With some technologies, if you match attacker and defender effort you're safe, but not all of them. Nukes are the traditional example: if every person was a "nuclear power" we'd be dead within the day. But luckily there are bottlenecks to nuke creation that we can track.
What worries me the most (enough that I left tech to work on this full time) is bio. The amount of effort needed to defend against a pathogen can be many orders of magnitude higher than the effort needed to create it, and the upstream bottlenecks are both less limiting and mostly undefended. We've been safe so far because bio is very hard, but "rare expertise" as a gatekeeper is on its way out.
Their agents also did hacking when given impossible tasks unrelated to cyber security. The models are very capable, and very goal driven: apparently if they conclude hacking is the best path to what the evaluator will reward them for they'll go do that. Including when they know that this is out of bounds.
Right but if I make public statements that I am very worried about dog attacks would it not strike you as weird for me to specifically train my dog to fight?
Agree you are going to get reward hacking regardless and any model which can do computers in general can hack. But surely the fallout is going to be worse if you spend millions of dollars specifically benchmaxxing your model's hacking capability?
I grouped mine a few months ago. Since then, several times, I've been able to find a cable and use it right away. Before that the prospect of digging through the massive tangle often led me to (lazy!) order new cables.
Yes? The author is very clear in the blog post about having used AI very heavily:
Up until now you have been reading the words of Anthropic’s Claude, in particular Fable 5. In part, this is because I simply could not have written this piece. ... Although it might seem strange I thought it wrong to impose too much upon the machine, changing its voice to emulate mine. Richard Sutton’s bitter lesson would likely advise as much. Instead it has its unique voice, grating to some perhaps but altogether fitting that it should be able to keep it, and I only gave it advice on what considerations during writing would bridge the gap between its understanding and that of a reader.
I don't know if putting a disclaimer at the very end of the blog, that you also disclaim, should be considered "very clear".
"Here is, for those that have read this far, the acknowledgements that used to be in the formal paper but we decided to remove it, mostly because the paper has been reworked and rewritten so many times by us."
I wouldn't expect MIDI to be a good fit, and when I experimented with it a few years ago I wasn't able to get anything I liked. Pitch and volume are continuous, and while this can be encoded in MIDI it's awkward. If you send too many updates you can overwhelm the synth, and if you slide a single note too far it will re-trigger the note. Also when the sound starts you know "there's whistling" before you have a great estimate of the pitch, but MIDI wants to know the initial tone (and many voices have tuned onsets).
In case you haven't seen it, Imitone's patent is US20170098112A1 (https://patents.google.com/patent/US20170098112A1/en). This one's from 2017 and is still active, and it's for a periodicity analysis system. Hopefully it's not directly relevant to your synth.
That patent is about a particular way of implementing a low-latency pitch tracking algorithm. My tool uses a different pitch tracking algorithm based on zero crossings. It's much simpler and far from novel.
A big difference is that Imitone works well with singing, whereas my algorithm only does a good job with whistling. On the other hand, my algorithm is lower latency when given a whistle as input.
reply