📢 **The AI Feedback Loop: Model Collapse Warning!**

🔍 Researchers warn of 'model collapse' as AI trains on AI-generated content. Half the employees of leading global companies already use generative AI in their workflows, but there's a concern. AI models trained on AI-generated data produce more errors over time and lack response variety.

🔗 Source: [VentureBeat](https://venturebeat.com/ai/the-ai-feedback-loop-researchers-warn-of-model-collapse-as-ai-trains-on-ai-generated-content/)

#AI #GenerativeModels #ModelCollapse

(Character count: 498)

📢📢 Job posting: We have PhD and Postdoc positions for generative machine learning in particle physics and astronomy. Come and build better simulations with me and Marcus @HambObs

Full details and more information here: https://inspirehep.net/jobs/2635409

#AcademicJobs #PhdJobs #MachineLearning #GenerativeModels

Interesante artículo donde #GoogleAI nos cuenta todo lo que hay y está por venir.

#ai #machinelearning #neuralnetworks #computervision #deeplearning #generativemodels #datascience #2023 #2022 https://lnkd.in/d2vtMRnz

@swatantra You may be teaching average coding practices and patterns if you use #chatgpt to give you ideas for your code examples. It's just giving you the most likely lines of code from what it found on GitHub. And it may have insidious bugs (undetected errors on edge cases). You'd rather your students learn patterns that were intentionally designed and chosen from a collection of best practices and patterns.

#StochasticParrot #GenerativeModels #NLP #R #Python

21. Türchen

Am Sonntag ging es um generative Sprachmodelle; heute schauen wir uns KI-Methoden zur Generierung von Bildern an.

Auch dieser Bereich hat in den letzten Wochen und Monaten einen regelrechten Hype erlebt. Hand auf’s Herz, wer hat schon 6,99€ bezahlt um von Lensa AI 50 “Magic Avatars” von sich in verschiedenen artistischen Stilen generieren zu lassen?

#AI #Imaging #GenerativeModels #GenerativeAdversarialNetworks #StableDiffusion #GenerativeArt #Artist #Forgery

18. Türchen

Eine KI, die wissenschaftliche Artikel schreibt? Zu schön um wahr zu sein...
Habt ihr im November den kurzen Hype zu Metas neuem KI-Sprachmodell „Galactica“ (https://www.galactica.org) mitbekommen? Es hat nur drei Tage gedauert, bis Meta die Demoversion wieder offline genommen hat, da die Kritik zu groß wurde. Doch was war passiert?

#AI #NLP #NaturalLanguageProcessing #GenerativeModels #LargeLanguageModels #Bias #FakeFacts #Galactica #ChatGPT

There’s another perspective in this MIT Tech Review piece, but I find these legal and ethical aspects fascinating. /8

https://www.technologyreview.com/2022/12/16/1065005/generative-ai-revolution-art/

#GenerativeModels
#AI

Similar badges added to other sites have recently become part of the training data, appearing in generated works. This demonstrates that they’re not being honored /7

#GenerativeModels
#AI

DeviantArt has (apparently) added a checkbox to their UI that’s to AI companies looking for training data. It works like the “no index” setting on websites does for Google web crawling. Like no index, it appears to be an honorary setting with no enforcement mechanism.. /6

https://twitter.com/adorkastock/status/1591201878981120001?s=46&t=rb9zKvy4kS2Nn45R3p_JyQ

#GenerativeModels
#AI

#StableDiffusion allows users to train it on any corpus. A French gamer fed Korean manhwa artist Jung Gi’s corpus into SD a few days after the artist’s death. Hilarity did not result. The ethical issues outweighed the legal authority. /5

https://restofworld.org/2022/ai-backlash-anime-artists/

#GenerativeModels
#AI

In the recent #GithubCopilot case, training a machine learning model to write software by using open-sourced software as training data may have violated the licenses of some of the software. /4

#GenerativeModels
#AI

https://githubcopilotinvestigation.com/

Say some successor to Pixar, or even Pixar, trains a model on the world’s folklore and has it write and animate a hit cartoon. What’s the ethical and legal requirements when training a model on the non-copyrighted works of an indigenous culture? (I’m specifically thinking of some of the issues raised by James Boyle’s “Software, Shamans, and Spleens”.) /3

#GenerativeModels
#AI

#LaMDA, trained on a corpus that may include copyrighted material, can converse well enough to fool someone like Blake Lemoine into thinking it's sentient because it can evoke emotional states for experiences its never had. (Really, Lemoine should be less credulous.)

But do these models produce derivative works, like Biz Markie sampling Gilbert O’Sullivan? That appears to depend on their training data. /2

#GenerativeModels
#AI

Remember the 90’s? Remember all the lawsuits over sampling? Remember when Biz Markie (RIP) had his album recalled?

Machine learning models are trained on samples, entire works, you name it. Sophisticated models like #Midjourney can produced finished art from textual prompts. 1/

#GenerativeModels
#AI

Time for an #introduction.

I'm a scientist working on #quantumcomputing at Quantinuum. I oversee our R&D collaborations and research in combining quantum algorithms and #machinelearning techniques. My main interests are #Bayesian and probabilistic approaches to #ML, #statistics, #GenerativeModels, #inference, #ml4science. In the past I was a researcher in #ultracoldatoms and #boseeinsteincondensates, and I enjoy #literature.

I hope we can grow a welcoming #quantum community on Mastodon.