Veedrac

40 ms·

481.
by Veedrac 6y ago
This comment doesn't make much sense given the announcement of the Digital Services Act and Digital Markets Act recently preceded this.
482.
by Veedrac 6y ago
Obviously it's impossible to prove definitive statements about every possible potential action, as as per the halting problem some of those actions are unprovable. It is as ridiculous to suggest that this means you can't contain
483.
by Veedrac 6y ago
Correction/clarification: 10k proof steps was the number of steps in the training dataset, which was for 100 such proofs.
484.
by Veedrac 6y ago
> Yet Metal optimization resources imply 16-bit arithmetic should be significantly faster, in addition to a reduction of register usage leading to higher thread count (occupancy). I believe this is a difference between the A14 GPU and th
485.
by Veedrac 6y ago
> The accuracy largely comes from the fact that addition rarely requires carrying more than a single digit. So it's easy to pattern match from single digit problems that it was previously trained on. Again, not at all true due to BP
486.
by Veedrac 6y ago
> So GPT-3 does very well with 2-digit addition and okay with 2-digit multiplication, but can't abstract to 6-digit arithmetic. This is false, GPT-3 can do 10-digit addition with ~60% accuracy, with comma separators. Without BPEs it
487.
by Veedrac 6y ago
Note that BOOM is now on v3[1], which claims to be 3.93 DMIPS/MHz, or about twice RSD. [1] https://people.eecs.berkeley.edu/~krste/papers/SonicBOOM-CAR...
488.
by Veedrac 6y ago
These issues would be a lot less troublesome if hardware offered basic security, like zero-cost poison-on-overflow and bounds-checked pointers, either of which would solve any risks of just letting it be. But right now there's the prag
489.
by Veedrac 6y ago
There clearly is a change in structure somewhere, even if the precise location is arguable, since extremely small brains are fixed in structure (every connection is deterministic) and seem to implement completely fixed programs. I imagine t
490.
by Veedrac 6y ago
> We are chasing the wrong things. Our conceptualization of the problem domain is fundamentally insufficient. What we really need is GOFAI. Back at ya mate :D.
491.
by Veedrac 6y ago
Tell that to Stockfish NNUE :P. A simpler neural network running at Stockfish speeds, all on CPU, crushes Lc0 (Lc0 is probably stronger than AlphaZero IMO).
492.
by Veedrac 6y ago
> It's not a net that you input a board into and get a move out, but a net you input a board into and get a score out. It's both actually. The former helps guide the search (MCTS, not an Alpha-Beta minmax), and the latter helps
493.
by Veedrac 6y ago
Google always had internal review processes. If your argument is not that the new review process is unreasonable, just that people were hired under the expectation that there would be none, then I get where you're coming from but it do
494.
by Veedrac 6y ago
You're confusing the ideas with their presentation. Google never allowed someone to post a paper under their name saying ‘datacentres generate CO₂, that's evil and we should focus on smaller datacentres’, but that doesn't mea
495.
by Veedrac 6y ago
I'm not going to claim it'll be without issue, but it seems way more robust than DCMA, since there's an appeals process that can boot people out the program if they flag incorrectly too much.
496.
by Veedrac 6y ago
I agree that fact checkers are a nasty can of worms, but here they're only expected to flag illegal content, which I don't think is a comparable concern.
497.
by Veedrac 6y ago
Bundling is explicitly allowed by the regulation. Some examples of other ‘clearly unfair’ practices are restricting the use of third-party software or stores, or preventing them from accessing them other than from the company's channel
498.
by Veedrac 6y ago
https://ec.europa.eu/info/sites/info/files/proposal_for_a_re... --- Article 19, Trusted flaggers 2. The status of trusted flaggers under this Regulation shall be awarded, upon application by any entities
499.
by Veedrac 6y ago
> Except, you know, the whole point of VLIW It's really indicative of the problem I'm talking about that so many people think the options are either a standard, Intel Core-style OoO or Itanium-style VLIW. Not only is VLIW a mas
500.
by Veedrac 6y ago
The context is very different if you're talking about the early 1980s. I recognize (and have said before in this thread) that my arguments about lock-in don't apply to then.
501.
by Veedrac 6y ago
They're also pointless once you do even a small amount of dynamic local reordering, which every fast CPU does.
502.
by Veedrac 6y ago
> Itanium was not one attempt It was one ISA , and a terrible one at that, which is my point. > Except, guess what, Cuda and OpenCL look exactly like C in many ways. You're missing my point, which is about rewriting the world. C
503.
by Veedrac 6y ago
And yet most non-games graphics workloads in industry were CPU rendered until 2010ish, I think. The draw of defaults is extremely strong.
504.
by Veedrac 6y ago
> I do believe there is a future beyond classic OoO, but dynamic scheduling is going to be part of it, although possibly not of instructions but of BBs or even microthreads. I agree, I find it hard to see how you could abandon dynamic sc
505.
by Veedrac 6y ago
> A lot of time and effort has been spent by major players This really isn't the case IMO. Itanium is one attempt 20 years ago, it didn't try to rewrite the world (compatibility with standard C code is a huge crutch; GPUs cou
506.
by Veedrac 6y ago
Good luck showing that when all graphics code is designed to be run on CPUs, nobody has spent more than trivial effort making GPU-compatible graphics algorithms, you can't afford new nodes, your GPU is first-generation (and so far from
507.
by Veedrac 6y ago
GPUs were only able to come about into the mainstream because they were brought about and developed alongside their use-case. Much the same can be said for camera ISPs or neural accelerators. GPUs only started tackling more general workload
508.
by Veedrac 6y ago
IMO a lot of the problem is simply that powerful players (eg. Intel) had a vested interest in tying things to the platform, and the megaliths we've built on top all assume this construction, so even if there are clear ways in which thi
509.
by Veedrac 6y ago
> This is a great post showing why you have to measure the specific tasks you care about rather than relying on general assumptions. If anything it turns out it's an argument that benchmarks mislead way too readily and first-princip
510.
by Veedrac 6y ago
Exploration is a major part of RL already, see AlphaGo for example.

More ›