nullc

6 ms·

61.
by nullc 2mo ago
It's a cheap to evaluate proxy for totally broken or not, which is a good start. It also has a lot of resolution and not a lot of noise. Better would be multi-turn benchmarks with tools but getting good precision and accuracy for that
62.
by nullc 2mo ago
I'm not having any looping. > --temp 0.2 Looping is a common symptom of changing the sampler settings from what it was RL trained with.
63.
by nullc 2mo ago
35Ba3b is usable in plain cpu inference on a fast server, the dense model is MUCH slower. On GPU the 35ba3b is still around 2x the tok/s single threaded, which can be a good tradeoff for some applications.
64.
by nullc 2mo ago
HUH? AgentWorld is a simulation of the world (e.g. tools and programs) for use by an agent!
65.
by nullc 2mo ago
> (1) make effective use of retrieval tools and ( A downside is that you can't just download a lot of that knowledge, vs with the weights the copyright infringement has been outsourced to the lab. Nor can you just search for the in
66.
by nullc 2mo ago
Diverging from the sampler used in RL training is not good for long multi-turn results-- it's a great way to knock models into reasoning loops that wouldn't otherwise.
67.
by nullc 2mo ago
I wouldn't find glimmer interesting except that it has much less memory usage per token of KV than Qwen. So I can get 24x concurrent glimmer on 2xRTXA6000 (with 128k context) where I can only get 6 Qwen 27b. This means I can get some
68.
by nullc 2mo ago
Private AI is practical by running the model locally, every much more so than any homomorphic encryption scheme. So essentially the headline sells this as work to keep your data private, but really it's work to keep the AI-- which was
69.
by nullc 2mo ago
for layer parallelism (e.g. to get more vram) the bandwidth between layers is essentially nothing (like 16kb per token I think), so I don't think x4 would even be a problem!
70.
by nullc 2mo ago
Modern benchmarks across the board really need to start severely penalizing disobedience and hallucination. A year ago models weren't really strong enough to justify this but they are now-- the frontier isn't in squeezing out the
71.
by nullc 2mo ago
Make cyber not Cyber.
72.
by nullc 2mo ago
> Labs have already used up internet-scale data Not really, but a lot of what isn't used isn't very good. More important is synthetic data. Use a teacher model with RAG with a huge reference library to write synthetic transcri
73.
by nullc 2mo ago
Did you know: general purpose computers are completely pointless, because programs can run forever without producing a result.
74.
by nullc 2mo ago
I've made comments on HN on this point a number of times e.g. https://news.ycombinator.com/item?id=44284083 I had some tedious debate on HN once where I asked if anyone had any pointers to good parallel SMT solvers,
75.
by nullc 2mo ago
That is proximal to one of my peeves in this space: People who misunderstand approximation results to be meaningful when often they're not. For example, the minimum set cover problem shows up in cases like "What minimal set of te
76.
by nullc 2mo ago
You're missing that modern CPUs substantially lock the users out of control of their own computer and include things like hidden additional network connected processors that run their own full on operating systems. ... and may well be
77.
by nullc 2mo ago
This is a great starting point to go looking for PSP / SMM backdoors.
78.
by nullc 2mo ago
Anyone else struck by the similarities between the description an "The Cookie Monster” by Vernor Vinge?
79.
by nullc 2mo ago
Might be an interesting motivation for looped LLMs to cut the gate count down. Perhaps even a collection of mixed programmable layers and baked layers in a loop.
80.
by nullc 2mo ago
Broad form car insurance might accomplish what you want: https://www.policygenius.com/auto-insurance/what-is-broad-fo... But I believe the intersection of states with highly private registration and where broad form is
81.
by nullc 2mo ago
You shouldn't link that without mentioning that it is a particularly intense S&M/rape/incest fetish piece, and that the squick content is entirely integral to the story. I'm generally dubious of "trigger warnin
82.
by nullc 2mo ago
Good luck even finding the non-directors cut. I liked the film noir voice-over.
83.
by nullc 2mo ago
Sorry to offend, my point is that the necessity of auto insurance will stick your name (as the insured party for that vin) in a place accessible to flock (or at least flock's users). And as such I don't think it provides the res
84.
by nullc 2mo ago
AI psychosis is alive and well, it seems.
85.
by nullc 2mo ago
Insurance demands your vin and reports it to LexisNexis, and anyone with a Lexis Nexis account can turn a VIN into a Montana plate so long as they self-certify that their use complies with the DPPA (which includes enough reasons to drive tw
86.
by nullc 2mo ago
Care to explain how you can insure a car without the insurance selling your data to data brokers? All states with remotely private registration still require insurance, including Montana.
87.
by nullc 2mo ago
I disagree with the underlying cause claimed in that analysis: https://news.ycombinator.com/item?id=49141886
88.
by nullc 2mo ago
> Perhaps the most compelling forensic observation is that the first vulnerable firmware was released on March 17th, 2021. From that date onward, we see stolen coins beginning to pour into the few shared addresses used by the attacker. t
89.
by nullc 2mo ago
Hopefully people are being a little circumspect right now with explotation instructions because there are still vulnerable coins out there (particularly from mk4+ wallets, as well ones with passwords that are not strong enough) which have n
90.
by nullc 2mo ago
This flaw would probably be easier to find from the binary: the source itself was misleading, but if you analyized the binary you'd find that there was no access to the hwrng at all (assuming the dead code got eliminated by the compil

More ›