Dan Luu’s new post is about a habit he started noticing in early 2025: handing an LLM a task — summarize this text, write this code — and simply assuming it worked. Back then it usually produced silly results. By September 2026 he says the practice has spread and improved to the point where “for loop meat proxying” (accept the model’s output; if it fails, ask the model to fix it) produces software that sort of works. He’s careful to say he’s impressed by how far that’s come.

Then he grants the optimistic case and asks an economic question instead of a technical one.

  • Suppose LLMs improve enough that brain-off development produces average-quality software, or even good software with no human involved at all. What reason would a company have to keep employing the meat proxy? It can just run the LLM in a loop and lay the person off.
  • His conclusion: “There’s no point at which this methodology will work for the employee.” Better models don’t rescue the person delegating to them; they eliminate the need for that person. The technique can pay off for the employer and still be a dead end for the employee.
  • The post borrows “meat proxy” from Niklas Gruhn, who coined it for relaying model output in conversation and code review. Luu extends it, in a footnote, to acting as the loop yourself — for and while loops, not just copy-paste.

The evidence lives in the footnotes, and it’s concrete:

  • Luke Burton’s walkaway rule: he’ll only leave work unattended if it’s low value and “can afford to fail.” On high-value work he has to act as QA, engineering manager, and architect at once, and says the loop “often feels like a crunch time.”
  • Burton’s Bazel migration — a task that sounds like ideal automation — took months with agents, because it was full of requirements nobody had written down. He puts “convert this to Bazel” and walk away at years out, “maybe not ever.”
  • Out-of-distribution work is where the gap shows. Agents are worse in obscure programming languages, and on a modern board game like Lost Cities or Dominion, a top model plus harness plays worse than a reasonable human who has never seen the game. The model sounds right to someone who doesn’t know the game and obviously wrong to someone who does.
  • Luu checked the receipts on a public claim that programming is solved. The GitHub he looked at “either didn’t work or worked very badly” — including a game bot weaker than what you get from prompting a model to write a simple minimax opponent, which should have been demolished by the bot that was posted.
  • Gary Bernhardt’s version, quoted in the post: reviews that “cut the diff to 25% of its original size,” and a one-line database-URL fix that arrived as roughly 20 hunks of conditional shell scripting. After correction: “+0 lines, +1 word.”

Luu is not arguing against the tools. He says he uses agents for data analysis in a loop, knowingly getting wrong results he then directs them to fix, and he builds personal software he’d call broken if it shipped. The line he draws is between producing work you know is wrong and cleaning it up, and mistaking fluent output for a finished artifact — or shipping it at that quality and declaring the field solved. Thomas Dullien, on reading the draft: people look at him like he’s crazy when he says LLMs don’t solve all programming problems, “and I look at them like they are.”

There’s also an incentives argument in the last footnote. “We’re all going to be obsolete soon, so why bother” only makes sense if obsolescence is certain and immediate. And the imagined window where you can coast as a meat proxy during the transition has it backwards — companies are treating AI as their excuse to cut, which makes this the worst time in decades to be collecting a paycheck without producing anything.

The 150-comment thread on Hacker News

What the thread adds

  • Arubis — relocates the important cognitive work. It isn’t in the loop: “The most important time to use your brain is before the prompt. Once you engage with your LLM and agent, you start biasing yourself, and its reasonable suggestions constrain your visibility into other options, other worlds.” Their prescription: “First, use your brain. Then write the prompt.”
  • Arainach — says the pressure is organizational, not individual: “any time I argue that AI isn’t the right tool for something management gets angry… There is a constant push to fully automate things, regardless of what those things are.” Two replies reinforce it — bunderbunder argues the executive sales pitch was headcount reduction, so whether the work is currently faster or better is beside the point, while awsglkhj reframes the whole arrangement as hiring someone to absorb blame: “At this point it should be clear that I’m not interested in quality work, but in having someone poised to absorb blame. I want a patsy.”
  • supermdguy — a practitioner’s calibration problem. Roughly 80-90% of tasks can be done “with ~10% brain power,” but identifying the minority that needs real thought is genuinely hard, and re-orienting into the code to do it “adds a lot of friction.” That matches Luu’s complaint more precisely than laziness does.
  • dr_dshiv — names the state “zombie mode” and offers a countermeasure: “saying ‘gnite’ and coming back in a new terminal window with a clear restatement of my purpose. Or ‘what are you working on and why’ can help.”
  • kixiQu — the pattern predates LLMs. They’ve known colleagues who wanted complex cross-system investigations collapsed into a runbook: “If a drinking bird on an Enter key could do it, I could rig something up so I wouldn’t need you!”
  • shadowgovt — sharpens the deskilling question rather than assuming an answer: is this the deskilling where you can no longer farm without understanding crops, “or the kind of deskilling where you can drive without knowing how to maintain your own internal combustion engine?”
  • gwbas1c — reads the whole thing as the outsourcing debate of twenty years ago with one change: “the feedback loop is within minutes instead of overnight.”

The answer the thread gives, and doubts

Luu asks what reason a company has to employ the meat proxy. The nearest thing to a consensus answer is liability — and three commenters arrive at it independently, which is why it reads as one position with three names rather than three arguments. 01284a7e proposes “the AI patsy”; science4sail calls the human “compliance ablative armor,” whose job is to watch the model, occasionally interfere, and “incur social/legal liability for anything that the LLM does wrong,” like the backup driver in a self-driving car; LZ_Khan generalizes it, noting airline pilots and train conductors are “basically meat proxies 95% of the time and they are still employed,” because they’re needed when things really go wrong. kg supplies the objection the others don’t answer: is being ablative armor “uplifting, fulfilling work?” and does personal competence change the outcome at all? “That sounds miserable.” No one in the thread shows a mechanism by which liability duty builds into job security rather than a slow exit.

Where the thread pushes back

  • vintermann disputes the premise’s vocabulary: “the person who forwards a machine-answer to someone else who didn’t ask for a machine answer” is the meat proxy they’ve heard of, and “Calling someone a ‘meat proxy’ for being the human in the loop is wild.” jadar answers with a distinction — if you’re genuinely thinking, you aren’t a meat proxy; the failure is “cognitive surrender,” and the people who get real value from these tools have the discipline to stay in the loop, even if “it’s still hard to find the joy and reward when you’re not the one solving the problems.”
  • dzink offers the strongest counter-thesis, via photography: the old photographer measures angles for one perfect shot while a tourist with a phone “spin[s] around and get[s] thousands of photos,” so “make and filter now may be a better approach than overthink and try only once.” Two rebuttals land. ssfdg says both photographers still had to choose a subject and recognize which frame was worth keeping, and that the AI version “would just vomit up endless permutations of statistically average images.” btrettel makes the empirical point: the credulous users Luu describes aren’t comparing thousands of attempts at one problem — “They take a couple photos at most, say ‘LGTM,’ and move on to taking a photo of something else.”
  • jochem9 and stephbook take the inertia argument further than Luu allows. jochem9 notes job security often depends more on being embedded in an organization than on output, and stephbook concludes against the title: “Sounds like a quiet few years and then effort once demanded. So I guess… being a meat proxy does work, after all?”
  • agentultra states flatly that “An AI system cannot replace professional people,” while krapp counters that this is already observable: “Developers are already losing jobs. LLMs and agents are already running in the hot loops of our technological civilization, unsupervised, with root.” Neither side produces evidence, and the thread doesn’t resolve it.

Handles here are pseudonymous, HN publishes no per-comment scores, and the ordering above follows HN’s own ranking — this is a slice of the thread, not a consensus.