On feeling rusty, DNA and random ML facts
Weird, but gotta start again. At least with the digest-style ones...

Search for a command to run...
Weird, but gotta start again. At least with the digest-style ones...

No comments yet. Be the first to comment.
Why investing in mechanism design can prove worthy

Simple desires beget simple utopias. So, I’ve recently completed the AGI Strategy course by BlueDot (highly recommended, some of the best peeps lurking there), and some things picked at my mind… First: will we be able to get to the said utopia at al...

What won't we do to feel better

My take on the nature of change

, or how I learned to love oh-so-big graphs

Teleogenic
61 posts
I specialize in a very specific type of everything.
Smaller posts - how many times I’ll promise myself those?
So, neurosymbolic stuff seems to be taking off. Or not - this time, at least.

After AlphaFold’s success, Deepmind is pursuing new ground with AlphaGenome - with a free non-commercial API! (Their tech report i-is here)
Uses CNNs + Transformers
In goes the DNA sequence (up to 1M base pairs, compared to <10k of OSS or ~200k for less precise nets)
Out goes the multimodal data on regulatory activity:

Bottom line? Now predicting how a single mutation may alter, say, cancer’s progression or a drug’s effifacy - is possible.
🆓 Other FOSS worth mentioning are:
ChromBPNet (predicting chromatin accessibility)
Borzoi (predicting RNA-Seq from DNA, i.e. predicting gene expression)
Boltz (affinity prediction with a generous NVIDIA inference service):


That’s called domination 😏 (or, having enough superstars, data and compute to realize your Matthew advantage.
Yeah, yeah - Claude Code, Gemini CLI, Cursor, wow, such code.
Recently Atlassian had entered the game with a Claude Code wannabe-competitor called Rovo Dev (a short overview here). Even more TLDR:
20M tokens/day allowance, but 10.8k tokens system prompt each request + cached tokens are calculated in
Likely uses a mix of Claude 3.5, 3.7 and 4, but Claude’s limits are different
From what I’ve seen, people are increasingly turning towards CC compared to Cursor, Gemini (2), etc. Some are using Claude as the brain and Gemini as the brawn (translating to English from “Claude using Gemini as its bitch”)
Someone had leaked their system prompt and made an awesome write-up of the Z/80 technique:

I’d reckon there are several parts to the strategic puzzle (given you want to have your own opinion, rather than relying on the next big influencer I’m not one! I don’t influence!):

ML’s ‘periodic table’ for realizing something from first principles
Some awesome explainers and analyses, e.g. an agent village, 2025-x forecast, and election disruption overview - the scary thing (for me) is that it’s only getting better and at-scale…


(The world has changed) And on a side note - capital seems to be shifting:
About Hilton’s hotel chain - it’s actually a bit scary to me that there are that many Hiltons:

While the company itself was founded by Conrad N. Hilton in 1919, the first hotel to bear the now iconic “Hilton” name wasn’t opened until six years later.
Now, a full century on, the business is still finding ways to grow its vast portfolio, which includes properties like Hilton Garden Inns and Hampton, its mid-range chain that has boomed to become the world’s largest lodging brand. Last year, Hilton opened more than two hotels on average every single day.
Welcome to Teleogenic❣️
Other places I cross-post (not always) to: