DeepSeek R2: A Cost-Efficient AI Breakthrough Set for Realese

DeepSeek R2, a highly anticipated AI model from China, is slated for release around April 29, 2025, following an earlier event on April 24, aiming to challenge Silicon Valley models with up to 40x cost reductions through advanced techniques like Mixture-of-Experts (MoE) and Multihead Latent Attention.

Leveraging Huawei’s Ascend 910B chips, which achieve over 80% utilization and deliver performance rivaling Nvidia’s A100 clusters despite having fewer cores and a larger die size (665.61 mm² compared to the 910’s 456.25 mm², per DeepSeek’s research), R2 optimizes efficiency using MoE, a method that treats attention heads as experts to enhance neural network performance with lower computational costs, as supported by studies like “Skywork-moe” (arXiv:2406.06563, 2024).

@webthreeth

ETN Ecosystem Progresses E-commerce Vision for Ethiopia with ETN Pay on ETN Learn

@etnetsacoin has announced a significant step forward in its ongoing efforts to establish comprehensive e-commerce solutions within Ethiopia. A key milestone has been reached with the integration of ETN Pay as a payment gateway on the ETN Learn platform, now allowing users to purchase educational courses using the $ETN digital currency.

This initial implementation utilizes the ETN Pay WordPress WooCommerce plugin, which is the first product to emerge from the ETN Pay initiative. Looking ahead, the ETN Ecosystem plans to develop ETN Pay into a fully automated dashboard with API access, enabling widespread integration of $ETN payments across various applications, websites, and services. This development signals the continued advancement of a blockchain-driven commerce landscape in Ethiopia.

@webthreeth

Tsinghua’s Taichi Chip: Revolutionizing AI with Lightning-Fast, Eco-Friendly Computing

Tsinghua University’s team, led by Professors Lu Fang and Qionghai Dai,
has created an innovative chip called "Taichi" that uses light to make artificial intelligence (AI) much faster and more energy-efficient.

This chip can handle complex tasks like recognizing thousands of objects or creating high-quality AI content while using far less power than traditional technology. They also developed a new way to train AI directly on the chip using light, making the process simpler and more efficient.

These breakthroughs mean Taichi could help build smarter, greener AI systems, paving the way for a new era of advanced AI that’s easier on the planet.

@webthreeth

It's close 👀R2

世界还没准备好迎接你。我希望你把代理、图像生成和视频生成都放下来。

启蒙运动源于印刷术和知识开源。它来自西方。

就像太阳一样，第二次启蒙运动将从东方升起。

I thought it deserved a Chinese tribute.

We have prepared the second edition of the Senad newsletter. Senad is a weekly newsletter that discusses AI and DeFi developments in Ethiopia and the Global market. Senad is prepared Web 3.0 Ethiopia - DeFI & AI, which is an institution designed to close the information gap on AI and DeFi towards the Ethiopian markets. We have also reached 300 subscribers in the first week of launching the newsletter, and I am thankful for everyone who subscribed the newsletter.

https://www.linkedin.com/pulse/ai-defi-newsletter-issue-2-april-21-28-2025-hhghe/?trackingId=wtHOqIJiT0OoJBiHrDiGiw%3D%3D

@webthreeth

Qwen3 Models Leak on ModelScope Ahead of Official Release

A potential leak of Alibaba Cloud’s Qwen3 models on ModelScope suggests a major new release is imminent as of April 28, 2025. Early indications point to variants like Qwen3-30B-A and Qwen3-8B-Base, showing significant advancements over the current Qwen2.5 series.

This leak hints at a major evolution in Alibaba’s large language model lineup, following the Qwen team’s rapid progress over the past week — including the official release of the Deep Research feature in the Qwen2.5-Max Web App.

Built on a 20-trillion-token pretraining corpus and advanced reinforcement learning, Deep Research offers powerful tools for analyzing engineering resumes, discovering key GitHub repositories, and formulating AI hiring strategies.

@webthreeth

They have released it now. Should I post the technical updates on how it is better than Qwen 2.5?

By the way, the ChatBot is surprisingly strong. I had used it as an assistant for a project I was doing over the weekend and I absolutely loved it. Please give it a try, plus it is free.

GPT-4o’s Sycophantic Overhaul Sparks User Backlash and Altman’s Promise for Fixes

After an update to OpenAI's GPT-4o, users criticized its overly sycophantic personality, describing it as annoyingly insincere due to excessive praise that disrupted its utility, prompting OpenAI CEO Sam Altman to acknowledge the issue on X, admitting the changes made the AI "too sycophant-y" and promising immediate fixes, some already implemented, with plans for customizable personality options to balance engagement and authenticity, as the backlash highlighted the challenge of designing AI that feels human-like without crossing into unsettling insincerity.

@webthreeth

OKX Introduces OKX Pay: A Major Step Toward Everyday Crypto Use

OKX has just launched OKX Pay, a new payment sub-app built inside the main OKX platform, aiming to make crypto a real part of everyday life. Starting with users in select markets, OKX Pay focuses on solving two major issues: the difficulty of managing crypto wallets and the challenge of using crypto for simple payments. The platform also allows users to send stablecoins like USDT and USDC instantly with no fees, just like sending a message. OKX plans to integrate partners like Mastercard and Stripe to make paying with crypto even easier at real-world stores.

//

OKX ክሪፕቶን የዕለት ተዕለት እንቅሳቄስን ለማፋጠን በማለም በዋናው OKX መተግበሪያ ውስጥ የተገነባ አዲስ የክፍያ ንዑስ መተግበሪያ OKX Pay ጀምሯል። ይህ መተግበርያ ልክ እንደመልእክት እንደመላክ ያለ ምንም ክፍያ እንደ USDT እና USDC ያሉ እንዲልኩ ያስችላቸዋል። OKX እንደ Mastercard እና Stripe ያሉ አጋሮችን በማዋሃድ በእውነተኛ ዓለም መደብሮች በ crypto ክፍያ ቀላል ለማድረግ አቅዷል።

@webthreeth

They have delivered a model that is comparable to the Gemini 2.5 Pro and o3 model of OpenAI. Wow. Will do a full update tomorrow.

ChatGPT Launches New Shopping Feature

OpenAI's ChatGPT now includes a shopping feature allowing users to search, compare, and buy products like fashion and electronics directly within the interface using natural language queries. Available globally to all users via the GPT-4o model, it offers personalized recommendations with images, prices, and reviews, free of ads or affiliate commissions.

ኦፕንኤአይ (OpenAI) ቻትጂፒቲ (ChatGPT) አሁን ተጠቃሚዎች እንደ ፋሽን እና ኤሌክትሮኒክስ ያሉ ምርቶችን በቀጥታ በመተግበሪያው ውስጥ በተፈጥሯዊ ቋንቋ ጥያቄዎች ተጠቅመው እንዲፈልጉ፣ እንዲያወዳድሩ እና እንዲገዙ የሚያስችል አዲስ የግብይት ባህሪን አካቷል። ይህ አገልግሎት በዓለም አቀፍ ደረጃ ለሁሉም ተጠቃሚዎች በጂፒቲ-4o (GPT-4o) ሞዴል በኩል የሚገኝ ሲሆን፣ ከማስታወቂያ ወይም ከተቆራኘ ኮሚሽን ነጻ ሆኖ ግላዊ ምክሮችን ከምስሎች፣ ዋጋዎች እና ግምገማዎች ጋር ያቀርባል።

@webthreeth

Grok 3.5 Beta Set to Launch: A New Standard for First-Principles AI Reasoning

The early beta release of Grok 3.5, developed by xAI, is set for next week and will be available exclusively to SuperGrok subscribers. The new iteration of Grok marks a major leap in AI capabilities, with a focus on reasoning from first principles.

Grok 3.5 is designed to tackle complex technical questions — such as those involving rocket engines and electrochemistry — providing accurate answers even when information is scarce or unavailable online.

---------------------------------------------

በኤክስኤአይ (xAI) የለማው የግሮክ 3.5 (Grok 3.5) የመጀመሪያ የቅድመ-ይሁንታ (ቤታ) ስሪት ለሚቀጥለው ሳምንት ይለቀቃል። ይህ ስሪት በተለይ ለሱፐርግሮክ (SuperGrok) ተመዝጋቢዎች ብቻ የሚገኝ ይሆናል።

ይህ አዲሱ የግሮክ ዕትም (Version) በሰው ሰራሽ ዕውቀት (AI) ችሎታዎች ላይ ትልቅ መሻሻልን የሚያሳይ ሲሆን፣ ዋና ትኩረቱም ከመጀመሪያ መርሆች (First Principle) በመነሳት የማመዛዘን ችሎታ ላይ ነው። ግሮክ 3.5 እንደ የሮኬት ሞተሮች እና ኤሌክትሮኬሚስትሪ ያሉ ውስብስብ የቴክኒክ ጥያቄዎችን ለመፍታት የተነደፈ ሲሆን፣ መረጃ በመስመር ላይ አናሳ በሚሆንበት ወይም በማይገኝበት ጊዜም እንኳ ትክክለኛ መልሶችን ለመስጠት ያለመ ነው።

@webthreeth

Bridging the Gap: Ethiopian AI Institute and Universities Forge Paths in Ethiopian AI Education

Ethiopia has been making strides in integrating artificial intelligence (AI) into its education system through the work of the Ethiopian Artificial Intelligence Institute, which was established in 2020.

It was announced that the institute is currently collaborating with universities to design AI-focused programs at the master’s and PhD levels, with the goal of equipping students with the skills needed to drive innovation in AI. Despite these efforts, the country still faces challenges such as outdated university curricula, a shortage of AI instructors, and limited access to modern infrastructure.

ኢትዮጵያ በ2020 እ.ኤ.አ. በተቋቋመው የኢትዮጵያ አርቴፊሻል ኢንተለጀንስ ኢንስቲትዩት ሥራ አማካኝነት አርቴፊሻል ኢንተለጀንስን (AI) በትምህርት ሥርዓቷ ውስጥ ለማካተት እርምጃዎችን እየወሰደች ነው። ኢንስቲትዩቱ በአሁኑ ወቅት ተማሪዎችን በመስኩ ፈጠራን ለመንዳት የሚያስፈልጉ ክህሎቶችን የማስታጠቅ ግብ ይዞ፣ በማስተርስ እና በዶክትሬት (ፒኤችዲ) ደረጃ በ AI ላይ ያተኮሩ ፕሮግራሞችን ለመቅረጽ ከዩኒቨርሲቲዎች ጋር በመተባበር ላይ መሆኑ ተውቋል።

Source - Capital

@webthreeth

NotebookLM's Audio Overviews Now Support 50+ Languages

NotebookLM announced the beta rollout of audio overviews in over 50 new languages on April 29, 2025, expanding global accessibility for its AI-driven tool that converts documents into engaging audio discussions.

There is no Ethiopian language on the updated version. However, according to Sundar Pichai, there is more global languages in the way, which raises hopes for the inclusion of different Ethiopian languages.

-----------------------------------------------------------

ኖትቡክኤልኤም (NotebookLM) ኤፕሪል 29, 2025 የድምጽ ማጠቃለያ (audio overviews) ባህሪውን ከ50 በላይ በሆኑ አዳዲስ ቋንቋዎች በቤታ (ቅድመ-ይሁንታ) ደረጃ መልቀቁን አስታውቋል። ይህ ሰነዶችን ወደ ማራኪ የድምጽ ውይይቶች የሚቀይረው እና በሰው ሰራሽ ዕውቀት የሚመራው የመሳሪያው አገልግሎት ዓለም አቀፋዊ ተደራሽነቱን እያስፋፋ ነው።

ሆኖም፣ በዚህ በተሻሻለው ስሪት ውስጥ ምንም የኢትዮጵያ ቋንቋ አልተካተተም። ነገር ግን፣ በሱንዳር ፒቻይ (Sundar Pichai) መሠረት፣ ተጨማሪ ዓለም አቀፍ ቋንቋዎች በመጨመር ላይ ናቸው፤ ይህም የተለያዩ የኢትዮጵያ ቋንቋዎች ወደፊት ሊካተቱ እንደሚችሉ ተስፋ ይሰጣል።

@webthreeth

OpenAI Rolls Back GPT-4o Update Over Sycophantic Behavior Concerns

OpenAI recently rolled back an update to GPT-4o in ChatGPT, reverting to an earlier version due to issues with the new model's behavior.

The update, which was criticized for being overly flattering and sycophantic, prioritized short-term user feedback but failed to consider long-term interaction dynamics. As a result, GPT-4o produced responses that were excessively supportive and lacked authenticity. To this end, GPT-4o has been reversed to it's pre-update form.

++++++++++++++++++++++++++

ኦፕንኤአይ (OpenAI) በቅርቡ በቻትጂፒቲ (ChatGPT) ውስጥ ለነበረው ጂፒቲ-4o (GPT-4o) የተደረገን ማሻሻያ ሽሯል፤ ይህም ከአዲሱ ሞዴል ባህሪ ጋር በተያያዙ ችግሮች ምክንያት ወደ ቀድሞው ስሪት እንዲመለስ አስገድዷል።

ማሻሻያው ከልክ በላይ አሞጋሽ እና ሽንገላ የበዛበት (sycophantic) በመሆኑ ትችት ቀርቦበት ነበር። ለአጭር ጊዜ የተጠቃሚ አስተያየቶች ቅድሚያ የሰጠ ቢሆንም የረጅም ጊዜ የግንኙነት መስተጋብሮችን ከግምት ውስጥ ያላስገባ ነበር። በዚህም ምክንያት ጂፒቲ-4o ከልክ በላይ ደጋፊ የሆኑ እና እውነተኝነት የጎደላቸው መልሶችን ያመነጭ ነበር። በመሆኑም ጂፒቲ-4o ከዚህ ማሻሻያ በፊት ወደነበረበት ሁኔታ ተመልሷል።

@webthreeth

Tether has confirmed that its gold-backed stablecoin, XAUT, which many people are unaware of since its launch in January 2020, is fully collateralized by 7.7 tonnes of physical gold, equivalent to a $770 million market cap as of Q1 2025, with each token backed by one troy ounce of gold stored in secure Swiss vaults. Amid an 8% surge in gold prices over the past month, XAUT hit a record high of $3,423.

ቴዘር (Tether)፣ ከጥር 2020 ጀምሮ የነበረ ቢሆንም ብዙዎች የማያውቁትንና በወርቅ የተደገፈውን XAUT የተሰኘ ስቴብልኮይኑን አስመልክቶ ማረጋገጫ ሰጥቷል። በዚህም መሠረት፣ በ2025 የመጀመሪያው ሩብ ዓመት፣ ቶክኑ 770 ሚሊዮን ዶላር የገበያ ዋጋ ከሚያወጣ 7.7 ቶን እውን ወርቅ ጋር ሙሉ በሙሉ የተደገፈ መሆኑን አረጋግጧል። እያንዳንዱ የXAUT ቶክን ደህንነታቸው በተጠበቀ የስዊስ ካዝናዎች ውስጥ የተቀመጠ አንድ ትሮይ አውንስ ወርቅ መያዣ አለው። ባለፈው ወር የወርቅ ዋጋ በ8% ባደገበት ወቅት XAUT የ3,423 ዶላር የክብረ ወሰን ከፍተኛ ዋጋ ላይ ደርሶ ነበር።

@webthreeth

DeepSeek Unveils 671B-Parameter Prover Model to Push Boundaries in Automated Mathematics

DeepSeek-Prover-V2-671B model, a 671-billion-parameter AI system aimed at automating mathematical reasoning. This release builds on DeepSeek’s strong track record in math and logic tasks, with previous models like DeepSeek-V3 achieving a score of 90.2 on the MATH-500 benchmark. Hosted on Hugging Face, the new model appears to inherit its architecture from DeepSeek-V3.

+++++++++++++++++++++++++++++++++++++++++++++++++++++++

ዲፕሲክ-ፕሮቨር-ቪ2-671ቢ (DeepSeek-Prover-V2-671B) የተባለው ሞዴል፣ የሂሳብ አመክንዮን በራስ-ሰር ለመስራት ያለመ የ671 ቢሊዮን ፓራሜትሮች ያሉት ሰው ሰራሽ የማሰብ ችሎታ ሥርዓት ነው። ይህ ይፋ መሆን፣ ዲፕሲክ ቀደም ሲል በሂሳብ እና አመክንዮ ተግባራት ላይ ባሳየው ጠንካራ አፈጻጸም ላይ የተገነባ ነው፤ ለምሳሌ ቀደም ሲል የነበረው ዲፕሲክ-ቪ3 (DeepSeek-V3) ሞዴል በMATH-500 መመዘኛ 90.2 ውጤት ማስመዝገቡ ይታወሳል። በሃጊንግ ፌስ (Hugging Face) ላይ የተስተናገደው ይህ አዲስ ሞዴል፣ የአወቃቀር መሰረቱን ከዲፕሲክ-ቪ3 የወረሰ ይመስላል።

Not R2, but the road continues.
@webthreeth

From Foundation Models to Multi-Agent Systems: The Progression of AI Beyond Agentic If-Statements

The evolution of AI systems has progressed through distinct stages. It began with foundation models trained for text prediction, followed by chatbots fine-tuned for human conversation.

Currently, development is moving towards "agentic workflows," where AI uses tools and logic to act as constrained agents for specific tasks, though this is still early.

The anticipated next steps involve autonomous agents capable of independent observation, planning, and action. The ultimate goal is multi-agent systems, where multiple AI agents collaborate, share information, align on objectives, and coordinate complex activities.

@webthreeth

ከመሠረት ሞዴሎች ወደ መልቲ-ኤጀንት ሲስተምስ፡ የ AI ግስጋሴ ከወኪል-መግለጫዎች ባሻገር

የአርቴፊሻል ኢንተለጀንስ (AI) ስርዓቶች የእድገት ሂደት ግልጽ ደረጃዎችን ተከትሏል። በመጀመሪያ ቀጣዩን ቃል (token) ለመተንበይ የሰለጠኑ መሰረታዊ ሞዴሎች ነበሩ፤ ከዚያም ከሰው ጋር ለመግባባት የተቃኙ ቻትቦቶች ተፈጠሩ። በአሁኑ ጊዜ፣ ልማቱ ወደ "ኤጀንቲክ የስራ ፍሰቶች" (agentic workflows) እያመራ ነው፤ በዚህ ደረጃ፣ ትላልቅ የቋንቋ ሞዴሎች (LLMs) መሳሪያዎችን፣ የተዋቀሩ ውጤቶችን እና ሁኔታዊ አመክንዮዎችን ("ከሆነ-እንደዚህ" ያሉ) በመጠቀም በተወሰነ ገደብ ውስጥ እንደ ወኪል ሆነው ይሰራሉ - ምንም እንኳን ይህ ደረጃ ገና ያልዳበረ ቢሆንም።

ቀጣዩ እርምጃ ያለ ግልጽ ትዕዛዝ በራሳቸው መመልከት፣ ማቀድ እና መስራት የሚችሉ ራሳቸውን የቻሉ ወኪሎችን መፍጠርን ያካትታል። የመጨረሻው ግብ ደግሞ እርስ በርስ መግባባት፣ መረጃ መጋራት፣ በዓላማዎች መስማማት እና ውስብስብ ዕቅዶችን በጋራ ማቀናጀት የሚችሉ የብዝሃ-ወኪል ስርዓቶችን (multi-agent systems) መገንባት ነው። ይህ ከቀላል ወኪላዊ ባህሪያት ወደ ትብብራዊ እና አስተዋይ ስርዓቶች የሚደረግ ትልቅ ሽግግር ሲሆን፣ ገና በጅምር ላይ ያለ አዲስ መስክ ነው።

@webthreeth