“Microsoft increased worldwide water consumption by a whopping 34%—up to almost 1.7 billion gallons annually—last year, which … is most likely due to increased AI training. That's dwarfed by Google, which used 5.6 billion gallons last year, a 20% jump.” https://t.co/9koIENoSkR
Every once in a while when discussing AI regulation I point out that the FTC has the power to force a company to delete an algorithm entirely.
They just used it for the fourth time. 3/4 are due to the provenance of the data used to train an AI (NOT a LLM or Diffusion model)
Is OpenAI following the law that exists *today*? Unlikely.
Let’s lose the focus on building “AI policy” from a blank slate. Look to data protection, consumer protection, copyright & competition law to capture broader concerns + take advantage of levers we already have (2/4)
I'm excited to share our recent paper on "Ɛ KÚ [MASK]: Integrating Yorùbá cultural greetings into machine translation" that recently got accepted at AfricaNLP @ #ICLR2023 (non-archival) & C3NLP workshop @#EACL2023.
paper: https://t.co/NpGe00qOfU
Project led by @alabi_jesujoba
✨ day 2️⃣ update ✨
we still have $750 left to pool up for C! we need everyone to pitch in so we can make sure that she and her kids can keep their housing!!
give $5 to the pool to keep the momentum going!!
v*nmo: deejayayers
c*shapp: $bomi97
p*yp*l: pchangepools
Some background in case you're wondering what this is all about:
Indirect prompt injection is when an LLM is asked to analyze some text on the web and instead starts to take instructions from that text. Right now there's no good way to defend against it! https://t.co/bXjjqd7Fh4
"We almost never use this"
"Ok, but we only use it in emergencies"
"Fine, also other times, but then we get court orders"
"I mean, we get them sometimes, when we feel like it"
Next step: "we constantly use them, that's life, why are you complaining"
https://t.co/1G3Qugmg2p
@rachelmetz absoLUTEly not and to be honest I'm flabbergasted that the Bing rollout happened at all: I see it as a serious failure of technical leadership and the erosion of public trust over the last couple days is well warranted
it's been a week since I launched my discord for Black folk in DevRel.
They've shared some great content they made in the Discord, and I wanted to bring more exposure to their work!⬇️
Hello friends! Big announcement: I won't be streaming or releasing videos regularly this year.
(Ko-fi supporters, 1) thank you! and 2) there's a post on my page already up for you. Appreciate yah! 🥰)
Happy new year! I recently talked with @katriadoodles about artists, AI art and how we in the ML community can do better & it's now up on my YouTube channel. https://t.co/RrY6KfSyo6
VA-04 friends, PLEASE make time to vote in the firehouse primary today (I'd personally strongly suggest for Senator McClellan). Polls are open until 7, and anyone currently registered to vote can cast a ballot.
@andywalters @CohereAI I would counter that it is directly serving raw generated text to users that is feckless. There are no known mitigations for entire classes of adversarial attacks on LLMs and ignoring the risks that poses to both users and business is, in my opinion, professionally irresponsible.