148 karma · joined May 4, 2024
Now as punishment go write this on the wall 100 times!
I had to int4 a llama model to get it to properly run on my 3060.
I'm curious, how much resolution / significant digits do we actually need for most genAI work? If you can draw a circle with 3.14, maybe it's good enough for fast and ubiquitous usage.
If AI can do this on 64k tokens, iteratively, fully multimodal... I don't think I've ever actually been scared of a super intelligence / singularity moment until just now.
Now this is AI!
On a side note, used to frequent a bar where one of the creators of Ledger also did. Was nice to spend various crypto freely!
When it comes to the realities of dual-booting, I had tons of problems with Win7/8/10 with suspend-to-hiberfile.sys issues and updates 10 years ago breaking grub. 10 years ago I finally decided, "You know what, I'm just going to run Linux, if I really need Windows or Mac, I can run a VM or use a separate spare computer."
Since then I have successfully setup Secure Boot for my distro, learned how to tweak QEMU for performance and passthrough, got a working QEMU macOS VM (although having to update every few months to keep XCode working is a pain), and generally pretty happy with the state of affairs.
The extent to which microplastics cause harm or toxicity is unclear, although recent studies associated MNP presence in carotid atheromas with increased inflammation and risk of future adverse cardiovascular events2,3. In controlled exposure studies, MNPs clearly enhance or drive toxic outcomes4–6. The mantra of the field of toxicology – “dose makes the poison” (Paracelsus) – renders such discoveries as easily anticipated; what is not clearly understood is the internal dose in humans.
Any issues or recommendations considering the Proxmox route? You do port forwarding or multiple ipv4?
Thanks for putting this idea in my head!
Sorry I missed this thread, I think it was meant for me.
I currently work for a small logistics company as a senior engineer. I am pretty happy and not looking for work (like many of the comments here), but I am happy to converse or schedule a call sometime if you're interested to discuss. 0x03e14@protonmail.ch.
Other commenters had I think the correct answer - you just need ONE REALLY GOOD IT co-founder or vice president to help coordinate that, who can both leverage outsourced solutions but also build things on their own, and most importantly who truly understands and cares about the logistics and cost savings.
I see a lot of dysfunctional and large tech teams building out over-complicated solutions with zero industry awareness, instead of lean solutions.
Did you see this article [1] here from a couple months ago about Temu's semi-managed delivery model? I was pleasantly surprised to see logistics on YC. We don't talk about it enough but it's the backbone of global economy.
Best of luck! Let us know what you decide to do and shoot me an email if you want to connect.
[1] https://news.ycombinator.com/item?id=40497472 - Temu's semi-managed model could change everything
(On a side note, pleasantly surprised to see a Linux kernel android bug today)
With inference tasks it'd be nice to have something not as performance heavy and instead like 24gb vram or more.
I know there are tricks in CUDA to allow access to system ram as a proxy but limited success in duplicating this across various setups.
The people of Puerto Rico are fed up with the government and energy company here (LUMA) and do want change, current governor lost the primary and the candidate replacing him will likely win in November (even though it's the same party, business as usual) is partially because she's very, very anti-LUMA and appeals to the frustrations of people with the situation and is advocating change.
Unfortunately, decades of corruption have really impacted our politics and the energy and infrastructure here, and with the recent $4 million transformer issue, it doesn't surprise me that they are suing to try and get a windfall of money.
There's a difference between the educational-industrial complex (or the ed landscape in terms of market), and motivated learner-centric models which can be done independently.
As much as degrees and certs are important in the current paradigm, it's long been the case in tech that credentials matter less than output, skills, and soft skills.
As much as I personally believe in a stronger credential system, the shift is toward individual learning. The popularity of coding bootcamps in the tech space I think demonstrates that.
Personal AI tutor? Without the pesky SME credentialing? Sign me up!
Besides I need .sh scripts not just cli completion.
But this reminds me of warp. Gonna have to give it a spin in the morning.
I have an extensive MP3 collection, favorite albums I try to keep FLAC, while others I'm fine with MP3/320 or for stuff I collect but don't listen to all the time, VBR 256 is fine to keep size small.
Although it did turn my relatively high-end laptop into the new GE hypersonic ramjet[1] and I couldn't hear the music over the fans... the visualizers were slow (wondering if it doesn't have 3d acceleration?) on Firefox / Linux. (Worked fine on Chrome.)
+1 for Jonathan Coulton though.
Quantitative Criteria for Defining Planets (UCLA) https://arxiv.org/abs/2407.07590
I did a not-for-public writeup on BIMI recently. It boggles my mind that there are no widely used ID services (that is, avatar management and name consistency), especially for something as fundamental as email.
Turns out, you can absolutely setup a BIMI record without paying the ludicrous $1500 cert fee. Turns out, Yahoo and AOL even show those images to senders (at least when I tested last year).
Turns out, Google marks you as spam on an otherwise passing email setup (full alignment, policy = reject all, etc) when you do that. >:(
Some interesting analysis here and how to approach the problem domain (pawn structure as a concern) but that first example, there is just no way to get that knight there in any reasonable way. Talking about unrealistic moves.
Just wait until they have a fully intelligent automated pipeline for lifestyle ingestion (e.g. pervasive analysis of all communication and AV) into AI management (Timeline / Memories) backposted into web 2.0 feeds like Facebook.