AI Pulse News Roundup (February 2025 Edition)

[PaulBellow]

Welcome to the February 2025 AI Pulse News Roundup Thread!

This is your space to:

Post breaking news about AI research, applications, policies, product launches, ethical debates, and more.
Join the conversation by asking questions, sharing insights, and debating the implications with fellow members.
Review the highlights as this thread becomes a snapshot of January’s key moments in AI.

Whether it’s a groundbreaking paper, a policy shift, or an exciting new tool, everything AI is welcome. Let’s keep the momentum going and make this month another great one for AI discussions.

Have a story or topic to share? Drop it below and let’s get started!

Happy New Year, and here’s to an exciting February in AI!

[PaulBellow]

OpenAI is reportedly in discussions to raise up to $40 billion in a funding round led by SoftBank, potentially valuing the company at $340 billion. wsj.com

Google has announced the release of the Gemini 2.0 Flash AI model for its Gemini app, promising faster responses and improved performance for tasks such as brainstorming, learning, and writing. theverge.com

Krea AI has introduced Krea Chat, a new tool powered by DeepSeek that integrates all Krea features into a chat interface, offering a brand new way of using the platform. x.com

Mistral has unveiled Small 3, a latency-optimized 24B-parameter open-source model released under the Apache 2.0 license. mistral.ai

Sakana AI has released TinySwallow-1.5B, a small-scale Japanese language model trained with a new method called TAID, achieving top performance among similar-sized models. gigazine.net

ElevenLabs has secured $180 million in a Series C funding round, tripling its valuation to $3.3 billion, to advance its AI audio technology and expand its research. reuters.com

AI2 has introduced Tülu 3 405B, an open-source model that surpasses the performance of DeepSeek V3 and GPT-4o on certain benchmarks. allenai.org

[PaulBellow]

Microsoft and OpenAI are investigating whether a group linked to DeepSeek improperly obtained data from OpenAI’s technology. Security researchers detected large-scale data extraction, raising concerns about potential violations of OpenAI’s terms. (Source: Bloomberg)

Microsoft AI CEO Mustafa Suleyman announced that the ‘Think Deeper’ feature is now free for all Copilot users. The feature, powered by OpenAI’s o1 reasoning model, allows Copilot to analyze queries from multiple perspectives and provide detailed responses. (Source: The Verge)

Luma Labs has introduced an ‘Upscale to 4K’ feature for its Dream Machine platform. Users can generate AI-powered videos in lower resolutions and enhance them to 4K with a single tap. (Source: Medium)

The U.S. Navy has banned the use of DeepSeek AI due to security and ethical concerns. Officials worry that the chatbot could expose sensitive user data to foreign entities. (Source: New York Post)

The Bulletin of the Atomic Scientists moved the Doomsday Clock closer to midnight, citing AI-powered military threats. Other factors include nuclear risks from global conflicts and climate change. (Source: Reuters)

Scientists at the Ragon Institute and MIT unveiled MUNIS, an AI tool designed to speed up vaccine development. The model accurately identifies viral targets and outperforms traditional lab methods. (Source: Ragon Institute)

[PaulBellow]

The U.S. Copyright Office reaffirmed that AI-assisted tools do not undermine copyright protection when used to support human creativity. However, works solely generated by AI, such as images created from simple text prompts, remain ineligible for copyright.

The office emphasized that copyright applies when human authors “select and arrange” AI-generated elements, reinforcing its stance from previous guidance. It also rejected additional copyright protections for AI-generated content, citing potential threats to human creators.

The Motion Picture Association welcomed the ruling, highlighting AI’s benefits in post-production tasks like de-aging and object removal. However, the Copyright Office reiterated that AI models trained on copyrighted works without permission remain a key issue for future review. (Source: Variety)

[platypus]

Safety Research Alert

Constitutional Classifiers: Defending against universal jailbreaks (Anthropic)

Anthropic’s recent research introduces “Constitutional Classifiers,” a method designed to defend AI models against universal jailbreaks—inputs crafted to bypass safety measures. By training classifiers on synthetically generated data, the system effectively filters most jailbreak attempts with minimal over-refusals and moderate computational costs. In human red teaming tests, participants were unable to universally jailbreak the model despite extensive efforts. Automated evaluations further confirmed the system’s robustness, with a significant reduction in successful jailbreaks compared to models without such defenses.

Source

[platypus]

Infrastructure Alert

Introducing Data Residency in Europe (OpenAI)

OpenAI users/customers in Europe can now opt to have their API projects and ChatGPT workspaces (Enterprise and Edu only) sandboxed in Europe. This data residency complies with GDPR, DPA and various privacy laws, as well as SOC 2.

Sources: [1]

Research Alert

Open LLMs for Transparent AI in Europe

On February 3, 2025, the OpenEuroLLM project was launched, bringing together 20 leading European research institutions, companies, and EuroHPC centers to develop open-source, multilingual large language models (LLMs) tailored for commercial, industrial, and public services. Coordinated by Jan Hajič of Charles University, Czechia, and co-led by Peter Sarlin of AMD Silo AI, Finland, this consortium aims to enhance Europe’s digital competitiveness and sovereignty by providing transparent and compliant AI technologies. The initiative emphasizes collaboration with open-source communities such as LAION, open-sci, and OpenML to ensure the models are fully open and can be customized for specific industry and public sector needs. The project has been awarded the Strategic Technologies for Europe Platform (STEP) seal and is funded by the European Commission under the Digital Europe Programme.

While it’s welcome news for the development of state of the art AI in Europe, there are question marks over the efficacy of such a setup, considering the sheer number of very different organizations and institutions involved - what can be perceived as “death by committee”.

Sources: [1, 2]

[PaulBellow]

• Perplexity AI Open-Sources R1 1776: Perplexity AI has released R1 1776, an open-source version of DeepSeek’s R1 reasoning model, removing prior censorship filters to provide unbiased and accurate information. perplexity.ai

• Microsoft Unveils Majorana 1 Quantum Chip: Microsoft introduced Majorana 1, its first quantum processor utilizing a topological qubit architecture, potentially scaling to one million qubits on a single chip. azure.microsoft.com

• Apple Introduces iPhone 16e with Apple Intelligence: Apple announced the iPhone 16e, priced at $599, featuring the A18 chip, Apple Intelligence capabilities, and the company’s first in-house designed 5G modem, the C1. apple.com

• Convergence AI Releases Proxy 1.0 Web Agent: Convergence AI launched Proxy 1.0, a free AI-powered web agent that can navigate websites, click, scroll, and execute tasks on behalf of users, turning AI from conversation to action. producthunt.com

• Clone Robotics Unveils ‘Protoclone’ Android: Clone Robotics unveiled Protoclone V1, the world’s first bipedal musculoskeletal android, featuring over 200 degrees of freedom, more than 1,000 Myofibers, and 500 sensors, closely mimicking human anatomy. readwrite.com

[PaulBellow]

Microsoft Unveils Muse AI for Game Development: Microsoft introduced Muse, an AI model trained on Xbox gameplay data to generate video-game scenes, potentially reducing development costs for studios. (Bloomberg)

[PaulBellow]

• xAI Releases Grok-3 Model: Elon Musk’s xAI has launched Grok-3, a new AI chatbot available for free to all users for a limited time, with Premium+ subscribers receiving enhanced access and early features. x.com

• OpenAI Reaches 400 Million Weekly Users: OpenAI’s weekly active users have surpassed 400 million, with over 2 million paying enterprise customers, doubling since September. reuters.com

• NVIDIA and ASDC Launch ‘Signs’ Platform: NVIDIA, in partnership with the American Society for Deaf Children, has introduced ‘Signs,’ an AI-driven platform offering real-time feedback for American Sign Language learners. blogs.nvidia.com

• Pika Labs Introduces Pika Swaps: Pika Labs has unveiled Pika Swaps, a feature enabling users to replace elements in videos using image or text prompts, enhancing video editing capabilities. opentools.ai

• Spotify Integrates ElevenLabs’ AI Voice Technology: Spotify has partnered with ElevenLabs to incorporate AI-generated voice narration, allowing authors to produce and distribute audiobooks in 29 languages. theverge.com

• MIT Unveils ‘FragFold’ for Protein Binding Prediction: MIT researchers have developed ‘FragFold,’ an AI system capable of predicting protein fragments that can bind to and inhibit target proteins, advancing drug discovery efforts. news.mit.edu

[PaulBellow]

OpenAI has extended its Deep Research tool to Plus, Team, Edu, and Enterprise users, granting 10 deep research queries per month, while Pro users now have 120 queries monthly. (Source: help.openai.com)

Anthropic is finalizing a $3.5 billion funding round, bringing its valuation to $61.5 billion, following the release of its latest AI model, Claude 3.7 Sonnet. (Source: WSJ)

DeepSeek is accelerating the launch of its next AI model, moving up the timeline from May to capitalize on the success of R1. (Source: Reuters)

OpenAI has rolled out its Advanced Voice feature, powered by GPT-4o mini, to free ChatGPT users, offering a conversational experience similar to Plus and Pro subscribers. (Source: OpenAI on X)

Microsoft has removed usage restrictions on Copilot’s Voice and Think Deeper functionalities, granting unlimited access to free users while Pro subscribers get priority access during peak times. (Source: Msoft Blog)

Over 1,000 musicians, including Kate Bush and Damon Albarn, have released a silent album, Is This What We Want?, to protest the UK’s proposed copyright reforms allowing AI companies to train on creative works without explicit consent. (Source: Reuters)

IBM has announced plans to acquire DataStax to enhance AI-driven data access and expand its NoSQL database offerings. (Source: IBM Newsroom)

[PaulBellow]

Hume AI Unveils Octave: An Emotionally Intelligent Text-to-Speech Model

Hume AI has introduced Octave, a groundbreaking text-to-speech system that comprehends the emotional context of text, enabling creators to generate custom voices with precise control over emotion and delivery. Unlike traditional models that merely read words, Octave interprets their meaning, allowing for human-like expressiveness in applications such as audiobooks, podcasts, and voiceovers. Users can direct Octave with natural language instructions to modify emotional delivery and speaking style, offering unparalleled versatility.

hume.ai

Perplexity Enhances iOS App with Redesigned Voice Mode

Perplexity has updated its iOS application to include a redesigned voice mode featuring an interactive user interface and six distinct voice options. This enhancement allows users to ask questions and receive real-time, spoken answers, integrating search functionality that displays results directly within the voice interface. The update aims to provide a more engaging and efficient user experience.

testingcatalog.com

Poe Introduces Poe Apps for Building Visual AI Interfaces

Poe has launched Poe Apps, a new feature that enables users to create visual interfaces utilizing a combination of over 100 text, image, video, and audio models available on the platform. This development allows for the construction of applications with functionalities beyond traditional chat interfaces, supporting workflows, games, multimedia generation, and document editing. Poe Apps aim to simplify the process of building AI-driven applications with customized visual components.

poe.com

Vevo Therapeutics Releases Tahoe-100M Dataset via Arc Virtual Cell Atlas

Vevo Therapeutics has open-sourced Tahoe-100M, the world’s largest single-cell dataset, as the inaugural contribution to the Arc Institute’s Virtual Cell Atlas. Tahoe-100M encompasses data from approximately 100 million cells across 60,000 drug perturbation experiments, mapping responses of 50 cancer models to over 1,100 drug treatments. This extensive dataset is freely accessible to the scientific community and is expected to advance research in drug discovery and cellular biology.

arcinstitute.org

Exa Launches Websets: An AI-Driven Search Tool

Exa has introduced Websets, an innovative search product that employs AI agents to deliver highly accurate and comprehensive search results. Websets utilizes Exa’s web-scale vector search technology, allowing users to input natural language queries and receive verified information curated by AI agents. This approach has demonstrated superior performance, reportedly outperforming traditional search engines like Google by over 20 times and OpenAI’s Deep Research by 10 times on complex queries.

lsvp.com

IBM Expands Granite Model Family with Granite 3.2

IBM has unveiled Granite 3.2, the latest addition to its family of AI models designed for enterprise applications. Granite 3.2 introduces enhanced reasoning capabilities, a vision-language model for multimodal tasks, and specialized time series models for forecasting. These models are open-source and optimized for efficiency, aiming to provide businesses with scalable AI solutions that balance performance and computational cost.

ibm.com

Microsoft Launches Phi-4 Multimodal and Phi-4 Mini Small Language Models

Microsoft has expanded its Phi family of small language models with the release of Phi-4 Multimodal and Phi-4 Mini. Phi-4 Multimodal is capable of processing text, vision, and audio inputs simultaneously, facilitating context-aware applications and innovative solutions. Phi-4 Mini, a compact model with 3.8 billion parameters, excels in text-based tasks and is optimized for chat applications, offering high accuracy and scalability in a smaller form factor.

azure.microsoft.com