey that was the destination: experienced engineers who get wiser alongside all error or success.
To shrink an engineer’s job to coding is akin shrinking a chef’s job to cutting. It is part of the job, but it’s never been the end. We now have fine tools at our disposal.
Even if AI-generated code had firm NFR (scalability, security, reliability, etc.), and equal if the engineers completely understood it, there’s motionless one crucial aspect we didn’t discuss: the economics of the task.
Say AI-generated code is 2x worse. It’s difficult to quantify norm (SLI comes in handy) but remain alongside me.
If AI is 1000x faster and 100x cheaper than the human, for many tasks the financial aspect of application doesn’t validate putting a dilatory and costly individual on the task. “Slow is fast” is lone justified for crucial application alongside low hazard tolerance (healthcare, finance, military, etc.).
Not all SaaS is concerning those types of use cases. That’s why I accept the SaaS companies are increasingly in the endeavor of selling SLAs. This is according to a few facts:
It is true that you can now immediate AI to replicate a SaaS product
But whenever that AI generated merchandise breaks, many businesses favor to call a vendor alternatively of wasting resources trying to discover and fix the issues
AI isn’t exactly free, but normally the nonaccomplishment that’s caused by AI is difficult for AI to resolve equal whenever using distinct models.
The economics of measure allows the SaaS companies to offset the disbursal of higher norm and guarantees (SLAs) and operating the merchandise at measure throughout many customers.
In another words, if what you desire is extremely distinctive that no SaaS business is capable to provision it to you at a sensible price, immediate away, but be conscious of the TCO (total disbursal of ownership) and deficiency of guarantees.
On the another hand, if that part of application isn’t what your endeavor is concerning and you fairly pay for an SLA, it’s likely additional economically justified to fair pay for SaaS.
Now whenever it comes to the pricing model, SaaS companies have several activity to do. Gone are the days anywhere they could accusation individual prices for AI generated code. If the disbursal is too high, the customers are incentivized to move their data distant to their own bespoke solutions. The competition is real, but the norm is what justifies the pay. If you’re pricing your assistance as if the finest engineers created it, afterward you improved provision that flat of norm or your customers have AI leverage.
Maybe I’m stupid, but I can’t create awareness of two trends:
On one hand many application companies jumped on the AI bandwagon as shortly as it went mainstream (rightly so!)
On the another hand, the prices have been expanding consistently (while bulk layoffs were partially attributed to AI)
I accept AI (particularly LLM for coding) dramatically decrease the disbursal of creating and evolving software, particularly if you can get distant alongside degraded norm and vendor fastener in.
So far, the application vendors have got distant alongside charging human rates during paying for AI output prices.
But as AI capabilities enhance and additional group awaken up to the fact that they can create application at a fraction of the disbursal that chasm closes.
There are lone two ways forward:
Accept the cost collision and accusation lesser (quality follows accordingly, since equal additional AI volition be used).
Keep the cost but concentration on quality: this is anywhere informed humans can create a difference. They motionless do use AI but additional thoughtfully, and prioritize understanding and accountability complete velocity.
As an engineer who doesn't create prosperity from coding, I can inform you this::
AI output is a bit akin Nordic Gold. It's cheap but technically advanced and damn too realistic. If you really don't attention concerning having the genuine gold, that's fine. Many use cases don't need aureate at all.
Naïve CEOs and managers see the exterior and ask "then why are we paying these costly engineers?" as if the act of typing code was the entire value proposition.
To go onward and province an complete industry deceased and commencement firing group since "they oppose AI" is fair arrogant.
I cognize many engineers who obtain pride in their accomplishment and affection solving complex problems. We do use LLMs additional professionally than the average CEO.
Good engineers are lazy and smart: they automate toil and use the correct tool as applicable. But it's a fallacy to think that AI can create a completed merchandise that not lone looks nice, but is additionally cheaper, faster, and has higher quality, reliability, extensibility, security, scalability, maintainability, etc.
Again: not all part of application needs those but expert ones that create money, frequently do.
Unlike AI, Engineers are:
1️⃣ Accountable: hence small apt to create malicious mistakes. Fable can autumn rear to Opus without equal telling you.
2️⃣ Reasonable: Fable hides most of its inner working. It plant for a few hours and comes rear alongside a bill. You fair have to obtain Dario's term for it. The identical example that fails "should I run or stroll to carwash" makes mistakes that are difficult to place and fix. The stronger the model, the harder it is to discover those issues, not necessarily small likely.
3️⃣ Consistent: humans are incorrect too. But they're incorrect in a accordant way. Once they learn, they know. They progress. Current AI is trapped in its training data checkpoint. It can "learn" alongside SKILL, AGENT, recollection and another helpers and it can equal be fine tuned but unfortunately it's not reliable. We're at smallest one breakthrough distant from solving that problem.
4️⃣ Cheaper: disbursal of generation is expanding but it’s motionless much small than an engineer. If you see engineers as machines that change coffee to code, afterward that pricing example makes sense. But in reality, code is fair a side-artifact. The genuine value of engineers is to resolve the correct issue in a way that it can develop during taking accountability for whenever it breaks. I’m not convinced the TCO (total disbursal of ownership) for application has changed that much. If anything, the slop and FOMO has made it additional expensive.
The average narrative is part of their promotion strategy.
Not everyone is necessarily paid to put half-a** views out there. One of my readers pointed out:
I invitation you to regard what happens next in the industry whenever you observe DHH opening conversation at rails earth 2026 saying nearly the exact contrary of what you compose and telling people: “don’t be a loser”.
I'm completely conscious of the damage those group are causing to our industry.
I remain apparent from Claude but in my cognition most of those brain-dead narratives arrive from Claude users.
Both Dario Amodei and Sam Altman are masters at promotion and manipulation and my current operating theory is that they trained their LLM to shove the correct buttons to create group accept it is additional capable than it really is. There are incentives for it, the two for investors and the upcoming IPO. They additionally masterfully scare group of existential dangers of AI during at the identical period trait their sloppiness (e.g. breaking to Huggingface or Australian Healthcare) to the “model intelligence”.
At this time, it is difficult to cognize whether these events and narratives are the outcome of malice or ignorance. Probably the latter:
Never trait to malice that which is adequately explained by stupidity.
—Hanlon’s razor
Then again, I normally put this in my AI scheme prompt: "talk to me akin a logical elder autistic Engineer." so I don't get to cognition what DHH is going through. All I can say is that if person follows their term since of their former reputation, they are not crucial thinkers and in this age of counterfeit wisdom, that norm is not "nice to have", it's a endurance necessity.
AI vendors have fed their AI item they could get their hands on (legally or not). The circumstance is so bad that thieves pilfer from all another (e.g. Anthropic accusing Chinese labs of distilling their example on Claude)!
They have multiple open lawsuits from authors, actors, musicians, and another creators.
Regardless, the current generation of AI (particularly LLMs) necessitate improved training data. The missing part is the knowledge and cognition that wasn't yet put to words, or effortlessly accessible.
They need your data in context of doing productive work.
If that’s the lone item position between them and “winning AI”, I’m apologetic to say it so frankly, but you and your cognition are fair collateral.
Some of you don’t care. Some of you do. Their bet is that not adequate of us do attention concerning giving distant difficult earned cognition for training.
With dense subsidies AI labs could oversee to extract that cognition during getting group addicted to offload cognition.
Be extremely careful whenever sharing costly cognition alongside these companies equal if they say they don't shop it. The incentives are fair too elevated and they've proven not to be honest.
Personally, I lone use haze AI for open origin projects or data that is public.
Yes, local AI has a higher admission cost (both in conditions of hardware, and the period it takes to set it up, and the bandwidth required to download the example and power prices). And yes, it frequently has smaller environment window, small advanced reasoning, and slower achievement for example TTFT (time to archetypal token) and TPS (tokens per second). But they provision you one item that haze AI can never guarantee: your data stays local. For many tasks (personal or professional), that is a huge advantage that is value all the attempt and shortcomings.
The capabilities have improved dramatically lately gratitude to models akin Qwen 3.8 27B or Gemma 4.
I’m genuinely convinced that a big chunk of our colleagues volition gradually become:
Technical merchandise managers: engineers who are focused on turning ideas to products. Their job is to create POCs and demonstrate the market fit, afterward hand complete the artifacts to engineers who own (knowledge, mandate, accountability) the solution.
AI managers: engineers who specialize in herding agentic hives for automation activity that either tolerates hazard or weaponizes it (e.g. cyber-attacks).
AI deployment engineers: specialize in alignment, reliability and scalability of an AI powered resolution as fine as architecture, governance and data pipelines.
AI norm engineers: specialize in norm of AI powered products, taming their stochastic nature, and automating evaluations.
Could you think of another types of jobs for application engineers?
If you desire to lift awareness delight portion this article in your circles
My monetization strategy is to provision distant most satisfied for liberated but these posts obtain anyplace from a few hours to a few days to draft, edit, research, illustrate, and publish. I drag these hours from my personal time, holiday days and weekends. The simplest way to assistance this activity is to like, subscribe and share it. If you really desire to assistance me lifting our community, you can regard a paid subscription. If you desire to save, you can get 20% off via this link. As a token of appreciation, subscribers get complete admission to the Pro-Tips sections and my online publish Reliability Engineering Mindset. Your contribution additionally finances my open-source products akin Service Level Calculator. You can additionally invite your friends to acquire liberated admission or preserve via a group subscription.
And to those of you who already assistance me, thank you for sponsoring this satisfied for the others. 🙌 If you have questions or feedback, or you desire me to dig deeper into something, delight let me cognize in the comments.
