AI Breakfast
Posts
OpenAI's Plan To Kill Google

OpenAI's Plan To Kill Google

AI Breakfast
May 08, 2024

Good morning. It’s Wednesday, May 8th.

Did you know: On this day in 1946, Sony was founded as Tokyo Tsushin Kogyo, Ltd? The company changed their name to Sony in 1958.

In today’s email:

Microsoft’s New AI
OpenAI’s SearchGPT
“im-a-good-gpt2-chatbot’
Big Prompts Outperform Fine-Tuning
10 New AI Tools
Latest AI Research Papers
AI Creates Comics

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

Get our book: Decoding AI | Upgrade to ad-free

_{In partnership with}

Freepik acquires Magnific, the image upscaler that feels like magic

Combine your legendary AI image generation skills with Freepik and Magnific AI, and you've got the formula to break the AI world. Can’t wait to see how you elevate your Freepik creations upscaled with Magnific AI

_{Thank you for supporting our sponsors!}

Today’s trending AI news stories

Microsoft Is Building A Large AI Model That Could Rival OpenAI

Microsoft is about to throw down the gauntlet with MAI-1, their very own in-house AI language model. This 500-billion parameter monstrosity, helmed by Google DeepMind co-founder Mustafa Suleyman is set to make a splash at the upcoming Microsoft's Build developer conference May 21-23, 2024.

MAI-1 will significantly outsize its predecessors but still fall short of OpenAI's GPT-4, which boasts one trillion parameters. MAI-1's scale positions it alongside OpenAI's GPT-4, hinting at Microsoft's ambition to compete in the generative AI space independently.

Microsoft is throwing serious resources at this, with monster datasets and cutting-edge Nvidia GPUs to juice up MAI-1's learning. Snagging Suleyman and the Inflection crew signals a strategic shift but this might all be just about keeping their competitive edge sharp, weaving AI into the very fabric of their productivity suite, and staking their claim as a generative AI frontrunner. Read more.

OpenAI's "SearchGPT" Might Launch Soon and Feature GPT-4-Lite

OpenAI appears to be on the brink of introducing "SearchGPT," internally dubbed "Sonic," a multifaceted search engine poised to redefine online discovery. Insights from web developer Tibor Blaho suggest that alongside conventional text search, "SearchGPT" will offer image search functionality and an array of widgets covering weather, finance, sports, and more, along with interactive follow-up prompts reminiscent of LLM search engine Perplexity.

The system's design, showcased in a tantalizing preview shared by Blaho, presents search results within a familiar chat interface akin to ChatGPT, emphasizing user engagement and accessibility. SearchGPT will distill web content into succinct summaries of up to 300 characters, potentially mitigating concerns regarding unauthorized content usage.

Blaho's revelations hint at the utilization of GPT-4 Lite, GPT-4, or GPT-3.5 as the underlying language model, echoing previous hints from OpenAI CEO Sam Altman and earlier leaks from The Information. Read more.

"Im-a-good-gpt2-chatbot" And Its Sibling Hint At OpenAI's New Product Launch

OpenAI's rumored search platform and next-gen language models (think "GPT-4 Lite" or GPT-4 upgrades) might be getting the deep six, according to The Information. Two mysterious variants of the "gpt2-chatbot" model – "im-a-good-gpt2-chatbot" and "im-also-a-good-gpt2-chatbot" – just popped up on the LMSYS testing platform. Initial feedback suggests these models perform comparably or even surpass the performance of GPT-4 Turbo, with one benefiting from more recent training data.

Meanwhile, OpenAI's COO, Brad Lightcap, isn't sweating the rumors. He calls today's AI "laughably bad" and throws down some serious next-gen goals for GPT-5. We're talking complex tasks, seamless teamwork-like interactions, and a total shift towards voice interfaces and multimodality. Read more.

Massive Prompts Can Outperform Fine-tuning For LLMs, Researchers Find

Researchers from Carnegie Mellon and Tel Aviv University have found that incorporating numerous examples directly into the input (prompt) can significantly improve the performance of large language models (LLMs), a method termed In-Context Learning (ICL). This approach is particularly effective for tasks with abundant potential answers, as the context window of LLMs expands, allowing for the inclusion of hundreds or thousands of examples. Further enhancing performance, retrieval methods are employed to select relevant examples, especially when fewer examples are used.

While fine-tuning typically requires more data compared to ICL, it can sometimes outperform ICL with extensive contexts. Nevertheless, the study indicates that ICL with long contexts could emerge as a potent tool for various tasks, especially as LLMs become more adept at handling extremely lengthy texts.

These findings align with earlier research by Google Deepmind on the effectiveness of many-shot prompts. Read more.

🖇️ Etcetera: Stories you may have missed

> YouTube's AI-powered 'Jump Ahead' feature rolling out widely to Premium users

> Stack Overflow and OpenAI Partner to Strengthen the World’s Most Popular Large Language Models

> Alphabet-owned Intrinsic incorporates Nvidia tech into robotics platform

> Google’s AI plans now include cybersecurity

> Meta's new multi-token prediction makes AI models up to 3X faster

> Hugging Face launches LeRobot open source robotics code library

> USA invests $285 million in AI-powered chip development

> Apple Pushes Into AI With Launch of New iPad Pro Powered by M4 AI Chip

> OpenAI to Launch Media Manager for Content Creators to Opt In/Out of AI Research and Training

> Microsoft launches AI chatbot for Classified Data

> Microsoft developed an 'air-gapped' AI for spies that's isolated from the internet

> Bedrock Studio is Amazon’s attempt to simplify generative AI app development

> Red Hat integrates generative AI in OpenShift, RHEL and a host of developer tools

> How Schrödinger's cat could make quantum computers work better

> Sperm whale ‘alphabet’ discovered, thanks to machine learning

10 new AI-powered tools from around the web

AFFiNE AI leverages GPT Vision model for advanced text and image processing, facilitating tasks like mind mapping and presentation creation with non-linear capabilities.

Studio Neiro AI x Adobe Express empowers Adobe users with AI-driven avatars for lifelike video creation, enhancing content engagement and accessibility.

GeoSpy AI is an AI-driven geospatial intelligence tool for analyzing satellite imagery. It aids in environmental monitoring, urban planning, agriculture, disaster response, and military reconnaissance.

Zebracat is an AI-driven video platform, enabling marketers to effortlessly convert text into polished videos for marketing, ads, and social media.

Secret Llama is a browser-based chatbot emphasizing privacy, processing conversations locally to safeguard user data. Ideal for sensitive discussions, it operates offline and is open source for transparency.

Eraser AI is a pioneering copilot for technical design. Rapidly generate, edit diagrams, and draft tech docs with ease. Seamlessly integrates with existing workflows.

QRev AI is an open Source AI-driven alternative to Salesforce. Boost sales efficiency with Qai AI Agent, automating tasks and prospecting. Scale sales effortlessly.

Touch AI facilitates sales with AI-curated leads from 100M+ US contacts. Utilizes Bayesian models for targeted email campaigns via Gmail, ensuring deliverability.

Logome enables rapid logo and brand kit creation using AI. Customize and generate brand materials instantly, simplifying brand development. Seamlessly capture and convey your brand identity.

Bahama is an AI-driven platform facilitating team building and management of personalized AI assistants for diverse knowledge-based tasks, enhancing productivity and efficiency.