New Bard Release Surpasses GPT-4

Author

AI Breakfast
January 29, 2024

Good morning. It’s Monday, January 29th.

In Partnership with Blaze AI

Did you know: On this day in 1988, the first version of Tetris to be published in the United States was released by Spectrum HoloByte?

In today’s email:

  • Advances in AI Technology and Models

  • Corporate AI Strategy and Financial Decisions

  • AI Impact on Industry and Market Dynamics

  • 5 New AI Tools

  • Latest AI Research Papers

  • ChatGPT Creates Comics

You read. We listen. Let us know what you think by replying to this email.

Today’s trending AI news stories

Advances in AI Technology and Models

> Google's latest iteration of its Bard Gemini AI model has achieved a notable milestone by matching the performance of OpenAI's GPT-4, as assessed in human evaluations. Jeff Dean, the head of Google's AI division, announced this new model, intriguingly dubbed "scale," indicating a possible expansion of the previous Pro model's capabilities. The Gemini Pro "scale" has impressively secured the second position in the Chatbot Arena leaderboard, outperforming earlier versions of GPT-4, yet trailing slightly behind the most recent GPT-4 Turbo version. Oriol Vinyals of Google emphasized the nuanced challenges in assessing language models, advocating for human evaluations over academic benchmarks. The forthcoming release of the even more potent Gemini Ultra model is eagerly awaited, with expectations that it might surpass the achievements of the Pro-Scale version.

> OpenAI's latest feature in ChatGPT, the "@GPT" function, introduces the innovative concept of multi-GPT conversations within a single chat window. This beta feature significantly enhances user interaction by allowing the integration of different GPTs' expertise and responses in real-time. This advancement is a key step in realizing OpenAI's ambition of creating a versatile, universally applicable assistant, tailored to individual preferences and needs. It embodies Sam Altman's vision of a customizable and personal AI, promising to revolutionize the user's digital interaction experience.

> Midjourney has rolled out its V6 update, enhancing the user interface with innovative features like pan, zoom, and vary (Region), offering a more intuitive and detailed image creation process. The updated pan function now closely resembles zoom, delivering results with greater clarity and variety, while minimizing repetitive elements. These improvements are integrated with existing upscale, vary (region), and remix tools, available on the V6 bot via Discord and Midjourney's alpha website. Significantly, the alpha website is now accessible to prolific users who have generated over 5000 images, a milestone that can be checked using the "/info" command on Discord. In a move to prioritize development based on user experience, Midjourney has also introduced a feedback feature, encouraging user engagement and shaping future updates.

> Researchers at Stanford University and OpenAI have developed a new technique called meta-prompting, which significantly boosts the problem-solving abilities of large language models. This innovative approach involves breaking down complex challenges into simpler elements, which are then addressed by specialized segments of the same model, each following distinct guidelines. The primary model functions as a coordinator, harmonizing the efforts of these segments and integrating their contributions. Although meta-prompting enhances performance in logical and creative tasks, it does come at a higher computational cost and is hindered by its linear process, limiting its efficiency. Despite these challenges, this method represents a notable leap forward in the field of AI, particularly in models such as GPT-4.

> Apple is reportedly experimenting with an AI-enhanced version of Siri, potentially to be unveiled at WWDC 2024. This update follows Apple's alleged $1 billion investment in integrating AI into Siri. The beta code reveals references to OpenAI and hints at Apple testing four different AI models, including two versions of its in-house Large Language Model, AjaxGPT. Additionally, a model named FLAN-T5 is under evaluation. These advancements suggest Apple's efforts to bolster Siri's capabilities, using ChatGPT as a benchmark. This initiative indicates a significant enhancement in Siri's functionality, aligning with Apple's broader AI strategy and commitment to improving user experience with its virtual assistant.

Corporate AI Strategy and Financial Decisions

> Elon Musk has publicly refuted claims that his AI firm, xAI, is seeking investor funding. This clarification came after a report by the Financial Times suggested xAI was aiming to raise up to $6 billion, valuing the startup at $20 billion. Musk's denial followed a previous dismissal of rumors that xAI had garnered $500 million towards a $1 billion funding goal. xAI's launch of "Grok," a chatbot rivaling OpenAI's ChatGPT, aligns with Musk's known interest in AI safety. Despite stepping down from OpenAI's board in 2018, Musk remains vocal about strict AI controls, especially regarding Tesla's AI endeavors, where he seeks significant voting influence.

> Tesla has announced a $500 million investment to construct a "Dojo" supercomputer at its Buffalo, New York factory. This initiative, revealed by New York Governor Kathy Hochul, is part of Tesla's broader ambition to develop self-driving cars. The Dojo supercomputer, first introduced at Tesla's 2021 AI Day, is designed to process extensive video data from Tesla vehicles to train the AI for its Full Self-Driving Beta software. This move marks a significant pivot in Tesla's utilization of the Buffalo facility, originally intended for Solar Roof tile production. Elon Musk has described Dojo as a high-risk, high-reward project with the potential to significantly impact AI, emphasizing its importance despite the substantial investment in Nvidia hardware.

AI Impact on Industry and Market Dynamics

> Netflix, in its latest annual report to the SEC, has identified generative AI as a new risk factor, reflecting the technology's escalating influence and potential disruption in the media industry. The company acknowledges that advancements in AI, particularly generative AI, could give competitors a technological edge, possibly impacting Netflix's ability to compete effectively. Additionally, Netflix highlighted concerns about the ambiguous legal territory of intellectual property rights in AI-generated content, indicating a shift towards acknowledging the complex challenges AI poses to traditional content creation and copyright norms.

> China has rapidly advanced in the field of artificial intelligence by authorizing more than 40 AI models for public deployment. This move is part of a broader strategy to challenge the United States' dominance in AI technology. The recent approvals, which include various large language models, highlight significant participation from major Chinese tech firms like Xiaomi and 4Paradigm. Initiated in August, this development underscores China's ambition to not only foster AI innovation but also to ensure these advancements remain under regulatory control, reflecting a cautious yet ambitious approach to harnessing AI's potential.

In partnership with BLAZE AI

Blaze is an AI tool that helps teams of one create better content in half the time — all in their brand voice. Blaze is the "missing piece of Canva" because it helps people who have to do marketing, but who aren't professional marketers.

With Blaze, teams of one can:

  • Brainstorm and outline content ideas with AI.

  • Go from idea to unique content in minutes (using their own Brand Voice).

  • Turn a blog post into 10+ asset types—ads, social posts, emails, and more.

And unlike other simple AI content tools, Blaze is built on a truly modern document editor — with real-time collaboration, tracked changes, guest-sharing, approvals, mobile editing, and more.

At just $20/month (with a 7-day free trial), you'll instantly make your money back producing high-quality content in 50% less time.

Thank you for supporting our sponsors!

5 new AI-powered tools from around the web

Sune AI is a collaborative AI operating system (aiOS) for team productivity, enhancing desktop workflows across documents with AI integration. Enables seamless collaboration, document editing, automation, and workflow creation.

ChatDox is an AI tool for extracting information from PDFs, DOCXs, TXTs, CSVs, and YouTube videos. Users can ask specific questions for instant answers.

XspaceGPT, leveraging GPT-4.0 AI, converts Twitter Spaces into MP3s, text transcriptions, summaries and mind maps, enhancing accessibility and content repurposing for users and creators, simplifying engagement with live discussions.

HyperHuman is an advanced tool for generating animatable 3D faces via text. It is compatible with Blender, Unreal Engine, Unity, and Maya.

Atlabs is an AI-driven creative studio with features that include AI Script Writer, Motion Graphics, and multilingual voiceovers, streamlining video production for enhanced business growth.

arXiv is a free online library where researchers share pre-publication papers.

Developed by Microsoft and various universities, EAGLE accelerates Large Language Models (LLMs) using speculative sampling at the feature level, addressing next-feature prediction uncertainties. This innovative approach maintains the original LLM output distribution while tripling generation speed compared to standard auto-regressive decoding. EAGLE outperforms methods like Lookahead and Medusa, showing effectiveness across LLMs, including Vicuna and LLaMA2-Chat, in tasks like dialogue and mathematical reasoning. Its low training cost and simple integration make it practical for accelerating LLMs efficiently.

Research teams from the University of Toronto and Google Deepmind introduces GenEM, a method leveraging LLMs for creating adaptive, expressive robot behaviors. GenEM translates human language instructions into parametrized control code, considering social norms and human preferences. It facilitates robots to exhibit behaviors like nodding or using light patterns to communicate. User studies showed GenEM-generated behaviors, enhanced with user feedback, were competent and understandable. This approach promises flexible, rapidly-created robot behaviors, adaptable to different embodiments and user feedback, potentially revolutionizing human-robot interaction.

A team of researchers from ETH Zurich and Microsoft Research introduces a novel post-training sparsification scheme named SliceGPT for large language models (LLMs). It reduces computational and memory demands by replacing each weight matrix with a smaller matrix, thereby diminishing the embedding dimension of the network. Tests on models like LLAMA-2 70B, OPT 66B, and Phi-2 showed that SliceGPT can remove up to 25% of model parameters while retaining up to 99% task performance. The sliced models run faster on fewer GPUs without needing extra code optimization. This method offers a unique approach to computational invariance in transformer networks, enabling significant reductions in memory and computation requirements for pre-trained models, promising efficiency improvements in deploying large-scale AI models.

This paper from Tencent AI Lab presents a revolutionary 3D scene editing tool that skillfully combines text and image prompts with a 3D bounding box for precise editing localization. The core innovation lies in its stepwise 2D personalization strategy, which includes a unique localization loss for accurate object placement and a dedicated content personalization phase for the reference image, utilizing LoRA. This approach leads to significantly improved editing accuracy and fidelity, marking a substantial advancement in the field of 3D scene editing.

In this paper from Google DeepMind, the focus is on exploring the limits of meta-learning by integrating Solomonoff Induction (SI) into neural networks. The researchers leverage Universal Turing Machines (UTMs) to generate a broad range of data patterns, facilitating the development of versatile representations in neural networks for general problem solving. They present a theoretical analysis of UTM data generation and meta-training protocols, applying these methods to various neural architectures like LSTMs and Transformers. The study demonstrates that UTM data can effectively train networks in universal prediction strategies, pushing the boundaries of AI’s capability in quick, data-limited task learning. This work opens new avenues in AI research, especially in understanding the potential of neural networks to approximate complex, universal predictors.

ChatGPT Creates Comics

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, apply here.