Hakuna Matata !!!

Last statement is too bold

Gemini Diffusion is super fast. 564 tokens/s. It means it can write almost a 200 page book in ~3 minutes.

If AI can, why shouldn't it take the entire software engineering jobs or a job of a research scientist?

We don't care about AI, AGI, or whatever for tonight, let's cook this chicken, let's go United🔥

We are back to AI

I hate Lofi and i don't really get them. But if you need something while working just check Mulatu's Jazz. He's simply the best.

Here are my favorites

Yekermo Sew
https://youtu.be/jwdBRqIsVUY?si=X-7T9QIUiMsO4a5a

Tizita
https://youtu.be/sXLfV2kegUI?si=cjZdWw_FKUXhmXLi

Religious benchmarks for LLM evaluation seems cool, I've not seen much work towards this. Are the best models of today biased against one religion, teaching, how would they interpret things.

Recommend me a paper if you've seen in this area, I'll be happy to read it.

Building LLMs from scratch has to be one of the challenging things I was in and very underrated. Pretraining data, how many parameters is enough, instruction fine tuning, making them generalize and alignment, all under resources constraints, even in big tech companies compute budget exists, this all is really hard.

So when ever a new model is out and they beat others on some areas is a huge W.


So one suggestion, if you don't have to, don't start from scratch and also expanding toknizer and updating model weights per user or something should be well studied to adapt models to new langs and tasks.

Big W, single author, simple and nice paper, look at that citation num

So I randomly discovered an Ethiopian cat Instagram account with over 100 followers(same as mine lol) and casually following a bunch of other cats and dogs. And an entire childhood documented online. Honestly, I didn’t think I'd ever be jealous of a cat's social life😂. Now I'm thinking turning our cat into the next big influencer.

Btw this is the level of creativity LLMs won't get😂

We got another AI Research oriented channel by Biruk. He's a Masters student at Paris Saclay and will share cool AI research papers, lectures, codebases, experiments and opportunities from around the world. Join his channel here.

@ethio_sota

For anyone interested in computational neuroscience.

Apply to Simons Computational Neuroscience Imbizo which will happen in Cape Town, South Africa. They cover all your costs to attend.

https://imbizo.africa/

every new model is a combination of two things, training algorithms and data

open-source algorithms are up-to-date. the things that work are usually fairly simple

but the data is complex, massive, gatekept, ever-changing. here open models are way behind, and probably hopeless

Source

Hellooo, what do you think about this? Will make the code opensource if people like to play around with it etc, maybe for anything interesting, AI integration, bible verses or anything. For now it just displays some random proverbs.

PS. it's for Mac

For people who love tokens, two great papers that came out recently.

Dynamic Chunking for End-to-End Hierarchical Sequence Modeling, i saw this yesterday and they created a hierarchical network (H-Net) that learns dynamic, content-aware chunking directly from data, and it replaces tokenization, finally :)


Tokenization is NP-Complete
This one is mostly a theoretical paper and they prove that finding the optimal tokenization (either by direct vocabulary selection or merge operations) is NP-complete, shows the inherent computational difficulty of the task.

In the past, most ML research was driven by academia and there was a circle from people at UoToronto, UdM, Stanford, Berkeley etc. Almost everything that came out or even most pioneers were here. Now academia isn't dead but lost it's place.

Now the circle is OpenAI, Anthropic, Meta, Deepmind, XAI, Microsoft Research etc. Most of the people at the academia circle in the above are running these labs now but with less freedom to explore ideas over time.

Here comes the downfall in the second circle, in academia people spend 5 years to make some cool things and they did, most of the things were initiated in uni labs than industry. But now everyone is chasing benchmarks and 1% gain over the SOTA is almost enough to lead the race. Ofc industry made so many great things, but the same way scaling is converging, it'll be for industry labs, unless they have teams to do foundational research that might take years, e.g FAIR(Meta).