1/6 Introduce MoCa, a new method for continual pre-training of multimodal embeddings! 🚀 MoCa is the first to effectively scale with unlabeled interleaved image-text data, marking a paradigm shift in multimodal embeddings. Paper, code, & checkpoints! 👇 #AI #Multimodal #ML #NLP

"Chromebooks are the new Baby Einstein – an untested educational panacea with laughably dubious evidence of any positive outcomes." Putting computers on students desks was a terrible mistake. By John Allen Wooden johnallenwooden.substack.com/p/ok-google-ma…

"In spite of decades of reform and billions spent on improvement initiatives, it’s not clear that teachers are any more effective today than they were fifty years ago." Pleased to share this important piece by Nidhi Sachdeva, PhD and Jim Hewitt with readers of my Substack.

But the odds that a criticism that you came up with after having thought about homeschooling on and off when you see a tweet about it will resonate, be useful, be meaningful or even be literally true are just not good.

'you're not qualified to teach your kids' I'm familiar both with the large scale literature on homeschool outcomes and on the actual test scores of the homeschooled/alt-schooled kids I know. There just isn't a productive conversation to be had here until you acknowledge that.

I couldn't believe whisper was SOTA and then found out there is actually a better model from nvidia (WER around 6 vs 9-10 for whisper)

I built this during my PhD! you can embed *documents* using the OpenAI embeddings, do arithmetic on them, and then decode them back to text e.g. E = emb(“Elon Musk bought Twitter in 2022”) - emb(“Elon Musk”) vec2text(e) = “Someone bought twitter in 2022”

Interesting post. However, it seems to be in conflict with the most central problem in theoretical computer science: P vs NP ,which is exactly the question: is it fundamentally easier to verify a solution rather than solve a problem. Most people believe that verification is

It is a major policy failure that the US cannot accommodate top AI conferences due to visa issues.

I've got a newsletter going out tomorrow about screwworms and it got a little spicy

Casey Handmer "Why would they be suppressed?" Society has a lot of ways to award social status and if you publish rankings which fail to agree with existing ladders you will be regulated, in both an informal distributed fashion and then, if you are actually successful, more formally.

If you are not constrained on organizational complexity, if meetings with yourself are free, then a lot of the standard stack that BigCo uses is both overkill and underkill at the same time.

Is it "bad" that everyone is distilling from / training on Chinese models? While not directly bad, there is a large soft power component. Many completions that soapbox about Chinese socialist ideals / PRC values that filter into future AI models / spread all over the internet.

going through GitHub issues, it's clear ds*y is meant to be a programming model, not an optimizer. and it's decent at the former, but not so great at the latter. I guess I had such high expectations for the latter that I was bound to be disappointed lol github.com/stanfordnlp/ds…

eugene Appreciate you sharing your take. Seems like no one on my feed is willing to share an even partly negative take on this library or write out the full abbreviation when they do for fear they’ll summon those that disagree. While I don’t have the good fortune of meeting all that

my thoughts exactly

reasoning models are great but man it's confusing if you're not an expert o3, o4-mini, gp2.5, ds r1, grok 3 mini, grok 4: always think, you can't disable it sonnet 3.7, sonnet 4, opus 4, flash 2.5 lite: only think when you enable it flash 2.5: thinks by default but you can

Cassandra Team at Apple is searching for a fresh grad / person early in their career to join our ranks in SF/Bay Area! Come work on super interesting problems with world class team. Help us build better Cassandra! Ping me if you’re interested! jobs.apple.com/en-us/details/…

So basically you're free to say anything you're not forbidden to say. Right.

Thinking about this observation from tylercowen again: