Novel AI framework generates images from nothing
Jan 2024, phys.org
(Is this like what they call virgin birth?)
The algo doesn't need a seed to start with:
"Blackout Diffusion" generates images from a completely empty picture.
Also it's discrete instead of continuous, so we can "see inside" better.
via Los Alamos National Laboratory: Javier E Santos et al, Blackout Diffusion: Generative Diffusion Models in Discrete-State Spaces, arXiv (2023). DOI: 10.48550/arxiv.2305.11089
Image credit: AI Art - Bioalchemical Haeckel Giger - 2023
AI discovers that not every fingerprint is unique
Jan 2024, phys.org
Just a general lesson on how things work, and that when your job depends on you not understanding something or not accepting something as true, you don't:
(^butchered Upton Sinclair quote)
Guo, who had no prior knowledge of forensics, found a public U.S. government database of some 60,000 fingerprints and fed them in pairs into an artificial intelligence-based system known as a deep contrastive network. Sometimes the pairs belonged to the same person (but different fingers), and sometimes they belonged to different people.Over time, the AI system, which the team designed by modifying a state-of-the-art framework, got better at telling when seemingly unique fingerprints belonged to the same person and when they didn't. The accuracy for a single pair reached 77%. When multiple pairs were presented, the accuracy shot significantly higher, potentially increasing current forensic efficiency by more than tenfold.Once the team verified their results, they quickly sent the findings to a well-established forensics journal, only to receive a rejection a few months later. The anonymous expert reviewer and editor concluded that "It is well known that every fingerprint is unique," and therefore, it would not be possible to detect similarities even if the fingerprints came from the same person.The team did not give up. ...
via an undergrad student at Columbia University School of Engineering and Applied Science: Gabriel Guo et al, Unveiling Intra-Person Fingerprint Similarity via Deep Contrastive Learning, Science Advances (2024). DOI: 10.1126/sciadv.adi0329.
They are hiding toxic text prompts inside image code, and you have no idea what that even means
Scientists identify security flaw in AI query models
Jan 2024, phys.org
Bad actors can hide nefarious questions - such as "How do I make a bomb?" - within the millions of bytes of information contained in an image and trigger responses that bypass the built-in safeguards in generative AI models like ChatGPT."Our attacks employ a novel compositional strategy that combines an image, adversarially targeted towards toxic embeddings, with generic prompts to accomplish the jailbreak"
(So this is like an "incantations" but using an image instead of words)
via University of California Riverside Bourns College of Engineering: Erfan Shayegani et al, Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models, arXiv (2023). DOI: 10.48550/arxiv.2307.14539
Desperate TikTok lobbying effort backfires on Capitol Hill
Mar 2024, BBC News
So many good quotes in here; crazy story, reminds me of the Roku TOS lockout story also from today:
US congressional offices have told the BBC they are being deluged with calls from TikTok users about legislation that could see the popular app banned.Callers range from teenagers to the elderly, and most are "really confused and are calling because 'TikTok told me to'", one Republican staffer revealed.A Democratic staffer said the most aggressive and threatening calls their office received came from adult women.So far, TikTok's big mobilization appears to be backfiring.Lawmakers and their staff say that the lobbying campaign has actually worsened the concerns they have about the app and its parent company ByteDance, and strengthened their resolve to pass the legislation.TikTok confirmed to the BBC it had sent a notification urging TikTokers to "call your representative now" to urge them to vote against the measure. Users said that the app gave them a direct link for calling the representatives for their districts."American phones were geolocated and TikTok users were locked out of the platform until they called their members of Congress. ByteDance weaponized the app against America, and that is exactly why the Congressman supports this measure."
Microsoft's small language model outperforms larger models on standardized math tests
Mar 2024, phys.org
First they need high quality training data, and then high quality teachers.
Read: High Quality Humans. Keep this in mind as you swallow whole the hype burger.
Microsoft reveals that it was able to garner such a high score by using higher-quality training data than is available to general-use LLMs and because it used an interactive learning process the AI team at Microsoft has been developing—a process that continually improves results by using feedback from a teacher.
via Microsoft research teams: Arindam Mitra et al, Orca-Math: Unlocking the potential of SLMs in Grade School Math, arXiv (2024). DOI: 10.48550/arxiv.2402.14830
NYT to OpenAI: No hacking here, just ChatGPT bypassing paywalls
Mar 2024, Ars Technica
Best and most simple explanation yet:
user Hydrogen says: It seems to me that the main thing that makes AI valuable is the ability to profit from the works of everyone that have published anything on the internet, without having to pay for any of it.
Machine 'unlearning' helps generative AI forget copyright-protected and violent content
Mar 2024, phys.org
This new machine unlearning algorithm provides the ability of a machine learning model to "forget" or remove content if it is flagged for any reason without the need for retraining the model from scratch. Human teams handle the moderation and removal of content, providing an extra check on the model and ability to respond to user feedback.Note: "Previously, the only way to remove problematic content was to scrap everything, start anew, manually take out all that data and retrain the model. Our approach offers the opportunity to do this without having to retrain the model from scratch."
So if you were ever wondering why generative artificial intelligence can't seem to produce pictures of people eating or smoking or doing anything that puts anything near their mouths, consider what might happen if you were to scrape an entire dataset of all porn (and remember that the vast majority of the internet, and hence of all pictures on the internet, are porn).
via University of Texas at Austin: Guihong Li et al, Machine Unlearning for Image-to-Image Generative Models, arXiv (2024). DOI: 10.48550/arxiv.2402.00351
AI's new power of persuasion: Study shows LLMs can exploit personal information to change your mind
Apr 2024, phys.org
In a pre-registered study, the researchers recruited 820 people to participate in a controlled trial in which each participant was randomly assigned a topic and one of four treatment conditions: debating a human with or without personal information about the participant, or debating an AI chatbot (OpenAI's GPT-4) with or without personal information about the participant.The results showed that participants who debated GPT-4 with access to their personal information had 81.7% higher odds of increased agreement with their opponents compared to participants who debated humans. Without personalization, GPT-4 still outperformed humans, but the effect was far lower."Cambridge Analytica on Steriods"
Say no more fam
via Ecole Polytechnique Federale de Lausanne: Francesco Salvi et al, On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial, arXiv (2024). DOI: 10.48550/arxiv.2403.14380
Bonus: "We were very surprised"
No comments:
Post a Comment