Skip to main content
Back

Google vs OpenAI: Two Incredible Breakthroughs in 24 hours

GET ACCESS

February 16, 2024 4 min read

Hey 

Welcome to this week's edition of AlphaSignal. 


Whether you are a researcher, engineer, developer, or data scientist, our summaries ensure you're always up-to-date with the latest breakthroughs in AI. 


Let's get into it!


Lior

IN TODAY'S EMAIL

  • Google's Context Breakthrough

  • Trend Github Repos

  • OpenAI's Sora T2V

Read Time: 4 min 45 sec

NEWS

Gemini 1.5: The First 1 Million Token Context Window?

What's New?

Yesterday, Google AI unveiled Gemini 1.5 Pro, just one week after the release of Ultra and 70 days after Pro 1.0. 


This multimodal model has a context window of 1 million tokens, and 10 million tokens in research. It’s the largest context window ever for a foundational model.


To put things in perspective, 1M tokens is:

  • 1 hour of video

  • 11 hours of audio

  • >30k lines of code

  • >700k words

This capacity positions Gemini Pro 1.5 way ahead Anthropic’s Claude, which previously led the race with a 200 thousand token context window. 1.5 Pro outperforms 1.0 Pro on 87% of the benchmarks used for developing LLMs on all data modalities (text, code, image, audio and video).


Despite its unprecedented context window, Pro 1.5 excels at retrieval even for small details in a large amount of information. It achieves near-perfect “needle in the hay stack” retrieval (>99%) in all data modalities.


Gemini Pro 1.5 also demonstrates remarkable capacity in generalizing zero-shot to long instructions. The easy-tuning feature to customize Gemini for specific needs, based on a set of examples, will roll out in the next few days.


For developers, this enhances the utility of Gemini for coding by providing enough context to the model to understand complex intricacies of the codebase. This would allow to generate new code that would encompass all the specificities of the existing codebase, but also to find, explain, modify and refactor existing code.


Plans for a broader release with standard (128k tokens) and expanded context window options, are underway.


Access the limited preview through AI Studio and Vertex AI.

Get access

On-Demand GPU Instances with Special Pricing at Latitude.sh

Start today and accelerate your training and inference workloads with the power of NVIDIA’s H100 Tensor Core, deploy your instances in seconds with Latitude.sh!


High-performance networking and enterprise-grade hardware designed for Machine Learning engineers. Our super intuitive dashboard and pre-installed AI software gets you from zero to training with just a few clicks.


Multiple clusters with 8 x H100s available today, use the code G3OFF10 and get 10% off during your first 3 months.

Get started

partner with us

TOP OF GITHUB

crewAI (☆ 7.5k)

---------------------

CrewAI helps you orchestrate AI agents to work together, simulating a crew. You can build smart platforms or automated services by leveraging role-based collaboration and shared goals. It simplifies integrating multiple AI models and supports customization with open-source tools.

apple ml-mgie (☆ 3.3k)

---------------------

MGIE helps you edit images through natural language instructions, leveraging multimodal large language models for pixel-level manipulations. It combines expressive instruction derivation with visual imagination for effective editing, optimized in an end-to-end training scheme. It supports Photoshop-style modifications, global photo optimization, and local edits.

shell_gpt (☆ 7.4k)

---------------------

ShellGPT, powered by GPT-4, helps you generate shell commands, code snippets, and documentation directly from the command line, streamlining your workflow. It supports major shells and OSes, with options for using OpenAI's API or local models.

metavoice-src (☆ 2.1k)

---------------------

MetaVoice-1B helps you generate human-like, emotionally expressive speech in English, including zero-shot voice cloning for American and British accents with just 30s of reference audio. It's built on a 1.2B parameter model trained on 100K hours of speech data, supporting cross-lingual voice cloning and long-form synthesis. 

phidata (☆ 2.1k)

---------------------

Phidata helps you build AI assistants through function calling, integrating tools like Streamlit, FastAPI, and databases for AI applications. You can create, serve, and extend functionalities with pre-built or custom functions, streamlining the development of AI-driven solutions.

NEWS

OpenAI Announces Sora: The Most Impressive Text-To-Video Model Yet

What's New?

OpenAI has introduced Sora, its latest innovation designed to transform text prompts into detailed videos up to a minute long, marking a significant evolution from the existing text-to-image models.

Key Highlights:

  • Advanced Video Creation: Sora stands out by its ability to craft videos featuring multiple characters, diverse motions, and intricate backgrounds directly from textual instructions.

  • Innovative Technology: The model leverages a combination of diffusion model techniques and Vision Transformers (ViT), embedding video data into a latent space for enhanced processing. This method allows video segments to be treated as tokens, similar to the way language models process text.

  • High-Quality Resolution: Sora is engineered to support various resolutions, including high-definition formats like 1920x1080p, pushing beyond the conventional 256x256 resolution ceiling seen in many multimodal models.

  • Enhanced Realism: By applying re-captioning techniques initially developed for DALL·E 3, Sora significantly improves the generation of realistic video patches from noisy inputs, demonstrating a deeper understanding of the physical world.

  • Scalability and Effectiveness: Thanks to its transformer-based architecture, Sora ensures scalability and effectiveness, underlining the technical sophistication of the model.

Despite its groundbreaking capabilities, Sora faces challenges in accurately simulating complex physics and specific cause-and-effect scenarios, which could affect the precision of its generated videos. OpenAI has yet to release specific performance benchmarks or accuracy rates, choosing instead to focus on the model's broad potential and innovative aspects.

Potential Applications: Sora is primarily aimed at creative professionals in filmmaking, visual arts, and design, offering a novel tool for content creation that bridges the gap between imagination and video production.


Access and Development: For now, OpenAI is granting access to Sora exclusively to a select group of red teamers and creative professionals for risk assessment and feedback. This strategic, limited rollout reflects OpenAI's commitment to refining Sora through practical application and ethical consideration.

Read the report

Not Great   Good    Amazing

Thank You.

Want to promote your company, product, job, or event to 170,000+ AI researchers and engineers? You can reach out here.