What's rarely addressed in these articles is the question: if the product/service is so bad relative to the cost, where's the competition? Specifically in this article about fire trucks, they say that margins have tripled... ok, why isn't anyone jumping on that?
> Society gains massively from future workers/tax payers
It seems to me that, all other things equal, future workers/tax payers will lead to economic increases proportional to their costs.
A reasonable forward looking plan / budget scales with the population size. Therefore there would be no need for these special one off exceptions and nudges.
All these little bandaids add up to complexity that necessitates more bandaids.
If your populations shrinks quickly, you end up needing to run the infrastructure (and elder care!) for a whole country with too few working-age people.
This is a massive problem, and some incentive complexity to avoid it is certainly worth it.
> Under the US Cloud Act, American companies are required to hand over all information they store to the government if requested to do so, even if it is stored abroad.
Hrm. It's my understanding that a US company is required to give almost no data to any government without a warrant.
True, but a US warrant. So if data is stored on a system in another country owned by an American company, they can be compelled to hand the information to the US government even if it is illegal to do so according to the law of the country where in the information is stored.
So, for example, a lot of medical information stored on AWS by the NHS could be obtained by the US government. So could a lot of financial and government data around the world. Zoom calls, Teams meetings, emails sent to GMail. Google Drive and one Drive document. Lots and lots more.
> Replace ‘CTF’ with ‘high school’ or ‘university’ and you’ve described the total slow motion collapse of education; the only saving grace is that most of it requires in person presence.
So something like, "Frontier AI has broken the 'high school' or 'university' format"?
The hype surrounding AI is just pervasively exhausting: you've got the folks talking about an entire new age for humanity where we're shortly going to take over the entire universe. And you've got the folks talking about how our entire society is crumbling.
Education is one place folks seem to throw up their hands and say nothing can be done.
The fix is simple: students are to be evaluated on their performance in person. That's it.
Any other "collapse of education" isn't due to AI, it's something else.
I would love for the standard to be to ALWAYS report the required amount of memory to load and run a model in bytes of RAM alongside any other metrics. I'd love to see time to first token, token throughput, token latency as well but I'd settle for memory size as described above.
Essentially, many people want to know what the minimum amount of memory is to run a particular model.
Parameter count obscures important details: what are the sizes of the parameters? A parameter isn't rigorously defined. This also gets folks into trouble because a 4B param model with FP16 params is very different from a 4B param model with INT4 params. The former obviously should be a LOT better than the second.
This would also help with MOE models: if memory is my constraint, it doesn't matter if the (much larger RAM required) MOE version is faster or has better evals.
I'm waiting for someone in anger to ship the 1 parameter model where the parameter according to pytorch is a single parameter of size 4GB.
As a proxy for the total size of the parameters, you can just look at the download size of a model on Huggingface.co.
Because for most models the weights are provided in many *.safetensors files of approximately the same size, you can estimate the total size without adding all file sizes by multiplying the number of *.safetensors files with the approximate size of one file.
For quantized models, estimating the size is simpler, because there is just one GGUF file, which also includes metadata, but most of the file is occupied by the parameters.
While there are models where the native size of all parameters is BF16, there are also models that use multiple parameter sizes, e.g. a large number of parameters with a small size, even down to 4 bits, together with a small number of parameters with a bigger size, up to FP32. Therefore, as you say, the number of parameters is much less informative about memory requirements than the file sizes.
While the download size of the *.safetensors files or GGUF files is not the same as the total memory requirement, it can give an approximate estimate and it can be used to assess which of 2 models will need more memory. It becomes more complicated when you must use multiple kinds of memory, e.g. GPU memory and CPU memory, or even SSDs, when you must know more about the structure of the model to determine how much of each kind of memory is needed.
The KV cache size is a joker though. Different models use very different amounts of memory per token in the KV cache. The VRAM requirements for say 64k context can vary almost by an order of magnitude. So the download size might indicate you should have room for the model, how much context you can fit in the leftover VRAM budget is harder to predict at a glance.
That some models like Qwen3.6 27B seems to not be very affected by Q8 quantized KV cache while others degrade heavily doesn't make it easier.
I'd never heard of it either. A comment further down suggests it is Japanese.
Digging deeper, the kyu -- or Q for quarter millimeter -- is apparently a foundational distance measurement in Japanese typesetting, which is metric and operates on a millimeter grid.
It's probably the sanest adaptation of the point to the metric system. A traditional point is close to a third of a millimetre, but that's too weird.
Since the Q is close to 3/4 of a traditional point, it's also quite easy to convert from traditional multiple-of-three point sizes: 9 pt -> 12 Q, 12 pt -> 16 Q, etc.
Although it's even easier just to call those 3 mm and 4 mm!
When the mcdonalds quarter pounder with cheese came out, the europeans came out with this government sponsored measurement for fast food chains to standardize and compete without adopting american standards. (like usb-c over lightning)
Caveat: brain-computer interfaces are not quite my field, but I think the consensus is (judging from some conversations with folks who know more):
Neuralink is doing interesting BCI research, with decent hardware, but it's not really a step-change above and beyond the rest of the field.
There's definitely a lot of promise in using BCIs for rehabilitation of patients with brain injuries but their input-output capabilities are still incredibly crude: for example, we can't reliably "write" to the brain to make people perceive things beyond very simple stimuli (e.g. a phantom touch sensation, or a visual phosphene).
This is understandable: the brain has a bajillion neurons and we only have ~1,000 electrodes that aren't particularly precise in how/where they zap the brain---and even if they were, we don't really know well enough how the brain works to "control" perception finely.
Other problems for BCIs include (i) "representational drift", where the brain's code changes over time, so you need to keep fine-tuning your interface in some sort of closed loop fashion and (ii) damage/scarring to neural tissue.
> Is there enough signal for this to really work?
I'm not quite sure what Neuralink's marketing claims are, so I'm not sure what you mean by "this" here. But intracranial electrodes do have a surprising amount of signal, especially relative to non-invasive methods (I'm currently collecting some iEEG data myself!)
I really want the sci-fi future where we have brain-computer interfaces that augment our cognition and perception, but we're nowhere close---though we're getting better.
> Have you never met a bad doctor? A shoddy lawyer? A barista with a PhD?
I presume the implication is that bad doctors and shoddy lawyers exist and just because they have advanced degrees doesn't make them good at what they do. This seems reasonable.
BUT, I find it fascinating that people who aren't doctors or medical experts think they can spot a "bad" doctor or people who aren't lawyers or experts in law think they can spot a "shoddy" lawyer.
A good doctor/lawyer makes good decisions and executes beneficial actions given the facts surrounding a situation. It's pretty hard to judge whether those decisions and actions are good or bad if one isn't an expert.
That's a huge motivating factor for professional licenses.
Well, yeah, 99% of arXiv papers were not written for me or you. They were written for someone who works in a niche within a niche. That's (in my view) the beauty of research.
Agreed. There was already too much human generated slop in academia.
And I’m not talking about good faith research that didn’t pan out, I mean research that is completely useless for any other purpose other than convincing a casual observer that the authors are doing research.
reply