AI Hackers, Weird Watch Phenomena, & More

TLDR;

This week, we have a lengthy list of important news worth covering. From Google’s landmark release of AlphaGenome, an AI super hacker, and the rise of model merging, to Replit’s insane revenue growth and Zuckerberg’s shopping spree at OpenAI.

Finally, we also examine new products like Gemini CLI, the unusual behavior of AI models with watches, and how video tools can revive memories.

Enjoy!

HEALTHCARE

Google Open-Sources AlphaGenome

Google has released another DNA-based model, called AlphaGenome.

The system ingests stretches of DNA up to one million base pairs long (meaning the model’s token granularity goes up to the nucleotide level, with every three consecutive nucleotides giving you an amino acid, which are the basic construction units of proteins, essential for life).

It simultaneously predicts thousands of regulatory properties, such as gene expression levels, chromatin accessibility, 3D contacts, splice junction usage, and more, at single-base resolution.

The advance matters because most disease-associated variants lie in the 98 % of “non-coding” DNA whose function has remained opaque. For reference, Nature’s news coverage hails it as a way to probe the genome’s “dark matter”, likening the leap to what AlphaFold did for protein structure. And let’s not forget that work led to Nobel Prizes.

TheWhiteBox’s takeaway:

Before we answer what the primary goal of AlphaGenome is, let’s first understand how it works because, surprisingly, the similarities with ChatGPT-type models are uncanny.

In fact, they are, in essence, the ‘same’ model just applied differently, or at least both are based on the same algorithmic principles.

AlphaGenome, like ChatGPT, is a sequence-to-sequence model that maps a sequence of input tokens to a sequence of output tokens. Here, tokens are the base-pairs in a DNA genome, while in the case of ChatGPT, the tokens are words/subwords.

Thus, tokens are the simplest unit of semantic information that the sequence carries.

Just like words carry a ‘meaning’, amino acid base-pairs do too. But just like words share information with each other to build the sentence’s meaning, base-pairs do the same. Put another way, the order and presence of these tokens affect other tokens, just like ‘red’ in ‘the red carpet’ affects the meaning of ‘carpet’ (makes it red).

Hence, what AlphaGenome does (in the same way as ChatGPT) is to make these base-pairs talk to each other, as to find patterns such as ‘when this base-pair follows this base-pair, this happens.’

In the case of AlphaGenome, ‘this happens’ could refer to illnesses and other aspects we are very interested in understanding.

This is why Google describes the model as a ‘sequence-to-function’ because by understanding the properties and dependencies of the genome sequence, it maps them to biological functions.

There’s a reason AlphaFold gave the AI industry the first-ever Nobel in Chemistry (to Demis Hassabis, Google DeepMind’s CEO). Now, AlphaGenome tries to deepen our understanding of a new area of biology and, with that, opens the door to curing many illnesses. And who knows what else.

CYBERSECURITY

XBOW’s AI Surpassed All Humans in HackerOne

Another day, another ‘job’…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.