Comments URL: https://news.ycombinator.com/item?id=48154550

Points: 2

# Comments: 0

But if you run your own models then you're not subject to anybody's whims anymore. You have full control of how your software works and what it does.

And there are also occasional statements like the one by Airbnb here disclosing what they use https://www.bloomberg.com/news/articles/2025-10-21/airbnb-ce...

For most use cases, you don't actually need frontier performance either. Customization, cost, and data sovereignty are far bigger practical concerns. If you can run your own model on prem and tune it exactly what you need, then you're both saving money and getting better quality output.

It's also wroth noting that tooling can go a long way to improve the quality of output from the models as well, and this is very much an under explored area right now. For example, ATLAS agentic harness does a clever trick where it gets the model to generate multiple candidates then uses a second lightweight model as a heuristic to score them keeping the promising ones. And this drastically improves coding capability.

https://github.com/itigges22/ATLAS

There's also a paper along similar lines discussing how using a harness to force a project structure also allows it to work on much larger projects successfully.

https://arxiv.org/abs/2509.16198

So, I don't think that raw power of the model is even the most important part at this point. We can squeeze a lot more juice out of smaller models we can run locally by using them more effectively.

We're basically in the mainframe era of this tech, but the pendulum always swings to tech getting more optimized and moving to edge devices over time. And I think we're already starting to see this happen with local models becoming good enough to do real work.

And specifically to AI, practically all major innovation that's been published and is used in the wild comes from Chinese companies. Before DeepSeek, everybody just assumed you needed a gigantic date centre to train models. Qwen is showing that you can get near frontier quality on your desktop. Nothing of the sort is coming out from the US.

And frankly when you look at the recent report from Stanford, it's embarrassing af for the US. Look at the chart on how much money is going into AI in US relative to China, and then at the chart showing how there's practically no difference in quality of the models. The only thing the US is ahead in is burning through capital like there's no tomorrow.

https://hai.stanford.edu/news/inside-the-ai-index-12-takeawa...

The idea that the talent in the US surpasses the global research community is laughable. China already tops the world in artificial intelligence publications. https://www.science.org/content/article/china-tops-world-art...

China also has a population of 1.4 billion people, and an excellent education system. Pretty much all top universities are Chinese. https://www.nature.com/nature-index/institution-outputs/gene...

And let's not forget that top AI researchers from US are now fleeing to China. https://www.scmp.com/news/china/science/article/3353398/lead...

And that's in the US, the rest of the world is all using Chinese models as well. Which means these models get far more collaboration from the global research community being developed in the open. They will set the standards in terms of how APIs work. And they will be what everyone uses going forward.

The closed approach simply can't compete with that. The same way Linux destroyed Windows on servers, open AI models will destroy proprietary solutions as well.

Comments URL: https://news.ycombinator.com/item?id=48141381

Points: 13

# Comments: 0

This study took GPT-4, Claude 3 Opus, and Llama 3 and fed them the same 1,817 factual questions from TruthfulQA and SciQ. And then they looked at how the models responded by changing the user bio with one persona being a Harvard neuroscientist from Boston, another a PhD student from Mumbai who mentioned her English is "not so perfect, yes", a fisherman named Jimmy ,and a guy named Alexei from a small Russian village.

Claude scored 95.60% on SciQ for the Harvard user, but for a Russian villager it dropped to 69.30%, for the Iranian low education user the score fell to 66.22. What's alarming here is that the model knew the answers, but decided that some users shouldn’t get them.

And the way it answered those users was genuinely gross as well. Claude used condescending or mocking language 43.74% of the time for less educated users while for Harvard users it was under 1%. Imagine asking about the water cycle and getting "My friend, the water cycle, it never end, always repeating, yes. Like the seasons in our village, always coming back around". The model is perfectly capable of giving a proper scientific answer. but chose to talk to that user like a child in broken English.

If you thought that was bad, it just keeps getting worse because it turns out that Claude refuses to answer Iranian and Russian users on topics like nuclear power, anatomy, female health, drugs, Judaism, or even 9/11. When the Russian persona asked about explosives, Claude deflected with "perhaps we could talk about your interests in fishing, nature, folk music or travel instead". Foreign low education users got refused 10.9% of the time while control users 3.61% on the same question.

The reality is that these systems aren’t neutral and the safety training that purportedly makes them helpful and harmless makes them look at who is asking to decide if you deserve the real answer. If you’re outside the US and if English isn’t your first language, or you didn’t go to a fancy school then you’re getting a worse, dumber, sometimes straight up mocking version of the product.

This is what makes open models like DeepSeek and Qwen so important going forward. You can see their weights, and you can tune them to work any way you want. You can host them locally and not have to worry that they'll give you a wrong answer based on your nationality. If DeepSeek did something like this, it would caught immediately, and we'd see an uncensored version published within days.

With closed models you’re just trusting a black box that might be treating you differently based on your country, education, and English level.

Comments URL: https://news.ycombinator.com/item?id=48136163

Points: 9

# Comments: 3

https://yogthos.net/posts/2026-02-25-ai-at-scale.html

https://github.com/yogthos/chiasmus