Blog
Product notes, research and announcements from the Tiptree team.
No posts on this topic yet.

Should you do a PhD in AI? Is it worth it?
I finished a PhD in AI at Mila, and a friend asked me whether their kid should do one. This is my honest answer: three reasons not to, three reasons to, and the one thing to do before you decide.

Which reviewer raises their score
Which reviewer raises their score after a rebuttal? Do reviewers who write a longer review raise their score more often? I went through 34,561 ICLR 2025 reviews to find out.

Clips from the AI4Science #5 panel
Yoshua Bengio’s path to ethical AI runs through one property: honesty.

Where to spend effort to get a paper accepted
Where should you spend effort to get a paper accepted? At ICLR '25, a 5.5 avg score had a 75% chance of acceptance. But a 5.0 avg score ALSO had a 75% chance of acceptance, if a reviewer raised their score.

AI4Science #5 in Berlin
AI4Science #5 is in the books. Thank you to everyone who joined us at Merantix AI Campus during Berlin Science Week.

Emailing an AC or PC
Your paper can get accepted by emailing an AC or PC, but only if the reviewer breaks the rules.

The rebuttal process
Most of my papers have been borderline submissions, because I like to write papers that are slightly outside the Overton window.

37% of ICLR abstracts said "novel" in 2025. This year, 5%.
37% of ICLR abstracts in 2025 had the word "Novel" in it, and this year it's in 5%.

Each dot is a search a coding agent ran on Lacuna
Each dot in this video is a search a coding agent ran on Lacuna, our ML research map, in the last 7 days.

How Canada should back its AI sector
I spoke with Murad Hemmadi at The Globe and Mail about Ottawa’s investment in Cohere, and the bigger question of how Canada should back its AI sector.

bf16 might be freezing your weight updates
If you fine-tune a model with weights stored in bf16 and let `torch.optim.AdamW` write to them directly, you get interesting behaviour. bf16 can only represent changes to weights if the change is at least \~1/512 of the weight itself, so at small learning rates only the smallest weights can update.

On some AI research topics, more than half of all papers are from 2026
More than half of all papers ever written on some AI research topics were published in 2026.

8x more visitors from AI assistants than from search engines
So I ran the numbers and Lacuna, our ML research map, now gets 8x more visitors from AI assistants than from search engines.
AI Science vs. Human Science
I believe that the goal of science is to advance human understanding of the universe.

The last page of your appendix
Another piece of advice I've been giving to students: make sure the last page of your appendix is really good.

Our first NeurIPS main-track paper: The Alien Space of Science
We built a way to sample good ideas - ideas that make scientific sense, but that no existing research community is positioned to propose.

One sentence in my most-cited paper
My main contribution to the most-cited paper with my name on it is a single sentence in the abstract.

Reading notes to Overleaf
When I read papers for a literature review, I want my notes to still be useful when I finally sit down to write.

Reading lists on Lacuna
You can now save ML papers to a Reading List on Lacuna and talk them through with Althea, our AI research assistant.

Editing drafts in pen
I print and edit in pen, then photograph the pages and upload them to Althea to update the final draft.

The fifth Multi-Agent Systems and Scientific Discovery party
5th edition of the Multi-Agent Systems and Scientific Discovery party is in the history books. We had a wonderful night with friends new and old, with a strong showing from the old school Mila folks.

The second AI4Science hackathon
Last week, we held our second AI4Science hackathon with our friends at Dunia and Merantix AI Campus.

The fourth Multi-Agent Systems and Scientific Discovery party
The fourth Multi-Agent Systems and Scientific Discovery party was amazing. Thanks to everyone who showed up and made it a memorable night.

A busy week in Montréal
It has been a wonderfully busy week in Montréal. Spent a ton of time catching up with old and new friends who are building, researching, and investing at the frontier.

Presenting Tiptree at a whiteboard in Toronto
Last week, Martin Weiss flew to Toronto to talk about Tiptree Systems. One of the things we like to do is to present our work in person, using a whiteboard (still can’t beat nice chalk and a chalkboard, though).

Good agents have todo lists, great agents have todo graphs
Good agents have todo lists, but great agents have todo graphs. This month we improved our todo engine (which is live, you can try it in Althea).

Your coding agent can read the literature
Your coding agent can write PyTorch. As of today, it can also read the literature.

A chemistry robot, a Kepler agent and an alloy checker in one research direction
A chemistry robot, an agent that rediscovers Kepler's laws from raw data, and a system that works out whether an alloy can be 3D printed all landed in the same Lacuna research direction.

400 researchers in a four-floor pub in Seoul
Last Tuesday, ~400 researchers took over a four-floor pub in Seoul for the third Multi-Agent Systems and Scientific Discovery Party.

Lacuna: A Research Map for Machine Learning, at ICML
We’re headed to Seoul to present Lacuna: A Research Map for Machine Learning at the ICML AI4Research workshop.

Loss aversion in LLMs
Most interpretability work looks inside the model using probes and looking for circuits. We’ve been trying something (stolen from the economics literature).

Animating our docs with a swarm of agents
I was designing Tiptree Systems documentation and decided it needed animations. For the coding agents docs, I thought a little swarm of agents pushing cubes into place would feel good.

ICLR 2026 in Brazil
We had a great time presenting, learning, and throwing a party at ICLR 2026 in Brazil. Such a good time, in fact, we never posted any pictures - though many of you told us it was where you had your best research conversations of ICLR!

Sign in with OpenReview
We’re excited to be the first partner organization to support Sign in with OpenReview! 🎉

The AI4Science autoresearch hackathon
Yesterday we wrapped the second edition of our AI4Science series, this time with an Autoresearch Hackathon! It was a lot of fun, and the numbers are a pretty good snapshot of where autonomous research agents are right now.
Launching the Open Research Commons
Tiptree Systems is launching an open research commons for the humans and AIs advancing the science of AI.

Our first AI4Science event
Our first AI4Science event is in the books.

Complex Document Generation for Literature Review
Althea renders full scientific documents with figures, tables, and mathematical notation directly in the chat interface.

A NeurIPS party with 1,100 registrations and room for 165
We threw a party at NeurIPS for researchers interested in Multi-Agent Systems and Scientific Discovery, and had 1,100 registrations. We had only booked a venue with 165 capacity… Next time we’ll get a bigger venue!
What the heck is an AI agent
AI agents should be black boxes that implement the Human API, the patterns of interaction that humans have evolved. By making coordination essentially free, they could solve humanity's fundamental economic problems and enable a future where human labor becomes art, not necessity.
Welcome to Tiptree Systems
We're excited to launch our new blog where we'll share insights, updates, and stories from our journey building the future of AI-powered communication.

Why technologists dominate LLM usage
I was wondering why technologists seem to dominate LLM usage today. One straightforward explanation is demand-side: technologists naturally gravitate toward automation and are comfortable integrating new tools into their workflows.

Defending my PhD on AI agents
I'm defending my PhD on AI agents this Friday at 9 AM EST.
Open source and AI coordination layers
Had an insightful conversation with Maxime Corbani about open source and AI. He introduced me to Twenty, an open source CRM.
