storus

27 ms·

541.
by storus 1y ago
M3 Ultra GPU is around 3070-3080 for the initial token processing. Not great, not terrible.
542.
by storus 1y ago
RTX Pro 6000 can't do DeepSeek R1 671B Q4, you'd need 5-6 of them, which makes it way more expensive. Moreover, MacStudio will do it at 150W whereas Pro 6000 would start at 1500W.
543.
by storus 1y ago
If the rumors about splitting CPU/GPU in new Macs are true, your MacStudio will be the last one capable of running DeepSeek R1 671B Q4. It looks like Apple had an accidental winner that will go away with the end of unified RAM.
544.
by storus 1y ago
They pay hardware rates to software engineers (principal engineer at the salary level of a decent fresh graduate) so I won't be too optimistic about them attracting software people that would propel them forward.
545.
by storus 1y ago
People = upper management in this case. They are people too (at least until they get automated away by some LLaMA in the future).
546.
by storus 1y ago
ARM is not an open platform like IBM PC was. See Android phones and their custom Linux kernels with undocumented parts...
547.
by storus 1y ago
I meant the section of the plane was between front and center, behind business class.
548.
by storus 1y ago
Seat 11A. That's interesting, typically most survivors are located in the back, this one was in the front-center.
549.
by storus 1y ago
I've just had a "fun" with DeepSeek R1 when I asked it about the Pope Leo XIV's views on AI, and it started spewing that he doesn't exist and Francis is still the pope. After sending it multiple URLs its conclusion
550.
by storus 1y ago
After covid I noticed that drinking tap water was making me noticeably worse but buying packaged mineral water didn't (this was done over a few months). I later bought reverse osmosis and never hit that problem since. However, this yea
551.
by storus 1y ago
IIRC you can download OS update/installation DMG from Apple, put it on a USB key and run it on airgapped system. I don't think you even need Apple ID. MacOS with homebrew works more-less like Linux, at least tooling is basically t
552.
by storus 1y ago
MacStudio with 512GB RAM starts at around 10k and quantized DeepSeek R1 671B needs around 400GB RAM, making it usable for your needs. It produced some outstanding code on many tasks I tried (some not so outstanding as well).
553.
by storus 1y ago
This feels different; I asked DeepSeek R1 to give me an autoregressive image generation code in pytorch and it did a marvelous job. Similar for making a pytorch model for a talking lip-synced face; those two would take me weeks to do, AI di
554.
by storus 1y ago
Soon all coding will look like L3 support - debugging something you've never seen before, and under pressure. AI is really taking away the fun parts from everything and leaving just the drudgery in place.
555.
by storus 1y ago
If you want to understand how it works, take first Stanford's CS223A and then MIT's 6.832 Underactuated Robotics courses. They have lectures online as well. Udacity's Self-driving Car Nanodegree might be a good idea after tho
556.
by storus 1y ago
While this is an impressive achievement, I am wondering if this prompts Russia to actually use nukes as it looks like they might be in front of a dilemma "use it or lose it" given what one can do with drones. Wiping out one third
557.
by storus 1y ago
I had that feeling with Andor - Season 1 was great, Season 2 was... not sure how to even name it, it was like a completely different show.
558.
by storus 1y ago
I thought that fine-tuning is no longer being done in the industry, instead transformer adapters like LoRA are being used? Having 1000 fine-tune models for each customer seems too heavy when one can have instead 1000 transformer adapters an
559.
by storus 1y ago
This problem is happening even on Apple MPS with PyTorch in deep learning, where fast math is used by default in many operations, leading to a garbage output. I hit it recently while training an autoregressive image generation model. Here i
560.
by storus 1y ago
How is Google going to survive? Their search quality tanked, the AI is too expensive to run for every single query and they cut off people that made their advertisement model possible. Are the advertisers themselves going to form a consorti
561.
by storus 1y ago
Given EU is going to ban all privacy-preserving cryptocurrencies in 2027, what are the options for EU citizens?
562.
by storus 1y ago
Stanford requires pen & paper exams for their remote students; the students first need to nominate an exam monitor (a person) who in turn receives and prints the assignments, meets the student at an agreed upon place, the monitor gives
563.
by storus 1y ago
The same complaints come from my doctor friends - they are now also being run as a factory where each procedure has a certain allotted time and they need to fit each patient into it. Like you have 10 minutes for a GP task, 15 minutes for ps
564.
by storus 1y ago
Cool! How does it differ from Animatron that had SVG animations for ~10 years already?
565.
by storus 1y ago
Google seems to be firmly in the Kraftwerk-like definition of beauty where some band members believed that the more beautiful a sound looks on an oscilloscope, the better music it is.
566.
by storus 1y ago
The main issue with learning German is that each grammar rule has a bunch of heavily used exceptions one needs to learn, making the language very illogical and more idiomatic. This leads to a sort of a neurosis when many native Germans are
567.
by storus 1y ago
No wonder, the CEO basically said they'll use the free labor of love of all member devs to sell training data for AI in order to replace such devs so why would anyone keep contributing there?