CMay · HackerTrans

Iraq was never really about WMDs. WMDs were a surface political mechanism to achieve the aims, which were different. If you look at all the individual events that transpired before then it was clear something had to be done, but it wasn't a neat one liner that's easy to rally people around like WMDs.

CMay

ontemdiscuss

That's a propaganda line. Iran has lost most of its capacity to make war, but it can still harass.

The money is also not as relevant as people think. The amount of money is largely political fluff, not much in the way of serious economic impact in a $32 trillion economy.

CMay

ontemdiscuss

I suppose you have reliable sources for this?

CMay

anteontemdiscuss

Can't name any strikes Israel did that the US didn't support?

CMay

anteontemdiscuss

[flagged]

CMay

anteontemdiscuss

The US didn't do that, Israel did. I suspect behind closed doors, Trump thought Bibi went a little overboard there

CMay

anteontemdiscuss

[flagged]

CMay

há 5 diasdiscuss

Well, many Swiss people came to the US and got land or started businesses, so it's not like Swiss people were left out of it. :)

CMay

há 5 diasdiscuss

[flagged]

CMay

há 8 diasdiscuss

If you think those were about colonization, you would be well served to examine history closer.

CMay

há 8 diasdiscuss

Most countries have some kind of story they tell about themselves. New Zealand doesn't have any illusions that it is a superpower. It doesn't have the resources, the talent pool or any of that to even begin to dream of it. That is natural. If New Zealand was powerful, then it would find itself in a position of greater responsibility.

You cannot compare countries that barely have the option of ambition with something like the US and even begin to imagine that it is meaningful. You have the UK, France, Germany, Japan, Russia, China and the US. You can rewind history to name other civilizations.

These kinds of countries are the only ones that matter, because they're the ones that have to answer about what people were thinking when they chose to make use of their power in a way that is relevant on a larger scale. They have to answer about what the reasons were when things went wrong and whether they agree things went wrong. If people are even allowed to talk about it.

It's very cheap to label anything as propaganda without taking the time to appreciate whether it has any merit in terms of the overall behavior of a country or its people. You can always find counter-examples, but how influential are they in the larger picture?

CMay

há 9 diasdiscuss

I didn't make an exceptionalist argument, but if any country's behavior and values can be measured compared to others, you will always be able to make some kind of decision about where those fall in terms of goodness or badness. Do you not believe in good and bad?

CMay

há 10 diasdiscuss

The US doesn't have black sites anymore and when it did, the interrogation techniques were chosen to avoid physical harm. The results were bad, we didn't like it here in the US even if they were extreme measures for extreme times and so we shut it down. It had a high error rate and generally didn't reflect what we thought was right.

Meanwhile the CCP regularly abducts its own citizens and executes more people than the entire world combined.

CMay

há 10 diasdiscuss

In some ways it's an offshoot of the honest benefit of search engines already crawling all this content. That has its own conflicts, like just how much of a page's content should you reproduce in the results before it's basically considered stealing their content without benefiting the site itself.

There is a balance to strike, both in search engine fair use cases and AI fair use cases. The major cloud LLMs do double as web search engines now, though they didn't originally. In many cases there's no reason left to click the links they sourced from.

That is a legitimate concern. At least within the US, I think there are nuances around fair use and contract law. A lot of companies are getting paid for having their content used in these models, but many websites had no particular rules you had to abide by and the content was simply public. I think if you're operating under an agreement, then even if there is fair use or public domain content being reproduced by the site you are still bound by that agreement.

Similar to old paintings digitized and hosted on some museum website. It's 300 years old, right? It should be public domain, yet the people who digitized it or provided a service to give you access have some say in how their reproduction can be used. These AI services are obviously very different, but there are laws that can govern how you are allowed to use a service if that service has laid out acceptable usage.

I'm not exactly comfortable with the mass scale that everything was soaked up to train these models even within the umbrella of search services, but I also admit that a lot of the usage was probably quite legal. The potential displacement caused by the resulting trained models on artists or writers is almost its own facet. In practice, whether they ONLY trained on strictly legally acquired fair use content with no errors and paid agreements to acquire even more content than they already do or not, there was enough legally accessible information for fair use that there was no escaping some kind of impact on artists, writers, etc.

With any luck, artforms and skills impacted by technology will adapt and continue to be valuable instead of complete displacement or the dilution of opportunity.

CMay

há 10 diasdiscuss

Certainly the world is full of actions and reactions, nothing is happening in a vacuum. You don't have to be from a country to take sides, but presumably you have some kind of moral compass, some kind of values around personal freedom or the worth of a human life.

There can be a very real cost, because one side comes from an ideology with a history that wants to conquer the entire Earth which caused World War 2 while the other side is trying to prune the planet like a bonsai to prevent it from descending into total chaos to preserve some sense of international order.

Europe was constantly at war, and we helped stabilize it. Middle East as been constantly at war, and if Iran can be sorted then it will be the closest to some sense of peace it's been in a long time.

We used to be in Japan, Philippines, Germany, Vietnam, South Korea, Iraq, Afghanistan and so on. How many are US territories? None. We aren't out there to conquer the globe and take land. We're usually fighting other people's wars for them, because they're up against better resourced opponents. Meanwhile China is over there building artificial islands, ramming other country's ships, creating ideological police stations in countries around the world to harass people and engaging in the most widespread international interference campaigns in human history.

They do not treat their people well and they do not have free speech. The internet is flooded with their propaganda now, because they have a human numbers advantage.

It's true that given time most advantages are temporary, but there's always that slim chance we could slow them down until the CCP collapses and they could become a more normal country.

CMay

há 10 diasdiscuss

The usage of the output is probably considered legal. The usage of the service for that purpose may not be, and using it at scale in a dishonest way is not, which is what China has been doing. Countless thousands of separate requests abusing the service (which is not a simple static HTML feed, but an AI service request) for every kind of query to soak up the results.

The post is about what's in the local code, but for a long time there has already been modifications made to the request outputs from the major cloud services as they work together to both curb adversarial distillation and to degrade the quality of training China can get from that distillation.

It's likely not to make the answer wrong or bad, but to make it so that any model trained on the output would not gain the benefit of the model's reasoning generalization skills as easily and also identifying markers that might even link back to request IDs.

The techniques talked about in this post are naive and simplistic, largely because they are released publicly.

It's not as much about protecting IP as much as it is about slowing China down or being able to track the effects of abuse. So many people are talking about greedy company this, greedy company that. The world is not made up of caricatured giant money pigs wearing suits with monocles and gold watches. That is a children's view of Marx's exaggeration on free markets. Bad, greedy people do exist, but if that is your only hammer for every nail then you have a problem.

The upper-bound for how good these models can be is so crazy that it is essentially dual-use military applicable to an extent most other technologies are not. It's not only cyber attacks or biological weapons. Most people are not even built to understand the possible threats.

Why does it matter if China gains those capabilities? I invite you to begin to learn about China's behavior around the world. The CCP is darkside material.

CMay

há 10 diasdiscuss

Llama.cpp implemented some rotation optimizations for quantized kv cache to improve the preservation of attention quality or similar, after everyone was talking about TurboQuant. It's not perfect and when you're talking about long form reasoning, little differences can make or break the results so it is situational.

CMay

há 12 diasdiscuss

The MoE models hold up better on old hardware, but the dense models like this post promotes are in fact better. This isn't unique to Qwen. Are the dense models better-enough to use given the performance costs? It depends on what you are doing.

If a model runs fast enough for your use case and does exactly what you need it to, then you don't need a much slower model that might be more accurate. If you do anything more complicated, the dense models become more necessary and they are much more computationally heavy by comparison.

On your hardware an Unsloth quant of Gemma 4 26BA4B QAT would likely give you better results, but because it has 4B active parameters instead of Qwen's 3B active parameters, it will probably run slower.

CMay

há 12 diasdiscuss

This is my experience too. Qwen optimizes for a lot of scenarios which masks their weaker generalization compared to US frontier models.

Never go below an fp16 kv cache unless you've already tested it in advance with your model on a verified task that you know it can successfully complete. People should also test the difference using the exact same seed value so they can see how the tokens diverge. If you have memory constraints, sometimes you can still use an fp16 kv cache and use storage for an agentic buffer to work your task with mixed abstractions rather than having everything in memory.

For 4-bit weight quants, Gemma 4 31B QAT is where people should be looking instead of Qwen 3.6.

CMay

há 12 diasdiscuss

At 24GB, Gemma 4 31B QAT will be better and give more concise answers. This post is mostly about unquantized results, so it's less relevant and I can't say much about as I haven't tested Qwen or Gemma via cloud API or unquantized locally. All I can say is locally, quantized in a 24GB scenario, Gemma 4 31B is better in my tests which are mostly reasoning or C programming related.

Gemma 4 is the only model series at this parameter scale I've seen correctly answer some of these. One of the answers even made me re-evaluate what I thought the correct answer was, which I did not expect.

When I look at the Artificial Analysis numbers, I can see that some things about Qwen 3.6 look inflated as a result of either metrics that weren't measured yet for Gemma 4 31B, or for metrics that just aren't going to be relevant in a lot of the essential tasks. In a lot of the relevant metrics, Gemma 4 is either better or on par.

Then once it's all quantized all those benchmark results will be hurt, and Gemma 4 QAT has better quantized performance. I think it's more competitive unquantized than people give it credit for and way better quantized than people give it credit for.

Qwen 3.6 clearly isn't legitimately bad and maybe it's quite nice at fp16, but it was a disaster quantized in a 24GB scenario by comparison.

comments