Sunday, December 17
Daily News Stuff 17 December 2023
Holly Jolly Whatever Edition
First it was Hololive.
And now Phase Connect.
Who knows what Nijisanji or VShojo could have chained up down there?
Disclaimer: Well, probably Pomu and Kson respectively, but they're not talking. I mean they're talking a lot, but not about this. Yet. Actually, they're mostly talking about frogs. I don't know why.
Holly Jolly Whatever Edition
Top Story
- Small is the new Large: Microsoft has announced Phi-2, which it terms a "small language model". (VentureBeat)
With 2.7 billion parameters it generally outperforms Llama-2 and Mistral with 7 billion parameters, while being small enough to run happily on graphics cards with just 4GB of RAM.
It's comparable to Mistral on language tests, but significantly better on mathematics and programming.
If that's too big for you, Microsoft recently released Phi-1 and Phi-1.5 which will run on any reasonably specced potato.
I'm happy to see this progress in making improvements to small models you can run yourself, that have potential to do some limited set of things well. The push to ever larger models at astronomical expense is going to fail unless fundamental changes are made to the designs - and to the culture of the companies building them.
- Mistral also announced what they call Mixtral last week. (Mistral)
This a similar idea. Rather than building one huge model that tries to handle everything, Mixtral uses eight small (7 billion term) models, each able to fit on a commodity graphics card, and each tuned to a specific kind of task.
It outperforms the largest version of Llama-2 (70 billion terms) while being 30% smaller overall (for all eight models combined) and working on hardware at one twentieth the price.
Tech News
- Speaking of commodity graphics cards Nvidia is expected to launch "new" models in January. (Tom's Hardware)
The models we're looking at are the 4070 Super, 4070 Ti Super, and 4080 Ti.
The only one of interest is the 4070 Ti Super, which is a cut-down 4080 and lowers the price point for a good 16GB Nvidia card. I mean, there is the 4060 Ti, but it's kind of crap.
- MongoDB got hacked. (MongoDB)
They believe this is limited to the corporate systems and does not extend to their cloud offerings. If it turns out it does extend to their cloud offerings, that would be catastrophic.
I run MongoDB at my day job - just finished migrating a huge cluster from one cloud to another - but that's using regular cloud servers rather than cloud databases.
- All these worlds are yours, except Europa. And Enceladus. (NASA)
In fact, all these worlds are ours. Piss off back where you came from.
- Intel, Samsung, and TSMC have now all shown off working CFETs - CMOS transistors with the n-type and p-type transistors stacked on top of each other, rather than adjacent. (IEEE Spectrum)
This is not as profound a change as stacked flash cells, which changed the entire storage industry, but is equivalent to a free process node advance.
Given that process nodes are going to run out of room for improvement in ten years, that's a good thing.
- Hasbro is continuing layoffs at Wizards of the Coast, makers of Magic the Gathering and owners of the Dungeons and Dragons franchise. (Geekwire)
Despite Wizards of the Coast repeatedly fucking over its own customers with woke bullshit, those two product lines are major money earners for Hasbro. It makes little sense for the company to fire core staff working on both products, and yet the announcements from those affected are all over Twitter.
Okay, maybe it does make some sense. (Twitter)
If your team's major accomplishment after a year of work is agreeing not to use AI-generated art, you should expect to be fired. At a minimum.
- Apple's M3 Max is faster than an RTX 4090 in AI-based transcription tasks if you use an utterly broken benchmark. (WCCFTech)
If you use the correct benchmark, the 4090 is twelve times faster.
- If you need to run AI models on the go, here's a laptop with a 64-core CPU and a desktop RTX 4080. (Tom's Hardware)
You may need a bigger lap though. This beast is not small.
VTuber News
- Congratulations to Fishman and all the Sad Girls at Sad Girls, Inc for winning the rising star category at the 2023 VTuber awards. (Twitter)
It's amusing to see people complaining that Sad Girls should be competing with the titans like Hololive and Nijisanji. The entire reason they won this award is because they've grown so much this year that people compare them to the titans.
- And congratulations too to Vedal and Neuro-sama on the Best Tech VTuber award.
Neuro-sama is a home-made AI VTuber who loves nothing more than roasting her poor creator. As the creation of a single person - even building off open-source tools - she's truly impressive.
Secure Your Damn Basements, People Videos of the Day
First it was Hololive.
And now Phase Connect.
Who knows what Nijisanji or VShojo could have chained up down there?
Disclaimer: Well, probably Pomu and Kson respectively, but they're not talking. I mean they're talking a lot, but not about this. Yet. Actually, they're mostly talking about frogs. I don't know why.
Posted by: Pixy Misa at
04:12 PM
| No Comments
| Add Comment
| Trackbacks (Suck)
Post contains 817 words, total size 8 kb.
54kb generated in CPU 0.0137, elapsed 0.1121 seconds.
56 queries taking 0.1025 seconds, 339 records returned.
Powered by Minx 1.1.6c-pink.
56 queries taking 0.1025 seconds, 339 records returned.
Powered by Minx 1.1.6c-pink.
