i’ve heard of the dataset poisoning and degradation caused by llm-generated content present in the dataset myself, but i’m not sure whether it was a practical observation, or a mere experiment. And I still fail to see how new datasets are really useful for developing a new llms, or how it’s a problem for the devs to switch back to the older datasets.

And the cornerstone stays the same: to have any significant effect on the final LLM quality, shouldn’t the poisoned (either by llm-produced content, or by intentional poisoning) data portion be… well, statistically significant?

source?

Doesn’t Ken’s virus depend on the fact that the compiler (a pre-existing binary version thereof) was used to bootstrap itself? Had a different compiler been used to compile the compiler, the virus would have been killed. No. Ken notes “I could have picked on any program-handling program such as an assembler, a loader, or even hardware microcode.” Put it in the linker, say, and all your compilers are affected. You could use a debugger to inspect the compiler’s operation. Perhaps a debugger toggled in via the patch panel on the front of a PDP-10 … otherwise s/login/debugger/ in what Ken said and paste in another trojan or two. How about modification of the compiler sources for another reason? Might have fooled the “am I compiling myself” algorithm, or might have caused the Trojan to have been inserted incorrectly. There are no C compilers out there that don’t use yacc and lex. But again, the really frightening thing is via linkers and below this hack can propagate transparently across languages and language generations. In the case of cross compilers it can leap across whole architectures. It may be that the paranoiac rapacity of the hack is the reason KT didn’t put any finer point on such implications in his speech … These examples are not given to rebut Thompson’s point (which is a valid one), but instead to remind of good practices to prevent this sort of thing. What practices? I can’t think of any way to do it without the resources of a major TLA. KTH demonstrates that we are not able to fully trust any binary even when we have compiled it ourselves from trustworthy sources on a system we have compiled ourselves. The OpenSourceMovement might have an off switch … but your cell phone probably does not.

It’s been more than twenty years since I read Thompson’s marvelous paper, but I believe I correctly recall his fundamental point: UNIX, and every system like it, can NEVER be “secure”. It doesn’t matter how many layers of anti-virus software, “internet worm protection”, “firewall” or any other buzzword – systems like UNIX (including all versions of Linux, Macintosh OSX, and all versions of WinXP) will NEVER be secure. Thompson published his paper and revealed his hack in order to demonstrate this point. Whereas what systems are secure? Ones that are not compiled from source code, presumably. And certainly not, say, Windows which you left of your list by sheer coincidence.

In August 2009 a virus utilizing the Ken Thompson hack was seen in the wild. It infected Delphi 4 through 7 and applications genereated with it. http://www.h-online.com/security/Virus-infects-development-environment--/news/114031

Although non-trivial, an effective counter has been identified: http://www.schneier.com/blog/archives/2006/01/countering_trus.html. That’s only (and trivially) effective if you already have a trusted compiler executable. The problem is generating one of those in the first place. Living in the era of PRISM, we can guarantee you can’t get one online - KTH in the network will see to that. You sure didn’t get one shipped with your OS - all the commercial OS mfrs are compromised and the open source ones all went through PRISM’s internet. The only way you’re going to get a clean compiler executable now is to build your computer yourself out of TTL logic you salvage from some museum somewhere, then code up a set of basic binary tools sufficient to assemble assembler. Then bootstrap from the source of all the modern tools. This actually sounds like a worthy project until you realize that the only person who could possibly trust your clean compiler executable is you. No one else will know for sure that you’re not secretly working for PRISM. Or that you’re not just too careless, or that black hats didn’t hack their way into your garage and add the KTH while you weren’t looking. And even if you got far enough to prove something, the great conspiracy would dose you with mind-altering chemicals or blackmail you into silence or discredit you or similar. Face it, little dude, against TheKenThompsonHack there simply is no defense. Little dude? Why the put down? Who are you putting down, all of us? Please see GoodStyle. {I don’t think it was intended to be a put-down so much as a colloquial generic diminutive used as a rhetorical invective to imply weak superiority. It suggests the writer was not confident in his/her claim that “against TheKenThompsonHack there simply is no defense”, and subconsciously sought to blunt the full force of the assertion by diminishing the reader.}

• [Dude. Print out the hexdump of the disassembler, disassemble it manually, see that it works, then use it as the first step in analyzing the compiler. Oh, and TheKenThompsonHack is trivially visible in the .s file if you’re looking for it.]
• {You’re assuming the disassembler, hexdump tool, editor you use to read the .s file aren’t all compromised.}
• Yep. The weakness in TheKenThompsonHack isn’t there. It’s in the fact that it will have to make mistakes in identifying the disassemblers, hexdump tools, editors, compilers, etc. Somewhere, there’s a program whose output it will change but it shouldn’t have, or a program whose output it should have changed but didn’t. See TuringIncompletenessTheorem.

‘’‘I’ve never been able to understand why simply doing some programmatic transformation of your compiler first isn’t considered a good defense. Basically just obfuscating the code. For such a backdoor to work the injection point needs to be recognized by the injector. So if one … Started with a corrupted compiler and then obfuscated it’s source. Then compiled the obfuscated but “clean” source with the corrupted compiler. One should end up with a clean and uncorrupted compiler. Then compile the unobfuscated source with the “washed” compiler (if the obfuscation significantly degraded the performance of the compiler). If in my paranoia I needed to bootstrap an entire toolchain this way. I’d start with an FPGA. Use it to emulate an FPGA (at horrendous loss of efficiency). Then use the emulated FPGA to run a custom simple soft-CPU. Preferably hand written. Ad then bootstrap a build toolchain using some expensive obfuscations for the first pass after which the tools are considered clean. One would probably always need to be careful of the CPU the toolchain runs on though, but similar obfuscation arguments should apply…’’ How do you know that the compromised compiler won’t recognize the obfuscation you used? If it does, your “clean” compiler is no longer clean. (The FPGA and CPU solution has the same problem.)

Wouldn’t it be extremely hard (in any sense I can think of) to actually be able to recognize every program that does a particular operation? Say you wrote a program that runs a JVM implementation of a MIPS emulator that can make host system calls that runs a program that implements login by reading whatever files contain the passwords in backwards and converting them to EBCDIC before doing processing on them. I find it hard to believe that the KTH-implementing compiler would actually recognize anything that program is doing. Yes, it will have to make mistakes somewhere. This is because any implementation that was always correct could be used to solve the HaltingProblem. But this doesn’t make any particular trick undetectable. What they detect is mostly a matter of whether or not they thought of it and whether or not they considered checking for it worth the effort.

Can’t you just check network connections for suspicious activity? At some point the thing affected by the KTH would send an outbound signal either of data collected, or inbound signals if it’s updating the backdoor.

With sufficient (usually stolen) training data they could probably do it to a state where it’s kinda usable in an assistive manner.

Though since these LLMs often struggle with anything that doesn’t exist in truly abhorrent abundance in its training data as well as the actual engineering part (i.e. structuring the code sensibly and consistently, which I’d wager is crutially important with Assembly to not end up with spaghetti) trying to have it generate something from scratch would probably end up in a huge mess.

That’s just a reflection of the real life and not unique to Lemmy, dw

Pretty effective because of its high density, too. It’s a shame it’s brittle and would give you heavy metal poisoning.

Your comment is so fit for purpose, displaying such a massive willful ignorance about the world in general. Because the biggest problem with asbestos wasn’t using it for home insulation. It killed millions of workers in manufacturing plants and mines. Its biggest industrial use case was for electrical insulation of high voltage transmission lines. Where it also killed hundreds of thousands of technicians in the us alone. Before that it was used for lamp wicks. Where it killed at least a couple of millions more over a century. Today, asbestos kills 250 thousand people a year, 12 to 15 thousand in the us every year.

I supposed any tool can be considered useful if you’re ignorant enough about it or if you’re willing to lie without remorse.

Pmuch this.

How do they feel about machine language translations? It’s been “AI” (ML model) since the late 90’s.

Traffic prediction?

Medical scan data processing?

And so so many more specific uses of ML models that have been a part of daily life for literal decades.

Call out LLMs specifically.

Yes I do 😊

You are probably also one of those insane ideologues that refuse to hammer in a screw for some reason, although you know how well that hammer worked on nails… 😂