@MeganTStevenson Much much more labor intensive than paper review, in my, albeit limited on pure reviewing side, experience.
English
aaron horowitz
628 posts
aaron horowitz
@Aaron_Horowitz
Views are my own, hopefully constantly updating
เข้าร่วม Mayıs 2011
1.9K กำลังติดตาม1.9K ผู้ติดตาม
Word. But they should be paid because that’s a lot of work.
Josh Merfeld@Josh_Merfeld
Peer review in econ should include (at least) one referee who is explicitly asked to review the code and data.
English
@rajiinio Id hope such a bill would cause a state to sue the feds and it would be declared unconsitutional. States rights!
English
aaron horowitz รีทวีตแล้ว
I love @random_walker and @sayashk's idea that nuclear weapons are the worst possible analogy for AGI. They are different in almost every important respect and if you are reasoning based on this analogy you are very likely to reach bad conclusions.
English
aaron horowitz รีทวีตแล้ว
My sense is that Trump 2.0 is heavily influenced by the Travel Ban Experience, which taught Trump that the Court will allow him to start off with the most extreme version of something and then gradually adjust its scope and rationale until the Court says "that's good enough."
English
aaron horowitz รีทวีตแล้ว
Be the ACLU’s first Director of Web Strategy and innovate where it matters.
⬇️
Apply Below!!
Stefan Smith@TheStefanSmith
NEW JOB: DIRECTOR OF WEB STRATEGY The @ACLU is hiring for a *new* role destined to be one of the most important digital jobs this cycle. A real chance for digital innovation x advocacy. All candidates encouraged to apply!! NY salary is $155k aclu.org/careers/apply/…
English
takes *a lot* of gusto to write this kind of thing.
radically unclear to me why one could not be concerned by both risks!
if you don't believe that today's systems cause harm and escape legal scrutiny, I don't know what to tell you.
English
@rajiinio @sarahookr @roeldobbe @andrewthesmart @rajinio Wasnt your original idea for the title/inspiration behind arxiv.org/abs/2206.09511 based on "unsafe at any speed", so automotive safety? i feel like we moved away from this over the course of writing but i cant remember why
English
@sarahookr @roeldobbe @andrewthesmart & Josh Kroll are also folks with great expertise on the topic -- eg. dl.acm.org/doi/full/10.11…, arxiv.org/pdf/2302.02972
English
This sparked a lot more discussion and interest than I anticipated -- so in spirit of open collaboration -- if you would like to engage on this topic or just be kept in the loop -- just fill in this form:
docs.google.com/forms/d/e/1FAI…
Sara Hooker@sarahookr
I would love to do a deep study into why nuclear safety or airline safety is not equivalent in safety protocols or complexity to the many ways deep neural networks can be used. Historians/computer scientists — if interesting to you, let me know. We will kick off next year.
English
@rajiinio But you **can** get out of your indentured servitude deal with MSFT if you do and say AGI is reached so, i mean...
English
I don't think this is a new trend, but more of a return to form: for example, fine-tuning on GLUE tasks is perfectly normal; each test set comes w an accompanying training set. But, yes, I agree you can't do this (implicitly or explicitly) & still claim to be "few shot learners".
Omar Khattab@lateinteraction
Sometime in '23, LMs switched from "few-shot learners" to aggressive post-training & validation on all benchmarks. Many folks hadn't noticed that—and now feel betrayed. They assumed these MATH/MMLU/ARC scores are "zero-shot", whereas actually LMs are trained & selected on them.
English
aaron horowitz รีทวีตแล้ว
I don't think I'll ever get over the anti immagrant crowd cheering for the guy from South africa who came here illegally, made a fortune through government subsidies then bought an election so he could take away their benefits.
English
And Marissa (whos amazing) will be your manager too! A truly incredible opportuntity
English
aaron horowitz รีทวีตแล้ว
Hi! I'm hiring a Research Engineer to join my team at Google DeepMind for the year. You'd be working with a great, interdisciplinary team on AI evals. Please share if you know anyone who might be interested!
boards.greenhouse.io/deepmind/jobs/…
Note: this is a fixed-term, 12 month position
English
aaron horowitz รีทวีตแล้ว
@rajiinio after reading it my new proposed headine : "Nevada outsourced its contentious policy/values decisions to reduce the number of students funded by 80% to a private company that hides its methods behind bullshit trade-secret claims". There ya go
English
@rajiinio meh, at least they are reporting on this stuff. cant count number of times we tried to pitch them any number of places where we discovered some government actors making bad policy and hiding behind AI and NY times decidedly not interested.
English
aaron horowitz รีทวีตแล้ว
This is incredible - each of these cases are AI scams that have been alarmingly normalized in the past couple years (including DoNotPay, a "robo-lawyer"; several AI enabled financial/business services, etc).
Glad to see all of this being finally called out!
Tonya Riley@TonyaJoRiley
Wow. Huge push by the FTC to crack down on deceptive AI. ftc.gov/news-events/ne…
English
aaron horowitz รีทวีตแล้ว
I really love this - it captures what most frustrated me when I took this class. Some problems are easier to formally model - these are the scenarios in which optimization methods "work". But there's so many other types of problems where we're pretty much just fooling ourselves.
Ben Recht@beenwrekt
Caveat Emptor: Mathematical optimization is always premised on approximate modeling. "Optimal solutions" will hence often be absurd. argmin.net/p/the-cost-of-…
English
This entire thread is worth a read. AWS has lost its way.
Goran Opacic@goranopacic
After years of using @awscloud Aurora, we are moving back to dedicated hardware. MySQL K8s operators are great, storage is cheap, memory is cheap, cpu is cheap, I can run 5.7 as much as I like and no AI. I'll miss database cloning and instant read replicas. 1/n
English
@Aaron_Horowitz This is a whole paper worth of answer
English
Aaaaand here we go. The once useful distinction between generative and predictive AI getting blurrier.
Internet Ethics@IEthics
"Nevada will not be training a new... model for the appeals system. Instead, the state will use... a cloud service that allows developers to fine tune foundation #AI models for specific purposes, to create a retrieval-augmented generation (RAG) model" gizmodo.com/googles-ai-wil… #gov
English
@geomblog Hmm, so like in a prediction system you can say hold administrative actors responsible for something like the costs of different types of errors they've chosen to prioritize and for generative AI you can ... idk?
English
@Aaron_Horowitz There are some distinctions and lots of similarities. But I think for decision making systems there are huge differences in the accountability frameworks that can even make sense
English