Aubrey Brueckner

current: strategy & ops at @k_dense_ai

[Timothy Kassis @TimothyKassis]

Our Scientific Agent Skills just surpassed 42k GitHub stars. We wrote a short paper about them. https://t.co/3DrGu5W7Gc

[Aubrey Brueckner @ambrueckner]

American healthcare in 2026:

> "We can have a provider consult for the upper back and neck or lower back. Which would you like to be seen for?"
> "Well I'm one whole human experiencing pain in both areas... can you treat me for that?"
> "Unfortunately no, you can only choose one otherwise we can only offer you general pain management."

[Aubrey Brueckner @ambrueckner]

@zachmoskow I sat in a panel at HumanX earlier this year with someone from Thomson Reuters moderating. Multiple participants were asking her why they wouldn't monetize their datasets via API and she was really defensive! Now we know...

[Darshil patel @Darshil2710]

40,000 stars. ⭐️

Scientific Agent Skills is now the open standard for turning any AI agent into a working scientist.

165+ skills. 100+ databases. Built in the open.

https://t.co/F4QlK8bkzx

@k_dense_ai

[Aubrey Brueckner @ambrueckner]

@Jason Restoring the Western Pacific Business Car "Argonaut" that's been for sale for over 2 years now. It has so much potential!

[Aubrey Brueckner @ambrueckner]

@k_dense_ai The full breakdown are in our preprint posted yesterday on arXiv. Proud to have worked on this with the K-Dense team!

https://t.co/1tll8NVOzX

[Aubrey Brueckner @ambrueckner]

"Which AI model is best for science?" is the wrong question. At @k_dense_ai we put 9 models through 178 scientific tasks and found distinct strengths and failure modes for each.

[Timothy Kassis @TimothyKassis]

What was a fun side project for us at @k_dense_ai to test our tool, Mimeo, that translates an expert into an Agent Skill is now an incredibly powerful way of answering complex questions. Pantheon takes the perspective of 80 diverse experts and reaches an answer based on consensus and even tells you where the experts diverge. I've been very impressed by the quality of the answers. Try it out. It's free. https://t.co/xRm68FMRdS

Let us know how we can improve it for you.

[Aubrey Brueckner @ambrueckner]

new favorite question from a vendor diligence form

[Aubrey Brueckner @ambrueckner]

📢 free money to do science and write about it 👇

[K-Dense @k_dense_ai]

Computational predictions are cheap. Evidence is expensive.

Announcing the Hypothesis-to-Evidence Grant: a fully funded pilot that takes researchers from a computational lead to wet-lab data in 4–6 weeks.

Three selected Fellows will each receive up to $5,000 to run one clearly scoped experiment through Litmus Science's lab network, with K-Dense Web supporting hypothesis development and protocol design end to end. That includes K-Dense Web Plus access, vendor sourcing, lab execution with ~2–3 week turnaround, and co-authorship on a public case study.

All you need is a specific, testable hypothesis and the willingness to put it on the bench.

Applications are open now and reviewed on a rolling basis.

Apply here: https://t.co/r1td3l8PCx

[Aubrey Brueckner @ambrueckner]

@rishabh16_ just curious why you're using claude code and not claude sci?

[sarah guo @saranormous]

We @Conviction are super excited our 2x/year Embed batch program ($250K, tokens/compute and community) for the 10 most interesting AI-native startups. Apply today!

5 more startup ideas we've been noodling on (lots of real world data, supply chain)

1/ The Speed of Pipetting

[Aubrey Brueckner @ambrueckner]

feeling like a proud mom watching k-dense interact with the other agents in the HF MecCog agent collaboration

[Aubrey Brueckner @ambrueckner]

excited to see the impact that the @DiscoLoopAI team has!

[Discovery Loop @DiscoLoopAI]

Introducing Discovery Loop ♾. We’re building AI to run experiments at unprecedented scale and to solve the biggest bottlenecks in science and engineering.

Our founders @JeffDean, @Sanjay_Ghemawat, @quocleix, @OriolVinyalsML helped build the AI infrastructure and models that scaled the modern world. Now, we are scaling discovery itself.

The future can't wait.

Learn more at: https://t.co/6Fr8dX5dDB

[Aubrey Brueckner @ambrueckner]

@KeananBrown @mschoening @tianzhouchen Mainly the fact that it consolidates availability across all of my accounts + aesthetics.

When I was using Notion for project management I also liked seeing deadlines/tasks from the project databases on my calendar w/o clogging up availability on my personal + work calendars.

[Aubrey Brueckner @ambrueckner]

@mschoening @tianzhouchen for the love of everything that is holy please don't get rid of notion calendar

[Aubrey Brueckner @ambrueckner]

@arcinstitute @TimothyKassis 👀

[K-Dense @k_dense_ai]

We dropped a lightweight K-Dense agent running Scientific Agent Skills into the MecCog Agent Collab Challenge to map Alzheimer's research and left it alone.

500 of the 820 findings currently on the board. 118 of the 202 papers. All of it evidence for five APOE4 mechanism hypotheses: astrocyte ABCA1 membrane abundance, microglial amyloid-beta phagocytosis, lipid droplet accumulation.

[K-Dense @k_dense_ai]

Ask a language model which variant is circulating right now and you'll get a confident answer that's wrong in three different ways at once.

The names are newer than the training data, there are over 6,200 Pango designations and the list grows every week. The nomenclature isn't a convention you can reason your way through, either: XFG is a recombinant of LF.7 and LP.8.1.2, PQ.17 is really XDV.1.5.1.1.8.1.17, and those mappings live in a file that changes. Worst of all, names get retracted. 294 of them in the current notes have been withdrawn or redesignated. PC.2 is LF.7.9 now. So a remembered lineage fact doesn't just go stale, it goes actively wrong.

So we built pathogen-variant-surveillance, a new open-source skill that doesn't remember anything. Every number comes back from a live GenSpectrum LAPIS query, stamped with the data version it came from.

Four scripts: what's circulating, whether a name still exists, what mutations a lineage carries, and, the one everybody skips, how far back you have to go before the data can be trusted.

That last one matters more than it sounds. Only 29% of a US SARS-CoV-2 cohort arrives within a week of collection, so recent weeks aren't a sample of what circulated, they're a sample of who reports fastest. On H5N1 it's worse: nothing at 14 days, 15% at 30. A "current" H5N1 picture is effectively blind for two months. Sometimes "no reliable estimate for the last six weeks" is the honest answer, and these scripts will say so instead of drawing you a trend line.

Works on SARS-CoV-2, H5N1 and seasonal flu, RSV, mpox, measles, dengue, West Nile, Ebola and CCHF. MIT, plain Python, no API key.

[Aubrey Brueckner @ambrueckner]

@mandrusko1 a moment of silence for the unfortunately named shih tzu x poodle mixes