I'll quote you the rest of that paragraph that you -somehow- failed to read:

  The first point is the question of whether an LLM can do work not markedly inferior to that of a human doing the same task or not. This one is debatable. An LLM absolutely cannot produce work to the highest human standards, or even a good human standard at the moment, and it's unlikely that it ever will. Most human work, however, is not done at a good standard. At best it's mediocre, and at worst it's downright horrible. The bulk of human knowledge work done in an awful of settings, then, is unfortunately more or less interchangeable with that of an LLM (seriously, so much human work is just awful: I wish I could say otherwise). On this point, then, LLMs probably pass the criterion.

Regardless, I'd recommend that you take 15->30 minutes and read the entire essay. It's a good read.

For analysis that justifies the claim that "[a]n LLM absolutely cannot produce work to the highest human standards, or even a good human standard at the moment, and it's unlikely that it ever will", check out [1]. In order to understand that essay, you will need to read it carefully. If you skim that essay or have a lossy compressor "summarize" it, you will not understand it.

If you're prepping a "The author just hates LLMs and has never ever used one, let alone used one seriously... they're just a hater!" retort, carefully read [2]. That essay also provides further support to their claim that LLM-based tools don't currently produce good work.

[0] <https://news.ycombinator.com/item?id=50032485>

[1] <https://deadsimpletech.com/blog/llms-on-the-boundary>

[2] <https://deadsimpletech.com/blog/engineering-judgement-claude...>

And there are so many indicators that it's not currently economically viable and will only be if retail price is massively increased and strong regulatory barriers to competitors are erected. ;)

"And where would she move to?" -- Either the graveyard or the incinerator? She's dead... the bill doesn't come due until after her death and is settled from the value of the house.

I agree!

> ...and you're breaking the HN rules by insinuating it.

It's a good thing that I'm not, then.

> ...just because I'm rebuking a flawed argument with basic facts.

cough

Eyuuup. Even on The Red Site (which usually has fairly sane commentary... as long as simonw stays out of the conversation thread), the first commenters on the submission about this essay have very obviously either not read the entire essay, or are pointedly ignoring the parts of it that either conflict with or directly address the claims in their comments.

As is sometimes said by idiot Twitch streamers: Fuck my chungus life, I guess.

I see that the 1.5 million USD State Department grant is listed on that return. I do hope that they weren't obligated to return that. The Award History section here [0] suggests that they were, but I don't know much at all about how these sorts of grants work.

Anyway. This is where Tor Project seems to officially publish their tax documents [1], but I can only find docs up to 2023->2024...

Notable is that in 2023->2024 (the year prior to the document you link to) Mullvad's funding was 1.4 million USD, whereas in 2024->2025, Mullvad's funding was 306 thousand USD. In that year, the Swedish International Development Cooperation Agency provided nearly twice the funding that Mullvad did! Perhaps Mullvad donated a enormously large amount of money in 2025->2026, but, yeah, the numbers don't support the "the Tor Project will start censoring because Mullvad is paying them so much" story.

Again, thanks for digging up that tax return.

[0] <https://www.usaspending.gov/award/ASST_NON_2534285_049#award...>

[1] <https://www.torproject.org/about/reports/>

Wow, this was ejected off of the front page. It's currently sitting at page 10 two hours after posting. Based on what I've seen over the years, carefully-written criticism about the products of the LLM industry rarely does do well here... for whatever reason.

I'll include this excerpt from the essay as an attempt to shame folks who feel the urge to continue to claim that the author is critical of "AI" because they hate the tech and never ever have anything good to say about it:

> I will say that Claude Code is a noticeable step change over the Opencode-based GLM models that I tried a few months ago. Those were bad enough to not be usable for anything serious or of real complexity. Claude Code, by contrast, can do some relatively complex things: it was able to perform the bulk of my website migration from Nuxt 3 to Nuxt 4 and to the newest version of Nuxt-Content for instance, which the GLM models had utterly barfed on. It wasn't perfect: I had to do quite a bit of tinkering after the fact to make things work, but unlike the GLM attempt, Claude did actually save me some time. Claude also did pretty well with adding Schema.org markup to my website and performing a number of other stupid-but-necessary SEO-related tasks on the same website: none of these tasks are really groundbreaking, but they're the kind of relatively complex pain-in-the-ass task that you'd hope we could automate: at this point in my practice, there's little of value that I can glean from doing those tasks by hand. If it's possible to make this kind of tool accessible to people at an acceptable, sustainable cost and without burning too many resources, then, I think that'd be a good thing.

> Claude Code also does quite a bit better than Opencode tools did at CI/CD tasks, IaC in OpenTofu, building Dockerfiles and that kind of task. It's still something I'm quite iffy on, but unlike the GLM models I tried earlier, what Claude Code produces basically works, at least initially. This probably makes it a lot more immediately useful than the GLM models were, as this kind of task is a large part of what tech workers actually do day-to-day, and it's the kind of task that a software engineer might have to perform while not having a very good grasp of the underlying technology.

I recommend folks take 15->30 minutes to actually read the essay, rather than either skimming it or having some LLM "summarize" it. There's a lot of subtlety and nuance that the current generation of lossy compressors is incapable of reliably compacting.

[0] <https://deadsimpletech.com/blog/engineering-judgement-claude...>

Also, I'd correct your opening sentence like so:

> This is the thing: they probably AI "summarized" it.

I've seen many "summaries" that have come out of those lossy compressors... when tasked with compressing inputs of any appreciable complexity, the output is often too poor to serve as a summary. What you get sure is summary-shaped, though.

> Most of the comments in this thread are, unfortunately defending the thing that will make them less useful and less relevant...

If your claim here is that most of the comments so far are astroturfers or similar, ignore the rest of this line. If it isn't, I'd recommend that you read TFA in its entirety... or go back and read it more carefully if you already did. A sizeable section of the essay talks about what I think you're gesturing at.

  yep. This is the choice quote.
  
  Even most humans cannot produce work to the highest human standards. The fact is, most work does not need to be done to the highest human standards, it simply needs to be done to the lowest possible standard that someone would pay for, at the cheapest price.

  The first point is the question of whether an LLM can do work not markedly inferior to that of a human doing the same task or not. This one is debatable. An LLM absolutely cannot produce work to the highest human standards, or even a good human standard at the moment, and it's unlikely that it ever will.  Most human work, however, is not done at a good standard. At best it's mediocre, and at worst it's downright horrible. The bulk of human knowledge work done in an awful of settings, then, is unfortunately more or less interchangeable with that of an LLM (seriously, so much human work is just awful: I wish I could say otherwise). On this point, then, LLMs probably pass the criterion.

glances over at their supporters page [0]... I mean, that's not literally impossible.

While it is true that Mullvad supports at "Shallot Onion" -and has since mid 2023- [1] and "Shallot Onion" is Tor Project's highest support tier, we don't know how much Mullvad is paying. Six years ago, the smallest amount one could donate to get into that tier was 100k USD [2]. In contrast, just one of their eleven (or twelve, depending on how you count the two listed US State Department entities) supporters provided 500k USD per year in 2022->2024, and "committed" to providing 750k USD per year in 2025->2027... until -one assumes- DOGE canceled that grant in its "efficiency" sweeps.

FUTO is another of the eleven-ish entities on that page. While FUTO doesn't seem to mention how much it granted to the Tor Project, their grants page [3] claims that they've granted five million USD to twenty entities in their "Legendary Grants" -er- program, for an average of 250k USD per entity.

There's a lot that we don't know about the details of the financial situation of the Tor Project... but while it's possible that they become captured by Mullvad's money, it seems unlikely to me.

[0] <https://www.torproject.org/about/supporters/>

[1] <https://mullvad.net/en/blog/2023/3/3/mullvad-becomes-highest...>

[2] <https://archive.torproject.org/websites/lists.torproject.org...>

[3] <https://futo.tech/grants>

What specifically is being done to persecute this hypothetical pig you refer to? As always, the details matter.

In regards to the Tor Project's public statement that is TFA, they've chosen to stop advertising with Mullvad [0] because they strongly disagree with specific political positions that a Mullvad co-founder has publicly endorsed. They're not backing out of working on the Mullvad Browser because they figure that stopping that work -and losing the new feature testbed and (presumably) the money they get from it- would hurt more people than the implicit endorsement that some folks will see by continuing to work with Mullvad.

[0] ...this is referred to as "proactive co-branding" and "joint promotional activity"...

If you're talking about US courts, then no. If they had, then we wouldn't see -for example- SCOTUS ruling that "geofence warrants" were searches that require Fourth Amendment analysis, rather than being equivalent to picking up a recording of events that happened in plain sight. We also wouldn't see at least one Federal judge openly noting that the judges that cite US vs. Knotts really need to read the entire opinion, rather than just the bit that agrees with what is convenient for them.

US v. Knotts is the forty-year-old case that is the source of the "You've no right to privacy when you're on public roads" idea that both people and judges have been parroting for quite a while now. In addition to establishing that principle, it also considered a possible future where the electronic surveillance that -at the time-

* had to be manually installed into cargo destined for the target's vehicle

* was only capable of tracking a single target vehicle

would become sufficiently advanced as to permit 24/7 dragnet surveillance... at which time, courts would need to reconsider what was just and right in light of such dreadfully advanced mass surveillance capabilities.

Pages 28 (starting at section C.) through 38 of this judge's opinion and order [0] are good reading... it can be hard to tell, but the judge really takes other judges who rely on Knotts but clearly haven't actually read it to task. If you have a half hour free, you should really read the whole thing.

[0] <https://storage.courtlistener.com/recap/gov.uscourts.oknd.76...>

One can publicly declare a speaker to be a wrongheaded pig that all decent people should ignore, but that's a wholly different thing from ensuring that said pig is unable to utter its wrongheaded ideas in public.

You've never heard statements such as "I disapprove of what you say, but I will defend to the death your right to say it."?

That is, you don't draw a real distinction between repeatedly, loudly, and publicly calling someone a cotton-headed slop-spouting pig to whom no one should pay any mind and actively working to ensure that said pig is prohibited from spouting its slop?

Second, why are you doing this check manually? Every major consumer OS and web browser does this for you:

Firefox does it automatically regardless of OS [0] by hitting 'firefox-portal-detection.com'. Windows does this [1] by hitting one of a few domains, depending on OS vintage. OSX and iOS are said to do this by hitting 'captive.apple.com', and Chrome by hitting one of a few domains, but I can't find corroborating official docs. Moreover, this report [2] published by the WBA ~six years ago [3] confirms that every major OS except for Linux -as well as some minor ones- does this automatically for you... and both Firefox and Chrome do it regardless of OS, so the lack of OS-level detection on Linux doesn't matter. I'd be shocked if things have gotten worse in the past six years, especially with the publication of relevant RFCs back at the turn of the decade.

Don't be a gaping asshole. If you must manually hammer on something, hammer on <http://cloudflarestatus.com> or similar. Don't hammer on example.com.

[0] <https://support.mozilla.org/en-US/kb/captive-portal>

[1] <https://learn.microsoft.com/en-us/troubleshoot/windows-clien...>

[2] <https://captivebehavior.wballiance.com/>

[3] <https://github.com/wireless-broadband-alliance/captive-behav...>

As someone who has done this, it's a one-time cost (as long as you're not the sort who simply can't stop tinkering with it and ends up totally rebuilding it like once a quarter (don't ask me how I know)) and -if you have even just a shaky understanding of how to do it- it's not _that_ large of a cost.

> ...and IT knowledge though.

I definitely agree that doing this requires quite a bit of IT knowledge... but it's all stuff that's pretty easily learnable for anyone who's interested in technical stuff and/or technically-inclined.

For folks who are looking to do this on their home LAN, I have some hardware manufacturer recommendations:

All of this VLAN work will be entirely pointless if your switches can't be programmed to enforce the separation, so one will need "managed" switches of some kind. I'd recommend anyone who wants to try to do this to have a look at Mikrotik switches... they are inexpensive and definitely more than good enough for a fancy home LAN.

Mikrotik also sells routers and WiFi APs. I can't comment on the quality, as I have slapped together my own router PC and use OpenWRT Ones for my APs... but I've found their switches to be more than good enough for my fancy home LAN. Perhaps their routers and AP are equally good?

Mikrotik publishes pretty comprehensive documentation here [0]. If you want to dick around with the Mikrotik management CLI for RouterOS -which is their name for their fancy management software- you can install the x86 version of RouterOS in a VM by booting a VM from one of the x86 install images at [1]. They also have a much simpler management software that you can run on all of their switches called "SwOS" -documented here [2]- but that doesn't have any x86 installation media so you can't play with it on your PC.

[0] <https://manual.mikrotik.com/docs/introduction>

[1] <https://mikrotik.com/download?architecture=x86>

[2] <https://manual.mikrotik.com/docs/bridging-and-switching/swos...>

...homeboy here is behaving as if there aren't experts in this world who are paid to do and publish research and other experts who are paid to collate, analyze, vet, and publish that research for other domain experts to learn from. [0]

I'm fine with my doctor reading from expert-vetted documents. Hell, I'm more than fine with my doctor going through expert-vetted checklists; most of the time what's wrong with you can be easily figured by going through a few good checklists. I'm not fine with my doctor relying on a lossy-compressed database with a absentminded librarian that has a penchant for people-pleasing bolted on top.

I'm absolutely not against the use of ML systems in safety-critical domains like medicine, but I am absolutely against the use of LLMs in those domains.

[0] One might choose to retort with some variant of "But all that expert work is so expensive!". I would retort: "First, without that work the data fed into the LLM is catastrophically unreliable. Second, have you bothered to look at how much the major LLM manufacturers have spent over the past five years or so? I suspect that it's more money than has been spent on medical research 'meta analysis' over the past fifty years."

  I am afraid that ship has already sailed ... These companies might not be the ones, my expectation is that they will be scrapped for their pieces ...

  We *absolutely* should be telling people to stop using the LLMs from the major LLM manufacturers.

Eh? As Greg K-H mentioned in that video from a few days back about how very disappointing Fable was when compared to its astronomical hype, LLMs that one can run locally are quite good enough for a great many tasks... including bug hunting in the Linux kernel. And -as LLM boosters keep saying- they're only keep getting better, right?

We absolutely should be telling people to stop using the LLMs from the major LLM manufacturers. Those manufacturers have spent so many billions of dollars on this project, made so many promises that they're going to have no way to keep, and now that they're running into resistance, [0] they're holding all of humanity hostage unless they're permitted to get intimately involved in the creation of new laws and regulations especially for them. [2]

Even if one only considers their recent threats to humanity, it's clear that these are not companies that deserve any of our hard-earned money.

[0] Some of that resistance comes from their ever-more-sharply-increasing cost to produce the next performance increment, some from folks asking the probing questions about their promises, claims, and business practices that should have been asked years ago, some from ongoing State AG's court cases in regards to their illegal conduct, and some from folks who -unsurprisingly- don't want enormous warehouses that suck up quite notable amounts of power [1] but provide dreadfully little revenue to the areas that house them in their communities, and still others who are starting to realize that benefits of cloud LLMs aren't worth the tradeoff of being unable to afford a new personal computer.

[1] Note carefully that I made no reference to water usage. The only counterargument you can make here is that 500MW -> 2GW is not a quite notable amount of power.

[2] <https://news.ycombinator.com/item?id=49951653>