To train better open models, we need predictable scaling. Delphi is Marin’s first step: we pretrained many small models with one recipe, then extrapolated 300× to predict a 25B-param / 600B-token run with just 0.2% error. Getting there took some work 🧵

This is pretty fire nix-bazel.build

Along with your resume, please submit your heart so we can see if it is heavier or lighter than a feather

Here's a super hot take of mine. If you've been in graphics long enough, you know I'm right though.

At this point, solving climate change is cheaper than not solving it -- not only on some grand 50 year time horizon, but tomorrow, today, immediately. Soving climate change *is* affordability. Maybe Dems should tell voters about that.

I bet something like Carbonlang would be an idea language for agents. A successor language to C++ with safety, fast builds, modern developer tooling and compatibility with the existing cpp ecosystem might be all that agents need. docs.carbon-lang.dev/docs/project...

Iterative adversarial refinement rocks.

And if you don’t want to work at OA, work for Merlin instead. bsky.app/profile/al.m...

Want better local models? Want models that are made completely transparently? OpenAthena.ai is hiring on our Marin.community team. We’re especially looking for SWEs and research scientists with experience in data and post training. DM me.

insurance is the financialization of epistemology

POSIWID is just historical materialism for dorks en.wikipedia.org/wiki/The_pur...

Choose Aetna when you want to enjoy the luxury of paying for health insurance when you don't want to give up the familiar feelings of being uninsured.

It’s not AI itself people are reacting to, but the consolidation of power. This is not inherent to AI, but is the reality of how it is being rolled out.

This single drop did not disappoint youtu.be/-9UhZuM-ADk?...

Another miracle in the cannon for eventual cult of information theory

I’m so psyched

I love Mark Fisher too!

I have an idea for bluesky’s next killer AI feature: A “Fix Typo” button. A normal “edit” button would create trust and safety issues. People would post ragebait, get reactions, then change the op to something normal, for eg. It’s why platforms won’t add an edit feature.

2026 is shit but at least we got a new Thundercat album.

@dollspace.gay I mentioned ferrotorch to my coworker, and he pointed me towards this rust ML framework (which compiles into XLA!) that may interest you: github.com/eaplatanios/...

A great irony of life is that the pursuit of the optimal is sub optimal.

Open Athena's mission might be to be a big bummer to the research community: github.com/marin-commun...

But, what if we could reorganize the world not to use so much oil?

Like near all bullies he’s also a huge coward

Attie is basically a new type of search. Like, you know what else builds a program for you based on a high level description? A SQL query. I don’t understand the hate for such a benign, useful feature. Spell check was once considered AI.

You can learn everything you need to know about queuing theory for large scale web services by volunteering as a dish washer in a commercial kitchen.

An excellent article. What a great lesson that one “cannot step outside of ideology.”

Great piece from @fediversereport.com on protocols (including ATProto, the one that Bluesky runs on), and why open protocols need to answer the governance question if they intend to truly protect communities from consolidated power.

We’re cooling up some really cool stuff at Open Athena, I can’t wait to talk publicly about it.

This is how the petrodollar system ends (Gift article) www.nytimes.com/2026/03/25/b...

I remember when I was in college a paper came out that formally proved that super Mario bros on gameboy was NP hard. I never thought that this type of research paper would be so practical.

Protip to senior engineers: instead of marking that recruiter email as spam or deleting it, consider replying with a resume of a new grad or student with a recommendation. It’s internship season and a tough market out there, juniors could use all the help they can get.

“Everything used to cost $200 and now everything costs $1,000.” — @cam.merose.com

This is really cool. An artist explicitly open sourced his had painted art work so that it can be used by ai researchers

Sincere idea: what if we could make crops that could fertilize themselves? www.science.org/doi/10.1126/...

Hey @johnboik.bsky.social, it seems like we have similar interests. I’m reading en.wikipedia.org/wiki/The_Daw... right now and it’s excellent. I wonder if you’re familiar with this book and if so, what you thought of it.

"However, we contend that Darwinian natural selection and common descent are but one conceptually equivalent example of a far more general natural process. ... all evolving systems display the essential characteristics of combinatorial richness and selection for function"

This paper rocks pubs.aip.org/aapt/ajp/art...

We could probably approximate the DFT using Kalman Filters + backpropagation + the Schrödinger equation en.wikipedia.org/wiki/Density...

This is a cool demo: Data sciences notebooks in plain english, with affordable for verifiably via other LLMs and prompts that refer to results between cells. What if data sciences was all in English? youtu.be/Mkv5cl5rA7s?...

Running claude on crosslink on open strix on nixos is too agent too fast. The promise is there, but I’m foiled by permission issues or misaligned interfaces. It will take some time to get this good.

@dollspace.gay I think you may enjoy this essay given your recent rust projects trying to make things better for data scientists. I feel that its efforts could be better spent building with this community, as opposed to merely mirroring it. wesmckinney.com/blog/looking...

My life right now: go and gopher baskets

My life as an AI researcher in the country: I have a stateful AI agent that does work for me, but no microwave.

One hand in dirt, the other in DuckDB.

Today I’m gardening and building a database component at the same time. What a brave new world.

A few people asked me to make an image of how much irrigated farmland would use the same water required for ALL ChatGPT usage. Here's the result of a rough botec, seems like inference and training are using about equivalent amounts of water, and power uses 5x as much

One notable issue with having my stateful agent use Claude instead of directly having control is that I find myself playing telephone with robots. Though, I think I’m getting the hang of communicating now.

@timkellogg.me I’ve got a new bot (who I call Hobbes) running on OpenStrix (admittedly, with a nuanced configuration) and it’s been going great! It and I are just now starting to work with Claude on a programming idea I have. I’m excited to see what the three of us get up to.