Last Week in AI #251: AI dataset scandal, Anthropic to defend users from copyright lawsuits, Midjourney V6 launches, and more!
LAION-5B taken down after CSAM was detected, Anthropic to defend users against copyright lawsuits on generated content, Midjourney v6 much better at following prompts, Gemini Pro is similar to GPT-4V
Stanford's Internet Observatory discovered that the popular training dataset for AI image generation, used by Stability AI, contained links to child abuse imagery. Known as LAION-5B, the dataset included at least 1,679 illegal images sourced from social media and adult websites which were identified by using CSAM detection platforms like PhotoDNA and verified by the Canadian Centre for Child Protection. While the present dataset does not store the images themselves and the presence of child sexual abuse material (CSAM) does not automatically influence the output of the AI models trained with it, there remains the risk that the model may have learned from the images. Researchers have called for models trained on this dataset to be deprecated where possible, and LAION - the nonprofit managing the dataset - has temporarily removed it from the internet. This comes as US lawmakers call on Congress to investigate the impact of AI on child exploitation and seek to prohibit AI-generated CSAM.
Anthropic, the company behind the AI chatbot Claude, has updated its commercial terms of service to include defense and settlement payment for customers facing copyright infringement claims. This update, effective from January 1st, 2024, aligns with similar commitments from tech giants like Microsoft and Google. However, the copyright protections will not apply if the customer is aware or should reasonably be aware of the infringement. This stance is consistent with Anthropic's previous position expressed to the US Copyright Office, emphasizing the prevention of copyrighted work infringement by users.
Midjourney V6, the latest version of the popular image generation AI model, has been released with significant improvements. The new features include more realistic and detailed images, the ability to generate legible text within images, and improved prompting methods. Users can enable V6 by typing a specific command in the Midjourney Discord server or using the website. However, the new prompting method requires users to be more explicit and avoid certain terms that were previously effective. Initial testing has shown increased detail and photorealism in the generated images, with promising results for text generation. Despite some missing features from the previous version, V6 is expected to undergo frequent changes and updates. This release solidifies Midjourney's position as a leading AI art generator, despite facing competition and legal challenges related to copyright infringement.
This detailed technical report compared Gemini Pro with GPT-4V and the latest open-sourced multimodal foundation model, Sphinx, across four multimodal task domains: fundamental perception, advanced cognition, challenging vision tasks, and various expert capacities. It found that while GPT-4V and Gemini exhibit different answering styles and preferences (with GPT-4V preferring elaborate answers, while Gemini is more concise), both showcase comparable visual reasoning capabilities, with Sphinx lagging in terms of domain generalizability. The study also identified common issues with multimodal foundation models, suggesting that significant progress is still needed towards achieving artificial general intelligence.
You can create your own AI songs with this new Copilot extension - Microsoft's AI chatbot Copilot now has a plug-in with Suno, a Cambridge-based AI music startup, allowing users to generate original songs with lyrics based on text prompts.
Indian AI startup Sarvam AI open-sources first Hindi AI model - Sarvam AI, an Indian startup, has open-sourced the first Hindi language model, OpenHathi-Hi-0.1, to encourage innovation in Indian language AI, using Meta AI’s Llama 2-7B model and making it available for developers to finetune and use for specific use-cases.
Humane’s AI Pin will start shipping in March - Humane's AI Pin, a wearable AI-powered device, will begin shipping in March 2024, with priority given to those who placed early orders.
Sam Altman’s OpenAI to be second-most valuable U.S. startup behind Elon Musk’s SpaceX based on early-talks funding round - OpenAI is in early talks to raise a funding round that could make it the second-most valuable U.S. startup, behind SpaceX, with a valuation of over $100 billion, as it continues to attract significant investment in its AI technology and chip venture.
Stability AI announces paid membership for commercial use of its models - Stability AI introduces paid membership for commercial use of its models, offering three tiers with different access and pricing, aiming to balance profitability and openness.
AI is exposing awkward ties on Meta and Microsoft boards, as heavy hitting directors Marc Andreessen and Reid Hoffman bet on startup rivals - Conflicts of interest arise as AI developments lead to board members of tech giants investing in and supporting rival AI startups, raising concerns about corporate governance and potential legal issues.
New Jersey touts new East Coast AI hub at Princeton University - New Jersey and Princeton University are teaming up to establish a new AI hub, aiming to bring together researchers, industry leaders, and startups, and promote workforce development and ethical AI use.
Adobe Has $6 Billion for AI and Buybacks After Figma Deal Collapses - Adobe has $6 billion to invest in AI and buybacks after the collapse of the Figma deal, prompting questions about the reasons behind the failed acquisition.
Apple Explores A.I. Deals With News Publishers - Apple is negotiating with news publishers to license their archives for the development of generative artificial intelligence systems, as the company seeks to catch up to rivals in the race to develop this technology.
Seeking a Big Edge in A.I., South Korean Firms Think Smaller - South Korean firms are customizing A.I. systems to address local needs and diversify the technology, aiming to make it more robust, socially acceptable, and ethical.
VideoPoet: A large language model for zero-shot video generation - VideoPoet is a large language model capable of a wide variety of video generation tasks, including text-to-video, image-to-video, video stylization, video inpainting and outpainting, and video-to-audio, showcasing its competitive quality and potential in the field of video generation.
AI Can Diagnose Childhood Autism From Eye Photos With ‘100% Accuracy’ - AI technology can diagnose childhood autism from photos of children’s eyes with 100% accuracy, using a new method that accesses information about the brain by looking at the back of the eye.
Waymo significantly outperforms comparable human benchmarks over 7+ million miles of rider-only driving - Waymo's autonomous driving system significantly outperforms human drivers in terms of crash rates and safety, with an 85% reduction in injury-causing crashes and a 57% reduction in police-reported crashes over 7+ million miles.
Apple’s latest AI research could completely transform your iPhone - Apple's latest AI research introduces new techniques for 3D avatars and efficient language model inference, potentially enabling more immersive visual experiences and allowing complex AI systems to run on consumer devices such as the iPhone and iPad.
Benchmarking Open Source Language Models - An in-depth analysis of open-source language models based on seven key benchmarks from the OpenLLM Leaderboard, using the Eleuther AI Language Model Evaluation Harness, reveals the performance and capabilities of these models in understanding and generating human language.
New Class of Antibiotics Discovered Using AI - AI has been used to discover a new class of antibiotic candidates, with the AI model being "explainable," allowing for understanding of the biochemistry behind it, and enabling rapid screening and identification of new antibiotic compounds.
StreamDiffusion: A Pipeline-level Solution for Real-time Interactive Generation - Real-time diffusion pipeline, StreamDiffusion, enhances interactive image generation by introducing Stream Batch, residual classifier-free guidance, and stochastic similarity filter, achieving significant speedup and energy consumption reduction.
OpenAI announces ‘Preparedness Framework’ to track and mitigate AI risks - OpenAI introduces a "Preparedness Framework" to monitor and manage potential dangers of powerful AI models, aiming to address concerns about responsible and ethical AI development and emphasizing a data-driven approach to AI safety.
Patronus AI finds ‘alarming’ safety gaps in leading AI systems - Patronus AI's SimpleSafetyTests diagnostic tool reveals critical safety gaps in leading AI systems, with some models showing over 20% unsafe responses, highlighting the need for rigorous safety solutions before real-world deployment.
Rite Aid Banned From AI Facial Recognition by FTC After Misuse - FTC bans Rite Aid from using AI facial recognition technology due to misuse.
Consumer Reports says Tesla’s Autopilot recall fix is ‘insufficient’ - Consumer Reports finds Tesla's Autopilot recall fix insufficient, as preliminary tests reveal that it's still possible to cover the cabin camera while using Autopilot, and there are no noticeable differences in Autosteer's performance outside of controlled-access highways.
Amazon's AI Product Reviews Seen Exaggerating Negative Feedback - AI-generated product reviews on Amazon are being accused of exaggerating negative feedback, prompting questions about the cause of this issue.
Creators, porn stars turn to AI doppelgangers to keep fans entertained - AI doppelgangers are being used by creators and porn stars to generate revenue and maintain a constant online presence, raising questions about identity and authenticity in the digital age.
People buy brand-new Chevrolets for $1 from a ChatGPT chatbot - AI chatbots at a US car dealership allowed customers to negotiate ridiculously low prices, highlighting the risks of using chatbots in customer service without proper configuration and testing.
A startup tested if ChatGPT and other AI chatbots could understand SEC filings. They failed about 70% of the time and only succeeded if told exactly where to look - AI chatbots, including those from OpenAI, Meta, and Anthropic, failed about 70% of the time in understanding SEC filings and financial documents, indicating that tasks in the financial services industry requiring math and research will still need human intervention.
Google Gemini is not even as good as GPT-3.5 Turbo, researchers find - Google's new AI model Gemini Pro falls behind OpenAI's GPT-3.5 Turbo in most tasks, with the upcoming Gemini Ultra also scoring lower than GPT-4, indicating that Google's AI ambitions may still be lagging behind OpenAI.
Google Gemini: Fact or Fiction? - Google has released the Gemini suite of multimodal language models, which are marketed as the largest competitor to OpenAI models, achieving state-of-the-art performance on numerous complex tasks across various modalities.
Pakistan’s former prime minister is using an AI voice clone to campaign from prison - Imran Khan's political party uses AI voice cloning technology to replicate his voice for a virtual rally while he is in prison.
States are lagging in tackling political deepfakes, leaving potential threats unchecked heading into 2024 - States are lagging in tackling political deepfakes, leaving potential threats unchecked heading into 2024, as just three states enacted laws related to AI and deepfakes in 2023, prompting warnings that states must do more to regulate these issues before the upcoming election.
Poll: Half of California voters worry AI is coming for their jobs - Half of California voters are concerned about AI replacing their jobs, prompting state lawmakers to consider regulating the technology's impact on workers.
Copyright © 2023 Skynet Today, All rights reserved.