drvdevd

I think this language is fascinating, but immediately reached a similar conclusion: it might trip up an LLM consuming the text directly but once an interpreter/compiler/decompiler has been written for the language it's trivial for an agent to consume these programs and produce them.

Now on the other hand, as a class of techniques - "anti-LLM languages" - it might be a useful method of emitting code that some agents with the correct tools can consume and others cannot, by design. Evolving such techniques with differing specifications might make for an interesting line of investigation.

also I believe it helps you track save games. I have multiple Linux boxes I play GOG games on using Heroic launcher and save game tracking is a big issue (maybe there's a way to do this with Heroic, idk). But I think Galaxy would help here.

If you find macro shots of their faces too they kinda look like dogs with a few extra eyes

> Whether February 22, 2024 comes in like a lion or a lamb, that day will always be September 11131, 1993.

> And one thing is sure: If September is eternal, then Usenet is eternal, too.

Amazing. Love this.

I don’t have a good list of which games I validated in parallels and which I didn’t (since I’ve stopped gaming for the last 6 months or so).

And I do heavily play boomer shooters so it’s likely I was impressed by performance of less demanding games.

Doom Eternal is one AAA game I definitely would want to play in Parallels but can’t because the Vulkan version is not new enough. I’m confident though that Microsoft could bring these APIs up-to-date on an “ARM PC” though, if it has capable 3D hardware.

I haven’t done any objective tests but I frequently run Windows 11 on ARM via Parallels on macOS and - it’s kinda insane (subjectively) how fast it is. Emulated x86 code is pretty fast there as well - I can easily get a solid 60 FPS on some x86 steam games as well.

This appears to be an architectural vulnerability where a speculative execution side channel similar to Spectre can be utilized within Safari or any other browser. The specifics of which environment is exploitable comes down to the specifics of the JavaScript-based gadget they use to trigger/measure this side channel. It may be in the linked paper which I haven’t read yet.

From a cursory review of the FAQs on the page it appears one mitigation might be to only keep one browser tab open at a time? They appear to be using timers and a cache eviction gadget to infer the state of other browser tabs/processes so it’s unclear what they can recover if you are not concurrently having a session to a particular site outside the gadget execution context. ???

Don’t forget these systems are trained or aligned to respond to common niceties because they are meant for public consumption.

I need a self driving Segway, yesterday.

The song is Fu Man Chu, by Desmond Dekker and the ACES.

So from an attackers perspective nowadays it sounds like focusing on encryption weaknesses and/or hardware issues (e.g. glitching) or firmware vulnerabilities is the right place to look?

I was wondering what the newly exposed vector of exploitation was here and I think you nailed it.

This is an interesting thought and I have to agree it’s a legitimate issue. I have also considered something similar - that in the near future, if all software is designed with AI, or “generated” rather than explicitly engineered, we would certainly reap some benefits from that but we would also cease understanding how anything actually works. Obviously no one person fully understands the abstractions we have currently engineered when considering the totality.. but at least currently we have them.

Yes it seems like any possible physical side channel (eg Tempest as well) is now amenable to machine learning approaches. Very interesting indeed.

I share this opinion as well, I think. I’m looking at any CRUD app I’ve worked on and thinking: this is just a specification over a database. Same with most other software but the web and most mobile apps seem ripe to just become generated patterns.

I think I agree with you on how I felt at the end of the post, but at the beginning and most of the way through, I found myself generally agreeing. But I want to look past the “bitcoin” switch after the bait because I generally agree with the author on most points.

Especially as I have delved more into how LLMs actually function and are trained, a lot of the magic in e.g. ChatGPT has evaporated for me, but in a good way. The performance of larger models like GPT-3.5 or 4 is still consistently impressive, to me, but as I’ve increased my understanding I do believe that censorship poses a greater threat than the models themselves.

Perhaps I’m misunderstanding the law or article, but wasn’t the author implying that the act of inference on an LLM model itself is a sort of “statistical experiment”:

> In probability theory, the law of large numbers (LLN) is a theorem that describes the result of performing the same experiment a large number of times.

And so the mean in this case is “perceived like a real human”?

This is an interesting idea but is it feasible? What does “filter these vectors” mean? In the context of deep models are we talking about embedding specific models, weights, parameters, etc at some point in memory with the hardware? Are we taking about filtering input generally and globally (on a general purpose system)?

Gonna leave my footprint on this one too. Here’s my take since my favorite app on my Meta Quest 2 is actually Immersed (for virtual desktop / multi monitor setups): I’m less interested in either Apple or Meta here because I think Apples Vision Pro ideas just reinforce what we’re being told repeatedly: the industry wants to move away from screens, and the future is not screens. We have two big things on the near horizon now: AR/VR and AI. The industry as a whole will push these technologies until they are the norm. Smartphones, PCs, tablets and laptops (maybe even TVs?) will become legacy technology, but they will of course live on in emulated forms (a “window”). Conversational and ubiquitous computing will be the future, like it or not.

Yeah this is it. And we get to reuse the existing roads hopefully without having to build more. We can make better use of suburban sprawl.

Indeed any AI defense capabilities will necessarily need to include some offensive abilities, which for the public are mostly censored right now (eg try asking for any of the obvious public PoCs which it almost certainly was trained on) I find it to be one of the worst aspects of the current alignment….

By driving down the costs of training and inference, and then encouraging experiments. For LLMs, QLoRA is arguably a great step in this direction.

Using ctags for this purpose is genius. Thanks for posting!

"Arguing on Hacker News isn't real arguing. Prove me wrong." (sips coffee cup)

The great job market rebound of 2024.

Yeah and one of the fundamental assumptions around statements like “let’s make AI a licensed regime” is the idea that we know what AI even is. This idea is banking on current technology being the best algorithm or method to produce “AI” and the whole lesson from the “we have no moat” crowd is that this is actually quite uncertain. Even if they succeed in getting some class of model like LLMs under “regulatory capture” - the technology they are working with today is likely to be undermined still by something cheaper working on weaker hardware and with smaller datasets and it’ll probably happen faster if they seek this market capture.

So yes it is quite comparable to the export restrictions of the 90s.

But since Microsoft is involved and we are all of course thinking about Windows vs Linux, I think another good comparison is the worst assumption Microsoft made in the 90s: “we know what an operating system is and what it is for.”

While I agree that is a bit of an extreme statement (many companies are also adopting LLMs from what I've been reading here on HN), this part just below that does seem insightful:

> Unless you work for OpenAI, your GPT leverage is likely wasted on trivial business code, and even more likely that due to the overwhelming amount of poor code, GPT with its 4/8k context window is unable to optimize, further weakening the leverage.

If you are sitting at a company in this situation it is of course arguable that you were already experiencing a negative return pre GPT-4, but certainly that productivity feeling from GPT-4 probably makes it seem even worse.

Perhaps this is one of the greatest paradigm shifts for traditional software engineers as “AI” becomes more commonly integrated: maybe we will have to use anthropomorphic language and maybe it will be considered the most technically correct way of talking about these systems due not only to their non-deterministic nature but also their emergent properties.

(This was a great post btw - highly recommend if you’re just perusing the comments)

I can say I read about 1/4 of the way through the article and although I’m interested in the authors research, I did get the feeling the author hadn’t actually driven an EV - which may not be true, but I agree the author didn’t touch on the idea that EVs may simply offer a superior driving experience.