Tuna-Fish

6 ms·

31.
by Tuna-Fish 1mo ago
Because as China has shown, people get angry when you drop hundreds of tons of steel and propellant on people. Rockets sometimes fail. SpaceX has one of the lowest failure rates in the industry, but that's still ~0.5%. When they fail n
32.
by Tuna-Fish 1mo ago
People have been pointing out that various european fringe groups got funded by Russia for a long time. Like, they have actually been caught at it. I always thought it was a weird theory, why would they even bother? That was until the morni
33.
by Tuna-Fish 1mo ago
The 27B one is great on a 5090. This one is basically aimed at macs, Strix halo and DGX Spark.
34.
by Tuna-Fish 1mo ago
It has no internet access and only 8B 3-bit parameters. That's simply not enough for it to compress all that much knowledge.
35.
by Tuna-Fish 1mo ago
The version I've had argued to me didn't change any chronology, they just insisted that the Song dynasty started earlier so the Tang shouldn't be counted. The context was weird, the person pushing it was very passionate abo
36.
by Tuna-Fish 1mo ago
It's terrible because it's Llama 3.1 8B. It's such a crappy model because HC1 was a relatively low budget proof of concept. The team that built is working on a better implementation.
37.
by Tuna-Fish 1mo ago
Please explain why you think cheaper/faster is not coming? All current devices used to run AI are very far from an efficient solution to the problem. What you really want is a pure dataflow architecture, instead of a von Neumann machin
38.
by Tuna-Fish 2mo ago
Chips are made from silicon wafers that are ~0.8mm thick. The transistor layer is built on one side of the wafer (because the transistors need to be built from epitaxial silicon, that is, where all the atoms are neatly arranged, and you can
39.
by Tuna-Fish 2mo ago
It's an interesting option for infrastructure that's not performance critical. I agree with Domen Kožar that I want an extern "fil-c" in Rust. https://domenkozar.com/2026/08/13/i-want-exter
40.
by Tuna-Fish 2mo ago
The current state of the market is primarily funneling VC cash to Nvidia. NV is also holding lots of stock in various companies, most of which is probably going to go to zero, but that won't hurt the company in any meaningful way. Afte
41.
by Tuna-Fish 2mo ago
You forgot the variable that RISC-V chose to maximize: 5) Weird principles that are completely detached from anyone's actual needs and that are carried to a length similar to religious convictions. My biggest personal pet peeve about t
42.
by Tuna-Fish 2mo ago
They burn. But generally speaking, if the plane's structural integrity is degraded enough that the water gets into the battery casings, everyone onboard is probably already dead.
43.
by Tuna-Fish 2mo ago
It has a diesel generator onboard, which provides the reserve.
44.
by Tuna-Fish 2mo ago
> Like why can't the CPU compete against a GPU like that for GPU-y tasks - or can it? Others have taken a stab at the actual differences, but there is a deeper fundamental reason. A CPU is optimized for low latency of operations .
45.
by Tuna-Fish 2mo ago
Power is possible, Thorium-230 has been pointed out as the ideal fuel for such a scenario. Very nice, long and energetic decay chain, 75ka half-life, many of the daughter products are relatively long-lived which helps to flatten out the pow
46.
by Tuna-Fish 2mo ago
That is not a given. A database server can run faster on better hardware because it more effectively and aggressively caches things in memory, which can hurt it on lower-spec systems. Or it can better utilize SIMD instructions that are not
47.
by Tuna-Fish 2mo ago
... Yes. Quadratic is really bad for large enough n, and you need that big context for useful work.
48.
by Tuna-Fish 2mo ago
Note the k in his figure. Taalas HC1 was clocked at 17000 tokens/s.
49.
by Tuna-Fish 2mo ago
You can make any LLM deterministic by dropping the temperature hyperparameter to zero. This will generally make them suck, though, a little bit of randomness is necessary for proper function.
50.
by Tuna-Fish 2mo ago
It's not even just that. If you just built the rom chips separately and swapped them for the RAM of a normal accelerator, it would not help at all. The trick is that every compute element in their system has it's own small pool of
51.
by Tuna-Fish 2mo ago
Taalas exploits the low cardinality to store one 4-bit weight with one transistor. (They are using metal layer traces for the ROM, and connecting an access transistor to light up one of 16 options.) Their system is honestly very efficient f
52.
by Tuna-Fish 2mo ago
The system I'm comfortable with is to set the agent up as an unprivileged unix user, with no ability to change system configuration and no access to any files I didn't specifically give it access to. Need to let it access a file o
53.
by Tuna-Fish 2mo ago
Note that the orbital perturbation of Juno in 1839 was of rather low magnitude, and could have been caused by an asteroid as small as ~1km diameter. Based on current models, there are potentially more than a million rocks that big in the be
54.
by Tuna-Fish 2mo ago
Flash only got competitive after both HDDs and DRAM scaling slowed, allowing it to catch up, and end up in an interesting position between them. There are about a dozen upcoming memory technologies with better theoretical headroom waiting i
55.
by Tuna-Fish 2mo ago
The ceiling is not 100% efficient, when you have a heat pump. Also, the US residential gas infrastructure leaks ~1% of the gas it moves. That sounds like nothing, until you account for how unburnt ch4 is ~80 times worse than co2 on a timesc
56.
by Tuna-Fish 2mo ago
Electrons require significant voltage to move through free space or insulators. They generally move between places by hopping between atoms with free outer electrons. Electrolytes are designed to not have any.
57.
by Tuna-Fish 2mo ago
HBF will probably save the small inference nodes. (Eventually, the initial price will obviously be in the stratosphere.) Long term, HBF shouldn't be more than 4x as expensive as commodity flash, and Kimi K3 is an interesting target for
58.
by Tuna-Fish 3mo ago
Yes, others should be able to do this too, and the change is larger memory chips. The next big step is Medusa Halo, which will have a 384-bit LPDDR6 interface. Those should be able to support 256GB at release, with 512GB coming later with l
59.
by Tuna-Fish 3mo ago
Do not base products on models that are not open-weights. Doing it is like building a product on someone else's platform, you are entirely at their mercy, and even when they don't have any reason to hurt you, you are tiny enough t
60.
by Tuna-Fish 3mo ago
Yes. Not because rom is better than flash, but because the critical part is distributing the memory with the compute. Mask ROM is just the densest way of embedding memory with logic. Instead of having a large pool of memory connected to the

More ›