Top Story: Chinese National Arrested for Stealing AI Trade Secrets
Linwei Ding, also known as Leon Ding, was indicted by a U.S. federal grand jury on four counts of trade secret theft for allegedly attempting to transfer proprietary AI technology from Google to companies in China.
Ding, a Chinese national residing in California, is accused of moving over 500 confidential files from Google's network to his personal account while maintaining undisclosed affiliations with Chinese AI firms. The stolen secrets relate to Google's supercomputing data centers, which are crucial for AI model training. If convicted, Ding faces up to 10 years in prison and a $250,000 fine per count.
This case is part of a broader effort by the U.S. Disruptive Technology Strike Force to prevent the appropriation of sensitive technology by entities that threaten national security.
AI trade secrets are as important as nuclear secrets, many argue, and a broader debate about closed vs. open-source is underway. In fact, I just posted a lengthy video going over the different positions of the top minds in AI:
AI Trust Report
The Edelman Trust Institute just dropped a comprehensive report about the trust in artificial intelligence (among other topics). It includes dimensions by country, political group, income level, and more. Here are some of the most interesting slides:
US Army tests AI chatbots as battle planners in a war game simulation - Researchers in the US Army are experimenting with commercial AI chatbots as battlefield advisers in war game simulations – but experts caution that such AI should not be used in high-stakes situations
Biden wants to ban AI voice impersonation - He mentioned it during last night's State of the Union address but did not provide further details. Recently, robocalls that used an AI-generated voice clone of the President attempted to dissuade New Hampshire voters from voting, leading to an FCC prohibition of such calls. Is this an indication that Biden is hinting at a wider issue? The entertainment sector, currently dealing with the challenge of cloned musicians, actors, and comedians, will undoubtedly be keen to find out if that is the case.
Researchers jailbreak AI chatbots with ASCII art - Researchers have discovered a method called ArtPrompt to bypass safety measures in large language models (LLMs) such as GPT-3.5 and GPT-4. Their paper details how ASCII art can be used to cloak sensitive words that would normally trigger content restrictions, thereby tricking LLMs into engaging with topics like bomb-making or counterfeiting. The two-step ArtPrompt attack involves masking words and generating a replacement using ASCII art, which does not set off the LLMs' ethical or safety alarms. This technique is highlighted as a notable challenge to the robustness of current AI safety protocols.
SPONSOR
Release high-quality LLM apps quickly without compromising on testing. Never be held back by the complex and subjective nature of LLM interactions.
Try Deepchecks LLM evaluation now for free! https://bit.ly/3SVtxLJ
Zapier Central - Zapier Central is an experimental AI workspace where you can teach bots to work across 6,000+ apps.
Facebook Jepa - Self-supervised learning architecture for building visual representations from video. The representations generated are robust and can be applied effectively to various video and image tasks without needing parameter modification. V-JEPA does not rely on usual sources like human annotations or pre-trained encoders but instead leverages unsupervised feature prediction. It includes a feature-space prediction rather than pixel-level reconstruction.
Stability AI Triposer - TripoSR's 3D reconstructions are achieved through enhanced training data preparation, using diverse rendering techniques for better real-world image representation. The Objaverse dataset was selectively refined for quality, and the model itself boasts optimizations such as channel tuning and mask supervision, along with a revamped crop rendering approach. Detailed comparisons with OpenLRM and extensive technical specifics are accessible in the technical report.
Yi-9b - The Yi series models are the next generation of open-source large language models trained from scratch by 01.AI. Targeted as a bilingual language model and trained on 3T multilingual corpus, the Yi series models become one of the strongest LLM worldwide, showing promise in language understanding, commonsense reasoning, reading comprehension, and more.
Inflection 2.5 - Inflection has launched Inflection-2.5, an enhancement to its personal AI, Pi, blending high emotional intelligence (EQ) with increased intellectual abilities (IQ) to challenge the leaders in the large language model (LLM) field, such as GPT-4. This upgrade, developed with 40% less compute, promises improved performance in technical areas like coding and mathematics, along with integrated real-time web search for timely information.
Awesome Research Papers
An innovative vision transformer based on the transformer-based architecture used in large language models like LLaMA. VisionLLaMA, which exists in both plain and pyramid forms, is designed specifically for image processing tasks. It represents a unified framework for a variety of vision-related applications, with a focus on image perception and generation. Comprehensive evaluations reveal that VisionLLaMA surpasses prior state-of-the-art vision transformers in performance. It is poised to become a robust baseline model for research in vision generation and understanding, with the source code made available on GitHub.
The Orca-Math project by Microsoft demonstrates that smaller language models (SLMs) can perform complex tasks like grade school math problem-solving, traditionally a challenge for AI, at levels comparable to much larger models. Orca-Math, a 7 billion parameter model fine-tuned from Mistral 7B, impressively scores 86.81% on the GSM8K benchmark, outperforming bigger general and math-specific models. Its success is attributed to training on high-quality synthetic data of 200,000 math problems, created through a multi-agent system called AutoGen, and an iterative learning approach that relies on continuous practice and feedback without the need for external tools or costly ensembling techniques. The research underscores the efficacy of SLMs in specialized tasks and the significant potential of teacher-student iterative learning to enhance model performance.
This work introduces Chain-of-Abstraction (CoA), a new method empowering large language models (LLMs) to improve multi-step reasoning by planning with abstract reasoning chains. CoA trains LLMs to create reasoning chains with placeholders before invoking specific domain tools to fill in details, fostering more general strategies and robustness to domain knowledge shifts. It offers dual benefits: enhancing QA accuracy by ~6% over current methods, and accelerating inference speed by ~1.4 times, as validated in math reasoning and Wiki QA domains. This approach mitigates the inference delay typically incurred when LLMs wait for external tool responses.
This research investigates the Design2Code task, where multimodal Large Language Models like GPT-4V convert visual webpage designs into code. A benchmark of 484 web pages was created for assessment using automatic evaluation metrics and human evaluations. Results showed GPT-4V outperforming other models, with about half of its generated webpages being visually and content-wise on par with, or better than, the originals. However, open-source models fall behind in element recall and layout accuracy. While certain attributes like text and color benefit from fine-tuning, overall performance suggests generative AI is increasingly capable of transforming design into functional code.
The paper identifies a significant safety concern in large language models (LLMs): the presumption that their interpretation of content is purely semantic, which is not always the case in the real world. For instance, ASCII art in online forums can convey non-semantic information, revealing a vulnerability. The authors introduce a novel ASCII art-based attack (ArtPrompt) and the Vision-in-Text Challenge (ViTC), a benchmark for evaluating LLMs on non-semantic prompt recognition. They demonstrate that leading LLMs like GPT-3.5, GPT-4, and others fail to process ASCII art accurately, making them susceptible to ArtPrompt. This attack, executed with only black-box access, can effectively compromise the models' safety measures to trigger undesired behaviors.
