{"slug": "a-letter-from-a-machine-learning-engineer", "title": "A Letter from a Machine Learning Engineer", "summary": "An anonymous email attributed to a machine learning engineer at a frontier AI lab claims the lab has hit limits on scaling models for programming tasks and that frontier labs lack a competitive moat, predicting an open-weights model as good as or better than \"Fable 5.1\" within a year. The letter, published by Nemin, also cites the 2023 Google memo stating \"We have no moat, and neither does OpenAI\" and argues AGI is not mathematically possible with LLMs. The author's identity and affiliation could not be verified, and the throwaway email address used to send it had been deleted.", "body_md": "## Table of Contents\n\nA couple of days ago, I received an email from a throwaway address, whose author claims to be an ML Engineer at one of the frontier labs. In this, they explain their beliefs about the future of the current LLM craze and how frontier labs are bound to eventually go under.\n\nI have no means of ascertaining whether the author was really who\nthey say they are, nor was I able to reply to them, because the address\nhad since been deleted.<sup>1</sup>\n\nI'm not one to air out my private mails normally and I deliberated\nfor a few days whether to make an exception here. However, considering\nthe importance of the topic in the current times and the fact that the\nauthor wrote it in a way that (per their own words) cannot be used to\nidentify them, I figured it'd be worth sharing with the greater\ncommunity.<sup>2</sup>\n\nI will first quote the mail verbatim for those who want to read it as a coherent essay. Then I'll repeat it while interrupting the text whenever I have something to say.\n\nBoth sections will contain the exact same text, so if you're fine with reading the cut up text with my annotations, feel free to skip to the second part.\n\n# 1 The letter\n\nHi Nemin,\n\nI work at a frontier AI lab as an MLE. I help train models and do RLHF for frontier stuff at one of the big companies you probably know.\n\nSorry this email is short. I have an NDA, so I need to keep this general.\n\nI saw your post about worrying about the future of programming. Honestly, I'm more worried about losing my job than yours. How is that even possible? Internally, we've hit limits on scaling models for programming tasks, which is why so many people are switching careers. At my company, a lot of executives have quit in a year, and there's real stress about the IPO. There is lots more I cannot share, but it does not look good internally.\n\nThe real threat to labs like ours is that there is no competitive moat. We have to keep advertising our math-solving abilities and overselling because things are changing fast. The proprietary diffusion models got largely replaced by open-weights ones. Image and video generation can now be done with open models without needing us. Sales tanked and that's why the frontier labs quietly stopped promoting image/video gen.\n\nLLMs are next. Even parts of the serious research community like LeCun have moved on from them. Most of the people who wrote the attention paper have left Google and scattered across new labs and startups. If you look at efficiency and size, the trend is clear. We'll see an open-weights model as good or better than Fable 5.1 within a year. There is research going on into parameter efficiency, so if it pans out, something like this could one day run on devices like high-end laptops.\n\nAGI is not mathematically possible with LLMs. There was a paper a couple of years ago arguing that, using Cantor's diagonalization argument. I can't remember the name for the life of me as I read it a long time ago. This realization changes everything. And a Google researcher leaked an internal memo in 2023 that said the quiet part out loud: \"We have no moat, and neither does OpenAI.\" feel free to read if you're interested.\n\nThe online discussion is poisoned. A lot of the voices driving the hype, the influencers who sold those online courses, are now just being sponsored by the AI companies themselves. This is why programmers are so stressed, I think. Don't trust anything you hear or read on mainstream sites. A few engineers have been deploying agents as \"fun projects\" to market models and scrape data (that's how I found your email), and you should absolutely ignore the forecasts from the CEOs. They are completely clueless about what's actually happening.\n\n# 2 My thoughts on the individual sections\n\nHi Nemin,\n\nI work at a frontier AI lab as an MLE. I help train models and do RLHF for frontier stuff at one of the big companies you probably know.\n\nMLE is Machine Learning Engineer, this I figured out on my own,\nbecause it was pretty obvious. However, I had no idea what RLHF is (only\nheard of RTFM and GLHF before :), so I looked it up. It means [Reinforcement\nLearning from Human Feedback](https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback), i.e. a particular variant of agentic\ntraining, where the reward function (what determines how \"good\" an AI's\nresponses are) is first manually tuned by human operators.\n\nSorry this email is short. I have an NDA, so I need to keep this general.\n\nHard to fault the author (who I'll refer to as \"G\"[<sup>3</sup>](#fn3)\nfrom now on, due to the fake moniker they used in the mail) for this.\nWhile [NDA](https://en.wikipedia.org/wiki/Non-disclosure_agreement)-s\nare [not\nuniversally enforceable](<https://www.law.cornell.edu/wex/non-disclosure_agreement_(nda)>), they are still a sword of Damocles above\none's head of \"Talk too much and you'll never work in this field\nagain\".\n\nI saw your post about worrying about the future of programming.\n\nBy \"my post\", G means my [commitment](../llm/index.html)\nto not using LLMs for hobby projects, due to their ethical issues, the\nperceived uselessness of AI, and the love of the craft.\n\nSince writing this declaration my opinion on LLMs has slightly shifted or, rather, I'd say became more nuanced:\n\n**For one,** I can no longer say that the output of the\nvarious coding agents is useless and I think anyone who keeps insisting\nthat they are still outputting pure gibberish has not tried an agent\nrecently.\n\nI still do not trust vibe-coding, as I believe that unsupervised development results in write-only apps, that cost exponentially more to maintain over time and are impossible to comprehend without the constant assistance of the agent.\n\nThis, in my opinion, is tying yourself to a kind of \"supply chain\" of its own, where the token prices are in constant flux and aren't based on any real, measurable metric. It also does nothing to improve your own understanding and mastery over computers. I'm sure, regardless of this, you can \"get rich quick\" by vibing up something, but I'm not very interested in that.\n\nHowever, after being strongly recommended at my workplace to try\nusing AI, I found that agents are great at being a sort of \"secretary\",\nwho can fetch you information about the codebase based on free-form text\nqueries. While `grep`, `find`, etc. can get you\nfar, \"vibe searching\" takes things a step further in a way, I couldn't\nreally replicate with conventional tools.\n\nI also found, that using an agent as a pre-reviewer before you send in your PR for a human to actually take a look at is a nice way of avoiding some unnecessary back and forth. While I still ultimately defer to human judgement (including my own), it is nice to catch issues such as \"hey, this line is potentially dereferencing null\", \"this comment wasn't updated to what the code actually does\", and similar nitpicks.\n\n**For two,** and this might sound a little paradoxical,\nafter I just conceded agentic output being better than I once thought,\nbut I no longer expect programming as a job to disappear. In fact, I\nthink people who understand how these mechanical parrots \"think\" have\nmore job security than ever.\n\nBack when writing my first post, I was genuinely somewhat worried,\nthat after almost five years of university (and before that more than a\ndecade of hobby programming) will just become *obsolete* once\nmodels become good enough to fully replace us.\n\nHowever, while I've seen a lot of improvement to the quality of agentic output, I still experience a ton of flukes as well. My company allows us to use top of the line models and they are both incredibly capable at times and terribly, horribly dumb.\n\nAmong other things: The model we use still occasionally hallucinates functions that simply don't exist, despite being plugged into LSP. It occasionally moves code into conditionals, that should happen in both branches. And sometimes it just starts using legacy code that's clearly marked as such as a style to follow. And that's just a small sample.\n\nYou simply need a human in the loop to make sure the machine doesn't do anything stupid. And, as mentioned before, I'm absolutely not the sort to let the agent loose. I rarely allow it to generate anything more than 10-20 lines. How could it then replace me as a whole?\n\nInstead, I believe the recent lay-off wave, that was conducted \"due\nto AI performance optimization\" was really just CEO-s buying into hype\nand the market experiencing a painful normalization in the post-COVID /\n[ZIRP](https://en.wikipedia.org/wiki/Zero_interest-rate_policy)\nera.\n\n**That being said,** as convenient as LLMs are, the\nmoral concerns and the fact, that I want to keep my skills sharp still\nmakes me want to relegate AI usage to my job only.\n\nHonestly, I'm more worried about losing my job than yours. How is that even possible? Internally, we've hit limits on scaling models for programming tasks, which is why so many people are switching careers. At my company, a lot of executives have quit in a year, and there's real stress about the IPO. There is lots more I cannot share, but it does not look good internally.\n\nFunnily enough, I assumed both Anthropic and OpenAI have already\nIPO-d, but it turns out neither of them have. I guess the numbers [don't look good](https://isaiprofitable.com/) enough to bring\nthis stuff onto the stock market.\n\nAs for the rest of the claims, I have no real means of checking the\nvalidity of them. There has been that story about a researcher [quitting](https://www.businessinsider.com/anthropic-researcher-quits-over-ai-safety-concerns-2026-9),\nbecause they fear AI will endanger humanity, but this person wasn't an\nexecutive and they have seemingly left for ethical reasons, not\nfinancial ones.\n\nThe real threat to labs like ours is that there is no competitive moat. We have to keep advertising our math-solving abilities and overselling because things are changing fast. The proprietary diffusion models got largely replaced by open-weights ones. Image and video generation can now be done with open models without needing us. Sales tanked and that's why the frontier labs quietly stopped promoting image/video gen.\n\nThis part feels particularly striking after the recent [controversy](https://simonwillison.net/2026/Sep/8/on-navier-stokes/)\nrelated to progress on the Navier-Stokes equations, which OpenAI\npotentially snatched from the hands of two researchers.\n\nFor clarity, G's mail arrived OpenAI published its findings, so this was a reaction, not a forewarning to it and it is no evidence that G may be from OpenAI.\n\nAs for the other part of this quote, it really is interesting how\nmuch AI ~~slop~~ videos became a cheap commodity, that is simply\nnot worth selling as a singular product. Back in the day we had things\nlike Sora and then it went belly up, when the math wasn't mathing about\nits economics. I wonder if one day selling tokens will be seen as a\npointless venture, because they'll be so commodified.\n\nLLMs are next. Even parts of the serious research community like LeCun have moved on from them. Most of the people who wrote the attention paper have left Google and scattered across new labs and startups. If you look at efficiency and size, the trend is clear. We'll see an open-weights model as good or better than Fable 5.1 within a year. There is research going on into parameter efficiency, so if it pans out, something like this could one day run on devices like high-end laptops.\n\nG is talking about [Yann LeCun](https://en.wikipedia.org/wiki/Yann_LeCun), known\nfor his work on ML and computer vision, and who in [April\nof this year](https://www.brown.edu/news/2026-04-01/yann-lecun-artificial-intelligence-pioneer) engaged in a lecture about LLMs being a dead-end and\nthat a new approach would be necessary, if humanity really wanted to go\nfor AGI.\n\nDespite this, he is also [involved in](https://thealliance.ai/projects/tapestry) a\n\"collaborative foundation for open and sovereign AI,\" so take that as\nyou will.\n\nAGI is not mathematically possible with LLMs. There was a paper a couple of years ago arguing that, using Cantor's diagonalization argument. I can't remember the name for the life of me as I read it a long time ago. This realization changes everything. And a Google researcher leaked an internal memo in 2023 that said the quiet part out loud: \"We have no moat, and neither does OpenAI.\" feel free to read if you're interested.\n\nI suspect G is thinking of [*Hallucination is\nInevitable*](https://arxiv.org/abs/2401.11817), a frequently-cited paper from 2024, which proved\nthat it is impossible to make an LLM, that doesn't hallucinate (i.e. act\nas a perfect general problem solver).\n\nAs for the other paper, it may be reached [here](https://newsletter.semianalysis.com/p/google-we-have-no-moat-and-neither).\nIt is an allegedly authentic memo of an anonymous Google employee, who\nargues that ultimately the big frontier labs are living on borrowed time\nand eventually open-weight models will not only catch up, but surpass\nclosed models.\n\nI'm not nearly educated enough to discuss either papers in detail. My\nonly input is that *We have no moat…* is from 2023 and so far (at\nleast based on my limited knowledge) it doesn't really feel to me like\nthe things that Google employee was so worried about came to pass.\nClosed models are still king and while there are now capable models,\nthat one may run on consumer hardware, it's still not on par and is also\nquite expensive in terms of hardware.\n\nI guess G did specifically touch on this in the previous segment, where they posited, that we'll see an open-weight Fable 5.1 till the end of 2027, but until that actually happens, we can only work with the assumptions we have.\n\nOn the other hand, it is also true, that Google's focus on LLMs has\nshifted hard from agentic stuff into making people engage with their\nsearch in a new way. So perhaps that is Google's *moat,* their\ndata and the fact that nobody else is connected so directly to it.\n\nThe online discussion is poisoned. A lot of the voices driving the hype, the influencers who sold those online courses, are now just being sponsored by the AI companies themselves. This is why programmers are so stressed, I think. Don't trust anything you hear or read on mainstream sites. A few engineers have been deploying agents as \"fun projects\" to market models and scrape data (that's how I found your email), and you should absolutely ignore the forecasts from the CEOs. They are completely clueless about what's actually happening.\n\nI have no doubts this is the case.\n\nAll I really see online is either doomerism or accelerationist\neuphoria, and the less 'aggressive' (be they pro- or anti-AI) voices are\ndrowned out by the hypebeasts and antihypebeasts. Some of these voices\nbeing paid shills doesn't feel like a massive leap of logic to me.<sup>4</sup>\n\nAnd I'm pretty sure any larger company's upper management is expected to buy into the Current Thing, because if it turns out to deliver, you don't want to be the chump who was left out by not jumping on the bandwagon.\n\n# 3 Closing thoughts\n\nSo, do I believe G? To an extent, sure.\n\n- I believe that there is no secret \"silver bullet\", that would suddenly catapult LLMs from a versatile, yet wonky tool into AGI.\n- I also believe, that CEO-s are absolutely hype-chasing, just like Big Tech was previously chasing the Blockchain, NFTs, dotcom, etc.\n- I even believe, that one day (perhaps in the near future) we'll be running these things on cheap hardware and they'll be about as magical to the average person as a smartphone is.\n- And I also believe some of the huge labs that have impossibly large valuations now will not survive the end of the decade.\n\nHowever, everything in this email is either post-facto or unverifiable by me, so I cannot adequately confirm if it the identity of the author is legit, and the mail wasn't written by a very convincing \"layperson\".\n\nI don't mean to insult G with my doubts. I understand, why they had their hands tied, NDA and all. In fact, if they ever read this, I'm very grateful to them. They saw a person clearly anxious about the future and tried their best to console them.\n\nMostly, I'm just trying to protect my own credibility and cool any\nreactions a reader of this article might potentially feel. I\n*think* the mail is legit, but I cannot know. And even if it is\nlegit, it is just one researcher's beliefs, we (sadly or thankfully)\nhave no means of scrying the future.\n\nIf I had one qualm with the mail, however, it'd be that I wish G\nexpressed it in more detail what they expect the future to look like,\nbecause I think \"everyone will be able to run LLMs\" and \"it wont be AGI\"\ndon't really address what *really* makes people nervous.\n\nI mean topics such as copyright washing, write-only apps and the devaluation of programming as a field, the environmental harm from data centres, and the onslaught of AI-generated ragebait / disinformation. None of these would be really fixed just with frontier labs going under and LLMs becoming even more widespread than they already are.\n\nAnd then there are the questions of, for example, how will these\nhomebrew models get their training data? Will such a proliferation lead\nto an even stronger focus on scraping the open web? What will happen to\nthe current money whirlpool of AI labs->NVIDIA->Chip makers->AI\nlabs? Where is the field expected to go next if LLMs truly are an almost\n2 trillion dollar[<sup>5</sup>](#fn5) mistake?\n\nAll questions an actual researcher with insider info could make far better guesses about. But I suppose we can't have everything and, at the end of the day, the point of this mail was to make me feel a little better. I found G's words interesting to digest and I hope you have too.\n\nThank you for reading!\n\n# 4 Footnotes\n\n1. I know this, because both of us used Protonmail, which tells you if the account associated with the address has since been deleted. [↩︎](#fnref1)\n2. If you ever considered writing me an email (I don't exactly hide my address too hard), please don't feel discouraged by this. This was an extraordinary case. If you write even with an identifiable alias, let alone your real name, unless you give me explicit instruction and permission, I'll consider your message confidential and my eyes only. [↩︎](#fnref2)\n3. G was using a name from a fictional series. Even if I wrote it out in full, it'd not be identifiable in any way, because it'd be as if I said \"I got an email from a Xenomorph\". However, I want to do my due diligence to further conceal their identity and thus won't even write out the whole word here. [↩︎](#fnref3)\n4. Truthfully, I don't really follow the \"latest\" beyond second-hand news from seeing what's posted on Hacker News and Lobste.rs. Therefore, even if I wanted (not that I want), I couldn't point fingers, because I intentionally avoid tech influencers. [↩︎](#fnref4)\n5. I originally wrote 1.5 trillion here, because that's the last number I remember, but it turned it *Is AI Profitable Yet* is now countin 1.9 trillion. Terrifying and unthinkable numbers,\ngenuinely.[↩︎](#fnref5)", "url": "https://wpnews.pro/news/a-letter-from-a-machine-learning-engineer", "canonical_source": "https://nemin.hu/llm-letter/index.html", "published_at": "2026-09-14 17:58:22+00:00", "updated_at": "2026-09-14 21:32:53.658622+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-safety", "ai-startups", "ai-research"], "entities": ["Nemin", "Google", "OpenAI", "LeCun", "Fable 5.1"], "alternates": {"html": "https://wpnews.pro/news/a-letter-from-a-machine-learning-engineer", "markdown": "https://wpnews.pro/news/a-letter-from-a-machine-learning-engineer.md", "text": "https://wpnews.pro/news/a-letter-from-a-machine-learning-engineer.txt", "jsonld": "https://wpnews.pro/news/a-letter-from-a-machine-learning-engineer.jsonld"}}