Andrea Volpini

We found Google's list of twiddlers, the programs that rework search results after ranking: 120 of them! 111 names are exclusive, among them AI Mode's fan-out plan, a filter still named after Bard or an Expertise twiddler think.resoneo.com/twiddlers/

LLMs accumulate a lot of declarative knowledge. That's how they can answer a large number of questions. That's also why they require so many parameters and hence so much memory. But one can replace the accumulation of knowledge by *understanding* That what JEPA world models do. That's one reason they can be useful with considerably less memory and training data than LLMs. This paper merely shows that the underlying similarity structure of the data, whether text or image (e.g. co-occurence of words in text or objects in images) uniquely determines the shape of the representation point cloud, when the encoder is trained trained with a common anti-collapse criterion like SIGReg or VCReg (not by reconstruction).

On this day 32 years ago, in 1994, Nirvana's "MTV Unplugged in New York" was released. It is considered the greatest Unplugged performance ever made.

Oh, so you asked an LLM for its favourite number and based on their training data it's 7? Well it's a little more nuanced than that: dejan.ai/favourites/

Are you an AI slopaholic? Read this before it's too late: thomasjiralerspong.substack.…

Reading and processing. aimclear.com/prompts-for-sal…

Here's an example of Connected Apps working from within Search or the Gemini App. Users can add apps they trust for Gemini to use via MCP and then that experience simply becomes a part of the conversation. If you missed it, yesterday Google announced the Connected Apps Platform which will allow any business to create your own MCP enabled app that ppl can use in Search and Gemini.

Our TPUs just phoned home... from space. ICYMI, on October 1, we sent four TPUs into orbit to test whether our AI hardware could operate in space. The initial goal was to see if our chips could turn on, process a prompt, and transmit a response back to Earth. This week, we got our answer. ✨ This milestone brings us one step closer to understanding if space-based compute is possible. We’re sending two more TPU-equipped satellites to orbit in 2027 for our next milestone. Watch this space 👀 and learn more about the first milestone of our long-term moonshot with Project Suncatcher ↓

Skills are so much more powerful if you start to think of them as memory/brains not static files

Seriously, this is such an insane release, I did not expect this: Kimi K3.1 Pro beats every listed US frontier model on 6 of 14 benchmarks in Moonshot’s evaluations. At $2 per million input tokens and $10 per million output tokens, its standard token rates are 50% below Opus 5.5 ($4/$20 !!) and 80% below GPT-6 Astra ($10/$50!!) In other words: SOTA much cheaper, local running (with enough compute)