Y Combinator 4. Wildmoose: founded by Yasmin Dunsky, Roei Schuster, and Tom Tytunovich. The mooose helps with debugging in production to solve issues in minutes instead of hours, reducing MTTR 100x. It helps you avoid costly downtime, save time, and ycombinator.com/launches/I29-w…

If you're using an LLM in any meaningful capacity, your prompts probably include untrusted inputs, and the models' outputs probably affect the control flow of your program. This means prompt injection is something you have to be keenly aware of. This is not a theoretical threat.

I am thrilled to announce that I will be joining @CISPA as a faculty. I am looking forward to lots of collaborations with lots of outstanding researchers over there.

When do we admit NeurIPS got too big? Not just in size but also in scope: as far as I see, almost *any* practical CS paper nowadays can be framed as a NeurIPS contribution. For the cost of getting totally random reviews and virtually no boost in visibility. Something is off here.

Reflecting again on how knowing all the architecture & equations of the Transformer model is really of no use at all in convincingly explaining to someone how an LLM like ChatGPT can write paragraphs of lucid text in response to a prompt. I guess I’m saying “Beware reductionism”.

is this a new “we lost the rigor” cycle? post-2017 CS grad students who “only did engineering not science” or “just tried things until the benchmark gave” are doing quite well today last I checked.. I think those today who are “only doing prompting” will do okay, too

Rik van Noord It’s possible that a problem exists and has no viable solution. One thing I’ve learned about the NLP community is that this fact makes NLP’ers extremely uncomfortable. Other CS communities, being somewhat more humble, can accept what they cannot change ☯️

Lord, give me the confidence of a legacy systems engineer explaining why there’s no way you could solve something as complicated as *their* industry with newer technology.

Techniques for fine tuning LLMs to "bake in knowledge" is a pure, nicely encapsulated technical research problem, with far reaching implications on the software industry. Does anybody know what's the academic NLP community's input here?

with all the recent attention to "long context", Mosh and Alon went to find out "does it really work?" and in particular in the case where you want to not only *locate* separate pieces of information in the context, but to also *reason* over them. Turns out, not really.

Survey: when onboarding a new dev, what types of things *about your system* do you have to teach them before they can get to work? Not asking about general technologies (e.g. "JavaScript"), but about your own app and infra.

!

ChatGPT o1-preview is scary good. This is 100% how a human reasons.

And that's a wrap! Thank you for everyone that attended our NeurIPS Conference workshop on Red Teaming GenAI ☃️ 1/5 🧵

I was asked what's new since 2022. Zcash and Aleo are on major US exchanges. Other credible Zcash-inspired chains like Penumbra and Anoma are live but lack US listings. Unlike Zcash, they support stablecoins. Will Zcash add them or remain the 'Bitcoin of privacy' is the question.

Google announced they will support privacy preserving age verification via zero-knowledge proofs. You prove you have a signed digital copy of a drivers license and it says you are over 18 without revealing anything about you (name, birthdate, etc) blog.google/products/googl…

אוקיי היה נחמד אפשר את o3 בחזרה?