nl

31 ms·

631.
by nl 7mo ago
> Most popular "VPN" companies only offer a proxy that merely runs over a VPN protocol. Well that doesn't seem true? Mullvad, Proton, Private Internet Access, NordVPN, ExpressVPN etc are all VPNs. You can use them for what
632.
by nl 7mo ago
But then you can use CPU/RAM offload, which already allows you to offload without a kernel module.
633.
by nl 7mo ago
I believe Lemonade is the AMD team right? But yes I agree with you about their lack of prioritization for software!
634.
by nl 7mo ago
This is really interesting engineering, but I agree with the other commentators that the benchmarking makes it hard to understand the contribution various factors are having. The ExLlamaV3 EXL3 2bpw (8 GB, full VRAM) row is an order of ma
635.
by nl 7mo ago
The Ryzen AI CPU/GPUs (Ryzan AI 395+ etc) seem to have increasing support - https://lemonade-server.ai/ now has support for the NPU as well as the combined CPU/GPU (which I guess is a APU but is different to the G
636.
by nl 7mo ago
I don't think RAG is dead, and I don't think NFTs have any use and think that they are completely dead. But the OP's blog is more about ZK than about NFTs, and crypto is the only place funding work on ZK. It's kind of a
637.
by nl 7mo ago
> One thing I noticed is that both this and another incredible film this year, Sirāt, were, at least in part, funded by a grants and state institutions. Aren't almost all films partially funded by grants and state institutions? Fo
638.
by nl 8mo ago
The 0.8B can do this pretty well. Actually pg's original "A plan for spam" explains how to do this with a Bayesian classifier.
639.
by nl 8mo ago
> Have they announced this? No and indeed they have said they never do this at all.
640.
by nl 8mo ago
/security-review really is pretty good. But your codebase is unique. Slop in one codebase is very dangerous in another.
641.
by nl 8mo ago
Things like the METR benchmark aren't sufficient? I mean Moore's law is just a rule of thumb but the curve fits METR just as well..
642.
by nl 8mo ago
Read the linked report - it discusses this. The short answer is that it's possible if the prompt has sufficient control but only the parts controlled by the human are eligible for copyright. Using AI doesn't automatically disquali
643.
by nl 8mo ago
This oversimplifies it. You can't copyright a work that is only generated by a machine: "In February 2022, the Copyright Office’s Review Board issued a final decision affirming the refusal to register a work claimed to be genera
644.
by nl 8mo ago
I hadn't seen Panoramax before, but I'm finding exploring it frustrating. If I look at Melbourne I see there is some data: https://api.panoramax.xyz/?focus=map&map=5.99/-37.126/147.65... But if I zoo
645.
by nl 8mo ago
This is a pretty interesting report. The TL;DR is that there is little measurable impact (and I'd personally add "yet"). To quote: "We find no systematic increase in unemployment for highly exposed workers since late 202
646.
by nl 8mo ago
> So, these speculators are like "oh no, more GPUs requires more RAM!", and then just start speculating on all RAM. Are you claiming that these speculators are buying DDR5 RAM and warehousing it somewhere? Or what exactly is th
647.
by nl 8mo ago
No, we are literally trying to find a use case where using a lower accuracy LLM makes sense for a vision task. But fine - what are these industrial processes where that prioritize latency over reliability and using a LLM - as mentioned by
648.
by nl 8mo ago
There is no way you can diagnose this like that. Correlation isn't causation and much more likely is a common source of reinforcement training data.
649.
by nl 8mo ago
I've never tried smart folders - I didn't know that feature existed. Search is mostly ok for me.
650.
by nl 8mo ago
I tried vibe-coding a CRDT app last year in the pre-Sonnet 4.5 era (based partly on this write-up), and it really struggled with timing issues. It would be interesting to try again now.
651.
by nl 8mo ago
Not often enough for me to really notice. I use my MBP around 10-12 hours a day, maybe more. I don't love the finder but never really have bugs exactly. Time machine isn't great but it's ok. It gets stuck if you don't ha
652.
by nl 8mo ago
Yes I agree. I have an M1 Max like the author of this piece and recently upgraded. It's fine. I don't like the look of it much and the drag targets are annoying but other than that it's been completely normal.
653.
by nl 8mo ago
Model distillation is very useful! Put it like this: Reinforcement Learning from Human Feedback (RLHF) is useful with hundreds of examples, and LLM distillation is basically the same thing.
654.
by nl 8mo ago
Why? Seems like a valid requirement to me? I build micro apps from 10-word prompts multiple times a day.
655.
by nl 8mo ago
Strix Halo machines are a good option too if you are at all price sensitive. AMD (with all the downsides of that for AI work) but people are getting decent performance from them. Also Nvidia Spark.
656.
by nl 8mo ago
Claude Sonnet can easily one-shot that without specifically asking for plan first.
657.
by nl 8mo ago
> Speculation is that the frontier models are all below 200B parameters Some versions of some the models are around that size, which you might hit for example with the ChatGPT auto-router. But the frontier models are all over 1T parame
658.
by nl 8mo ago
A 3 bit quant is quite a lot weaker than the OpenRouter version the OP is using.
659.
by nl 8mo ago
But thay doesn't imply they can't do business with say the German Federal Government for example?
660.
by nl 8mo ago
Yes I got that. But doesn't that mean that non-US customers would be the major customer segment still open to Anthropic in that scenario? I still don't see any way to read that as saying they could only do business with US custome

More ›