photog.social is one of the many independent Mastodon servers you can use to participate in the fediverse.
A place for your photos and banter. Photog first is our motto Please refer to the site rules before posting.

Administered by:

Server stats:

242
active users

#foundationmodels

1 post1 participant0 posts today

GREmLN: A Cellular Regulatory Network-Aware Transcriptomics Foundation Model
biorxiv.org/content/10.1101/20

Comments
* really brilliant work: LLM + neural transformer architecture
* a molecular biologist, I long-thought (no time to learn / program) flux-based analysis / ODE models
* this type of modeling is foundational to in silico modeling
* huge proponent of graphs - add LLM, attention heads - brilliant! 😀 👍️ kudos

#wwdc25 #FoundationModels #Xcode #LLM

Has anyone seen the Foundation Models actually working? Am I holding it wrong? I am not getting any responses either in the playground or app, and there is a bunch of this crap in the console when running the app.

Running Xcode 26 beta on macOS 26 beta.

Are there regional restrictions? Xcode AI chat also is “unavailable in my region”.

ChatGPT & Co für sicherheitsrelevante Einsatzfelder nutzbar machen
Die @Cyberagentur hat #Forschungsprogramm #HEGEMON zur Bewertung und Anpassung generativer #FoundationModels für sicherheitskritische Anwendungen gestartet.
Entwicklung neuer Benchmarks für KI-Modelle und Anwendung auf komplexe Aufgaben aus dem Geoinformationswesen.
Mehr Infos zur Vergabe: t1p.de/qcuq4
#Cyberagentur #HEGEMON #KI #Sicherheit #Benchmarking #PCP #Geoinformation #Forschung #KITransparenz

Can time series (TS) #FoundationModels (FM) like Chronos zero-shot generalize to unseen #DynamicalSystems (DS)? #AI

No, they cannot!

But *DynaMix* can, the first TS/DS foundation model based on principles of DS reconstruction, capturing the long-term evolution of out-of-domain DS: arxiv.org/pdf/2505.13192v1

Unlike TS foundation models, DynaMix exhibits #ZeroShotLearning of long-term stats of unseen DS, incl. attractor geometry & power spectrum, w/o *any* re-training, just from a context signal.
It does so with only 0.1% of the parameters of Chronos & 10x faster inference times than the closest competitor.

It often even outperforms TS FMs on forecasting diverse empirical time series, like weather, traffic, or medical data, typically used to train TS FMs.
This is surprising, cos DynaMix’ training corpus consists *solely* of simulated limit cycles & chaotic systems, no empirical data at all!

And no, it’s neither based on Transformers nor Mamba – it’s a new type of mixture-of-experts architecture based on the recently introduced AL-RNN (proceedings.neurips.cc/paper_f), specifically trained for DS reconstruction.

Remarkably, DynaMix not only generalizes zero-shot to novel DS, but it can even generalize to new initial conditions and regions of state space not covered by the in-context information.

We dive a bit into the reasons why current time series FMs not trained for DS reconstruction fail, and conclude that a DS perspective on time series forecasting & models may help to advance the #TimeSeriesAnalysis field.

Replied in thread

@Techmeme This is the danger of closed source

These are knowledge models, and they only output what they are fed with

And no, they won’t magically develop ’reasoning skills’ and be able to sift through propaganda. NOT when it’s part of the training

To think otherwise means you don’t know shit how they work

They obey statistics. Training data for #ai #foundationmodels should be subject to public #academic scrutiny. Otherwise the models are bound to fall for flooding attacks

Herausforderung an der Schnittstelle von KI und Geodaten!
Die @Cyberagentur gibt am 31. Juli 2024 beim Online-Partnering-Event erste Einblicke in das neue Forschungsprojekt: „HEGEMON“.
Die Gelegenheit, sich zu vernetzen und sich für die wegweisende Forschung fit zu machen.
Mehr Infos und Anmeldung: t1p.de/m9vpi
#Cybersicherheit #FoundationModels #GenerativeAI #KI #Benchmarking #LLM #Multimodalitaet #Geodaten #GIS #Sicherheitstechnologie #Forschung #Innovation

#FoundationModels - Open for science and open to the world: Eine neue Generation von KI-Modellen, die Foundation Models, soll eine ganze Reihe von großen Herausforderungen in der Wissenschaft angehen und bewältigen.

In einer Session auf der @republica am 29.05.2024 um 10 Uhr stellen wir vier Pilotprojekte der Helmholtz Gemeinschaft vor und diskutieren mit euch über die Möglichkeiten und Grenzen dieser KI-Modelle. #rp24 #HFMI

#GenerativeAI, #FoundationModels, #LLMs, and all of that hokey nonsense shall not appear in my #robotics roadmaps as anything other than a neat research item until it can demonstrate a feasible path to #FunctionalSafety or mathematical completeness.

I lead #Product on the largest mobile-#robotic fleet known to humankind. I will not entrust decisions that could maim or kill to a pile of nondeterminate math prone to “hallucinations” or confabulation.

What's the most efficient way of working with #FoundationModels and #LLMs? Generally I have found that I like to keep machines busy. So, something is continuously training and running validations, while I'm preparing for the next experiments and improvements.

This means that the machine isn't waiting idle for my work and my work isn't waiting idle for the machine to finish. It means that I have to mentally keep track of runs started a while ago, to make maximal use of their results, even if I am already working on the next thing which is a moving target.

It means I have to carefully design the experiments or runs so that I know what knowledge I gain from them, so that I can add that knowledge to the pile of learnings even if the actual codebase has progressed from that point already.

It also means that I have to make multiple training or validation runs in parallel in a way that doesn't stop me from working on something else, while keeping the information of the runs somewhere so that they can be retrieved later. Also, it means that I often need to incorporate multiple learnings from these runs to the subsequent runs all together, in a YOLO sort of a way, instead of trying to do slow but structured "change one thing at a time" type of more systematic progress.

In my experience this is the most effective way to work on these types of things.

Keep machines running, do not wait for results, but make sure the results are useful when you eventually get them.

Good morning everyone! Here's my latest #Connections #Introduction #Introductions #TwitterMigration post, where I curate interesting accounts for you to follow from across the #Fediverse :fediverse:

@maryrobinette is a #writer #author, and I am listening to her incredible #LadyAstronaut series at the moment. If you love #SciFi (esp hard scifi) you should read it, too! 🇺🇸

@sayashk is a #ComputerScience #PhD candidate at #Princeton, who is researching failures in #ML (he's also co-running a workshop on open #FoundationModels in about 15 hours, see my previous posts for more info) 🇺🇸

@michcampbell is Dr Micha Campbell and she is a #PalaeoClimate #PostDoc living on #Dharawal country 🇦🇺

@mthv is a #Research #Engineer who works in #GIS at #CNRS 🇫🇷

@astrolori is Lori and she is into #OpenSource, #fashion, #space and #tech #WomenInSTEM 🇨🇦

@pandas_dev is the official account for #pandas, the #Python #DataAnalysis tool 🐍 📊

@jessie is a lover of #languages and helps run #CommonVoice, @mozilla 's open #voice #data set, which now supports over 100 languages. She also teaches #WebDev and loves #hiking. She's awesome you should follow her 🇬🇧

That's all for now, please do share your own lists so we can create deeper connections, and a tightly-connected community here

I'm reminded here of @maryrobinette's short story - "Red Rockets" - "She built something better than fireworks. She built community."