- Luis Poveda's AI Newsletter
- Posts
- Luis Poveda's AI Newsletter: March 17, 2025
Luis Poveda's AI Newsletter: March 17, 2025
AI's New Frontier: Open Models and Autonomous Agents Lead the Revolution
Luis Poveda
March 17, 2025
Executive Summary
This week in AI brings significant advancements across multiple fronts: Google unveils Gemma 3, its most capable compact model yet; Cohere launches the highly multilingual Command A; and Block becomes the first North American company to deploy NVIDIA's cutting-edge GB200 systems. Meanwhile, Chinese AI agent Manus goes viral, Y Combinator reveals 25% of its latest cohort relies on AI-generated code, and geopolitical tensions rise as OpenAI labels DeepSeek "state-controlled".
AI Model Releases and Updates
Google Unveils Gemma 3: The Most Capable Single-Accelerator Model
Google has launched Gemma 3, its most advanced open model based on Gemini 2.0 technology. Available in four sizes (1B, 4B, 12B, and 27B), Gemma 3 is designed to run efficiently on a single GPU or TPU while outperforming much larger models. The model supports over 140 languages, handles a 128k-token context window, and includes advanced visual reasoning capabilities alongside function calling capabilities.
"These are our most advanced, portable and responsibly developed open models yet. They are designed to run fast, directly on devices — from phones and laptops to workstations — helping developers create AI applications, wherever people need them".
Google Gemini Updates: Deep Research and App Integration
Google has expanded Gemini with significant updates, including Deep Research powered by 2.0 Flash Thinking Experimental for higher-quality research reports. The upgrade includes a 1M token context window for Gemini Advanced users and integration with additional Google apps such as Calendar, Notes, Tasks, and Photos. Google is also introducing a personalization feature that allows Gemini to connect with various Google services for more tailored responses.
"Today, we're upgrading Deep Research with Gemini 2.0 Flash Thinking Experimental. This enhances Gemini's capabilities across all research stages — from planning and searching to reasoning, analyzing and reporting — creating higher-quality, multi-page reports that are more detailed and insightful".
Cohere Launches Command A: Multilingual Enterprise Model
Cohere has released Command A, a powerful new AI model designed for enterprise applications. The model stands out for its ability to operate on just two GPUs while outperforming competitors, supporting 23 languages, and delivering a 256,000-token context window. Command A is 1.75x faster than GPT-4o and 2.4x faster than DeepSeek-V3, with a focus on retrieval-augmented generation and tool use for enterprise customers.
"Command A is exceptionally good at generating responses in conversational tasks, attending over long inputs, and extracting and manipulating numerical information in financial settings".
SesameAI Opens Up Revolutionary Voice AI Technology
SesameAI, founded by Oculus co-creator Brendan Iribe, has released its base Conversational Speech Model (CSM-1B) under the Apache 2.0 license. This 1 billion parameter model powers their viral virtual assistant Maya, which features remarkably human-like voice interactions that include natural pauses, disfluencies, and the ability to be interrupted mid-sentence. The model generates "RVQ audio codes" from text and audio inputs and uses Meta's Llama as its backbone paired with an audio decoder.
"At Sesame, our goal is to achieve 'voice presence'—the magical quality that makes spoken interactions feel real, understood, and valued".
AI Infrastructure and Hardware
Block Deploys NVIDIA DGX GB200 Systems for Frontier Models
Block, Inc. has become the first company in North America to deploy NVIDIA's latest DGX GB200 systems, marking a significant advancement in AI infrastructure. The deployment at an Equinix data center will be used to develop frontier open-source AI models. This move highlights Block's commitment to leading-edge AI research and development for tackling complex, real-world challenges.
"With NVIDIA DGX GB200 systems, Block engineering and research teams can develop frontier open source AI models that can tackle complex, real-world challenges with state-of-the-art AI supercomputing".
NVIDIA GTC 2025 Set for March 17-21
NVIDIA's GPU Technology Conference (GTC) 2025 is scheduled for March 17-21 in San Jose, California. The event, often referred to as the "Woodstock of AI," is expected to feature major announcements in AI hardware and software. Industry leaders anticipate significant revelations about NVIDIA's next-generation AI systems and technologies that will shape the future of AI development.
NVIDIA
"GTC 2024 was the single most important event in the history of the technology industry".
AI Platforms and Tools
Hugging Face Expands LeRobot Platform with Self-Driving Capabilities
Hugging Face has partnered with AI startup Yaak to expand its LeRobot platform with a massive new training dataset for autonomous navigation. The "Learning to Drive" (L2D) dataset, over a petabyte in size, contains sensor data from German driving schools, capturing real-world driving scenarios. This expansion aims to empower the AI community to build end-to-end self-driving models with plans for real-world testing this summer.
"L2D aims to be the largest open-source self-driving data set that empowers the AI community with unique and diverse 'episodes' for training end-to-end spatial intelligence".
Perplexity Enhances Model Context Protocol (MCP) Integration for Sonar API
Perplexity has enhanced its Model Context Protocol (MCP) implementation for the recently launched Sonar API, strengthening the connection between its powerful search capabilities and various AI assistants. Building on the Sonar API released earlier this year, the MCP integration allows AI models like Claude to perform real-time web searches through Perplexity's platform without requiring custom development work. The implementation follows the open standard developed by Anthropic, designed to connect AI assistants with external data sources, and has already gained traction among developers building tools that leverage Perplexity's search technology.
"The Perplexity Ask MCP Server follows MCP's open standard, allowing any AI assistant or automation tool to connect to the Sonar API for live web searches. AI models can query the server for information retrieval, leveraging Perplexity's search capabilities to return the most relevant insights".
Browser Use, an AI tool that helps autonomous agents interact with websites, has seen explosive growth following its integration with viral Chinese AI agent Manus. Daily downloads of Browser Use quintupled from 5,000 to 28,000 in just one week. The tool extracts website elements to allow AI models to interact with them more easily, supporting multiple browser tabs and handling various inputs, positioning itself as a foundation layer for the growing web agent ecosystem.
"We wanted to create a foundation layer that everyone will build browser agents on. In our minds, there will be more agents on the web than humans by the end of the year".
AI in Business and Startups
25% of Y Combinator's Current Cohort Built on AI-Generated Codebases
According to Y Combinator managing partner Jared Friedman, a quarter of startups in YC's Winter 2025 batch have codebases that are 95% AI-generated. Despite being founded by highly technical people capable of building products from scratch, these entrepreneurs are leveraging AI coding tools to accelerate development. YC executives emphasize that even with AI-generated code, developers still need traditional coding skills to debug and maintain systems, especially as products scale.
"It's not like we funded a bunch of non-technical founders. Every one of these people is highly technical, completely capable of building their own products from scratch. A year ago, they would have built their product from scratch — but now 95% of it is built by an AI".
AI Education and Learning
Andrew Ng: Not Learning to Code Due to AI is "Worst Career Advice Ever"
AI researcher and educator Andrew Ng has strongly criticized those who discourage learning programming on the grounds that AI will automate it. In a recent statement, Ng argued that as coding becomes easier through AI assistance, more people should learn to code, not fewer. He explained that throughout computing history, from punch cards to modern IDEs, programming has consistently become more accessible, which has expanded opportunities rather than diminished them.
"Some people today are discouraging others from learning programming on the grounds AI will automate it. This advice will be seen as some of the worst career advice ever given... As coding becomes easier, more people should code, not fewer!"
AI Governance and Ethics
OpenAI Labels DeepSeek "State-Controlled," Urges Ban
OpenAI has labeled Chinese AI startup DeepSeek as "state-sponsored" and "state-controlled" in a policy recommendation to the Trump Administration's AI Action Plan initiative. OpenAI is urging the U.S. government to consider banning DeepSeek's models, citing security risks related to Chinese laws requiring companies to share user data at the government's request. This accusation represents a significant escalation in tensions between the two AI giants and highlights growing geopolitical concerns in the AI sector.
"As with Huawei, there is significant risk in building on top of DeepSeek models in critical infrastructure and other high-risk use cases given the potential that DeepSeek could be compelled by CCP to manipulate its models to cause harm".
Conclusion
As we approach the end of Q1 2025, the AI landscape is evolving rapidly along three critical dimensions: model efficiency, autonomy, and geopolitics. Google and Cohere's focus on delivering more capable models on less hardware points to a future where advanced AI becomes more accessible and deployable at the edge. Meanwhile, the explosive growth of autonomous agents like Manus and Browser Use suggests we're entering an era where AI not only generates content but actively navigates and interacts with digital environments. Against this backdrop of technological advancement, the intensifying rivalry between U.S. and Chinese AI companies highlights how AI development is becoming increasingly entwined with national security concerns and international relations. As these trends converge, we can expect both unprecedented innovation and complex governance challenges in the months ahead.
The Author
 Luis Poveda’s AI Newsletter | Luis Poveda is a technology optimist and passionate innovator, constantly exploring and researching the latest trends. Based in Barcelona, he is currently focused on AI and developing a modern AI-driven network observability tool. He is also the creator and maintainer of Luis Poveda's AI Newsletter, where he curates and shares key insights on the evolving AI landscape. |