New AGI definition: AI whose errors are no longer detected by human review of the output.
Mihai Christodorescu
- 76 Followers
- 155 Following
- 404 Posts
searchable
Nicholas WeaverAn interesting thing on EVs. For the rest of the country they are cheaper than a gas vehicle per mile.
But here in CA our electricity rates are SO outrageous that a gas burner can be cheaper per mile!
@krismicinski @shriramk @gwozniak @cross @steve @jfdm @csgordon @lindsey @jeremysiek Qwen models are quite good for coding, and they run decently on desktop-level GPUs.
scyIt's nice that music player software has "play next", i.e. "insert into the play queue after the currently playing track".
Now, can we please also get an extension to the pause button that does "pause after the current track"?
I don't know about you, but I'm fairly often like "I want the music to end, but at a natural moment, not now in the middle of the track", but I also don't want to camp in front of the player to push pause at the correct moment.
@krismicinski Credential gatekeeping is certainly one way to do ad hominem attacks.
@regehr I didn't see this paper mentioned, may be relevant:
https://www.ndss-symposium.org/ndss-paper/auto-draft-206/
https://www.ndss-symposium.org/ndss-paper/auto-draft-206/
Re: https://infosec.exchange/@mttaggart/115860061627547920
Privacy puritans are out in full force today.
Taggart (@[email protected])
Let me reiterate: ChatGPT Health is a dangerous product and a terrible idea. Keep it away from everyone you care about.
@rene_mobile Economics has lots of solutions to deal with externalities without assigning moral value to a transaction. So one can imagine a system of taxing the resource usage of (and thus increasing the costs of and reducing the demand for) LLM inference, regardless of whether the LLM inference is silly or stupid. Note that resource usage is a completely orthogonal problem to that of malicious LLM applications.
@krismicinski Re: "AI will write almost all code"
This is likely to be true, given that the vast majority of software is boring, run-of-the-mill data movement and formatting, typically hard to specify and verify formally (mostly because it solves underspecified problems). For every computational fluid dynamics implementation there are a billion `convert2JSON` ones.
@rene_mobile Trying to stop AI hype seems like a hopeless exercise. And frankly I do not see how stopping AI hype would help stop bad uses of GenAI.
The assumption I'm moving forward with is that bad people will use AI for bad purposes for sure, regardless of how much or little of the AI hype is realized. Thus we need to come up with ways to stop or at least mitigate those bad outcomes.