AI Digest #10
Good morning, ladies and gentlemen. Let's start today with two hot news...
Era of 1-bit VLMs: All major language models have 1.58 bits
The era of 1-bit VLMs begins. To understand what's happening, I highly recommend reading the link. In short, MS introduced BitNet b1.58, which allows representing model parameters in the form of {-1,0,1}, reducing the required amount of GPU memory (it's simpler than dragging along float values), and should positively affect the efficiency of new models. As a nice bonus, we get reduced energy consumption and possibly the next step in hardware development specifically for these models with additional optimization.
EMO
Meanwhile, researchers from Alibaba Group have introduced an absolutely wonderful model for animating faces with "emotions" as the name suggests. You can see the quality at the link, but I liked it. One of the pros is that this thing really impressively animates the entire face at once. Cool!
To other news:
In France, a scandal is gaining momentum regarding MS investments in the Mistral model developers. Europe in general and France in particular are concerned about the growing influence of the States in the AI sector. Although the French government claims they knew nothing and it's all just business.
Exclusive: Montana AG claims Google Gemini has 'political...
The passion around Google does not subside, this time Gemini is noticed in political preferences. In general, since the model is trained on human materials, I wouldn't call this strange, but the fact is unpleasant, a model that promotes political narratives is so-so.
Tech News: How ChatGPT Could Influence Voters and Candidates
Especially in light of the news that ChatGPT could have a similar influence. We talked about this in the podcast a year ago, but now this fact seems to become obvious to everyone. Especially against the backdrop of the approaching large number of elections around the world this and next year.
AI-Based Glaucoma Care - Glaucoma Today
Very good news from the world of medicine. AI helps in the diagnosis and treatment of glaucoma. Considering that this is a quite relevant disease, we are waiting for something good from the world of cancer treatment.
Tim Cook says Apple will 'break new ground' in GenAI this year | TechCrunch
Tim Cook stated that Apple will finally release something with AI this year.
To tech novelties:
Titan Takeoff Inference Server | Haystack
Haystack invites to use Takeoff inference server for local work.
AI platform behind new supermarket carbon footprint model...
3.6 million pounds for the carbon footprint in supermarkets.
Google Colaboratory
Test Notebook for comparing Gemma and Mistral.
Airbyte loader in Langchain. If someone doesn't know, Airbyte is such a good ETL tool with an open source version and support for a bunch of sources.
New models for reranking
According to the developers' statements and synthetic tests, they look very serious. Whether this is the case or not, maybe it's time to change Cohere rerank and bge-rerank to something new
——
Don't forget to donate to the Armed Forces of Ukraine (verified donations):
Continuous collection for the strike drones company "Taistra" of the 10th Special Forces Brigade, everything for their drone workshop: antennas, control remotes, repeaters, batteries for fpv and bombers. Don't forget to check out the video surprise for a donation on MonoBank from 150 UAH.
——
Support the project with:
Like, repost, and money: https://www.buymeacoffee.com/dkuzin
That's all for today.