| Favorite team: | LSU  |
| Location: | |
| Biography: | Played golf, then found math |
| Interests: | Finding ways to buy expensive fishing gear without telling my wife. |
| Occupation: | |
| Number of Posts: | 2303 |
| Registered on: | 1/6/2013 |
| Online Status: | Not Online |
Recent Posts
Message
re: Which one of you baws was on Peachtree Rd in Buckhead this morning?
Posted by CFDoc on 7/1/26 at 9:12 am to Hangover Haven
quote:
Is that a Ram truck?
Hard to tell from the pic but the keen eye will notice the tailgate opening handle has the convex curve shaped. Dead giveaway it’s a RAM.
0
re: WTB old Tundra
Posted by CFDoc on 6/28/26 at 8:50 pm to TheDrunkenTigah
I’ve heard the statement that emissions made them get rid of the V8 but I honestly have no clue what that means in detail.
I know the government was throwing around money like candy to get them to build EV’s but did some emission law change that would have made selling the 5.7 V8 illegal?
Several manufacturers offer multiple engine choices in a platform. Couldn’t they have done the turbo V6 and the 5.7 V8 as options?
I know the government was throwing around money like candy to get them to build EV’s but did some emission law change that would have made selling the 5.7 V8 illegal?
Several manufacturers offer multiple engine choices in a platform. Couldn’t they have done the turbo V6 and the 5.7 V8 as options?
I just sold my 2016 Tundra with 271k trouble free miles and bought a 2021 TRD Tundra with 50k miles from gulf states auto in Tuscaloosa.
Toyota is just so stupid for giving up that V8.
Toyota is just so stupid for giving up that V8.
quote:
Pretty well known spieth gets around
If his getting around is still the same as it was with the Nikki Amthor thing before the texts got nuked from the internet, that’s some of the most pathetic ‘getting around’ I’ve ever heard of.
But it’s exactly the way I’d imagine spieth doing it.
quote:
Frontier AI will be reserved for use by select few organizations.
People are going to assume all sorts of things as to why the AI future is going to be highly controlled and limited. In reality, the math (cost and energy) required to build and maintain AI Frontier models is insane.
The Efficient Compute Frontier for AI roughly estimates that we are within about 10-15 years of a single Frontier AI model (either JEPA or GPT based) requiring roughly a Terawatt to build/maintain.
For reference, the entire Earth requires about 18 Terawatts of continuous power. So 20 AI companies in 10 years could easily require an entire Earth’s power grid worth of energy to operate.
All that to say, AI is going to be highly controlled because there’s no form of government or political system that is going to allow a product to operate at that scale without massive control parameters.
On top of that, you can throw all the other stuff like security, control, and whatever other socioeconomic theory you can come up with.
quote:
I've never asked, but this conversation has curiously gone from "it's easy" to "it's a secret".
So I went back and read my original posts and I’ll own up to using poor wording when I said it was ‘easy.’
My intent was to describe how deterministic and low dimensional spaces are ‘easy’ to exploit. More of a generality.
At this point, you are correct that it is not easy to exploit a GPU. What I am seeing is progress in research that essentially breaks a GPU down into a low dimensional, deterministic space when operated in transformer architectures. This is news since, for some time now, the general consensus was more along the lines that LLMs were uncontrollable stochastic dart throwers.
Maybe the details will become more public in the near future. But for now, the cybersecurity firm I’m doing work for requires otherwise.
quote:
Your claim seems to be that there is something specific to GPUs that that reduces the diversity
Yes, that is precisely the claim.
And like I mentioned, from what I’ve seen to date, there’s exploits that work when running an LLM on a GPU that do not work when running the exact same LLM on another architecture. It is the very fundamental physical makeup of the GPU that allows for it.
And no, I’m not sharing IP with you on this topic.
quote:
They don't use CISC, they use SASS as the instruction for the tensor core, it's fundamentally different from RISC-V.
It's at the lowest level that the sheer volume of data (the transformer architecture) is what has been giving us the leaps and bounds improvement.
SASS is not the only instruction set for tensor cores.
My argument has nothing to do with improving LLMs.
At this point I think we’re just talking past each other. It’s prolly time to end the conversation and just go have a beer.
quote:
If it doesn't then there was nothing there.
This is categorically false.
quote:
They are also more diverse than you are willing to admit, we've seen huge differences in going from FP 16 to FP 4.
The diversity almost completely disappears when applied to LLMs though. That’s the link I’ve been trying to explain (and admittedly probably doing a poor job).
quote:
You say that but FP4 doesn't even use RISC, thats all tensor core. Tensor core is why the B200 is insane.
???
Maybe you know something I don’t but all the FP-4 I’ve tested still based their instructions off of RISC-V, or some other reduced instruction model. Does nvidia even support CISC at the chip level?
CUDA vs Tensor doesn’t really matter when it comes to the on board instruction set. So I’m confused as to what you’re referring here.
quote:
Look it's an interesting theory, and maybe there is something to it, but your statements are far too broad without data to back it up
From what I’ve seen to date, there’s a ‘there’ there. Whether or not it makes it into open literature, who knows.
quote:
what chipsets do people think this hardware determism is actually predictable
Chipsets don’t have free will.
And chipsets based on SIMT coupled with RISC-V controllers really really really don’t have free will.
And I’m completely ok with you not believing me.
re: U.S gov will get to decide who gets access to GPT-5.6
Posted by CFDoc on 6/27/26 at 9:24 am to trinidadtiger
quote:
I think he is looking for funding for his NGO...
Nobody in this area is having issues with funding.
quote:
It doesn't really have anything to do with exploitation of physical chip sets, that paper seems wholly unconnected to GPU architectures.
That paper doesn’t have anything to do with GPU architectures. That paper is more of an example of how recent literature is showing how to deterministically control heteroscedasticity of an LLM. Essentially drawing determinism out of what’s fundamentally a stochastic model.
I don’t know if the open literature is actively publishing exploits of gpu architectures yet or not. I’m just telling you it’s an active topic in the field. And yes, Ampere, Hopper, Blackwell, and Vera Rubin matter. However, at the transistor/ALU level, coupled with RISC, it makes all GPUs much more reverse engineer-able.
TPUs are a different story.
quote:
But I think it really is a cyber security issue
Did you see copy.fail ?
That was just a precursor and that is full root access to almost every single Linux distribution from the last 10 years all wrapped up in a 700 byte Python file.
There’s a good many rabbit holes you can go down here.
The best place to start is googling the latest developments on what’s known as ‘LLM Mapping’ or ‘LLM Fingerprinting’. Specifically the folks looking at how fingerprinting can allow LLM exploits and expose vulnerabilities.
From there, you can go down the latest research paths where groups are trying to really get into the nuts and bolts of where and how probabilistic outputs are governed/mapped by deterministic processes like encoding, attention, etc. LINK
In my current line of work, there’s a good bit of research focused on tying these concepts to the physical properties of GPUs with the initial goal being efficiency focused; however, it’s not a difficult jump between efficiency and exploitation.
The best place to start is googling the latest developments on what’s known as ‘LLM Mapping’ or ‘LLM Fingerprinting’. Specifically the folks looking at how fingerprinting can allow LLM exploits and expose vulnerabilities.
From there, you can go down the latest research paths where groups are trying to really get into the nuts and bolts of where and how probabilistic outputs are governed/mapped by deterministic processes like encoding, attention, etc. LINK
In my current line of work, there’s a good bit of research focused on tying these concepts to the physical properties of GPUs with the initial goal being efficiency focused; however, it’s not a difficult jump between efficiency and exploitation.
quote:
It takes 5-6 months for a major version to release, it takes less than that to rip it off.
Rip it off, exploit it, bypass guardrails, you name it. LLMs are a security nightmare both by what they can output and by their very fundamental nature of being easily jailbroken.
I wouldn’t be surprised if GPUs, in their physical form, got deemed a national security risk in the near future. There’s a good bit of emergent research happening on this topic that is very interesting.
It has to do with the heteroscedasticity and homoscedasticity of the manifolds generated by LLM architectures when computed by GPUs.
Even though LLMs produce probabilistic output vectors, the extremely limited and deterministic instruction set native to GPU ALUs means the map of an LLM can be easily discerned and manipulated. Fate, it seems, is not without a sense of irony.
In layman’s terms, imagine the most complex maze possible. Billions of turns, trillions of choices needing to be made. On the surface, this maze would seem impossible. But then if I told you the solution to the maze was just always choose two left turns followed by a right turn. Now imagine you are a bad actor trying to exploit the maze. Easy.
This is essentially a GPU. Billions of transistors but each ALU laid out exactly the same and only an extremely limited set of instructions available. Very easy to exploit.
According to Gen Stanton, they are MAJOR cyber security threats.
re: U.S gov will get to decide who gets access to GPT-5.6
Posted by CFDoc on 6/25/26 at 6:31 pm to hawgfaninc
Pretty expected after the Mythos debacle.
re: Clark is unconscious
Posted by CFDoc on 6/21/26 at 3:19 pm to JackandWater
Clark’s putter is cooling.
Dude hit 50% of the greens in a us open and shot par yesterday. That doesn’t hold up.
Dude hit 50% of the greens in a us open and shot par yesterday. That doesn’t hold up.
re: NASA’s $247M ‘Son of Concorde’ high-speed jet just broke the sound barrier, w/o the boom
Posted by CFDoc on 6/15/26 at 2:55 pm to cajunangelle
I worked this program. I remember when we told NASA is would fly in late 2021 and cost $400M.
Suckers.
Suckers.
Ended with a double and a bogey late to finish over par for the day. This when everybody else in the field decided golf was easy.
Crazy because it’s not that hard to create an agent that enforces factual information over helpful information.
You can get almost all the frontier models to not output a statement that can’t be proven factual and provide the sources by which it decided a statement was factual.
This is extreme laziness on the lawyers.
You can get almost all the frontier models to not output a statement that can’t be proven factual and provide the sources by which it decided a statement was factual.
This is extreme laziness on the lawyers.
What would you call (1/3)*3?
Popular
17
