OpenAI launches its Google challenger, ChatGPT Search

OpenAI has unveiled ChatGPT Search, a new feature intended to directly compete with Google’s search capabilities. Building upon the prototype SearchGPT released earlier, ChatGPT Search is designed to provide timely responses by drawing from a wide range of online sources. Integrated into the ChatGPT platform, it leverages a refined version of OpenAI’s GPT-4o model to supply information on topics like sports, news, and stock updates, complete with source links to help users further explore the information.

ChatGPT Search can autonomously decide when to search the web-based on the question asked or allow users to activate a search manually via a new web search icon. Results appear with both in-line and sidebar attributions to the news publishers and data sources OpenAI has agreements with, such as for election-related inquiries, which are directed to sources like AP and Reuters. Responses also include interactive follow-up questions, allowing users to refine and deepen their searches.

Currently, ChatGPT Search is available to ChatGPT Plus and Team users on both mobile and web platforms, with OpenAI planning to roll it out to enterprise, educational, and free users in the near future. Additionally, OpenAI has released a browser extension enabling users to set ChatGPT Search as the default search engine in Chrome, signaling its ambition to position ChatGPT as a central tool for online research and browsing.

OpenAI is mindful of concerns from publishers about potential traffic cannibalization, incorporating feedback on factors like article relevance and summary length. By continually enhancing ChatGPT Search, OpenAI aims to make it a comprehensive tool for various use cases, including shopping, travel, and voice-activated research. This development marks a significant step toward integrating AI into everyday information access while navigating the complexities of content sourcing and digital publishing.

Use AI As Your Personal Assistant

Ready to embrace a new era of task delegation? 

HubSpot’s highly anticipated AI Task Delegation Playbook is your key to supercharging your productivity and saving precious time.

Learn how to integrate AI technology into your processes, allowing you to optimize resource allocation and maximize output with precision and ease.

Oasis: A Universe in a Transformer

Oasis represents a groundbreaking real-time, open-world AI model powered by a transformer, introducing interactive video gameplay generated end-to-end by AI. Unlike traditional game engines, Oasis simulates an environment frame-by-frame based on user keyboard and mouse inputs, enabling actions such as movement, jumping, item collection, and breaking blocks. This real-time gameplay model relies on advancements in transformer and diffusion-based model architecture, leveraging Decart's inference technology to achieve minimal latency during live interactions. Trained on gameplay footage, Oasis demonstrates the potential for foundation models to simulate dynamic, user-driven worlds without relying on conventional game engines, marking a major step forward in the field of AI-driven interaction.

The Oasis model integrates diffusion and transformer architectures to manage complex real-time physics, rules, and graphics generation. By using diffusion training with temporal and spatial layers, the model can generate successive video frames based on prior contexts, maintaining a seamless interactive experience. In addition, Oasis benefits from the ViT VAE compression, which enables high-level video generation and keeps latency manageable for user interactions. This technology positions Oasis as a novel example of how real-time, AI-generated video could transform interactive media.

Achieving high-speed video inference required extensive optimization on Decart’s part, as traditional text-to-video models struggle to keep pace with real-time demands. Oasis leverages Decart's proprietary inference framework, which includes kernel optimizations for NVIDIA H100 Tensor Core GPUs and advanced communication strategies like NVLink and PCIe Gen 5. These optimizations enable Oasis to achieve a frame inference time of 47 milliseconds and smooth real-time rendering. Currently, Oasis runs at 20fps at 360p, but the model is designed for future compatibility with Etched's upcoming Sohu chip, which could support 4K rendering at similar costs and performance.

Though Oasis is already a significant step forward, its developers are exploring further improvements. These include optimizing the model’s memory to retain long-term details, reducing initial haziness in video output, and training the model to handle inputs outside its learned distribution. As the project progresses, scaling the model and training data will be essential, alongside advancements in inference technology, to maintain efficient and affordable real-time interactive video experiences.

Meta is making a robot hand that can ‘feel’ touch

Meta is advancing its robotics capabilities by developing a new type of robotic hand equipped with sensors designed to simulate human-like touch. In collaboration with GelSight, a sensor technology company, and Wonik Robotics from South Korea, Meta aims to commercialize tactile sensors tailored specifically for AI research. These innovative devices aren’t geared toward consumer markets but instead serve scientific research, enabling AI systems to interact with and understand the physical world more precisely.

One of the key products emerging from this partnership is the Digit 360, a highly sensitive robotic fingertip with human-level multimodal sensing capabilities. This sensor, an upgraded version of Meta's previous Digit model, features 18 sensing parameters and an onboard AI chip that processes touch signals, allowing it to detect and respond to environmental changes. The device includes an optical system that captures surface deformations from a wide field of view, enhancing touch perception through multiple modalities. This design enables Digit 360 to sense vibrations, temperature, and even odors, making it highly responsive to different surface profiles.

Meta’s partnership with Wonik Robotics is set to produce the next generation of the Allegro Hand, which incorporates these sophisticated tactile sensors. The Allegro Hand will be equipped with Meta’s newly designed control boards, which convert tactile data into digital feedback on a host computer, adding a layer of physical intelligence to robotic systems. Both the Digit 360 and Allegro Hand will be available for purchase next year, with Meta offering early access opportunities for researchers interested in exploring their potential applications.

By focusing on tactile feedback, Meta hopes to support advancements in AI that enable more nuanced learning about physical environments. These developments could significantly enhance AI-driven robotics, especially in fields that require a refined sense of touch, paving the way for more complex and adaptive machine interactions with the world.

Other stuff

• Gemini API and Google AI Studio now offer Grounding with Google Search

• In a recent Reddit AMA, OpenAI CEO Sam Altman says lack of compute capacity is delaying the company’s products

• D-ID launches new high-quality avatars capable of real-time conversations

• Perplexity’s CEO punts on defining ‘plagiarism’

• Google Maps is getting new AI features powered by Gemini

• The current and future landscape of AI on wheels with Jesse Levinson from Zoox

• U.S. Military Makes First Confirmed OpenAI Purchase for War-Fighting Forces

Automate Calls and Boost Conversions with AI Voice Assistants

Set up an AI receptionist (on 24/7) or an outbound lead qualifier (#Speedtolead). Book appointments, transfer calls, and extract info seamlessly. Integrates with HubSpot, GohighLevel, and more. Deploy in minutes, no coding needed.

Try AI Phone Calling

Hope you enjoyed today's newsletter

⚡️ Join over 200,000 people using the Superpower ChatGPT extension on Chrome and Firefox.