LLM-isms are much less prevalent in base models, which is what GPT-2 was. It had significant problems with maintaining coherence, but GPT-2 generated text did not have the obvious tells of today's LLMs.

They can certainly enforce that you answer the survey. But it's very difficult to enforce a requirement that people answer questions accurately, particularly when they perceive that doing so will expose them to danger.

I don't get what danger is being referenced here that exists only if the data is released to the public (in aggregate)?

The government is the primary and arguably only source of the danger, and they already have most of the data whether you answer the ACS correctly or not.

I'm skeptical that you're going to be able to reliably exfiltrate ~10TB of model weights using TEMPEST. Which is not to say weights are secure, just that this isn't the threat model I would be concerned about.

This is not legislation.

Uncle "Sam" is ironic here, alternative man one might say

Come on, no one was worried that GPT-2 would help people engineer viruses. The concern was generating misinformation and spam.

Moolenaar's quote: "The AI models these companies use are trained by China’s censorship regime and introduce hidden vulnerabilities that put Americans’ data and businesses at risk." That is, Americans using Chinese-trained AI models are exposed to some form of cybersecurity risk.

That's not really a threat model described in either of the Anthropic posts you share, which mainly talk about the risks of allowing authoritarian regimes to use powerful US-trained models, and the geopolitical risks of authoritarian countries developing strong AI before democratic/liberal countries do.

Word was originally released for the Mac in 1985, so the deal was not that Office would be ported, just that MS would keep developing Office for the Mac.

This is a neat little trick, but I wonder if you could do substantially the same thing by just prompting/LoRA finetuning the model to produce a single-token output ("yes" or "no"). This only requires a single model forward pass, you can use the same KV caching strategy for shared parts of the prompt, and isotonic regression should work just as well to calibrate the output logits. I guess if you use this method and probe on an internal layer you can skip all the remaining layers, which could be a nice inference speedup.

> you could do substantially the same thing by just prompting/LoRA finetuning the model to produce a single-token output ("yes" or "no")

You could probably achieve this with logit masking. Or equivalently, comparing the "yes" vs "no" logprobs in the final dis-embedded vector.

I appreciate the extremely low fuss interface, but I'm always a little disappointed by chord progression ear training that just plays triads one after another with no thought for voice leading. Generating a nice voice leading for an arbitrary chord progression is a little tricky to do automatically but far from impossible, and might be a fun exercise either for you or your favorite LLM.

Using only 3/2 ratios can sound pretty bad in just intonation as well. Major thirds tuned to 81/64 are off (by a ratio of 81/80) compared with the standard 5/4 tuning, and they don't sound great. This difference is called the syntonic comma and it's been a major issue in the history of tuning.