ODSC’s AI Weekly Recap: Week of January 12th

Artificial intelligence has been moving at the speed of light with all of the news that has come out. So let’s recap what we covered here at ODSC and other stories that we may have missed so you can stay in the know of all things AI.

Subscribe to ODSC’s AI Weekly Recap newsletter here!

Open Data Science Blog Recap:

Baidu’s Ernie Bot Has More than 100 Million Users According to CTO According to the chief technology officer of Baidu, its ChatGPT-like chatbot, Ernie Bot has gained over 100 million users. This milestone was announced during a deep learning summit in Beijing

New Method Hopes to Understand Multi-Attacks Against Image Classification Systems A recent study has delved deep into the complexities of these attacks, particularly focusing on the alarming trend of multi-attacks that can simultaneously manipulate the classifications of multiple images.

Judges in England and Wales Have Given the Green Light for the Use of AI in Writing Legal Opinions

The Courts and Tribunals Judiciary said that AI can now be used to help write legal opinions. But, though AI has been given the green light

Deloitte Provides Employees with New AI Chatbot, PairD

Deloitte is providing 75,000 employees across Europe and the Middle East with a new generative AI chatbot aimed at assisting with emails, PowerPoints, and more.

AI News Highlights

A 4,700-person list of famous artists revealed to be used by AI image generators has gone viral.

The New York Times reporting on how AI-power robots and Chatbots are poised to transform 2024

New research has found a way to utilize the power of artificial intelligence to bring efficiency to the crowdsourcing process by targeting ideas.

Is AI the new Manhattan Project? A leading AI expert calls on the EU to build a publicly funded AI system or risk foreign dependence.

Midjourney V6, released in alpha and marks a significant advancement in AI-driven art generation with a focus on better text rendition and a more literal interpretation of prompt. The updates in V6 may provide major leap in photorealism and the introduction of in-image text capabilities.

Anthropic has significantly raised its revenue projections, now expecting to surpass $850 million in annualized revenue by the end of 2024. This forecast marks a substantial increase from their earlier prediction of $500 million.

DeepMind’s Next Leap in Robotics: AutoRT, SARA-RT, and RT-Trajectory to Revolutionize Robot Data Collection and Navigation Skills.

Microsoft quietly launches dedicated Copilot app for Android.

GitHub Copilot Chat now generally available for organizations and individuals.

JPMorgan research release paper on DocLLM ,a new large language mode for visually rich multimodal document intelligence. Its specifically designed to understand the layout and content of complex financial documents like reports and analyses.

The Financial Times thinks it’s already time to think about an AI tax to offset job displacement and ensure AI companies share the social costs.

OpenAI says New York Times ‘manipulated’ ChatGPT in copyright feud

Fast Company — AI Pin startup Humane is already laying off staff before its much-hyped device is even released

Trending AI Open Source Projects

Fairy — Fairy, is a minimalist yet robust adaptation of image-editing diffusion models, enhancing them for video editing applications.

LM Drive — Researchers have developed “LMDrive,” a novel autonomous driving framework that combines multi-modal sensor data with natural language instructions, enhancing vehicle-human interaction and navigation in complex scenarios. A first-of-its-kind approach to integrate LLMs for autonomous driving.

TinyGPT-V, developed by researchers from Anhui Polytechnic University, Nanyang Technological University, and Lehigh University. represents a significant advancement in Multimodal Large Language Models (MLLMs). This AI model claims a combination of high performance and reduced computational demands. Built upon Microsoft’s Phi-2.

OpenVoice is a voice clone project built by researchers at MIT and MyShell. Focused on voice processing and voice recognition from short audio clips its has applications in areas such as automated speech recognition, voice-enabled user interfaces, and other voice-driven technologies.

Research From Around The Globe

SecFormer: Towards Fast and Accurate Privacy-Preserving Inference for Large Language Models

SecFormer is a new method to achieve fast and accurate PPI for Transformer models that address growing privacy concerns related to bank account details and investment plans.

TinyLlama: An Open-Source Small Language Model

TinyLlama, a compact 1.1B language model pretrained on around 1 trillion tokens for approximately 3 epochs that are built on the architecture and tokenizer of Llama 2,

Fairness in Serving Large Language Models

A team of researchers from UC Berkeley, Stanford University, and Duke University has introduced a groundbreaking fair scheduler that is specifically designed for LLM serving. This approach functions at the level of individual tokens, providing a more precise and adaptable solution.

Improving Text Embeddings with Large Language Models.

This paper introduces a novel method for generating high-quality text embeddings using synthetic data and less than 1,000 training steps. This approach differs significantly from existing methods that typically rely on multi-stage pre-training with large amounts of weakly-supervised text pairs and further fine-tuning with limited labeled datasets.

A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models

This paper addresses a critical challenge in the field of AI and natural language processing: the tendency of Large Language Models (LLMs) to produce hallucinated content. The paper conducts a survey of over 32 techniques developed to mitigate hallucinations in LLMs, highlighting key approaches like Retrieval Augmented Generation, Knowledge Retrieval, CoNLI, and CoVe.

Task Contamination: Language Models May Not Be Few-Shot Anymore.

This paper studies that LLMs tend to perform better on datasets created before their training data was compiled, compared to datasets released afterward. This finding suggests the presence of task contamination, where LLMs are indirectly exposed to similar tasks during their training, thus skewing their performance in zero-shot and few-shot evaluations.

Start-Up Funding News

Palo Alto-based Luma AI a generative AI startup that enables users to transform text descriptions into corresponding 3D models, raised north of $68M in a Series B round.

Paris-based Nabla developer of Nabla Copilot, an ambient AI that helps clinicians get enjoyment in their work again, raised $44M in a Series B round.

Milan-based Contents.com, a Generative AI company that has developed a SaaS platform to support any business in content creation, raised north of 26M in a Series B round.

Linthicum Heights-based Impact Analytics, a business intelligence company that utilizes AI to help businesses make decisions, raised north of $61M in a Series C round.

ExtraHop, a Seattle-based leader in AI-based Solution for Cybersecurity, raise $100M.

Hyperexponential, a London-based provider of pricing decision intelligence (PDI) software for the commercial insurance sector based in London, raised a $73M Series B round.

OpenAI Users Wishlist Per Sam Altman

Some amazing stuff is going on in the world of AI. Keep up with the latest news by stopping by ODSC.com as often as possible. Don’t forget, now is your time to get your passes to ODSC East. There, you’ll engage with the leaders driving the industry forward, and have opportunities to upskill and build your network so you can push your career forward.