Felipe Vecchietti
LightNews LightNews
lfelipesv.bsky.social
Felipe Vecchietti
@lfelipesv.bsky.social
79 followers 170 following 2 posts
Postdoctoral Researcher at the Max Planck Institute for Security and Privacy. Interested in RL, Graph Neural Networks, AI for Science, and Multi-Agent Systems. 🇧🇷
Posts Replies Media Videos
Reposted by Felipe Vecchietti
dystopiabreaker.xyz
everyone here is extremely confident they’ve never used AI but an enormous fraction of their existence is mediated, observed, or recommended by AI models over the last decade+. it’s just not advertised as a Product Name, so it doesn’t exist
doomboy.bsky.social doom boy @doomboy.bsky.social · 29d
the nyt posts this stuff like avoiding AI for 48 hours is hard and meanwhile i have literally never once intentionally used an AI program for anything ever
October 31, 2025 at 11:47 PM
Reposted by Felipe Vecchietti
markriedl.bsky.social
This is a very good clip of Guillermo del Toro on the value of art. The value of human-made art is, in part, based on what the artist experienced that made them want to make the art. That is what we pay for. That is what AI does not have.
October 24, 2025 at 6:50 PM
Reposted by Felipe Vecchietti
kevinkaichuang.bsky.social
Are you a PhD student interested in ML and biology or health? Come do an internship with me, @avapamini.bsky.social, Alex Lu, @lcrawford.bsky.social, or Kristen Severson at MSRNE!

Applications are due Dec 1: make sure you include a research statement!

jobs.careers.microsoft.com/global/en/jo...
Search Jobs | Microsoft Careers
jobs.careers.microsoft.com
October 21, 2025 at 7:32 PM
Reposted by Felipe Vecchietti
workshopmlsb.bsky.social
You asked and we listened... @workshopmlsb.bsky.social is excited to be expanding to Copenhagen, DK at @euripsconf.bsky.social 🎉

Two workshops (San Diego & Copenhagen) will run concurrently to support broader attendance. You can indicate your location preference(s) in the submission portal💫
September 12, 2025 at 12:43 PM
Reposted by Felipe Vecchietti
martinsteinegger.bsky.social
If you use Boltz1/2, BioEmu, Chai1, or other MSA-dependent models, you’re likely using our ColabFold server. Please be considerate! Avoid large submissions across many IPs instead generate the MSA locally. Our server is an old-timer from 2014 and can’t handle that load.
August 15, 2025 at 5:48 PM
Reposted by Felipe Vecchietti
stellaathena.bsky.social
Are you afraid of LLMs teaching people how to build bioweapons? Have you tried just... not teaching LLMs about bioweapons?

@eleutherai.bsky.social‬ and the UK AISI joined forces to see what would happen, pretraining three 6.9B models for 500B tokens and producing 15 total models to study
August 12, 2025 at 12:40 PM
Reposted by Felipe Vecchietti
astylianou.bsky.social
I have found ACing for NeurIPS this year a miserable experience and I've been struggling to articulate why without basically whining about it being too much work (although.....true. My god the amount of words I'm reading...). But I think I've figured out part of my dislike for this process.
August 13, 2025 at 2:53 AM
Reposted by Felipe Vecchietti
astrokatie.com
"PhD-level experts in your back pocket" is a completely nonsensical description of AI but a pretty good description of social media if you follow the right people
August 9, 2025 at 11:07 PM
Reposted by Felipe Vecchietti
maxfus.bsky.social
Now that OpenCRISPR is in nature and rekindled the 'what's-a-novel-sequence' debate, I'm happy to share an app to check this, which I built for fun some time ago.

fuerstlab.shinyapps.io/SeqNovelty/

quick 🧵
August 1, 2025 at 11:50 AM
Reposted by Felipe Vecchietti
yoakiyama.bsky.social
Excited to share work with
Zhidian Zhang, @milot.bsky.social, @martinsteinegger.bsky.social, and @sokrypton.org
biorxiv.org/content/10.1...
TLDR: We introduce MSA Pairformer, a 111M parameter protein language model that challenges the scaling paradigm in self-supervised protein language modeling🧵
Scaling down protein language modeling with MSA Pairformer
Recent efforts in protein language modeling have focused on scaling single-sequence models and their training data, requiring vast compute resources that limit accessibility. Although models that use ...
biorxiv.org
August 5, 2025 at 6:31 AM
Reposted by Felipe Vecchietti
jiikae.com
As someone who’s been using the internet since 1999. It is crazy to see how much more restricted it has become

The old internet felt more futuristic than what we have now
July 30, 2025 at 1:31 PM
Reposted by Felipe Vecchietti
eikofried.bsky.social
🇩🇪 3-year universal basic income (UBI) study, in which n=120 received UBI, n=1500 did not. Website is pretty neat, check it out for core findings.

www.pilotprojekt-grundeinkommen.de/en
Basic Income Pilot Project / Study results
What does a universal basic income really achieve? Find the results of the first long-term German study here.
www.pilotprojekt-grundeinkommen.de
July 25, 2025 at 7:54 AM
lfelipesv.bsky.social
I am happy to share that our paper "Artificial intelligence-driven computational methods for antibody design and optimization" is out in mAbs! 🧪🤖

We review AI methods across the multiple dimensions of antibody design, with a special focus on antigen-conditioned methods.
Artificial intelligence-driven computational methods for antibody design and optimization
Antibodies play a crucial role in our immune system. Their ability to bind to and neutralize pathogens opens opportunities to develop antibodies for therapeutic and diagnostic use. Computational me...
doi.org
July 19, 2025 at 8:48 AM
Reposted by Felipe Vecchietti
abeba.bsky.social
from a 2022 paper with @gyllingberg.bsky.social & @soccermatics.bsky.social

The lost art of mathematical modelling www.sciencedirect.com/science/arti...
We argue that currently, the universalist approach dominates and creation of new models, which is inherent to pluralism, is not sufficiently emphasised. This brings us to, in Sections 6 Machine learning cannot replace modelling, 7 Are hybrid models the answer?, a discussion of how mathematical biology has responded with the rise of machine learning. We argue that ML, which emphasises prediction (activity 3), is ill-prepared to deal with complexity without incorporating some form of mechanistic model building. But we also, more controversially for those working in mathematical biology, emphasise how some of the responses to the rise of ML have fallen into the trap of making models of models (or fitting models to data generated by models) rather than innovating by creating new models of biology itself. We conclude that mathematical biology needs less unification and less analysis of existing models, and more creativity and more creation of new models. We should be creative without fear of them being wrong or producing ideas that are mathematically intractable, with an aim of providing a multitude of tools for better understanding of biological systems. Instead, the radical definition of complex systems comes from, what is known as, critical complexity. Work by Paul Cilliers and Alicia Juarrero warned against aggrandising models (even supposedly complex systems models) [3], [4]. They emphasise the need to embrace the ambiguous, messy, fluid, non-determinable, contextual, and historical nature of complex systems. They describe complex phenomena as unfinalizible and inexhaustible, which means that we can never capture any given biological system entirety with models [5]. Fig. 1, adapted from Di Paolo et al. (2018), captures the interdependence, fluidly and interactivity of agents and environments in a complex system [2]. Complex systems are open-ended, which means there is no uncontested way of telling whether what we have included in a model is crucial or what we have omitted as irrelevant is indeed so. Models can, according to the critical complexity approach, be contradictory: we can accept two incompatible predictions as both describing the same system. This approach views a model as a snapshot of a system and no single snapshot tells the whole story. For modelling the human body, for example, “a portrait of a person, a store mannequin, and a pig can all be models” [6]. None is a perfect representation, but each can be the best model for a human, depending on whether one wants to remember an old friend, to buy clothes, or to study anatomy. The critical complexity view suggests that theoreticians should avoid specialising in any one modelling approach and try to find the right set of models to understand a particular system in a given context. There can, of course, be more than one definition of complex systems. Indeed, Cilliers and Juarrero’s approach to complexity encourages a plurality of definitions (after all, there is no single view of a system). We would, though, emphasise that it is the radical definition of complexity – in which systems always resist a complete description, are open and unfinalizable – which is least well understood by mathematical biologists today. It is therefore important to investigate how complexity should be approached in the study of biological systems.
June 7, 2025 at 3:01 PM
Reposted by Felipe Vecchietti
eleutherai.bsky.social
Can you train a performant language model using only openly licensed text?

We are thrilled to announce the Common Pile v0.1, an 8TB dataset of openly licensed and public domain text. We train 7B models for 1T and 2T tokens and match the performance similar models like LLaMA 1 & 2
June 6, 2025 at 7:19 PM
Reposted by Felipe Vecchietti
delalamo.xyz
Lots of work seems to have gone into the Boltz-2 data curation process to learn binding affinities
A very long description of how data was curated to learn affinity prediction
June 7, 2025 at 8:06 AM
Reposted by Felipe Vecchietti
neuripsconf.bsky.social
Read our blog post about updates from the Position Paper Track - submissions have closed and initial review is now complete:
blog.neurips.cc/2025/06/06/u...
Updates from the NeurIPS 2025 Position Paper Track – NeurIPS Blog
Communications Chairs 2025 2025 Conference
blog.neurips.cc
June 6, 2025 at 5:05 PM
Reposted by Felipe Vecchietti
andrewgwils.bsky.social
AI benchmarking culture is completely out of control. Tables with dozens of methods, datasets, and bold numbers, trying to answer a question that perhaps no one should be asking anymore.
May 30, 2025 at 9:55 PM
Reposted by Felipe Vecchietti
delalamo.xyz
An interesting post from the other site. These days I read methods comp bio papers w/ wet lab validation while wondering about unreported failures, which is ironic because a data-heavy discipline like comp bio needs the negative data. But comparing it to p-hacking seems totally over-the-top
Twitter user blaming biology's replication crisis on simultaneous exploration of many hypotheses
May 3, 2025 at 9:35 AM
Reposted by Felipe Vecchietti
hannes-stark.bsky.social
Online LoG conference 27th – 30th Nov!
All the talks will be live-streamed on Zoom and YouTube with poster sessions and sponsor sessions on GatherTown:

Schedule Zoom: logconference.org

Youtube: www.youtube.com/@learningong...

GatherTown: app.gather.town/app/etG6JBKc...
November 25, 2023 at 3:42 PM