MMitchell

insanely stupid gimmick here - an interview with amodei that the kicker reveals is AI generated. honestly I've no idea what the argument is for this: it's not novel or entertaining and seems to piss off both readers and writers. baffling vanityfair.com/news/story/dar…

@AmandaAskell @MatriceJacobine 💔

Perhaps I should get married again so that the media has a more recent man they can reference any time they mention me or my work.

Similar to publication venues, the world of open coding is now being inundated by slop. What solutions can we use here to address this new world issue?

@spectate_or My post was in response to evidence of reviewers not following policies.

@mmitchell_ai the paid versions don’t and the policies explicitly say which ones you’re allowed to use icml.cc/Conferences/20…

Unless I’m mistaken, common versions of ChatGPT and Claude use conversations for training—meaning they can pick up ideas in unpublished work and pass them on to other users; laundering idea plagiarism and creating “scooping” dynamics without any of the people involved knowing it.

If we pinpoint “idea” as “single training example”, then yes (pointers to more refs in nature.com/articles/s4159…) But beyond that: I’d guess conversational content is upweighted, or takes some other high-value form. It’s post-training to capture actual users’ conversational usage, so would be sort of silly not to treat that like gold.

@mmitchell_ai Is there evidence that this could actually occur if an idea was only included as a single sample in the LLM's training data? At least for text-to-image models I know there's some work showing that concepts have to be seen hundreds or thousands of times in web-scale data.

@natolambert Hanna did great! She remarked on the hallmarks of privacy, security, and control. Everyone seemed pretty aligned that open = good. The VC fellow remarked on how fully closed systems mean you have to delegate trust, as opposed to fully open, which you can in[tro]spect.

Any good quotes on the Nvidia GTC open models panel? Maybe they'll invite me to one some day 🥺

@rabbitandtheAI It is on for many common versions by default and most people do not pay attention.

@mmitchell_ai I’m not sure I understand the concern then. It’s in the user agreement to turn it off.

@rabbitandtheAI Yes, turning off the option presumably counteracts this.

@mmitchell_ai I am not certain that’s accurate if you toggle off the option for use my data in training

If I was going to claim that a finetuning methodology for machine unlearning “really worked,” what evidence would you like to see?

Someone built a Google translate for Linkedin 😭

I joined Hari Sreenivasan on CNN International and PBS to discuss the use of AI in warfare and the impacts we're already seeing of this fallible technology being used in Iran, and how it ultimately obscures accountability. Full can be found at youtube.com/watch?v=w16fT3…

If you live in the US and have worked or are working in data annotation for platforms like Outlier, Handshake, Mercor, or others, I'd love to hear about your experience. Please drop me a line: karendhao.com/contact.

@JAldrichPL I keep pondering whether noting the type and level of assistance is useful. To what ends? Perhaps there can be different tracks for different levels of assistance…

So apparently AI use (mostly undisclosed) is becoming ubiquitous in ACM paper submissions. I don't think this happens in my group so I was surprised to hear it. But if that's actually the case, perhaps requiring disclosure is pointless. Thoughts?

As a non-prof with way too many research objectives that would ordinarily be shared with students, I am fascinated by how agent-as-student would work. Agents can’t take accountability for work like students can. If I tell my agent my research idea, and it executes it, I can’t have it co-author. Very curious to learn more about this.

I definitely dont think so, i know lots of top faculty, probably overlapping with those Sayash has spoken with that I decidedly hope do not retire... In fact, this view includes some of the very best AI faculty in the world in my view. This view is a reaction to a profound technological shock, it is entirely unclear if this view they currently have will be sustained, which is part of Sayash's point. Why should every CS or AI professor have the same prioritization of students vs productivity vs other objectives? I dont really get why we need that homogeneity: I think what people are saying is perhaps not consistent with why I would be faculty, but it seems well within the space of reasonable views to me. Sure, maybe I want to avoid faculty wholly dispassionate about their students and their growth, but what Sayash and I said is well short of that.

I'm confused by the very premise of this question, though? My PhD students are not code robots I command to implement what I want. They are human beings I am mentoring to become independent researchers

In a meeting at the office one time, this dude fully repeated something I had said almost word for word, just slightly changed. After he finished, my coworker raised her hand and said, “I don’t have anything to add, but I just wanted to point out he basically repeated exactly what Natalie said like it was his own idea.” Happy Women’s History Month to her.

🤖 In policy? Thinking about the definition of "AI"? Led by @AspenDigital, a set of us at intersection of AI/ethics/law/policy put together this resource on the lineage of policy "AI" definitions, what they're getting right, what might be improved. aspendigital.org/report/definin…

@binarybits I'm with you!

People have started abbreviating Anthropic as "Ant" and I don't like it. The th is a single phoneme. You shouldn't split it up like that.

“you require us — the people harmed — to put brakes on the harmful things YOU are doing” 🔥 piece. moryan.com/an-open-letter…