Lightnews — Scholar-powered news

Austin Tripp @austinjtripp.bsky.social · Jun 25

(I wrote this because this was something I inferred over time, and I think it's helpful to new reviewers to explain acceptance criteria more explicitly)

Austin Tripp @austinjtripp.bsky.social · Jun 25

That is, a paper should provide either a unique result or a unique idea (ideally both), and on top of that should have no correctness issues.

Full post is here: www.austintripp.ca/blog/2025-06...

Happy to hear comments/feedback! I know my approach is just 1 of many!

1

Austin Tripp @austinjtripp.bsky.social · Jun 25

NeurIPS reviews are due in 1 week 😱

If it's your first time reviewing (or if you don't feel totally confident about accept/reject), I recently wrote a blog post where I explain how *I* approach reviewing. Essentially:

accept = correct AND (result OR idea) [...]

1 3

Reposted by Austin Tripp

Valence Labs @valenceai.bsky.social · May 20

1/ At Valence Labs, @recursionpharma.bsky.social's AI research engine, we’re focused on advancing drug discovery outcomes through cutting-edge computational methods

Today, we're excited to share our vision for building virtual cells, guided by the predict-explain-discover framework 🧵

2 6 13

Austin Tripp @austinjtripp.bsky.social · May 16

If none of this makes any sense to you but you think multi-objective optimization is relevant, check out my full post below (where I explain MOO in more detail too). Bonus: also has an interactive visualization (kudos to Claude 3.7)

www.austintripp.ca/blog/2025-05...

Chebyshev Scalarization Explained

I've been reading about multi-objective optimization recently.1 Many papers state limitations of "linear scalarization" approaches, mainly that it might not be able to represent all Pareto-optimal sol

www.austintripp.ca

Austin Tripp @austinjtripp.bsky.social · May 16

2. Unfortunately, maximizing the Chebyshev objective may produce points which are *not* Pareto optimal (so some filtering might be required)

...

1

Austin Tripp @austinjtripp.bsky.social · May 16

For anybody working on multi-objective optimization: I recently did a deep-dive on Chebyshev scalarization and wrote a blog post. Key findings:

1. Unlike linear scalarization, varying the weights of Chebyshev scalarization will find *all* points on the Pareto front (not just the convex part)

...

1 1

Reposted by Austin Tripp

Valence Labs @valenceai.bsky.social · May 13

(1/3)The poster submission deadline for MoML 2025 has been extended to May 20th, 2025.

Don’t miss an opportunity to share your work at this years conference.

Submit here: portal.ml4dd.com/moml-2025-po...

1 1 1

Austin Tripp @austinjtripp.bsky.social · Apr 24

Really interesting essay- disagreements about AI existential risk might *really* be disagreements about dual-use nature of future technologies (since this is the vector people think AI could cause extinction).

Michael Nielsen @michaelnielsen.bsky.social · Apr 14

New essay exploring why experts so strongly disagree about existential risk from ASI, and why focusing on alignment as the primary goal may be a fundamental mistake
michaelnotebook.com/xriskbrief/i...

1 6

Austin Tripp @austinjtripp.bsky.social · Apr 14

Claude and Gemini do a pretty good job at coding some niche python packages from just a prompt. Some editing required, but if you haven't tried it yet then I highly recommend!

www.austintripp.ca/blog/2025-04...

Coding python packages with AI

I tried using some new LLM tools to code 2 entire python packages (instead of editing a handful of lines at a time, which is what I did previously). It went well! These tools are not perfect, but they

www.austintripp.ca

1 2

Austin Tripp @austinjtripp.bsky.social · Feb 14

(also, with ICML reviewing starting, this post will probably be the first in a series of posts about peer reviewing, stay tuned! 👀)

Austin Tripp @austinjtripp.bsky.social · Feb 14

I wrote a blog post explaining this in more detail: www.austintripp.ca/blog/2025-02...

If you think I'm wrong, I'd genuinely like to hear why. Please comment in 🧵

Is offline model-based optimization a realistic problem? (I'm not conv

This is a "quickpost": a post which I have tried to write quickly, without very much editing/polishing. For more details on quickposts, see this blog post. Offline model-based optimization (OMBO in

www.austintripp.ca

1 1

Austin Tripp @austinjtripp.bsky.social · Feb 14

Can anybody explain to me why so many ML papers study "offline model-based optimization"? This is essentially "1-shot optimization".

My main concern is "are there 1-shot optimization problems in real life"? Papers mention "drug discovery (DD)" as an example, but 1-shot DD never happens, no? 😂

1 1

Reposted by Austin Tripp

M A Osborne @maosbot.bsky.social · Feb 7

People who are masking are smart for two reasons:

1. They do not want to get brain damage
2. They are not getting brain damage

3 29 170

Austin Tripp @austinjtripp.bsky.social · Jan 31

Second note: there are a lot of more standard topics too, eg AI for science stuff, I'm just not posting that here.

Austin Tripp @austinjtripp.bsky.social · Jan 31

Also funny:

- Position: ML researchers should try to ensure their code is not a heaping pile of dogsh*t

- Position: ML researchers should learn basic math (I'm talking to you, people who don't add error bars to their plots!!)

- Position: focusing on meaningless benchmarks is stupid

1

Austin Tripp @austinjtripp.bsky.social · Jan 31

Other abstracts:

- Position: what if we started holding ML papers to actual standards?

- Position: reviewers should actually read the papers they are reviewing

- Position: reviewers should *at least try* to judge whether paper's claims are true before accepting them

Austin Tripp @austinjtripp.bsky.social · Jan 31

(Note: titles are summarized/anonymized since I don't think I'm allowed to share)

Austin Tripp @austinjtripp.bsky.social · Jan 31

My bid screen for ICML position papers is basically:

- "Position: ML conference peer review is sh*t"

- "Position: Let's abolish conference reviewing"

- "Position: C'mon ML reviewers, surely we can do better than *this*"

Am I in a "review hate" echo chamber or is everybody else seeing this too? 😶

4 3

Austin Tripp @austinjtripp.bsky.social · Jan 26

Easy to get started on the antiviral challenge! I plan to submit some GP baselines from my PhD work (possibly with a collaborator).

Polaris @polarishub.io · Jan 21

It only takes a few lines of code to get started with the @asapdiscovery.bsky.social x @omsf.io antiviral challenge!

In this short tutorial, we’ll show you how easy it is to login, load the challenge data, train a simple baseline model, and submit your predictions 🧵

youtu.be/KBZJFA_arAg

Tutorial: Getting Started with the Antiviral Challenge

YouTube video by PolarisHQ

youtu.be

1 5

Reposted by Austin Tripp

Polaris @polarishub.io · Jan 14

🏁 The antiviral challenge is live! 🏁

Ready to test your skills on new data? Hosted in partnership with @asapdiscovery.bsky.social and @omsf.io, we've prepared detailed notebooks showcasing how to format your data and submit your solutions. 🧑‍💻

1 6 17

Austin Tripp @austinjtripp.bsky.social · Jan 13

My PhD thesis is finally online. Thanks @cambridgemlg.bsky.social for a wonderful 4.5 years learning about probabilistic ML 😍

Code: github.com/austint/phd-...

Thesis DOI: doi.org/10.17863/CAM...

GitHub - AustinT/phd-thesis: LaTeX code for my PhD thesis (https://doi.org/10.17863/CAM.114023)

LaTeX code for my PhD thesis (https://doi.org/10.17863/CAM.114023) - AustinT/phd-thesis

github.com

2 18

Austin Tripp @austinjtripp.bsky.social · Jan 10

Slightly longer version of this post is on my blog: www.austintripp.ca/blog/2025-01...

What ML researchers and users get wrong: optimistic assumptions

ML is often done poorly, both by "ML experts" (by which I mean people who understand the algorithms but not the data) and "ML users" (by which I mean people who understand their data, but not the algo

www.austintripp.ca

Austin Tripp @austinjtripp.bsky.social · Jan 10

A common issue I see in ML, both from ML "experts" and "users", is overly optimistic assumptions.

"experts" (people designing algs) usually assume the data is very simple

"users" (people using algs) usually assume that algorithms are more robust than they really are

Conclusion: always be careful!

1 4 19

Reposted by Austin Tripp

Polaris @polarishub.io · Dec 18

📊 Imagining the Future of ML Evaluation in Drug Discovery

Our recent paper discussed the limitations of static leaderboards—they never tell the full story. What if we had a better and easier way of evaluating methods?

A vision for the future, in the latest blog 🧵

polarishub.io/blog/imagini...

1 1 5