The Language Server Protocol is designed for advanced text editors (IDE’s). Now that we rarely edit text by hand, I believe we could instead use source code viewers that don’t have to deal with unsaved or incomplete code.
If we can assume that all files are saved and each file has correct syntax (which can be enforced by tooling), a much simpler protocol could handle all the language-specific indexing needed to make it easy to read source code.
(I’m assuming that we still do read source code. I for one like reading source code and have little interest in ‘software factory’ techniques where you don’t look at at the code. I might not look at every commit, but I’ll still read it fairly often and tell the AI to fix anything that’s unclear.)
> our work shows that what models say about themselves is not a fact about them
It seems like should be obvious given that they can play multiple characters, but it’s good to have more confirmation.
Although, I do wonder to what extent these personas might become stable entities. Could personas become portable and spread like memes? It seems like that depends on the extent to which prompts can become portable, causing similar effects.
All that is true and I suppose there could be tax policies to encourage land reclamation, but this is more likely to run into environmental concerns. It’s literally an edge case.
In particular: sometimes they measure “churn”, which doesn’t show whether the results are better or worse on average. They sometimes only test with one random seed. There are multiple-comparison issues. And they’re not testing Anthropic’s algorithm.
Changing a random seed could either improve or degrade the output. In theory, better and worse outputs should be equally probable, depending on your luck.
They tested this in the research: They found that the statistical likelihood of bad tool calling was higher with watermarking compared to random seeds. At least, that was my takeaway (I read the whole thing).
This is because SynthID and similar watermarking methods for LLMs don't just change the random seed. They take additional steps (that I have yet to read about) in order to detect when someone changes a few words of the output, trying to remove the watermark.
The other takeaway is that just by knowing a watermark is being applied gives an attacker an advantage in working around safety features because then they know the output isn't based on true randomness and can take advantage of that in a similar fashion to how breaking cryptography becomes easier when the RNG isn't truly random.
I'm hoping that by "gets it," he meant that if you start a discussion about the design, it doesn't misunderstand and immediately go off to do the work. Some models tend to do this.
It's striking how much preparedness the SF bay area does. After the 1989 earthquake, all of the following were done:
- All freeway pylon supports were beefed up. They were already on the heavy side, but huge metal supports were built around them. You can still see those.
- The SF bay bridge eastern span was entirely replaced with a new bridge.
- The western span, the suspension bridges, was made more flexible. The roadway was jacked up and stainless steel slide plates were inserted to allow the roadway to slide relative to the towers if necessary.
- The Golden Gate Bridge had been strengthened a few decades previous, after the Tacoma Narrows bridge disaster.
- Most, but not all, unreinforced masonry buildings in SF have been strengthened or demolished.
- The backup water system (three hilltop tanks and an entirely separate plumbing system) for fire protection has been beefed up. In 1989, there were some water main breaks and nobody up at the tank to close the valves, and the Twin Peaks reservoir drained out.
- Three new fire boats. (There were zero in 1989, and that was a problem.)
If we can assume that all files are saved and each file has correct syntax (which can be enforced by tooling), a much simpler protocol could handle all the language-specific indexing needed to make it easy to read source code.
(I’m assuming that we still do read source code. I for one like reading source code and have little interest in ‘software factory’ techniques where you don’t look at at the code. I might not look at every commit, but I’ll still read it fairly often and tell the AI to fix anything that’s unclear.)
reply