PS - I made it Indian Accent just for the fun of it, but here is how the video goes, I think it is very good quality.

Imagen 3: Redefining Photorealism in AI Image Generation

Imagen 3, Google’s latest text-to-image model integrated into the Gemini ecosystem, represents a significant leap in AI-driven image generation as of March 2025. Launched globally to Gemini users around October 2024, Imagen 3 is celebrated for its exceptional photorealism, delivering sharper details, richer textures, and fewer visual artifacts than its predecessors.

It excels at interpreting natural language prompts with remarkable precision, enabling it to create diverse outputs ranging from lifelike scenes to stylized artworks like oil paintings or 3D digital renders. Built with improved safety filters and trained on enhanced datasets, Imagen 3 reflects Google’s focus on quality and responsibility, making it a versatile tool for creators. Its seamless integration with Gemini’s multimodal capabilities allows users to generate high-quality visuals directly within a conversational AI experience.

@webthreeth

MeetGamma: Revolutionizing Digital Content Creation with AI

MeetGamma is an innovative AI-powered platform designed to revolutionize the creation of presentations and landing pages, enabling users to produce polished, professional content in a fraction of the time traditionally required.

By leveraging advanced artificial intelligence, MeetGamma allows users to input specific requirements—such as a desired page type or use case—and generates customizable structures that can be fine-tuned with a variety of themes or branded elements. The tool streamlines the process by offering intuitive interfaces for editing content, selecting designs, and publishing pages instantly on a Gamma subdomain.

For advanced users, features like custom domains are available through a Gamma Pro subscription, enhancing its utility for businesses and individuals looking to establish a strong digital presence.

@webthreeth

Gemini Unveils Deep Research with 2.0 Flash Thinking Model, Now Free for All

On March 14, 2025, Google rolled out a significant upgrade to its Gemini AI assistant, launching the Deep Research feature powered by the Gemini 2.0 Flash Thinking Experimental model—and making it freely accessible to all users, not just Gemini Advanced subscribers.

Initially introduced in December 2024 for premium users, Deep Research transforms Gemini into a powerful research companion, capable of scouring the web, synthesizing data from diverse sources, and delivering detailed, multi-page reports on complex topics in minutes.

With the 2.0 Flash Thinking model, the feature now boasts enhanced reasoning capabilities, offering real-time visibility into its research process and producing high-quality insights for tasks like market analysis, academic research, or product comparisons. Users can customize research plans, export reports to Google Docs.

@webthreeth

Baidu’s Launch of ERNIE X1 and 4.5: Pioneering AI Innovation and Affordability

Baidu’s recent launch of ERNIE X1 and ERNIE 4.5 marks a significant advancement in the AI landscape, positioning the Chinese tech giant as a formidable player in the global market. ERNIE X1, a deep-thinking reasoning model with multimodal capabilities, delivers performance on par with DeepSeek R1 but at half the price, featuring enhanced understanding, planning, reflection, and autonomous tool use.

Meanwhile, ERNIE 4.5, Baidu’s latest foundation model, excels in multimodal understanding, boasting superior language abilities, reasoning, memory, and hallucination prevention, while outperforming models like GPT-4o across various benchmarks. Both models are now freely accessible to individual users via the ERNIE Bot website, with API access for enterprises on Baidu’s Qianfan platform, signaling Baidu’s strategy to democratize AI and intensify competition in the industry.

@webthreeth

OpenAI's Vision for AI-Driven Automation in 2025

OpenAI’s Chief Product Officer, Kevin Weil, has outlined an ambitious vision for 2025, predicting that AI will automate coding by 99% by year’s end. This projection is supported by advancements in AI models, including enhanced pre-training and improved reasoning capabilities, as well as the development of tools like Deep Research, which offers insightful, specific information rather than generic responses. Beyond coding, OpenAI aims to democratize software access for everyone through its products and API, while also exploring the integration of AI into robotics.

This move reflects a broader goal to extend AI’s impact from digital tasks to physical applications, aligning with the company’s mission to create safe, beneficial, and universally accessible AI systems. Weil also expressed confidence in GPT-5, anticipating it will unify OpenAI’s O-series and GPT-series models, enhancing seamless AI interactions across various tasks.

@webthreeth

Mistral AI launches its new Mistral Small 3.1 model, which is able to be hosted and run locally

Mistral Small 3.1 is a cutting-edge language model developed by Mistral AI, designed to deliver exceptional performance in a compact size. With 24 billion parameters, it achieves an impressive 81% accuracy on standard benchmarks and processes 150 tokens per second. The model's efficiency is further enhanced by its ability to run locally on a single RTX 4090 GPU or a MacBook with 32GB RAM, which is particularly beneficial for hobbyists and organizations handling sensitive information.

Mistral Small 3.1 supports a wide range of languages and excels in both coding and generalist tasks, offering a cost-effective solution. Its advanced reasoning capabilities and strong adherence to system make it a versatile tool for various applications, from complex problem-solving to everyday tasks. The model's development without reinforcement learning or synthetic training data sets it apart from competitors.

@webthreeth

LG AI Research Unveils EXAONEDeep Today, Revolutionizing Agentic AI for Industry Applications

On March 18, 2025, Korean LG's AI wing, LG AI Research introduced EXAONEDeep, a groundbreaking next-generation AI model advancing "Agentic AI" with enhanced reasoning capabilities tailored for math, science, and coding tasks across professional and everyday applications.

Released as an open-source model at Nvidia’s GTC 2025, EXAONEDeep’s smaller variants—7.8B and 2.4B parameters—dominated major benchmarks, while the 32B version achieved the top spot on the AIME benchmark, outperforming a competitor at just 5% of its model size. This leap in efficiency and performance aligns with broader AI trends, such as Google’s recent focus on test-time compute scaling to improve model reasoning through self-verification, positioning EXAONEDeep as a leader in transforming real-world industry solutions.

@webthreeth

AI intelligence is becoming more efficient and accessible over time. In AI, 'parameters' refer to the adjustable variables within a model that determine its capacity to learn, not the amount of training data itself.

LG today developed a model, EXAONEDeep, which is said to rival OpenAI’s o1 in performance while using only 4% of the parameters required for o1. This suggests that powerful AI can be achieved with fewer resources. LG’s model is also open-source, making it freely available.

For comparison, DeepSeek’s R1 model, trained on 671 billion parameters, is a leading open-source example, while LG’s EXAONE, potentially with 32 billion parameters, outperforms it. Such advancements could drive significant growth in AI, potentially paving the way for artificial general intelligence and super intelligence in the future.

Should I simplify posts like this to make the explanation better?

Gemini Introduces Audio Overview for Engaging Learning

Gemini is enhancing its platform with Audio Overview, a feature previously popular in NotebookLM, now available to Gemini and Gemini Advanced subscribers globally in English starting March 18, 2025.

This tool transforms uploaded documents, slides, and Deep Research reports into podcast-style audio discussions featuring two AI hosts who summarize content, connect topics, and provide dynamic insights. Users can upload various files like class notes or research papers, click a suggestion chip, and listen to an engaging breakdown on the go via the web or mobile app, with options to share or download.

This rollout aims to make learning fun and productive, with more languages planned soon, solidifying Gemini as a versatile collaborator at gemini.google.com.

@webthreeth

Google Unveils Canvas: A New Interactive Workspace for Gemini AI

On March 18, 2025, Google unveiled Canvas for its Gemini AI platform, introducing an interactive workspace designed to enhance productivity and creativity. This new feature allows users to draft, refine, and edit documents or code in real time with AI assistance, offering tools to adjust tone, length, and formatting, as well as generate and preview HTML/React code for web app prototypes.

Aimed at rivaling similar offerings like OpenAI’s Canvas and Anthropic’s Artifacts, Gemini’s Canvas integrates seamlessly with Google Docs for collaboration and supports a variety of use cases, from writing reports to coding interactive applications.

Launched globally for Gemini and Gemini Advanced subscribers, Canvas positions Gemini as a versatile tool for creators, developers, and students alike.

@webthreeth

Public link: Audio Overview Being Generated

https://g.co/gemini/share/ee9393f9d32c

I generated this podcast from a document written by McKinsey & Co, which is a report on The State of AI in 2025.

Link to the document - https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai#/

I found the podcast very impressive

@webthreeth

The Length of AI Task Endurance Doubles Every 7 Months

The length of tasks that AI systems can successfully perform at a 50% success rate is doubling approximately every 7 months, as depicted in the METR scatter plot. Spanning from 2020 to 2024, the chart tracks various models, starting with GPT-2 and GPT-3, which could handle tasks lasting mere seconds, to more advanced models like GPT-4o and Sonnet 3.7, capable of managing tasks up to an hour by 2024.

This exponential growth highlights the accelerating capabilities of AI, with models like Sonnet 3.5, 3.6, and 3.7 showing significant leaps in performance within short timeframes, reflecting rapid advancements in AI efficiency and reliability over the years.

@webthreeth

Amharic Llama 3.2: Advancing AI for Ethiopian NLP

The Amharic Llama 3.2 model represents a significant step in natural language processing (NLP) for Amharic, one of Ethiopia’s most widely spoken languages.

Built on Meta’s Llama 3.2 transformer architecture, the model was trained from scratch using 300 million Amharic tokens and fine-tuned with high-quality datasets, including poems, stories, and Wikipedia articles. With 400 million parameters and a context length of 1024 tokens, it can generate fluent Amharic text, summarize content, and answer complex queries.

Its instruction-tuned version further enhances its capability to generate creative texts such as poems, jokes, and historical narratives, making it a valuable tool for research, education, and digital content creation.

Link to Try - https://huggingface.co/spaces/rasyosef/Llama-3.2-Amharic-Chat

Source - Yosef Worku Alemneh

@webthreeth

Perplexity AI Unveils Deep Research Update for Next Week

Perplexity AI is set to launch an enhanced version of its Deep Research feature next week, promising significant advancements in its analytical capabilities. This update will equip the tool with increased computing power, enabling it to think longer and deliver more detailed and comprehensive answers. The improved Deep Research will also incorporate code execution and the ability to render in-line charts, providing users with richer, data-driven insights.

Building on the foundation laid by its initial Deep Research launch in February 2025, this update aims to further streamline and accelerate in-depth research and analysis. The feature, designed to save users hours of work, will continue to autonomously conduct extensive searches, evaluate numerous sources, and synthesize the information into clear, actionable reports.

@webthreeth

Navigating the Future: Aravind Srinivas and Perplexity's Bold Leap into Agentic AI

Aravind Srinivas, as the CEO and co-founder of Perplexity AI, is steering the company toward an ambitious transformation by embracing the potential of agentic AI, as evidenced by his recent activities and statements on platforms like X.

In early 2025, Srinivas announced the launch of Perplexity Assistant, an agentic AI for Android devices capable of performing multi-step tasks autonomously, marking a significant shift from Perplexity’s origins as a conversational answer engine to a natively integrated assistant that can interact with apps and execute real-world actions.

@webthreeth

DeepSeek R2's Rumored to achieve 90% ARC-AGI Score

DeepSeek R2's rumored 90% score on the ARC-AGI benchmark represents a groundbreaking achievement, given that ARC-AGI—created by François Chollet—is one of the toughest tests for AI systems to demonstrate human-like reasoning and adaptability, focusing on abstract problem-solving without relying on cultural or acquired knowledge.

This score, far surpassing the 15-20% achieved by DeepSeek’s R1-Zero and R1 models (and other leading systems like OpenAI’s o1), suggests significant progress toward Artificial General Intelligence (AGI), potentially reshaping AI research, industry competition, and applications.

The implications extend beyond this technical milestone: DeepSeek’s advancements, as noted in web results, could lower barriers to AI adoption, disrupt proprietary model providers like those in the U.S., and accelerate innovation across fields.

@webthreeth