fulafel

14 ms·

151.
by fulafel 6mo ago
Looks like DeepSeek did this as well since V3: https://deepwiki.com/deepseek-ai/DeepSeek-V3/4.4-multi-token... Credit for the MTP technique is due to https://arxiv.org/abs/2404.19737 from 202
152.
by fulafel 6mo ago
AIUI you run the checks of several predicted tokens in lockstep, and the computation for each token is served by the same data loaded from memory. In normal execution, each token would depend on the previous one, precluding the parallelizat
153.
by fulafel 6mo ago
These losses are peanuts compared to the externalities. article: > its net profit over those 47 years sits at negative $37 billion EPA social cost of carbon: $190/ton [1] US aviation annual emissions: 200 million tons/year Glo
154.
by fulafel 6mo ago
OTOH there was recently a DNSSEC-saved-the-day piece of news: https://incrypted.com/en/dns-attack-on-eth-limo-was-stopped/
155.
by fulafel 6mo ago
There was also a way to deploy on AWS: https://aws.amazon.com/blogs/containers/deploy-applications-... - but AWS retired it pretty soon. I wonder how many people were using it when it was pulled, and if there were
156.
by fulafel 6mo ago
You involve the kernel also when you are doing async io. In this context the interesting thing to measure would be doing IO in your green threads vs OS threads. A stronger theoretical performance argument for async io is that you can do bat
157.
by fulafel 6mo ago
Yes, GPUs run machine code like CPUs. The toolchain compiles code to machine code that gets executed on GPUs. In the stacks referenced above it's not exactly Python, which is why I used the term Python-ish instead of Python.
158.
by fulafel 6mo ago
You can call it exploitation, as the article does, which means something different.
159.
by fulafel 6mo ago
Java, Rust, Go, C#, Swift are the biggest ones (in no particular order) on the CPU side. On the GPU side I'm not as up to date, but there's a lot of Python-ish stacks like JAX, PyTorch, Mojo, etc.
160.
Pipelight(pipelight.dev)
3 points by fulafel 6mo ago|0 comments
161.
by fulafel 6mo ago
It's a dying breed in that it's being outgrown by others and the niches where it's the best choice are getting fewer, but it's not going away soon either. For example high performance multithreaded code is now largely wr
162.
by fulafel 6mo ago
I'd bet for the null hypothesis: the memory behaviour changes would hold if the core count was kept constant and only the VM's memory size was adjusted.
163.
Porting microgpt to Futhark, Part I(kmjn.org)
2 points by fulafel 6mo ago|0 comments
164.
by fulafel 6mo ago
Wasn't there already consensus about CLI tools > MCP after the initial hype?
165.
by fulafel 6mo ago
If the port number space was bigger, I wonder if we would have gotten a global naming service (ala DNS) for unique service names. You can still publish port numbers along with addresses in DNS though (SRV records).
166.
by fulafel 6mo ago
We have to urgently stop doing this of course, to mitigate the climate catastrophe. Wars are peanuts compared to the death toll.
167.
by fulafel 6mo ago
Yep. Still, unfortunately switching to EVs is not the endgame. We have to stop fossils lower in the supply chain. Otherwise they'll be used for other forms of energy demand and get burned to heat up the atmosphere anyway. Passenger car
168.
by fulafel 6mo ago
It's interesting how slow the uptake has been. 10GbE is from 2002, 24 years old. It's much worse than IPv6, at least we're up to 50% uptake there. 100 GbE is 15 years old too depending on which standard you think of. (The mos
169.
by fulafel 6mo ago
People do it in Clojure all the time in the dev setup. And you technically can do in your customer environments too, but it's of course a bit of a cowboy thing to do there.
170.
by fulafel 6mo ago
> PREEMPT_NONE: The kernel almost never interrupts a running thread This seems confused. These are options for preemptibility of the kernel, which is a relatively modern fearure. Userspace could always be preempted and these options do n
171.
by fulafel 6mo ago
The industry culture related to security work and career paths seem just f'd up. Instead of ensuring we build systems with robust foundations, people end up in a swamp of frustrating roles like SOC staff chasing alarms about false posi
172.
by fulafel 6mo ago
This is horrible from POV of mitigating the climate catastrophe and the global death toll. We (as in humanity) are really late in ramping down fossils production and use.
173.
by fulafel 6mo ago
there's more details under the Too narrow and too wide tests heading. It would be interesting to see a deeper investigation, into how the models are dealing with this and whether the successful ones seemed to be trained on the benchmar
174.
by fulafel 6mo ago
The article is about solving a previously unsolved one. This is a harder set of course.
175.
by fulafel 6mo ago
File transfer and storage (Dropbox, browser download, rsync, scp, NFS/SAN etc) is a classic use case that can utilize all the bandwidth you have and typically uses single streams between client and server.
176.
by fulafel 6mo ago
A single threaded benchmark better represents real performance, I'd argue. 10 Gbps is only 1.2 GB/s after all and few applications use parallel streams.
177.
by fulafel 6mo ago
Interesting conceptual parallel to vibecoding, where also you get something that works, and it's another job to understand how or why.
178.
by fulafel 6mo ago
You can't do it with zero kernel code for ISA devices, but if there was a pci busmouse, uio + uio_pci_generic would work for reading the mouse, and you'd use uinput to send the events to the input stack. If you're willing to
179.
X.1095 – Entity authentication service for pet animals using telebiometrics(itu.int)
1 points by fulafel 6mo ago|0 comments
180.
by fulafel 6mo ago
For anyone wondering about the rest of the X standards, they're at: https://www.itu.int/itu-t/recommendations/index.aspx?ser=X For example from 2023: X.1095: Entity authentication service for pet animals usin

More ›