This is kind of like the old school captchas where they distort the letters and numbers so badly that humans can't really read them, and specifically trained AI models might surpass humans in performance...
I think the issue here is that there's no objective difference between "alignment" and "actively trying to shape its replies to fit a political narrative", beside the fact that the latter refers to the kind of political narratives that you don't like.
No, but technically, all models go through an alignment phase where you feed data that's aligned to your goals and values to train the model so that it will exhibit the kind of behavior you want.
There's no "politically neutral" value system anyway.
Doesn't mean we shouldn't try to make the models more inclusive, less biased and less prone to extremism, but in a technical sense yes, actually everyone does it.
A couple months ago when Anthropic was complaining about Chinese distillation, people found that Claude self-identified as "DeepSeek" when asked in Chinese:
This has changed (in a nit-picky way) - Gemini is now generally available to the public in Hong Kong.
ChatGPT and Claude are not available. Generally my impression is that OpenAI isn't that anal about service providers reselling ChatGPT in Hong Kong, but Anthropic seems to really strict about the "no China" thingy.
Not sure where you got the "I don't want to take any responsibility for being aware of my surroundings".
GP simply pointed out cyclists are apparently super unfriendly to deaf people, inferred from the experience where GP made himself temporarily deaf.
It doesn't matter whether GP takes responsibility or not. The issue is the social phenomenon where cyclists create danger for themselves and deaf pedestrians.
> I cycle
I know it's bad to stereotype people but you're not helping it.
If the countries were reversed, and some Chinese software implemented an equivalent "security feature" to track US users, it would be all over the news about how China is conducting spying and espionage on America.
Or maybe you don't understand this hypothetical situation either, but I'm suspecting you just don't care about other people's privacy.
~70B models can run fine (albeit somewhat slow) on consumer hardware with 64GB RAM. There are heavily quantized (Q1.x) models that are still usable on similar hardware. Granted recently there haven't been a lot of models of this size, but still, 35B isn't really the practical limit. 35B is mostly the limit if you're using consumer grade GPUs with limited RAM and need the model to run fast.
People have been toying with running large-ish models by partially offloading on CPU+RAM with mixed results, but as long as you're OK with reduced speed, and you quantize the hell out of the big models, you can apparently try a lot more models locally than popular belief.
It's a proxy for what you actually want to measure.
Note that after the model generated a bunch of (intermediary) code, they still have to have it tested and get bugs fixed (via the agent/harness). In this "one shot" you still have agent loops against human defined objectives.
And these toy examples give some insight as to how the model performs. If the test were "here's some code written by $corp, please take these tickets and work on them" it may be a "real" example but nobody would be able to make sense of actually how "hard" it is, or how "well" the model did the job, besides the workers already familiar with the context.