nixon_why69

It's about effort into thinking. If no effort went into either thinking or writing, then I'll plead guilty to sweat of the brow doctrine. Again, just send me the prompt and I'll feed it to AI myself.

That's a little metaphysical for my taste. If its the same bytes of text, provenance is destroyed anyways, so who knows if the tree fell in the forest.

"Substantially" to me would mean not your ideas and not your writing. I'm sure there are borderline cases that are hard to decide but a rule of thumb for me would be if it takes more time to read than it took the "author" to write.

If its substantially AI generated then it's definitionally rubbish, even if it was a good answer to the prompt. I can send a prompt myself, don't waste my time republishing AI answers.

Speculative decoding is software, though.

How do you work around the memory wall when you're going to have to stream all weights, no matter what? Latency-hiding tricks don't matter when you're bandwidth constrained.

A couple data points, I like Java but I've seen metrics of container fleets at multiple companies that were memory constrained with low CPU usage sitting around underutilized. The reason in both cases was a bunch of memory-heavy yet CPU-efficient Java processes.

Soviet Russia and early communist China both went from mostly-illiterate peasant economies to mostly-literate, partially industrialized, with home-grown heavy industry and weapons technology in the span of like 15-20 years.

Then the soviet system stagnated hard under Brezhnev while the Chinese pivoted under Deng. But those first gains definitely happened under full communism and no capitalist economy has modernized faster.

KV cache is per conversation, I'm getting 100% hit rate on my single tenant local set up.

The benefits of scale are on the token generation side, you can batch rounds and generate tokens for multiple conversations per pass instead of just one token per pass.

It's not even a person, it's a system. There's a million people behind you who will take your job if you don't grind, and that goes for your boss as well.

The usage of Neijuan as I understand it isn't about a complete lack of results, Chinese test scores are really good, but there's only 10% in the top 10% so competing even harder just means everyone is miserable competing for the same number of spots. Same for production if it's driving prices down, you did produce more, but so did everyone else and the end result is a race to the bottom.

In China, one could look at hours of homework, work hours etc and conclude there's a reason the term was coined.

Numbers aside, SpaceX has a technological lead. They're not strangling competitors in a commodity business with monopoly practices. I wouldn't call ASMR or TSMC monopolies either, anyone can compete with them in theory. Standard oil, by contrast, would cut their competitors off from rail transit and selectively dump to drive them out of business.

The other reason they're not a monopoly is that they may do 90% of total launches, but most are for themselves to support starlink (another "nonopoly" with no competitors). They have 60% of the US government launch contracts, a small percentage of ESA when Ariane is flaky, amd of course 0% of China and Russia.

Defending standard oil is an all-timer hot take. They only had 90% market share? In a commodity?

SpaceX isn't a monopoly because they have domestic and international competition. Before they showed up, ULA was a deliberately constructed monopoly because Boeing and Lockheed didn't want to compete.

Standard oil is the textbook example of an abusive monopoly and was broken up for that reason.

They improved efficiency by cutting exclusive deals with railroads and transport companies due to their size which froze out every other company.

Speaking for China specifically, and this is all word of mouth since my kids are at a posh international school (apply hate appriopriately):

Elementary school is 8-5, longish day but not too bad on homework.

Middle school (grades 6-9) there's been a recent drive to have less homework so they stay till only 7:00 (!) doing it at school and only a little take home done between dinner and bed.

High school is your life revolves around homework, everyone realizes it's counterproductive to learning but they're all stuck in this trap grinding for the big exams that will shape the rest of your life. Key word neijuan.

Actual Chinese (and other asian countries) can chime in where I'm mistaken.

Living in one Asian country, phone scrolling still happens with adults but kids get a lot of homework and definitely do it. Honestly excessive amounts, quantity over quality but at least studying is happening.

If some third type of country didn't do doomscrolling but ignored their homework to play outside, that's better than scrolling but it's still not homework.

Yup, and additionally said stooge would probably be optimizing for their own position and package in a way Zuck doesn't have to. There's a whole new category of selfishness and shortsightedness out there.

If colloquial usage of "anarchy" has a different definition than, an, I'm sorry, irrelevant niche political movement, then you should assume a given speaker meant the colloquial form.

If we abandon the moral "should", then the right move is to shake down Google along the lines I already said.

Google should, as a rational actor, pay out better for legitimate vulnerabilities that pose actual risk to them to avoid such a situation.

My beef here is that morality and ethics are only assigned to the researchers.

That's one of the points, yes, but the black market doesn't dictate the value of the exploit to Google.

A hardline bargaining position with Google would be more like "pay me what I want, or else I'll give it to all takers on the black market for maximum damage". That would be unethical and probably illegal to boot but it's a better definition of value than "1$ greater than max bid".

That's really informative but maybe a little overly capitalist-brained.

We shouldn't look to the black market as cost discovery for these vulnerabilities, most non-criminal researchers are not putting up an ask order and letting the black market compete with Google.

For qwen, it's an additional transformer layer at the very back, it ships as part of the model.

I find mtp=3 does well with that model, only at 4 it becomes unprofitable.

Check your quants, its worth having the mtp layer be a bigger quant if it leads to 2x throughput from more accepted tokens.

Yeah, well, Google under Sundar hasn't exactly been full of bold stances.

Google isn't a branch of the government, they can call the lake whatever they want.

It's a decision to comply that quickly and cravenly, presumably because Trump is vindictive, why die on this hill.

> For example, something they still struggle with is consistently translating terms coined within the text with consistent phrases.

Harness issue, just explicitly tell them to maintain a vocabulary as they go for proper nouns and novel terms.

But another 30-40% of SpaceX will come out of lockup in the coming months.

What happens when all of that tries to sell, after the market barely absorbed 5%?

If the results of the distillation are released freely for everyone to download, I wouldn't even call it a wrong.

Big models are built by scraping the recorded thoughts of everyone, so giving everyone a chance to run a distilled small model is just going full circle.

Obviously the underlying motivations aren't 100% altruistic, but I'll still take it.

Sorry for your friend. It sounds unfortunate all around.

I'm not a doctor but I was an athlete and the default suggestion of "my neck/knee/legs/arms hurt" -> massage is standard common wisdom. Maybe an MD would have stalled for an ultrasound but then you have all of the insurance and delay issues.

Chiropractors are cranks but a normal muscle massage is really unlikely to cause a stroke AFAIK. The risks from Chiros as I understand it involve the big "adjustments" where they go for big pops, particularly in the neck.

I'm actually curious too, haven't been there in a while. It's possible that the MBR pyramid has become LLM assisted on all sides but the cultural themes are surely all still there.

Honestly, the whole thing felt LLM to start with, tons of people lacking context and throwing words and summaries around.