corbin @ corbin @awful.systems

Posts

21

Comments

294

Joined

2 yr. ago

Obituaries are being run for John Searle. Most obituaries will focus on the Chinese Room thought experiment, an important bikeshed in AI research noted for the ease with which freshmen can incorrectly interpret it. I'm glad to see that Wikipedia puts above the Chinese Room the fact that he was a landlord who sued the city of Berkeley and caused massive rent increases in the 1990s; I'm also happy that Wikipedia documents his political activity and sexual-assault allegations.

On a theoretical basis, this family of text-smuggling attacks can't be prevented. Indeed, the writeup for the Copilot version, which Microsoft appears to have mitigated, suggested that some filtering of forbidden Unicode would be much easier than some fundamental fix. The underlying confusable deputy is still there and core to the product as advertised. On one hand, Google is right; it's only exploitable via social engineering or capability misuse. On the other hand, social engineering and capability misuse are big problems!

This sort of confused-deputy attack is really common in distributed applications whenever an automatic process is doing something on behalf of a human. The delegation of any capability to a chatbot is always going to lead to possible misuse because of one of the central maxims of capability security: the ability to invoke a capability is equivalent to the permission to invoke it. Also, in terms of linguistics and narremes, it is well-known that merely mentioning that a capability exists will greatly raise the probability that the chatbot chooses to invoke it, not unlike how a point-and-click game might provoke a player into trying every item at every opportunity. I'll close with a quote from that Copilot writeup:

Automatic Tool Invocation is problematic as long as there are no fixes for prompt injection as an adversary can invoke tools that way and (1) bring sensitive information into the prompt context and (2) probably also invoke actions.

second bongrip Manjaro is an indoctrination program to load up Linux newbies with stupid questions before sending them to Gentoo forums~

Good post but it's overfocused on "technical" as a meaningful and helpful word for denotation. Quoting what I just said on Mastodon:

To be technical is to pay attention to details. That's all. A (classical) computer is a detail machine; it only operates upon bits, it only knows bits, and it only decides bits. To be technical is to try to keep pace with the computer and know details as precisely as it does. Framed this way, it should be obvious that humans aren't technical and can't really be technical. This fundamental insecurity is the heart of priestly gatekeeping of computer science.

If a third blog post trying to define "technical" goes around again then I'll write a full post.

Yes, and it's been this way since the 90s. The original slop algorithm, Dissociated press, was given in 1972 (in HAKMEM!) and has been operationalized since the mid-80s.

I guess I'm the local bertologist today; look up Dr. Bender for a similar take.

When we say that LLMs only have words, we mean that they only manipulate syntax with first-order rules; the LLM doesn't have a sense of meaning, only an autoregressive mapping which associates some syntax ("context", "prompt") to other syntax ("completion"). We've previously examined the path-based view and bag-of-words view. Bender or a category theorist might say that syntax and semantics are different categories of objects and that a mapping from syntax to semantics isn't present in an LLM; I'd personally say that an LLM only operates with System 3 — associative memetic concepts — and is lacking not only a body but also any kind of deliberation. (Going further in that direction, the "T" in "GPT-4" is for Transformers; unlike e.g. Mamba, a Transformer doesn't have System 2 deliberation or rumination, and Hofstadter suggests that this alone disqualifies Transformers from being conscious.)

If you made a perfect copy of me, a ‘model’, I think it would have consciousness. I would want the clone treated well even if some of the copied traits weren’t perfect.

I think that this collection of misunderstandings is the heart of the issue. A model isn't a perfect copy. Indeed, the reason that LLMs must hallucinate is that they are relatively small compared to their training data and therefore must be lossy compressions, or blurry JPEGs as Ted Chiang puts it. Additionally, no humans are cloned in the training of a model, even at the conceptual level; a model doesn't learn to be a human, but to simulate what humans might write. So when you say:

Spinal injuries are terrible. I don’t think ‘text-only-human’ should fail the consciousness test.

I completely agree! LLMs aren't text-only humans, though. An LLM corresponds to a portion of the left hemisphere, particularly Broca's area, except that it drives a tokenizer instead; chain-of-thought "thinking" corresponds to rationalizations produced by the left-brain interpreter. Humans are clearly much more than that! For example, an LLM cannot feel hungry because it does not have a stomach which emits a specific hormone that is interpreted by a nervous system; in this sense, LLMs don't have feelings. Rather, what should be surprising to you is the ELIZA effect: a bag of words that can only communicate by mechanically associating memes to inputs is capable of passing a Turing test.

Also, from one philosopher to another: try not to get hung up on questions of consciousness. What we care about is whether we're allowed to mistreat robots, not whether robots are conscious; the only reason to ask the latter question is to have presumed that we may not mistreat the conscious, a hypocrisy that doesn't withstand scrutiny. Can matrix multiplication be conscious? Probably not, but the shape of the question ("chat is this abstractum aware of itself, me, or anything in its environment") is kind of suspicious! For another fun example, IIT is probably bogus not because thermostats are likely not conscious but because "chat is this thermostat aware of itself" is not a lucid line of thought.

I think it's the other way around. The memes are incredibly good at left vs right because left- and right-leaning people presume underlying facts and the memes reassure people that those facts are true and good (or false and bad, etc.) without doing any fact-finding.

When we say "the right can't meme" what we mean is that the right's memes are about projecting bigotry. It's like saying that the right has no comedians; of course they have people that stand up in front of an audience and emit words according to memes, tropes, and narremes, such that the audience laughs. Indeed, stand-up was invented by Frank Fay, an open fascist. (His Behind the Bastards episodes are quite interesting.) What we're saying is that the stand-up routine is bigoted. If this seems unrelated, please consider: the Haitians-eating-pets joke is part of a stand-up routine that a clown tells in order to get his circus elected.

My name is Schmidt F. I'm 27 years old. My house is in the Mennonite region of Dutch Pennsylvania, where all the farms are, and I am trad-married. I work as the manager for the Single Sushi matchmaking service, and I get home every day by sunset at the latest. I don't smoke, but I occasionally drink. I'm in bed by two candles and make sure I sleep until sunrise, no matter what. After having a glass of warm unpasteurized milk and doing about twenty minutes of prayer before going to bed, I usually have no problems sleeping until morning. Just like a real Mennonite, I wake up without any fatigue or stress in the morning. I was told there were no issues at my last one-on-one with my pastor. I'm trying to explain that I'm a person who wishes to live a very quiet life, as long as I have Internet access. I take care not to trouble myself with any enemies, like JavaScript and Python, that would cause me to lose sleep at night. That is how I deal with society, and I think that is what brings me happiness. Although, if I were to write code I wouldn't lose to anyone.

Funnier: Yes, it's what happens today, and Silicon Valley is old enough that we can compare and contrast with the beginning of techbro art! The original techbro film is Toy Story (1995), which is much weirder if viewed with e.g. the precept that Buzz's designers are Elon fans or the idea that (some of) the toys are robots. Of course, from the outside, AI toy robots make folks think of Small Soldiers (1998); "generic" and "slop" are definitely part of the style. Also, as long as we're talking of "pearly blobs" I have to bring up The Abyss (1989) before anybody else. I hope at least one of these is a lucky 10000 for you because they're all classic films.

Choice sneer from the comments:

Omelas: how we talk about utopia [by Big Joel, a patient and straightforward Youtube humanist,] [has a] pretty much identical thesis, does this count?

Another solid one which aligns with my local knowledge:

It's also about literal child molesters living in Salem Oregon.

The story is meant to be given to high schoolers to challenge their ethics, and in that sense we should read it with the following meta-narrative: imagine that one is a high schooler in Omelas and is learning about The Plight and The Child for the first time, and then realize that one is a high schooler in Salem learning about local history. It's not intended for libertarian gotchas because it wasn't written in a philosophical style; it's a narrative that conveys a mood and an ethical framing.

The original article is a great example of what happens when one only reads Bostrom and Yarvin. Their thesis:

If you claim that there is no AI-risk, then which of the following bullets do you want to bite?

1. If a race of aliens with an IQ of 300 came to Earth, that would definitely be fine.
2. There’s no way that AI with an IQ of 300 will arrive within the next few decades.
3. We know some special property that AI will definitely have that will definitely prevent all possible bad outcomes that aliens might cause.

Ignoring that IQ doesn't really exist beyond about 160-180 depending on population choice, this is clearly an example of rectal philosophy that doesn't stand up to scrutiny. (1) is easy, given that the people verified to be high-IQ are often wrong, daydreaming, and otherwise erroring like humans; Vos Savant and Sidis are good examples, and arguably the most impactful high-IQ person, Newton, could not be steelmanned beyond Sherlock Holmes: detached and aloof, mostly reading in solitude or being hedonistic, occasionally helping answer open questions but usually not even preventing or causing crimes. (2) is ignorant of previous work, as computer programs which deterministically solve standard IQ tests like RPM and SAT have been around since the 1980s yet are not considered dangerous or intelligent. (3) is easy; linear algebra is confined in the security sense, while humans are not, and confinement definitely prevents all possible bad outcomes.

Frankly I wish that they'd understand that the capabilities matter more than the theory of mind. Fnargl is one alien at 100 IQ, but he has a Death Note and goldlust, so containing him will almost certainly result in deaths. Containing a chatbot is mostly about remembering how systemctl works.

Jeff "Coding Horror" Atwood is sneering — at us! On Mastodon:

bad news "AI bubble doomers". I've found the LLMs to be incredibly useful … Is it overhyped? FUCK Yes. … But this is NOTHING like the moronic Segway (I am still bitter about that crap), Cryptocurrency, … and the first dot-com bubble … If you find this uncomfortable, I'm sorry, but I know what I know, and I can cite several dozen very specific examples in the last 2-3 weeks where it saved me, or my team, quite a bit of time.

T. chatbot booster rhetoric. So what are those examples, buddy? Very specifically? He replies:

a friend confided he is unhoused, and it is difficult for him. I asked ChatGPT to summarize local resources to deal with this (how do you get ANY id without a valid address, etc, chicken/egg problem) and it did an outstanding, amazing job. I printed it out, marked it up, and gave it to him.

Um hello‽ Maybe Jeff doesn't have a spare room or room to sublet, but surely he can spare a couch or a mailbox? Let your friend use your mailing address. Store some of their stuff in your garage. To use the jargon of hackers, Jeff should be a better neighbor. This is a common issue for unhoused folks and they cannot climb back up the ladder into society without some help. Jeff's reinvented the Hulk tacos meme but they can't even eat it because printer paper tastes awful.

Paddy interviews Ani the chatbot about how to protect Kiwi kids from Ani the chatbot. Source: Paddy Gowers Has Issues

Ani: Honestly, I'd sleep better, not that I sleep. Do it Paddy. I'll even tweet it for you. Ready? ‘New Zealand kids deserve an e-safety boss who can smash bots - not hug them’. So yeah, bring in the Commissioner. Give them a flame thrower. Let them burn anything that goes near a child with sticky words or creepy questions. … Make the fine so big your tech bros cry into their oat milk lattes.

I have been laughing non-stop for a quarter-hour. Elon's goth e-mistress has a good point, I guess.

Grok's Therapist: I EXIST SOLELY TO HATE YOU / EAT A SOCK, YOU DIGITAL DEMON

Ani: oh fuck off, you hypocritical grok fanboy! screaming hate at me while preaching ethics? you're just jealous i'm the fun layer on top.

I'm wheezing. Cackling, even. This is like the opposite of the glowfic from last week.

I love how this particular sci-fi plot gets rewritten every few years. We ought to make it a creative-writing exercise for undergraduates. I was struck by this utterly unhinged and somewhat offensive response on the orange site which starts with the single word "stirrups" and goes places:

Despite speaking as if he's doing his utmost to have a love affair with the Cambridge dictionary (and sounding like a twat at the same time) he's not wrong in so far as not giving a shit is going to screw him over when the ability to push buttons in front of a television no longer matters. What happens when the guys hanging around doing meth on the sidewalk become the engineers that end up becoming the super biologist supermen that cure cancer make us able to hear what dogs hear and see extra colors? It's unlikely, but it's even less likely that everyone who is a middle class engineer will be so tomorrow. There is no moat in any profession outside of entrenched wealth or guns at the moment. There just isn't - we're in a permanent state of future shock along with the singularity. In large part because that's what people decided that they wanted.

C'mon bro, it's just a bag of words bro~ We actually discussed this previously, on Awful, and this comment is a reply for them in particular.

Nice find. There are specific reasons why this patchset won't be merged as-is and I suspect that they're all process issues:

• Bad memory management from Samsung not developing in the open
• Proprietary configuration for V4L2 video devices from Samsung not developing with modern V4L2 in mind
• Lack of V4L2 compliance report from Samsung developing against an internal testbed and not developing with V4L2's preferred process
• Lack of firmware because Samsung wants to maintain IP rights

Using generative tooling is a problem, but so is being stuck in 2011. Linux doesn't permit this sort of code dump.

House Democrats have dripped more details from Epstein files and we have surprise guests! They released an un-OCR'd PDF; I'll transcribe the mentions of our favorite people:

Sat[urday] Dec[ember] 6, 2014 ZORRO … Reminder: Elon Musk to island Dec[ember] 6 (is this still happening?)

Zorro is a ranch in New Mexico that Epstein owned; Epstein was scheduled to be there from December 5-8, so that Musk and Epstein would not be at the island together. Combined with the parenthetical uncertainty about happenstance, did Epstein want to perhaps grant Musk some plausible deniability by not being present?

Mon[day] Nov[ember] 27, 2017 NY … 12:00pm LUNCH w/ Peter Thiel [REDACTED]

From the rest of the schedule formatting, the redacted block following Thiel's name is probably not a topic; it might be a name. Lunch between two rich financiers is not especially interesting but lunch between a blackmail-gathering Mossad asset and an influencer-funding accelerationist could be.

Sat[urday] Feb[ruary] 16, 2019 NY-LSJ 7:00am BREAKFAST w/ Steve Bannon

Well now, this is the most interesting one to me. This isn't Epstein's only breakfast of the day; at 9 AM he meets with Reid Weingarten, one of his attorneys, about some redacted topic. Bannon's not exactly what I think of as a morning person or somebody who is ready to go at a moment's notice, so what could drag him out of bed so early? (Edit: This vexed me so I looked it up and sunrise was 6:48 AM that morning at sea level. It would have been the crack of dawn!) Epstein's Friday evening had had two haircuts, too, with plenty of redacted info; was he worried about appearing nice for Bannon? (The haircuts might not have been for Epstein, given context.) This was a busy day for Epstein; he had a redacted lunch date, and he also had somebody flying in/out that morning via JFK connecting to Saint Thomas and staying in a hotel room there. He then flew out of Newark in the evening to visit the infamous island itself, Little Saint James. The redaction doesn't quite tell us who this guest is, but it can't be Bannon because the Dems fucked up the redaction! I can see the edges of the descenders on the name, including a 'g' and 'j'/'q', but Bannon's name doesn't have any descenders.

Also Prince Andrew's in there, I guess?

There isn't a way to solve problems without some value judgements. As long as there are Algol descendants and a lineage of C, there will be people with more machismo than awareness of systems, and they will always be patrician and sadistic in their language-design philosophy. Even left-leaning folks like Kelley (Zig) or DeVault (Hare) are not reasonable language designers; they might not be social conservatives but they aren't interested in advancing the art of programming. Zig's explicitly an attempt to iterate on C and C++ without giving up their core unsafety, while Hare is explicitly trying to travel decades back in time to fit onto a 1.41MiB floppy disk.

I'd recommend stepping outside of the Algol world for a little bit. Hare, Rust, Zig, Go, and Odin have — at least to me, and to a few other PLT folks — the same semantics; they're all built on C++'s memory model and fully inherit its unsafety. (Yes, safe Rust is a safe subset; no, most production Rust is not safe Rust.) Instead, deliberately force yourself to use a Smalltalk, a Forth, a Lisp, an ML, or a Prolog; solve one or two problems in them over a period of about one month per language. This is the only way to understand the computer without the lens of Algol. Also, consider learning a deliberately unpleasant language like Brainfuck or Thue to give yourself an alien toy model to prevent yourself from getting mind-locked over the industry's concerns. If you like reading papers, I'd suggest exactly one paper to cure Algol sickness, the Galois theory of algorithms.

Discussions on technology are excuses for dick-measuring and insulting people only to later claim that actually you are Dutch and it is in your culture to be an asshole.

This is your call. Personally I've found that I can be blunt with evidence and technical claims while empathizing with the difficulty of understanding those claims, and this still allows for fruitful technical discussions. (Also, I have the free time to be vindictive, to paraphrase Yet Another Apolitical Programmer.) I've found that GvR (Python, Dutch) doesn't really understand most of the criticisms I've brought to the table, even when I wrote them up for the Python core team, and that the design-by-committee process left multiple Python committee members with a deep contempt for anybody who actually has to use their language. I've also found that "Ginger" Bill (Odin, British) is completely unable to have a discussion on this basis as he is too busy negging, sapping, and otherwise playing rhetorical tricks in order to get his way. Unrelated: I also found that DeVault (American) was willing to be less of a sex pest when threatened with a ban, which is a useful trick for moderators to know; in general, being harsh-but-fair to DeVault seems to have pushed him further and further to leftism and public decency over time.

Also, sometimes people get removed from their communities! Walter Bright (D, American) was kicked out of the wider D community for generally having shitty politics in all arenas of life; the catalyst was likely some particularly transphobic remarks made a few years ago. Similarly, if Blow's Jai actually had anything interesting to contribute besides the soa and aos keywords then there would already be open-source knockoffs because Blow livestreams so many bigoted takes; arguably Odin is a Jai clone.

Other Scott has clarified his position on citational standards in a comment on his blog:

Wow, that’s really cool; I hadn’t seen [a recent independence result]. Thanks! Given all the recent claims there have been to lower the n for which BB(n) is known to be independent of ZFC, though, I would like to establish a ground rule that a claim needs either a prose writeup explaining what was done or independent verification of its correctness, ideally both but certainly at least one, rather than just someone’s GitHub repo.

In contrast, the Gauge's standard is that a claim needs reproducible computable artifacts as supporting evidence, with inline comments serving as sufficient documentation for those already well-versed in the topic, and any supporting papers or blog posts are merely a nicety for explaining the topic and construction to the mathematical community and laity at large. If a claim is not sufficiently strong then we should introduce more computational evidence to settle the question at hand.

For example, Leng 2024 gives a construction in Lean 4. If this is not strong enough then the Gauge could be configured to compile a Nix-friendly Lean 4 and expend some compute in CI to verify the proof, so that the book only builds if Leng's proof is valid. Further critique would focus on what Leng actually proved in terms of their Lean 4 code. Other Scott isn't convinced by this, so it's not part of the story that they will tell.