Web 3.0 Ethiopia - DeFi & AI
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
Link to the document - https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai#/
I found the podcast very impressive
@webthreeth
McKinsey & Company
The state of AI in 2025: Agents, innovation, and transformation
In this 2025 edition of the annual McKinsey Global Survey on AI, we look at the current trends that are driving real value from artificial intelligence.
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
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
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
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
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
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
Tencent's Hunyuan-T1: A Breakthrough in AI Reasoning
Tencent's Hunyuan-T1, launched by the Hunyuan team, introduces a pioneering Hybrid-Mamba-Transformer MoE architecture, establishing it as the first ultra-large-scale reasoning model of its kind, designed for exceptional speed, accuracy, and efficiency in AI processing.
Building on the foundation of the earlier Hunyuan T1-Preview released in February 2025, and enhanced by large-scale reinforcement learning, the model outperforms or matches competitors like DeepSeek R1 and GPT-4.5 across various benchmarks, including MMLU-PRO, CEval, and AIME, demonstrating superior performance in knowledge, reasoning, math, and Chinese language tasks, as illustrated in the performance charts provided.
@webthreeth
Tencent's Hunyuan-T1, launched by the Hunyuan team, introduces a pioneering Hybrid-Mamba-Transformer MoE architecture, establishing it as the first ultra-large-scale reasoning model of its kind, designed for exceptional speed, accuracy, and efficiency in AI processing.
Building on the foundation of the earlier Hunyuan T1-Preview released in February 2025, and enhanced by large-scale reinforcement learning, the model outperforms or matches competitors like DeepSeek R1 and GPT-4.5 across various benchmarks, including MMLU-PRO, CEval, and AIME, demonstrating superior performance in knowledge, reasoning, math, and Chinese language tasks, as illustrated in the performance charts provided.
@webthreeth
One of the most remarkable aspects of Artificial Intelligence has been the rapid pace of its development over the past six months. I would argue that the level of disruption during this period has been significantly higher than in the preceding 24 months.
I don’t believe the pace of development will remain this intense until 2027—or perhaps until a model achieves AGI—but what we’ve witnessed in these last six months has been truly impressive. This also serves as a notable acknowledgment of China’s role in entering the competition and accelerating disruption by choosing to open-source everything.
As a result, to justify their pricing, proprietary models now need to deliver superior performance compared to open-source alternatives.
Source - McKinsey & Co. Superagency in the Workplace
@webthreeth
I don’t believe the pace of development will remain this intense until 2027—or perhaps until a model achieves AGI—but what we’ve witnessed in these last six months has been truly impressive. This also serves as a notable acknowledgment of China’s role in entering the competition and accelerating disruption by choosing to open-source everything.
As a result, to justify their pricing, proprietary models now need to deliver superior performance compared to open-source alternatives.
Source - McKinsey & Co. Superagency in the Workplace
@webthreeth
AI Revolution Unveiled: OpenAI and Grok Launch Image Editing Innovations
OpenAI and xAI's Grok significantly advanced their AI capabilities by introducing image editing features, marking a new era of accessibility in digital creativity. OpenAI integrated its image editing plugin with DALL-E 2 and later DALL-E 3 within the ResourceSpace platform, allowing users to regenerate specific areas of an image via text prompts, simplifying tasks like inpainting without requiring advanced technical skills.
Meanwhile, Grok, leveraging its Aurora model, rolled out a feature in March 2025 that enables users to upload images and modify them by describing changes—such as adding objects, altering backgrounds, or adjusting lighting—directly through natural language on the X platform. While OpenAI’s approach caters to structured editing within a professional environment, Grok’s implementation emphasizes ease and flexibility.
@webthreeth
OpenAI and xAI's Grok significantly advanced their AI capabilities by introducing image editing features, marking a new era of accessibility in digital creativity. OpenAI integrated its image editing plugin with DALL-E 2 and later DALL-E 3 within the ResourceSpace platform, allowing users to regenerate specific areas of an image via text prompts, simplifying tasks like inpainting without requiring advanced technical skills.
Meanwhile, Grok, leveraging its Aurora model, rolled out a feature in March 2025 that enables users to upload images and modify them by describing changes—such as adding objects, altering backgrounds, or adjusting lighting—directly through natural language on the X platform. While OpenAI’s approach caters to structured editing within a professional environment, Grok’s implementation emphasizes ease and flexibility.
@webthreeth
AI-Powered Economic Boom: Predicting 30% to 100% Annual Growth by 2045
GATE, an AI and Automation Scenario Explorer built by Epoch AI, shows how AI automation might dramatically boost the global economy by replacing human labor with faster, more scalable computing power. It compares two scenarios: one where full automation leads to a 100% annual growth in Gross World Product (GWP), and a more conservative estimate with 30% yearly growth.
Starting around 2025, both scenarios predict an exponential rise in GWP, driven by a cycle where AI advancements increase economic output, which then funds further AI development. By 2045, this could result in a global economy 10 to 100 times larger than today, demonstrating AI's potential to transform economic growth over the coming decades.
@webthreeth
GATE, an AI and Automation Scenario Explorer built by Epoch AI, shows how AI automation might dramatically boost the global economy by replacing human labor with faster, more scalable computing power. It compares two scenarios: one where full automation leads to a 100% annual growth in Gross World Product (GWP), and a more conservative estimate with 30% yearly growth.
Starting around 2025, both scenarios predict an exponential rise in GWP, driven by a cycle where AI advancements increase economic output, which then funds further AI development. By 2045, this could result in a global economy 10 to 100 times larger than today, demonstrating AI's potential to transform economic growth over the coming decades.
@webthreeth
Not R2, But DeepSeek Made Stellar Improvements on the V3 Model
DeepSeek has made substantial improvements to its V3 model, a 671B parameter Mixture-of-Experts (MoE) language model, with a notable minor version update announced on March 24, 2025, as reflected in the official DeepSeek channels.
This update enhances the model’s reasoning and performance capabilities, building on its knowledge distillation pipeline that leverages reasoning patterns from the DeepSeek R1 series, incorporating advanced verification and reflection techniques to significantly boost its mathematical problem-solving and multi-task question-answering abilities.
@webthreeth
DeepSeek has made substantial improvements to its V3 model, a 671B parameter Mixture-of-Experts (MoE) language model, with a notable minor version update announced on March 24, 2025, as reflected in the official DeepSeek channels.
This update enhances the model’s reasoning and performance capabilities, building on its knowledge distillation pipeline that leverages reasoning patterns from the DeepSeek R1 series, incorporating advanced verification and reflection techniques to significantly boost its mathematical problem-solving and multi-task question-answering abilities.
@webthreeth
ARC-AGI-2 : New Rankings Instrument to measure the race towards AGI
The ARC-AGI-2 benchmark, launched alongside the ARC Prize 2025 competition, has introduced a new ranking landscape for artificial general intelligence (AGI) systems as of March 25, 2025.
Designed to be a more challenging iteration of the original ARC-AGI, this updated benchmark retains its core format—tasks that are easy for humans but difficult for AI—while raising the bar for machine reasoning capabilities. Current rankings reveal a stark contrast between human and AI performance: humans consistently achieve scores above 95% with minimal training, whereas leading AI models struggle significantly.
Notably, OpenAI's o3, which previously scored an impressive 75.7% on ARC-AGI-1’s semi-private evaluation set, is projected to drop below 30% on ARC-AGI-2, even with high compute resources. This new ranking underscores ARC-AGI-2’s effectiveness in exposing current AI limitations.
@webthreeth
The ARC-AGI-2 benchmark, launched alongside the ARC Prize 2025 competition, has introduced a new ranking landscape for artificial general intelligence (AGI) systems as of March 25, 2025.
Designed to be a more challenging iteration of the original ARC-AGI, this updated benchmark retains its core format—tasks that are easy for humans but difficult for AI—while raising the bar for machine reasoning capabilities. Current rankings reveal a stark contrast between human and AI performance: humans consistently achieve scores above 95% with minimal training, whereas leading AI models struggle significantly.
Notably, OpenAI's o3, which previously scored an impressive 75.7% on ARC-AGI-1’s semi-private evaluation set, is projected to drop below 30% on ARC-AGI-2, even with high compute resources. This new ranking underscores ARC-AGI-2’s effectiveness in exposing current AI limitations.
@webthreeth
I had an idea. So, I was thinking about uploading a ten minute discussion about the weekly updates of AI in the form of a podcast through Gemini Audio Overview. I use the Gemini Audio Overview daily to learn about new stuffs and I think people will like the idea of weekly update on AI market for ten minutes on Sunday Morning.
Should I try it out? Please like if you think this idea is cool. I will make it only once per week and it will help people digest the weekly updates easily.
Should I try it out? Please like if you think this idea is cool. I will make it only once per week and it will help people digest the weekly updates easily.
👍6
Unveiling Gemini 2.5 Pro: Google DeepMind's Experimental AI Trial for Advanced Users
Gemini 2.5 Pro, an experimental AI model by Google DeepMind, has been released on a trial basis to select Gemini Advanced subscribers as of March 25, 2025. This rollout targets advanced users and includes features like a "thinking phase" for enhanced reasoning, aimed at handling complex tasks more effectively.
The model offers a 2M token context window, advanced coding capabilities, and multimodal input support, but some users have reported bugs, suggesting it's still in early testing. Access appears limited to those with a Gemini Advanced subscription, often part of the Google One AI Premium Plan, with no official confirmation from Google yet.
@webthreeth
Gemini 2.5 Pro, an experimental AI model by Google DeepMind, has been released on a trial basis to select Gemini Advanced subscribers as of March 25, 2025. This rollout targets advanced users and includes features like a "thinking phase" for enhanced reasoning, aimed at handling complex tasks more effectively.
The model offers a 2M token context window, advanced coding capabilities, and multimodal input support, but some users have reported bugs, suggesting it's still in early testing. Access appears limited to those with a Gemini Advanced subscription, often part of the Google One AI Premium Plan, with no official confirmation from Google yet.
@webthreeth
Web 3.0 Ethiopia - DeFi & AI
Unveiling Gemini 2.5 Pro: Google DeepMind's Experimental AI Trial for Advanced Users Gemini 2.5 Pro, an experimental AI model by Google DeepMind, has been released on a trial basis to select Gemini Advanced subscribers as of March 25, 2025. This rollout targets…
I shared its early sighting on my Telegram channel on March 25, 2025, just hours before Google DeepMind's official announcement.
I gave my subscribers a first look at this experimental model, sparking discussions about its advanced features like the 2M token context window and multimodal capabilities.
Anyways, it is official.
I gave my subscribers a first look at this experimental model, sparking discussions about its advanced features like the 2M token context window and multimodal capabilities.
Anyways, it is official.
Gemini 2.5 Pro, recently released on a trial basis by Google DeepMind, showcases impressive advancements in AI capabilities, as evidenced by its performance across multiple benchmarks, including reasoning and knowledge on Humanity’s Last Exam, science on GPQA Diamond, and mathematics on AIME 2025, where it consistently outperformed models like OpenAI’s o3-mini and Claude 3.7 Sonnet.
This experimental model introduces a "thinking phase" to enhance reasoning for complex tasks, positioning it as a strong contender in the AI landscape. Additionally, Gemini 2.5 Pro aligns with the growing trend of Agentic AI, where
systems autonomously tackle intricate problems, adapt to user needs, and perform advanced functions like coding and personalization, potentially redefining how AI interacts with and supports human endeavors.
@webthreeth
This experimental model introduces a "thinking phase" to enhance reasoning for complex tasks, positioning it as a strong contender in the AI landscape. Additionally, Gemini 2.5 Pro aligns with the growing trend of Agentic AI, where
systems autonomously tackle intricate problems, adapt to user needs, and perform advanced functions like coding and personalization, potentially redefining how AI interacts with and supports human endeavors.
@webthreeth
OpenAI Unveils Next-Gen Image Model: A Leap in Visual AI Innovation
OpenAI launched a significant upgrade to its image generation capabilities, integrating a new model into its GPT-4o framework.
This release, announced via a livestream by CEO Sam Altman, marks a shift from the earlier DALL-E system to a more advanced, native image generation feature within ChatGPT and Sora. Unlike its predecessors, this model excels at blending text and imagery with high accuracy, leveraging GPT-4o’s knowledge base to produce detailed, context-aware visuals.
It’s designed for practical use, offering improved "binding" (correctly linking attributes to objects) and legible text rendering—capabilities demonstrated through examples like multi-panel comics and scientific diagrams.
The rollout began immediately for all tiers of ChatGPT users—Free, Plus, Pro, and Team—positioning it as a versatile tool for visual communication, though free users may face usage limits similar to those previously set for DALL-E 3.
@webthreeth
OpenAI launched a significant upgrade to its image generation capabilities, integrating a new model into its GPT-4o framework.
This release, announced via a livestream by CEO Sam Altman, marks a shift from the earlier DALL-E system to a more advanced, native image generation feature within ChatGPT and Sora. Unlike its predecessors, this model excels at blending text and imagery with high accuracy, leveraging GPT-4o’s knowledge base to produce detailed, context-aware visuals.
It’s designed for practical use, offering improved "binding" (correctly linking attributes to objects) and legible text rendering—capabilities demonstrated through examples like multi-panel comics and scientific diagrams.
The rollout began immediately for all tiers of ChatGPT users—Free, Plus, Pro, and Team—positioning it as a versatile tool for visual communication, though free users may face usage limits similar to those previously set for DALL-E 3.
@webthreeth
Grok's Expansion to Telegram: A New Frontier for AI Accessibility
Grok, an AI developed by xAI, announced its integration with Telegram, marking a significant step in expanding its accessibility beyond the X platform. This move follows Grok's recent rollout of a standalone iOS app in January 2025, which launched in regions like the U.S., Australia, and India, highlighting xAI's ongoing efforts to broaden its reach.
Specifically, Grok is now available on Telegram for Premium users via @GrokAI ,offering real-time answers by pulling data from X and the web, as noted in a related update (https://grok.x.com). While features like "Think" and "DeepSearch" remain exclusive to X or the Grok app, Telegram users can still access Grok's conversational AI with multimodal skills, making this an exciting tech integration.
@webthreeth
Grok, an AI developed by xAI, announced its integration with Telegram, marking a significant step in expanding its accessibility beyond the X platform. This move follows Grok's recent rollout of a standalone iOS app in January 2025, which launched in regions like the U.S., Australia, and India, highlighting xAI's ongoing efforts to broaden its reach.
Specifically, Grok is now available on Telegram for Premium users via @GrokAI ,offering real-time answers by pulling data from X and the web, as noted in a related update (https://grok.x.com). While features like "Think" and "DeepSearch" remain exclusive to X or the Grok app, Telegram users can still access Grok's conversational AI with multimodal skills, making this an exciting tech integration.
@webthreeth
Introducing Researcher and Analyst in Microsoft 365 Copilot
Microsoft has recently unveiled two innovative reasoning agents, Researcher and Analyst, integrated into Microsoft 365 Copilot, marking a significant advancement in workplace productivity tools.
Researcher leverages OpenAI’s deep research model alongside Microsoft 365 Copilot’s advanced orchestration and search capabilities to tackle complex, multi-step research tasks, such as crafting detailed go-to-market strategies or identifying emerging market opportunities by analyzing both internal work data—emails, meetings, files, and chats—and external web sources.
Analyst, built on OpenAI’s o3-mini reasoning model, operates like a skilled data scientist, transforming raw data into actionable insights through chain-of-thought reasoning and real-time Python execution, enabling users to quickly generate comprehensive reports or perform advanced data analysis.
@webthreeth
Microsoft has recently unveiled two innovative reasoning agents, Researcher and Analyst, integrated into Microsoft 365 Copilot, marking a significant advancement in workplace productivity tools.
Researcher leverages OpenAI’s deep research model alongside Microsoft 365 Copilot’s advanced orchestration and search capabilities to tackle complex, multi-step research tasks, such as crafting detailed go-to-market strategies or identifying emerging market opportunities by analyzing both internal work data—emails, meetings, files, and chats—and external web sources.
Analyst, built on OpenAI’s o3-mini reasoning model, operates like a skilled data scientist, transforming raw data into actionable insights through chain-of-thought reasoning and real-time Python execution, enabling users to quickly generate comprehensive reports or perform advanced data analysis.
@webthreeth
A Free Alternative to OpenAI’s New Image Model: Introducing Reve Image 1.0 by Reve Image
Launched on March 24, 2025, Reve Image 1.0, codenamed "Halfmoon," is a groundbreaking text-to-image model by Reve AI that rivals OpenAI’s latest offerings, delivering exceptional prompt accuracy, stunning aesthetics, and superior typography—all for free.
Available at preview.reve.art with no sign-up or hidden costs, this model empowers users to generate unlimited high-quality images in seconds, outpacing competitors like Midjourney v6.1 and Google’s Imagen 3 in the Artificial Analysis Image Arena rankings. Whether you’re crafting photorealistic scenes, abstract art, or text-heavy designs, Reve Image 1.0 provides a powerful, accessible alternative to OpenAI’s subscription-based tools, making advanced AI creativity available to everyone.
@webthreeth
Launched on March 24, 2025, Reve Image 1.0, codenamed "Halfmoon," is a groundbreaking text-to-image model by Reve AI that rivals OpenAI’s latest offerings, delivering exceptional prompt accuracy, stunning aesthetics, and superior typography—all for free.
Available at preview.reve.art with no sign-up or hidden costs, this model empowers users to generate unlimited high-quality images in seconds, outpacing competitors like Midjourney v6.1 and Google’s Imagen 3 in the Artificial Analysis Image Arena rankings. Whether you’re crafting photorealistic scenes, abstract art, or text-heavy designs, Reve Image 1.0 provides a powerful, accessible alternative to OpenAI’s subscription-based tools, making advanced AI creativity available to everyone.
@webthreeth