Protocells from three inorganic salts, some formaldehyde and water? They grow? They synthesise organic molecules of core biomolecular classes: amino acids, sugars, lipid-like motifs? And, there are similar structures in today's oceans? Yes! Read on. arxiv.org/abs/2601.11013

translation: South Texas will be solid red by 2030

wow! a beautiful follow up to the me of my absolute favorite papers. This stuff is so cool

listened to @jamellebouie.net and @lioneltrolling.bsky.social on Democrats' plan to make violating the constitution illegal. guess i'm convinced from a pragmatic standpoint that it's the right/best course of action from here, but i'm still so angry that we've gotten to this point.

who the fuck is giving these fucking amateurs an anti-drone laser system? o wait. it's me. i am the asshole paying these idiots to fire anti-drone lasers at party balloons.

I’ve been downing horse dewormer for the past five years and I‘m desperate for vindication. the Republican party thought I’d be the right person to make decisions for everyone

😡 everyone has forgotten, but the push for hydroxychloroquine started with in vitro studies showing inhibition of SARS-CoV-2 replication. later found to be limited to in vitro studies, zero translation potential. this is the same motivated quackery

Hey Amazon, got a fun new way for ai to kill you haha so funny

AI super bowl ads were so dystopic. Amazon: we can track anyone's movements Amazon: AI will kill you ("haha") Anthropic: everyone else is selling your personal information Meta: secretly film your neighbors Genspark: use AI and take the rest of the day off ("haha")

Let’s be very clear: Levi’s stadium is not in San Francisco.

unbelievable halftime show. So subversive. Can’t believe it

Many such cases

We could have a space elevator before the decade is out, but chemists decided it wasn’t worth it to share their raw data + metadata in machine readable format

custom tailored lab coat. gigaflex. necktie always looks good under a lab coat. this man has fucked around and he is about to find out in the best way

you think LLMs have powerful world models? no. no man has handwriting this good, and piss-scribbling of this quality is simply out of the question. a truly intelligent system would know that

why has machine learning been so much more successful in biology than chemistry? because biologists, in the face of staggering complexity, made their data publicly available and machine readable. chemists, meanwhile, decided a PDF of an NMR trace was good enough. a costly mistake for humanity

how do we feel about glass slides with a manufacturer's logo on?

The world would be a better place if everyone used this representation of DNA at all times

I’ve transitioned to using LLMs as interfaces to other tools. LLM +codebase, +docs, +software, +web search. Massive increase in utility over a chatbot The base LLM without context has many issues. An LLM with appropriate context and guidance is much more powerful

Tech companies insisting “there is no bottom.” regulators urging “keep looking”

Something I’ve often thought about but never really learned about. Will read with interest!

Accept All Cookies has evolved misleading options and color schemes and it's the worst of policy and technology and design all in one staggering numbers in this report

We’ve let law enforcement of all kinds brutalize protestors for far too long

voters want a vision they can affirmatively vote for. "choice" election is another way of saying compromise, and voters don't want that. given a choice, they lean republican, defaulting to distrust in government this is why Mamdani can be an example for any politician regardless of policies

end of week machining. looking good looking good

claude skills are nice but the error rate is way too high, usually in the form of taking shortcuts that the skill is designed to prevent

damn tell the verifier how you really feel

at some level i treat AI code like a physical system. i use experiments and controls to make sure it's behaving the way i expect. things I'm constantly doing: 1. smoke tests to make sure code runs 2. sanity checks with expected outputs 3. extensive testing (auto and manual) 4. modular building

i think about this often. 1. my personal philosophy of science is centered on outcomes, not process. i want to make inventions and solve problems that benefit humanity and push science forward. i use AI as a tool to accelerate toward those objectives. personal growth is great, but secondary

hah, TIL! one of the first plastics and an interesting history! en.wikipedia.org/wiki/Galalith

early morning grabs have long been a strategy used by ICE. this is their normal operating procedure. from 2023 www.aclu.org/news/immigra...

i'm using this system to help a collaborator on a bioinformatics project. it's an added challenge, as about half of my edits are on the AI system and half are on the repo we're testing. the system works OK but oh boy is this a competitive area right now. hard to know how much value i'm creating here

this is compounded by LLM context tunnel vision. context is expensive, so they try to make correct decisions with minimal context. this is totally opposite to humans, where context is relatively cheap and sustained.

huge tension working on autonomous research agents: how much autonomy do i trust them with? should they be able to install any package, download any data, restructure directories? or should i hold their hand to make sure nothing goes wrong, costing me time and slowing down the proces?

image alt text is an ideal dataset for training multimodal LLMs and I find it ironic that AI haters are so dutifully labeling all the data they freely share. I’m not saying don‘t do it, but shits complicated. There’s layers

????

This is not how I use AI in my science. I’m the scientist. AI is the tool. I must use the tool responsibly, and there should be consequences for irresponsible use

Machine learners love statistics, just … not like that

This reminds me of Inscripta and their open source CRISPR enzyme. They were partially successful, but made some bad business moves under pressure to generate revenue and pressure from commoditization of the product.

we're headed for an era of hyper-competition

Is kristi allowed to take orders from miller? Like, those wouldn’t be orders, coming from someone below her rank and outside her chain of command, right? don’t see how this absolves her at all

a little Gemini Blender adventure. I described a fluidic manifold i'm machining to Gemini. acrylic block with two screw ports for fluidic connections. here's Gemini's first attempt. initial prompt in alt text.

o bro Nicole is retracting her science paper with Jacob Steenwyk? a reanalysis points to the opposite conclusion? retraction notice isn't published yet, but that's a big ole yikes

Wow and I thought our govt had a legitimacy crisis

legend. one of the names to pop up most often in my beloved Voet and Voet biochemistry text

best and fastest way to regain international credibility is for congressional candidates to run on impeachment, win, and follow through. letting trump serve his last two years keeps us on the path to isolation

biggest lessons after two weeks setting up autonomous agents: 1. hardcoded orchestrators and verifiers are too brittle for general-purpose systems, even if you want structure 2. the filesystem is your best friend 3. let agents be decision-makers. give them guidance and, well, agency

seeing Jed and Astera start to pop up in quite a few AI advances recently

this is somewhat similar to the tool i'm working on now. hope to share some experiments on here soon. it's been a pain in the butt to figure out the right patterns and architecture. feel like it's starting to come together

nice commentary. I don’t think we need that much data for this to work. An individual laboratory produces enough of this stuff to get going on its own