Sam "wrong side of FOSS history" Altman must be pissing himself.
Direct Nitter Link:
Lol. Lmao even. "DeepSeek R1 reproduced for $30: Berkeley researchers replicate DeepSeek R1 for $30—casting doubt on H100 claims and controversy"
2 years ago by Snot Flickerman to c/techtakes
brucethemoose 20 points 2 years ago
Like many AI headlines, to be honest. Even in the open source world, there is tons of fudging and hype. Been that way since way before ChatGPT became popular.
Still, it’s cool the whole finetuning pipeline even runs, and theoretically scaling up is just a matter of money
MrPoopyButthole 2 points 2 years ago
Yup and it's not even testing general reasoning. They didn't have money for that.
BlueMonday1984 59 points 2 years ago
To reference a previous sidenote, DeepSeek gives corps and randos a means to shove an LLM into their shit for dirt-cheap, so I expect they're gonna blow up in popularity.
fallowseed 30 points 2 years ago
open source behaving like open source? couldn't be the evil scary chinese!
reallykindasorta 18 points 2 years ago
Non-techie requesting a laymen explanation if anyone has time!
After reading a couple of”what makes nvidias h100 chips so special” articles I’m gathering that they were supposed to have a significant amount more computational capability than their competitors (which I’m taking to mean more computations per second). So the question with deepseek and similar is something like ‘how are they able to get the same results with less computations?’ and the answer is speculated to be more efficient code/instructions for the AI model so it can make the same conclusions with less computations overall, potentially reducing the need for special jacked up cpus to run it?
mountainriver 16 points 2 years ago
Good question!
The guesses and rumours that you have got as replies makes me lean towards "apparently no one knows".
And because it's slop machines (also referred to as "AI", there is always a high probability of some sort of scam.
froztbyte 10 points 2 years ago
pretty much my take as well. I haven’t seen any actual information from a primary source, just lots of hearsay and “what we think happened” analyst shit (e.g. that analyst group in the twitter screenshot has names but no citation/links)
and doubly yep on the “everyone could just be lying” bit
manicdave 4 points 2 years ago
The article sort of demonstrates it. Instead of needing inordinate amounts of data and memory to increase it's chance of one-shotting the countdown game. It only needs to know enough to prove itself wrong and roll the dice again.
justOnePersistentKbinPlease -1 points 2 years ago
From a technical POV, from having read into it a little:
Deepseek devs worked in a very low level language called Assembly. This language is unlike relatively newer languages like C in that it provides no guardrails at all and is basically CPU instructions in extreme shorthand. An "if" statement would be something like BEQ 1000, where it goes to a specific memory location(in this case address 1000 if two CPU registers are equal.)
The advantage of using it is that it is considerably faster than C. However, it also means that the code is mostly locked to that specific hardware. If you add more memory or change CPUs you have to refactor. This is one of the reasons the language was largely replaced with C and other languages.
Edit: to expound on this: "modern" languages are even slower, but more flexible in terms of hardware. This would be languages like Python, Java and C#
V0ldek 25 points 2 years ago
This is a really weird comment. Assembly is not faster than C, that's a nonsensical statement, C compiles down to assembly. LLVM's optimizations will most likely outperform or directly match whatever hand-crafted assembly you write. Why would BEQ 1000 be "considerably faster" than if (x == y) goto L_1000;? This collapses even further if you consider any application larger than a few hundred lines of code, any sensible compiler is going to beat you on optimizations if you try to write hand-crafted assembly. Try loading up assembly code and manually performing intraprocedural optimizations, lol, there's a reason every compiled language goes through an intermediate representation.
Saying that C# is slower than C is also nonsensical, especially now that C# has built-in PGO it's very likely it could outperform an application written in C. C#'s JIT compiler is not somehow slower because it's flexible in terms of hardware, if anything that's what makes it fast. For example you can write a vectorized loop that will be JIT-compiled to the ideal fastest instruction set available on the CPU running the program, whereas in C or assembly you'd have to manually write a version for each. There's no reason to think that manual implementation would be faster than what the JIT comes up with at runtime, though, especially with PGO.
It's kinda like you're saying that a V12 engine is faster than a Ferrari and that they are both faster than a spaceship because the spaceship doesn't have wheels.
I know you're trying to explain this to a non-technical person but what you said is so terribly misleading I cannot see educational value in it.
iltg -6 points 2 years ago
your statement is so extreme it gets nonsensical too.
compilers will usually produce higher optimized asm than writing it yourself, but there is room to improve usually. it's not impossible that deepseek team obtained some performance gains hand-writing some hot sections directly in assembly. llvm must "play it safe" because doesn't know your use case, you do and can avoid all safety checks (stack canaries, overflow checks) or cleanups (eg, make memory arenas rather than realloc). you can tell LLVM to not do those, but it may happen in the whole binary and not be desirable
claiming c# gets faster than C because of jit is just ridicolous: you need yo compile just in time! the runtime cost of jitting + the resulting code would be faster than something plainly compiled? even if c# could obtain same optimization levels (and it can't: oop and .net runtime) you still pay the jit cost, which plainly compiled code doesn't pay. also what are you on with PGO, as if this buzzword suddenly makes everything as fast as C?? the example they give is "devirtualization" of interfaces. seems like C just doesn't have interfaces and can just do direct calls? how would optimizing up to C level make it faster than C?
you just come off as a bit entitled and captured in MS bullshit claims
bitofhope 9 points 2 years ago
GPU programs (specifically CUDA, although other vendors' stacks are similar) combine code for the host system in a conventional programming language (typically C++), and code for the GPU written in CUDA language. Even if the C++ code for the host system can be optimized with hand written assembly, it's not going to lead to significant gains when the performance bottleneck is on the GPU side.
The CUDA compiler translates the high level CUDA code into something called PTX, machine code for a "virtual ISA" which is then translated by the GPU driver into native machine language for the proprietary instruction set of the GPU. This seems to be somewhat comparable to a compiler intermediate representation, such as LLVM. It's plausible that hand written PTX assembly/IR language could have been used to optimize parts of the program, but that would be somewhat unusual.
For another layer or assembly/machine languages, technically they could have reverse engineered the actual native ISA of the GPU core and written machine code for it, bypassing the compiler in the driver. This is also quite unlikely as it would practically mean writing their own driver for latest-gen Nvidia cards that vastly outperforms the official one and that would be at least as big of a news story as Yet Another Slightly Better Chatbot.
While JIT and runtimes do have an overhead compared to direct native machine code, that overhead is relatively small, approximately constant, and easily amortized if the JIT is able to optimize a tight loop. For car analogy enjoyers, imagine a racecar that takes ten seconds to start moving from the starting line in exchange for completing a lap one second faster. If the race is more than ten laps long, the tradeoff is worth it, and even more so the longer the race. Ahead of time optimizations can do the same thing at the cost of portability, but unless you're running Gentoo, most of the C programs on your computer are likely compiled for the lowest common denominator of x86/AMD64/ARMwhatever instruction sets your OS happens to support.
If the overhead of a JIT and runtime are significant in the overall performance of the program, it's probably a small program to begin with. No shame to small programs, but unless you're running it very frequently, it's unlikely to matter if the execution takes five or fifty milliseconds.
justOnePersistentKbinPlease -9 points 2 years ago
I have have crafted assembly instructions and have made it faster than the same C code.
Particular to if statements, C will do things push and pull values from the stack which takes a small but occasionally noticeable amount of cycles.
khalid_salad 11 points 2 years ago
python, what are you doing?"
idk, I'm written in C, it does things push and pull values from the stack, have you tried assembly, it's faster
self 8 points 2 years ago
Particular to if statements, C will do things push and pull values from the stack which takes a small but occasionally noticeable amount of cycles.
holy fuck. llvm in shambles
froztbyte 20 points 2 years ago
for anyone reading this comment hoping for an actual eli5, the "technical POV" here is nonsense bullshit. you don't program GPUs with assembly.
the rest of the comment is the poster filling in bad comparisons with worse details
justOnePersistentKbinPlease -7 points 2 years ago
For anyone reading this comment, that person doesnt know anything about assembly or C.
froztbyte 14 points 2 years ago
yep, clueless. can't tell a register apart from a soprano. and allocs? the memory's right there in the machine, it has it already! why does it need an alloc!
fuckin' dipshit
next time you want to do a stupid driveby, pick somewhere else
fartsparkles 9 points 2 years ago
I’m sure that non techie person understood every word of this.
blakestacey 21 points 2 years ago
And I'm sure that your snide remark will both tell them what to simplify and explain how to do so.
Enjoy your free trip to the egress.
msage -1 points 2 years ago
Putting Python, the slowest popular language, alongside Java and C# really irks me bad.
The real benefit of R1 is Mixture of Experts - the model is separated into smaller sections, that are trained and used independently, meaning you don't need the entire model to be active all the time, just parts of it.
Meaning it uses less resources during training and general usage. For example instead of 670 billion parameters all the time, it can use 30 billion for specific question, and you can get away with using 2% of the hardware used by competition.
UndercoverUlrikHD -3 points 2 years ago
Putting Python, the slowest popular language, alongside Java and C# really irks me bad.
I wouldn't call python the slowest language when the context is machine learning. It's essentially C.
justOnePersistentKbinPlease -5 points 2 years ago
I used them as they are well known modern languages that the average person might have heard about.
fallowseed -2 points 2 years ago
i read that that the chinese made alterations to the cards, as well-- they dismantled them to access the chips themselves and were able to do more precise micromanagement that cuda doesn't support, for instance.. basically they took the training wheels off and used a more fine-tuned and hands-on approach that gave them some serious advantages
fallowseed -6 points 2 years ago
just something i read, this isn't the original source i read, but a quick search gave me: https://www.xatakaon.com/...
froztbyte 12 points 2 years ago
okay so that post’s core supposition (“using ptx instead of cuda”) is just fucking wrong fucking weird and I’m not going to spend time on it, but it links to this tweet which has this:
DeepSeek customized parts of the GPU’s core computational units, called SMs (Streaming Multiprocessors), to suit their needs. Out of 132 SMs, they allocated 20 exclusively for server-to-server communication tasks instead of computational tasks
this still reads more like simply tuning allocation than outright scheduler and execution control (which your post alluded to)
[x] doubt
e: original wording because cuda still uses ptx anyway, whereas this post looks like it’s saying “they steered ptx directly”. at first I read the tweet more like “asm vs python” but it doesn’t appear to be what that part meant to convey. still doubting the core hypothesis tho
techtakes
@awful.systems
2727
1524
1751
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
go to feed...
techtakes
@awful.systems
2727
1524
1751
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
go to feed...
They finetuned 1.5-3b models. This is a non-story
save