We put your money to work

Betterment’s financial experts and automated investing technology are working behind the scenes to make your money hustle while you do whatever you want.

Learn more

Today:

• Meta's Llama 3.1

• Mistral Launches Large 2 Model

• Nvidia's AI Chip for China

• Condé Nast Sues Perplexity

• Meta AI Unveils Imagine Me

• Bing Introduces AI-Powered Search

• OctoAI's New Llama 3.1 Models

• Stable Video 4D Unveiled

Meet Llama 3.1

Meta introduces Llama 3.1, an open-source AI model available in 8B, 70B, and 405B versions, suitable for diverse applications. Llama 3.1's flagship 405B model supports various tasks like multilingual translation, complex reasoning, and coding assistance. Users can fine-tune, distill, and deploy these models with provided resources and tools. 

Pricing varies by model and service provider, with the 405B model offering advanced capabilities at higher costs. The models excel in benchmarks, demonstrating significant improvements over previous versions. Meta's open ecosystem encourages customization and integration, enhancing AI-driven solutions across industries. Stay updated through Meta's newsletter and community resources.

META

Mistral shocks with new open model Mistral Large 2, taking on Llama 3.1

French AI startup Mistral has launched its new open-source model, Mistral Large 2, featuring 123 billion parameters. While it lags behind Meta's Llama 3.1 in size, it nearly matches its performance and excels in cost efficiency, speed, and multilingual capabilities. Designed for non-commercial research, it offers open weights for customization. 

Commercial use requires a separate license. Mistral Large 2, accessible via API and cloud platforms, supports 80+ programming languages and excels in reasoning, code generation, and handling complex tasks. The model aims to minimize errors and improve instruction-following, making it a strong contender in the AI race.

VENTUREBEAT

Nvidia preparing version of new flagship AI chip for Chinese market

Nvidia is developing a new AI chip, the "B20," for the Chinese market, ensuring compliance with U.S. export controls. This chip, part of the "Blackwell" series, offers significant speed improvements over its predecessors. Nvidia will collaborate with Inspur for distribution, with shipments expected in Q2 2025. 

The move aims to counter challenges from Chinese firms like Huawei and Enflame, which have advanced due to U.S. semiconductor export restrictions. Nvidia's revenue from China has declined due to these controls, but sales of its H20 chip are now increasing, with projections of over 1 million units sold in China this year.

REUTERS

Condé Nast Sends Cease-and-Desist Letter to AI Search Engine Perplexity

Condé Nast has issued a cease-and-desist letter to AI search engine Perplexity, demanding it stop using content from its publications like The New Yorker, Vogue, and Wired. This follows a similar move by Forbes, accusing Perplexity of copyright infringement. 

Condé Nast's letter, sent on Monday, alleges that Perplexity plagiarizes its content, adding to the increasing legal challenges AI startups face over the use of news content for training language models. This action reflects broader concerns about AI startups using protected content without permission as they develop their technologies.

THE INFORMATION

Meta AI gets new ‘Imagine me’ selfie feature

Meta AI has introduced new features, including support for more languages and the ability to create stylized selfies. The assistant now uses the advanced Llama 3.1 405B model for complex queries, which excels in math and coding tasks. Users must manually switch to this model and are limited in the number of queries. 

A new feature, Imagine Yourself, generates images based on user prompts. Meta AI will also replace Meta Quest’s Voice Commands feature, offering contextual assistance. Available in 22 countries, Meta AI supports various languages and aims to enhance user experience across Meta platforms.

TECHCRUNCH

Introducing Bing generative search

Bing has unveiled a new generative search experience that leverages generative AI and large language models (LLMs) to create dynamic responses to user queries. This feature integrates AI-generated content with traditional search results, providing detailed, easy-to-understand answers with links to sources. 

For instance, a query about "spaghetti westerns" will yield a comprehensive overview of the genre. The goal is to enhance search accuracy while maintaining website traffic and user engagement. Bing is gradually rolling out this feature, inviting user feedback to refine and improve the experience before a broader release.

MICROSOFT BING

Introducing the Llama 3.1 Herd on OctoAI

OctoAI has launched the Llama 3.1 herd, offering models with up to 405 billion parameters. These models feature improved developer tools, including extended context windows, multi-language support, and enhanced tool-calling capabilities. The 405B model excels in tasks like code generation and math, rivaling closed-source models such as GPT-4o and Claude 3.5 Sonnet. Enhanced security features, such as LlamaGuard, ensure safe and compliant use. 

The models are optimized for performance and cost-efficiency on OctoStack, allowing enterprises to self-host them securely. Users can explore these capabilities on OctoAI’s platform with a free trial and proof of concept.

OCTOAI

Introducing Stable Video 4D, Our Latest AI Model for Dynamic Multi-Angle Video Generation

Stability AI has launched Stable Video 4D, a model that transforms a single video into multi-angle videos, generating five frames from eight views in about 40 seconds. This technology enhances video realism and is beneficial for game development, video editing, and virtual reality. Users upload a video and specify camera angles to get dynamic views, optimizing 3D representations. 

Stable Video 4D improves consistency across frames and views without using multiple diffusion models. Available on Hugging Face, it promises significant advancements in creating realistic multi-angle videos, with ongoing enhancements and applications in various industries.

STABILITY AI

🧠RESEARCH

LazyLLM: Dynamic Token Pruning for Efficient Long Context LLM Inference

LazyLLM is a method that speeds up large language model inference by dynamically pruning non-essential tokens during the prefilling and decoding stages. This approach, unlike static pruning, adjusts token selection at each step, reducing time without needing fine-tuning. In tests, it notably accelerates the LLama 2 model by 2.34 times.

SlowFast-LLaVA: A Strong Training-Free Baseline for Video Large Language Models

SlowFast-LLaVA (SF-LLaVA) is a training-free video language model that effectively captures spatial details and long-range temporal context using a two-stream design. The Slow pathway extracts detailed features at a low frame rate, while the Fast pathway captures motion cues at a high frame rate. SF-LLaVA outperforms existing training-free methods and rivals fine-tuned models.

OpenDevin: An Open Platform for AI Software Developers as Generalist Agents

OpenDevin is a platform for developing AI agents that interact with the world like human programmers by writing code and using the web. It supports safe code execution, agent coordination, and task evaluation. OpenDevin is a collaborative project with over 160 contributors and is available under the MIT license.

VILA^2: VILA Augmented VILA

VILA^2 improves Visual Language Models (VLMs) by enhancing data quality and model performance through a two-step augmentation process. The model first improves its data and retrains itself, then uses specialized VLMs for further enhancement. This iterative approach boosts accuracy and achieves state-of-the-art results on the MMMU leaderboard among open-source models.

HumanVid: Demystifying Training Data for Camera-controllable Human Image Animation

HumanVid is a large-scale dataset designed for human image animation, combining real-world and synthetic data. It includes 20,000 high-quality videos and 2,300 3D avatar assets, annotated for both human and camera motions. The dataset supports the CamAnimate model, which achieves state-of-the-art performance in controlling human poses and camera movements. Code and data are available online.

🛠️TOP TOOLS

Supermemory - Ultimate hub for organizing, searching, and utilizing saved information with powerful tools like a search engine, writing assistant, and canvas.

Vozo AI - Rewrite, redub, and lip-sync your viral videos into new stories with prompts.

Wix Studio - Design exceptional sites, with full-stack business solutions, multi-site management and built-in AI.

Study Map - Generate your own learning plan with AI.

Codium AI - Quality-first AI code generation to help busy devs write, test and review code.

📲SOCIAL MEDIA

🗞️MORE NEWS

Elon Musk launches poll asking if Tesla should invest $5 billion in xAI

Elon Musk's poll on social media asks if Tesla should invest $5 billion in his AI startup xAI. With 70.2% of 322,000 users supporting the idea, Tesla faces its lowest profit margin in five years. Musk's AI spending plan includes $10 billion for Nvidia hardware and other AI projects. THE ECONOMIC TIMES 

Free Gemini users can finally chat in a flash

Google's Gemini chatbot now includes the fast-response model Gemini 1.5 Flash for free users, previously limited to developers. This update, announced at Google I/O, also adds more source links to reduce inaccuracies. Gemini 1.5 Flash offers a large context window and file upload capability, making it ideal for handling complex queries. VENTUREBEAT

Cohere raises $500M to beat back generative AI rivals

Cohere, a generative AI startup founded by ex-Google researchers, raised $500 million, boosting its valuation to $5.5 billion. Investors include Cisco, AMD, and Fujitsu. Cohere customizes AI models for businesses like Oracle and LivePerson and aims for growth with new funding, focusing on enterprise AI solutions and expanding its team. TECHCRUNCH

Google brings AI agent platform Project Oscar open source

Google announced Project Oscar, an open-source AI platform to help software teams manage bugs and issues. Launched at Google I/O Bengaluru, Oscar allows developers to create AI agents for various tasks like bug tracking and support. The platform is initially for open-source projects but may expand to closed-source in the future. VENTUREBEAT

Adobe rolls out more generative AI features to Illustrator and Photoshop

Adobe has added new generative AI features to Illustrator and Photoshop. Illustrator now includes Generative Shape Fill and improved Text to Pattern tools using the Firefly Vector AI model. Photoshop updates include a new Generate Image feature and an Enhance Detail tool for sharper images. These features aim to boost creative workflows while addressing concerns about AI's impact on jobs. THE VERGE

Snowflake ropes in AI21’s Jamba-Instruct to help enterprises decode long documents

Snowflake has integrated AI21 Labs’ Jamba-Instruct LLM into its Cortex AI service, enabling enterprise users to build generative AI applications for long documents. This model supports up to 256K tokens, making it ideal for tasks like financial analysis and clinical reports. The hybrid model offers cost efficiency and high performance, with Snowflake also providing serverless inference for cost-effective scalability. VENTUREBEAT