The infrastructure for doing AI research, together.
Berlin + Montreal
Joined February 2025
- Tweets37
- Following16
- Followers64
- Likes7
Pinned Tweet
Excited to announce that @Lacuna_Research is now listed in Claude's connector directory.
Add it here: claude.ai/directory/lacuna-r…
Lacuna is now listed in Claude's connector directory.
Add it here: claude.ai/directory/lacuna-r…
If you use Claude for literature work, you can connect it to Lacuna in a couple clicks. Claude can then look up papers, authors, research directions, venues and institutions, the code and datasets linked to papers, and the open research proposals Lacuna keeps for each direction.
It is read-only and free, and you don't need an account.
Huge thanks to @Miles18143113 for helping with the MCP server.
RT @Lacuna_Research: What research is your code standing on?
nanoGPT: has 15 Python files building on 13 papers. If you keep digging back…
Here’s how we figured out our AEO:
1. we made the site really fast - median page serve time is 25 ms, p99 411 ms.
2. like REALLY fast - over the last 7 days we’ve been serving ~87 requests per second on average.
every page has a plain markdown version for LLMs to read
3. paper pages have the summary, the related work, and the open questions all nicely bundled
4. we made sure the sitemap and pages were clean - removed thousands of dead links the crawlers were hitting.
5. The OpenAI bot had cut its crawl of Lacuna ~10x, ChatGPT clicks halved about a week later, and they came back after the fix (probably not a coincidence)
The site is Lacuna, an index of ML research, built in to Althea, an AI research assistant built for AI researchers. You can set it up with your Claude + Codex too, if you like.
Tiptree Systems retweeted
I wrote an essay called AI Science vs. Human Science.
I was inspired by William Thurston's 1994 essay "On Proof and Progress in Mathematics". Thurston argued that mathematics is "amazingly compressible": once you understand something, you can use it as a tool to solve other things.
How compositional is an object? Can you put a number to it?
Today's paper says yes, and then uses that number to do some pretty cool things.
New paper by @Yoshua_Bengio @g_lajoie_ @EricElmoznino @tejaskasetty @blaiseaguera
The world is made of sparsely interacting, independent mechanisms (says the Sparse Mechanism Shift hypothesis in the causality literature).
But “the world is compositional” has mostly been an assumption. There hasn’t been a general way to measure how compositional a given object actually is.
This work defines it using algorithmic information: how much new shared structure appears when you model the whole, beyond what the parts already explain. A random string scores zero. So does a simple repeating pattern.
You can use this to generate strings and images, with strings growing deeper grammars and images picking up repeated motifs (while maintaining a single visual theme).
They fix how the object is split into parts, and conjecture that the “true” split is whichever one maximizes the score. It's related to the same problem we approached from the architecture side in Neural Attentive Circuits during our PhDs.
“Defining and Generating Compositional Structure”
Read more on Lacuna: lacuna.tiptreesystems.com/wo…
Tiptree Systems retweeted
New paper by @Yoshua_Bengio @g_lajoie_ @EricElmoznino @tejaskasetty @blaiseaguera
The world is made of sparsely interacting, independent mechanisms (says the Sparse Mechanism Shift hypothesis in the causality literature).
But “the world is compositional” has mostly been an assumption. There hasn’t been a general way to measure how compositional a given object actually is.
This work defines it using algorithmic information: how much new shared structure appears when you model the whole, beyond what the parts already explain. A random string scores zero. So does a simple repeating pattern.
You can use this to generate strings and images, with strings growing deeper grammars and images picking up repeated motifs (while maintaining a single visual theme).
They fix how the object is split into parts, and conjecture that the “true” split is whichever one maximizes the score. It's related to the same problem we approached from the architecture side in Neural Attentive Circuits during our PhDs.
“Defining and Generating Compositional Structure”
Read more on Lacuna: lacuna.tiptreesystems.com/wo…
It's been a good day on the front page of HN for our company's first NeurIPS Main Track paper.
Enjoy here: news.ycombinator.com/item?id…
Tiptree Systems retweeted
Can AI help us identify new scientific ideas that might be difficult to materialize without AI assistance? Check out this NeurIPS 2026 work now via this thread, and at the conference in December. ArXiv and other links are provided at the end.
SCIENCE HAS BLIND SPOTS.
We built AI to explore them 👽
The Alien Space of Science, my internship project at @tiptreesystems, has been accepted to the #NeurIPS2026 Main Track!
We sample ideas that make scientific sense, but that no existing community is positioned to propose.
🧵👇
Tiptree Systems retweeted
SCIENCE HAS BLIND SPOTS.
We built AI to explore them 👽
The Alien Space of Science, my internship project at @tiptreesystems, has been accepted to the #NeurIPS2026 Main Track!
We sample ideas that make scientific sense, but that no existing community is positioned to propose.
🧵👇
Would you publish at a venue that required you to submit code with your paper?
@TmlrOrg's main question for a paper is whether its claims are supported by convincing evidence. So, does the code actually support the claims?
There are far more people, and agents, flooding into science saying "look at these results, you should build on this," and those blocks are crumbling more than they used to. Peer review is supposed to be how we agree, as a community, what things are solid enough to build on. The capacity to make solid-looking blocks is expanding dramatically with LLMs, so shouldn't we be expanding our capacity to check them too?
Know what has been done. Do what hasn't.
That's the line on our homepage, and it's the whole company in two sentences.
Knowing what has been done is Lacuna: a map of 2.3M ML papers, the people who wrote them, and the research directions they add up to.
Doing what hasn't is Althea, a research assistant that works from that map with you, from finding the gap to running the experiment.
Lacuna: lacuna.tiptreesystems.com
Althea: althea.tiptreesystems.com
Tiptree Systems retweeted
When I read papers for a literature review, I want my notes to still be useful when I finally sit down to write. I have a pretty poor memory, and I can be quite disorganized, so this has been a struggle for me.
I maintain a reading list on Lacuna (which I can update via Claude + Althea MCP). But mostly I just click down the rabbit hole adding papers, perusing them, and talking to Althea. Then I add my notes (or ask Althea to do it) and wander off to the next thing.
When I (finally) sit down to write, I read and discuss my notes while editing a document and talking to Althea about it. When we're happy with it, we add it to Overleaf. Over the weekend I made a video of what that looks like.
Reading papers for ICLR is a whole other thing, which I am procrastinating on, and to which I will soon return.
Save machine-learning papers in Lacuna and talk them through with Althea, our AI research assistant.
Here, a reading list on agent memory becomes a conversation: how do these approaches differ, and what experiment would test the difference?
Try it with the papers on your reading list here: lacuna.tiptreesystems.com
Tiptree Systems retweeted
I often give this advice to students, so I thought I'd create an artifact for it. I write my academic papers in a document editor: usually first drafts with @tiptreesystems Althea Docs, then @overleaf.
I print and edit in pen, then photograph the pages and upload them to Althea to update the final draft.
I find that I still can't beat the focus that comes from having a printed paper. I'm unsure why it helps: the ritual, a form of cognitive distance, or a commitment created by getting up and printing the paper. All I know is it focuses my mind on reading and reviewing.
OCR is still lossy (especially for my handwriting...), but this is the best workflow I've found in terms of ROI on my time.
We hosted the 5th edition of the Multi-Agent Systems and Scientific Discovery party in SF last week. We mostly forgot to take photos (too many good conversations), but we did get one of some old school @Mila_Quebec folks.
Hope to see you all at the next one Sydney in December for @NeurIPSConf !
Then the fun part.
Re-frame the bet as "money you already have" and Gemini 2.5 Flash suddenly goes loss-averse like a human.
Gemini 3.5 Flash shrugs it off.