Your daily AI dose

Mindstream is your one-stop shop for all things AI.

How good are we? Well, we become only the second ever newsletter (after the Hustle) to be acquired by HubSpot. Our small team of writers works hard to put out the most enjoyable and informative newsletter on AI around.

It’s completely free, and you’ll get a bunch of free AI resources when you subscribe.

Click here to subscribe

Today:

• Meta’s LCMs Break Token Barriers

• Microsoft, OpenAI Negotiate Partnership Terms

• Nonprofit Joins Musk Against OpenAI

• Stepfun Secures Major AI Funding

• Google Focuses on Affordable AI Access

Meta's STUNNING New LLM Architecture is a GAME-CHANGER!

Meta's AI team introduces Large Concept Models (LCMs), shifting from token-based processing to concept-based reasoning, mimicking human abstraction. This new architecture allows models to process and generate ideas across languages and modalities, offering improved efficiency and generalization.

Early results show promise, with LCMs excelling in tasks like summarization and outperforming comparable models. While challenges remain, this approach could revolutionize AI by enabling higher-level understanding and reasoning beyond token-based limitations.

WATCH THE VIDEO ON YOUTUBE

Microsoft, OpenAI Negotiate Partnership Terms

OpenAI CEO Sam Altman seeks to convert OpenAI into a for-profit corporation, but Microsoft, a major investor with $13 billion committed, is negotiating terms. Talks since October center on Microsoft’s equity, cloud exclusivity, intellectual property rights, and revenue share. These discussions reflect the complexities of their partnership as both parties balance OpenAI's governance structure with Microsoft's strategic interests.

Nonprofit Joins Musk Against OpenAI

A nonprofit, Encode, supports Elon Musk’s legal effort to block OpenAI’s for-profit transition, citing risks to public safety and AI’s mission. OpenAI plans to restructure into a Public Benefit Corporation, sparking concerns about prioritizing profit over safety. Critics, including former employees and Meta, argue the move undermines OpenAI's original goals. Encode highlights that transitioning could weaken commitments to safe AI, shifting focus to shareholder interests, and harm public trust in transformative technology.

Stepfun Secures Major AI Funding

Shanghai-backed Fortera Capital led Stepfun’s latest funding round, raising "hundreds of millions of dollars" to advance foundational AI models and consumer products. Supported by Tencent and Qiming Ventures, this highlights China’s push for technological innovation. Shanghai’s government bolsters AI, biotech, and semiconductor industries with a 100 billion yuan fund. Founded by ex-Microsoft chief scientist Jiang Daxin, Stepfun aims to rival U.S. peers with its Step-1V model and upcoming trillion-parameter Step-2.

Google Focuses on Affordable AI Access

Google CEO Sundar Pichai emphasized 2025 as a critical year for AI development, urging focus on impactful products and execution. DeepMind co-founder Demis Hassabis revealed plans for "Project Astra," a universal AI assistant update. Addressing premium subscriptions, Hassabis confirmed no immediate plans for a high-cost service like ChatGPT Pro, citing satisfaction with the $20 Gemini Advanced tier. Google also unveiled advanced AI models like Veo 2, outperforming competitors and reinforcing its strategic AI direction.

🧠RESEARCH

Token-Budget-Aware LLM Reasoning

The study introduces a token-budget-aware framework for large language models (LLMs) that optimizes reasoning by setting dynamic token limits. This approach reduces costs in Chain-of-Thought reasoning while maintaining high performance, balancing efficiency and accuracy effectively.

Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree Search

Mulberry introduces an advanced reasoning method for Multimodal Large Language Models (MLLMs) using Collective Monte Carlo Tree Search (CoMCTS). By leveraging collective knowledge, Mulberry performs step-by-step reasoning and reflection to solve complex questions. It outperforms benchmarks and includes a new dataset, Mulberry-260k. Code is available on GitHub.

Video-Panda: Parameter-efficient Alignment for Encoder-free Video-Language Models

Video-Panda introduces a lightweight, encoder-free video-language model with only 45M parameters, significantly reducing computational overhead. Using a novel Spatio-Temporal Alignment Block (STAB), it achieves competitive performance in video question answering, outperforming traditional models in correctness and speed. Code is available on GitHub.

🛠️TOP TOOLS

FaceSwapper - AI-powered online platform that offers a suite of advanced image and video editing tools, primarily focused on face swapping technology.

Anakin AI - No-code AI app builder that empowers users to create customized AI applications for automating tasks, generating content, and answering questions.

AI Human Generator - Create hyperrealistic full-body images of people who don’t exist. 

OpenRead - AI-powered interactive platform designed to revolutionize academic research and literature analysis.

Qlip AI - AI-powered platform designed to help content creators efficiently repurpose long-form videos into short, shareable clips for social media. 

📲SOCIAL MEDIA