GitHub Copilot Evolves: Open-Source Integration and Autonomous Coding Capabilities

GitHub Copilot is now integrated directly within GitHub, enabling developers to collaborate more effectively by allowing the AI to autonomously handle tasks such as bug fixes and feature additions. Accessible from both GitHub and Visual Studio Code, this integration enhances the coding workflow with minimal manual input. As part of a broader initiative, Microsoft is open-sourcing Visual Studio Code and its AI capabilities—including the GitHub Copilot Chat extension—under the MIT license.

ጊትሀብ ኮፓይለት አሁን በቀጥታ በጊትሀብ ውስጥ ተዋህዷል፤ ይህም አርቴፊሻል ኢንተለጀንስ የሶፍትዌር ክፍተት ጥገናዎችን እና የአዳዲስ ተግባራት መጨመርን በመሳሰሉ ተግባራት ላይ በራሱ እንዲሰራ በማስቻል የሶፍትዌር አልሚዎች በበለጠ ውጤታማነት እንዲተባበሩ ያስችላቸዋል፡፡ ይህ ውህደት ከጊትሀብም ከቪዥዋል ስቱዲዮ ኮድም ተደራሽ ሲሆን፣ በአነስተኛ የእጅ ግብዓት የኮዲንግ የሥራ ሂደቱን ያሻሽላል፡፡ እንደ አንድ ሰፊ ተነሳሽነት አካል፣ ማይክሮሶፍት ቪዥዋል ስቱዲዮ ኮድን እና የጊትሀብ ኮፓይለት የውይይት ተጨማሪን (GitHub Copilot Chat extension) ያካተቱትን ሰው ሰራሽ አስተውሎት ችሎታዎቹን በ MIT ፈቃድ መሰረት ምንጫቸውን ለሕዝብ ክፍት እያደረገ ነው፡፡

@webthreeth

https://www.youtube.com/live/o8NiE3XMPrM?si=Kid84pGi6tnzShKj

Google ain't joking around at all. They are not here to compete, but to take over.

Should I do a summary of all they have done in one Telegram Post, or should I post every feature specifically here for tomorrow? Like if you think I should post every feature?

Feature #1 - Lyria 2.0

Lyria 2, developed by Google DeepMind, represents a significant advancement in AI-driven music generation, offering high-fidelity, professional-grade audio outputs across a diverse range of genres and intricate compositions. This model enables musicians to compose complex passages, suggest harmonies, and draft longer arrangements with unprecedented speed, thereby accelerating the creative process and helping artists overcome creative blocks.

ሊሪያ 2፣ በጎግል DeepMind የተሰራ፣ ከፍተኛ ጥራት ያለው ኦዲዮን በተለያዩ ዘውጎች የሚያመርት የላቀ AI ሙዚቃ ትውልድ ሞዴል ነው። ሙዚቀኞች የተወሳሰቡ ክፍሎችን እንዲያዘጋጁ፣ ስምምነትን እንዲጠቁሙ እና ረጅም ዝግጅቶችን በፍጥነት እንዲፈጥሩ፣ ፈጠራን በማጎልበት እና የፈጠራ ብሎኮችን ለማሸነፍ ይረዳል። አዳዲስ ዘይቤዎችን እና ቴክኒኮችን የማሰስ ችሎታው ለሙዚቃ ባለሙያዎች ጥበባዊ እድሎችን ያሰፋል። ከሊሪያ 2 ጋር፣ በሰዎች ፈጠራ እና በ AI የታገዘ ጥንቅር መካከል ያለው ድንበር ከጊዜ ወደ ጊዜ እንከን የለሽ ይሆናል።

@webthreeth

Feature #2 - Veo 3

Veo 3, introduced by Google DeepMind, represents a significant advancement in generative AI technology, capable of creating videos with synchronized soundtracks that include dialogue, sound effects, and ambient noise. This model enhances the realism and cinematic quality of generated content, making it a powerful tool for storytelling and creative expression. Veo 3 builds on previous research in generative models, such as GANs and diffusion models, showcasing the progression of AI's ability to simulate complex sensory experiences.

ጉግል ዲፕማይንድ ያስተዋወቀው ቪኦ 3፣ በጄኔሬቲቭ AI ቴክኖሎጂ ውስጥ ትልቅ መሻሻል ያሳያል፤ ይህም ንግግርን፣ የድምፅ ውጤቶችን እና የአካባቢ ጫጫታዎችን የሚያካትት የተመሳሰለ የድምፅ ትራክ ያላቸው ቪዲዮዎችን መፍጠር ይችላል። ይህ ሞዴል የተፈጠረው ይዘት ተጨባጭነትን እና ሲኒማዊ ጥራትን የሚያሳድግ ሲሆን፣ ለታሪክ ፈጠራ እና ለፈጠራ መግለጫ ኃይለኛ መሳሪያ ያደርገዋል። ቪኦ 3 እንደ GANs እና diffusion models ባሉ ቀደምት የጄኔሬቲቭ ሞዴሎች ላይ በመገንባት፣ AI ውስብስብ የስሜት ህዋሳት ልምዶችን የማስመሰል ችሎታ መሻሻልን ያሳያል።

@webthreeth

Feature #3 - Gemini 2.5 Pro Deep Think

Gemini 2.5 Pro Deep Think, unveiled by Google at the Google I/O 2025 event, represents a significant leap in AI reasoning capabilities, leveraging parallel thinking techniques to explore multiple hypotheses before generating responses.

This enhanced reasoning mode has enabled the model to achieve remarkable benchmark scores, including an 84% on the multimodal reasoning test MMMU, a 49% on the challenging 2025 United States of America Mathematical Olympiad (USAMO), and a 81.4% on LiveCodeBench v6.0 for code generation, surpassing competitors like OpenAI’s o3.

ጎግል በGoogle I/O 2025 ዝግጅት ላይ ይፋ ያደረገው ጀሚኒ 2.5 ፕሮ ዲፕ ቲንክ፣ ምላሾችን ከማመንጨቱ በፊት በርካታ መላምቶችን ለመመርመር ትይዩ አስተሳሰብ (parallel thinking) ዘዴዎችን በመጠቀም፣ በ AI የማመዛዘን ችሎታዎች ላይ ከፍተኛ መሻሻልን ያመለክታል።

@webthreeth

Feature #4 - Gemini Diffusion

ጎግል ዲፕማይንድ (Google DeepMind) በGoogle I/O 2025 ላይ ይፋ ያደረገው ጀሚኒ ዲፍዩዥን (Gemini Diffusion)፣ የ AI የጽሑፍ ማመንጨት ዘዴን እንደገና የሚገልጽ የሙከራ ጽሑፍ ዲፍዩዥን ሞዴል ነው። ጀሚኒ ዲፍዩዥን ጽሑፍን ቶከን በቶከን እንደሚያመነጩት ከተለመዱት autoregressive models በተለየ፣ በዘፈቀደ ድምጾች (random noise) ይጀምርና ቀስ በቀስ ወደ ትርጉም ባለው ጽሑፍ ወይም ኮድ ያጠራዋል። ይህ ዘዴ ሙሉ የይዘት ስብስቦችን በአንድ ጊዜ እንዲፈጠር ያስችላል።

Gemini Diffusion, unveiled by Google DeepMind at Google I/O 2025, is an experimental text diffusion model that redefines AI text generation. Unlike traditional autoregressive models that generate text token by token, Gemini Diffusion starts with random noise and iteratively refines it into coherent text or code, producing entire blocks of content simultaneously.

@webthreeth

Feature #5 - Gemini Agent's Mode

Gemini's Agents Mode, as part of Google's AI ecosystem, enables the creation of autonomous, task-oriented AI agents capable of executing complex workflows with minimal human intervention. This mode leverages Gemini's multimodal capabilities, integrating text, image processing, and potentially other data types to perform tasks like data analysis, content generation, or automated decision-making.

የጉግል ኤአይ ሥነ-ምህዳር አካል የሆነው የጀሚኒ "ኤጀንትስ ሞድ"፣ በትንሹ የሰው ልጅ ጣልቃ ገብነት ውስብስብ ሥራዎችን ማከናወን የሚችሉ ራሳቸውን የቻሉ፣ ተግባር-ተኮር የሆኑ የኤአይ ወኪሎችን (AI agents) ለመፍጠር ያስችላል። ይህ ሞድ የጀሚኒን ባለብዙ-ሞዳል (multimodal) ችሎታዎች በመጠቀም ጽሑፍን፣ ምስልን የማቀናበር እና ሌሎች የዳታ አይነቶችን በማቀናጀት እንደ ዳታ ትንተና፣ የይዘት ማመንጨት ወይም አውቶማቲክ ውሳኔ አሰጣጥ የመሳሰሉ ተግባራትን ያከናውናል።

@webthreeth

Devstral: Mistral’s Open-Source AI Model for Advanced Code Generation

Devstral is a new open-source AI coding model released by Mistral in collaboration with All Hands AI, licensed under Apache 2.0 for unrestricted commercial use. Positioned as a powerful tool for software engineering agents, Devstral outperforms other open models like Google’s Gemma 3 27B and DeepSeek’s V3 on the SWE-Bench Verified benchmark for coding tasks.

ሚስትራል ከተባለ ኩባንያ ከኦል ሃንድስ ኤ አይ ጋር በመተባበር ዴቭስትራል የተሰኘ አዲስ የክፍት-ምንጭ አርቴፊሻል ኢንተለጀንስ (ኤ አይ) ኮዲንግ ሞዴል ይፋ አደረገ። ይህ ሞዴል በአፓቼ 2.0 ፈቃድ ስር የተለቀቀ ሲሆን ይህም ለያልተገደበ የንግድ አጠቃቀም ነጻ ያደርገዋል። ዴቭስትራል በተለይ ለሶፍትዌር ምህንድስና ወኪሎች (software engineering agents) እንደ ισχυρό መሳሪያ የተቀረጸ ነው። በኮዲንግ ተግባራት አፈጻጸም በታወቀው ኤስ ደብልዩ ኢ-ቤንች የተረጋገጠ መመዘኛ (SWE-Bench Verified benchmark) ላይ እንደ ጉግል ጌማ 3 27ቢ (Google’s Gemma 3 27B) እና ዲፕሲክ ቪ3 (DeepSeek’s V3) ያሉ ሌሎች የክፍት ሞዴሎችን በበላይነት ማሸነፉ ተነግሯል።

@webthreeth

AI Safety Level Framework by Anthropic

The AI Safety Level (ASL) framework categorizes AI systems based on their potential for catastrophic risk. ASL-1 includes systems with no meaningful risk, such as a 2018 language model or an AI that only plays chess. ASL-2 covers systems with early signs of dangerous capabilities—like giving instructions on bioweapon creation—but where the information lacks reliability or isn’t more harmful than what’s already accessible through a search engine; current large language models, including Claude, fall into this category.

ASL-3 involves systems that significantly increase the risk of catastrophic misuse compared to traditional sources like search engines or textbooks, or that display early autonomous behaviors. ASL-4 and ASL-5+ have not yet been clearly defined, as they represent future, more advanced systems likely to exhibit a qualitative leap in misuse potential and autonomous capability.

Today, we are having the first ASL-3 level AI from Anthropic.

@webthreeth

Stargate UAE: A 1-GW AI Superhub

Stargate UAE is a joint initiative by OpenAI, Oracle, NVIDIA, Cisco, and SoftBank Group Corp. to build a 1-gigawatt high-performance computing facility at the center of the planned 5-GW UAE–US AI Campus. Purpose-built for the most demanding AI workloads, the hub will supply massive computational power to accelerate breakthroughs in healthcare, energy, finance, and advanced research, making the region a key global node for next-generation artificial-intelligence innovation. Under the partnership, the UAE will become the first country in the world to enable ChatGPT nationwide, giving people across the country the ability to access OpenAI's technology. Stargate UAE has the potential to provide AI infrastructure and compute capacity within a 2,000-mile radius, reaching up to half the world’s population.

@webthreeth

Anthropic Unveils Claude 4 Opus and Sonnet Models

Anthropic launched Claude 4 Opus and Claude 4 Sonnet, two advanced AI models designed to push the boundaries of coding and agentic tasks. Claude 4 Opus excels in complex, long-duration coding projects, capable of autonomously handling tasks that take humans up to seven hours, while Claude 4 Sonnet offers frontier performance for practical AI use cases. Cursor, an AI-powered code editor, announced the integration of Anthropic's Claude Sonnet 4 model on May 22, 2025, just 11 minutes after Anthropic's release.

አንትሮፒክ የተባለ ድርጅት፣ የኮዲንግ እና የተወካይ ተግባራትን ወሰን ለማስፋት ታስበው የተሰሩ ሁለት የተራቀቁ የሰው ሰራሽ የማሰብ ችሎታ ሞዴሎችን ማለትም ክላውድ 4 ኦፐስ (Claude 4 Opus) እና ክላውድ 4 ሶኔት (Claude 4 Sonnet) በተከታታይ የቀጥታ ስርጭቶች ይፋ አድርጓል። ክላውድ 4 ኦፐስ ውስብስብ እና ረጅም ጊዜ የሚወስዱ የኮዲንግ ፕሮጀክቶችን በማከናወን የላቀ ሲሆን፣ ሰዎች እስከ ሰባት ሰዓት ሊወስድባቸው የሚችሉ ተግባራትን በራሱ የማስተናገድ ችሎታ አለው። ከርሰር (Cursor) የተባለ በሰው ሰራሽ የማሰብ ችሎታ የሚንቀሳቀስ የኮድ አርታኢ፣ አንትሮፒክ ይፋ ካደረገ ከ11 ደቂቃዎች በኋላ፣ የአንትሮፒክን ክላውድ ሶኔት 4 ሞዴል ማካተቱን አስታውቋል::

@webthreeth

All of this is done by Veo3, according to Ehud AI Studio. It is in Amharic Videos.

Full Video - https://youtu.be/DXF-mnlvlEE?si=xBd9PL4j9xvaebuU

Wow. Especially the audio part is already done very well
@webthreeth

Seifu on EBS by Veo3

Source - Abenezer Alemayhu (Ehud AI Studio)

Join the channel - @webthreeth

AI Approaches Human-Level Math on FrontierMath Benchmark

A recent competition at MIT, highlighted by Epoch AI, showcased the o4-mini-medium AI model outperforming most human teams on FrontierMath.O4-mini solved 22% of the problems, exceeding the average human team's score of 19%. While o4-mini outperformed all but two of the eight human teams of mathematicians and undergraduates, the combined human score across all teams reached 35%, suggesting AI has not yet fully surpassed human capabilities. However, Epoch AI anticipates that AI could achieve superhuman performance on FrontierMath by the end of 2025.

በቅርቡ በማሳቹሴትስ የቴክኖሎጂ ኢንስቲትዩት (MIT) በተካሄደና በኢፖክ ኤአይ (Epoch AI) ትኩረት በተሰጠው ውድድር፣ “ኦ4-ሚኒ-ሚዲየም” (o4-mini-medium) የተሰኘው የሰው ሰራሽ አስተውሎት ሞዴል በፍሮንትየርማዝ (FrontierMath) ላይ ከአብዛኞቹ የሰው ቡድኖች የተሻለ ውጤት አስመዝግቧል። ፍሮንትየርማዝ ውስብስብ እና አዲስ ዓይነት የሂሳብ ችግሮችን መፍታትን የሚለካ መድረክ ነው። "o4-mini-medium" ከቀረቡት ችግሮች ውስጥ 22 በመቶውን መፍታት የቻለ ሲሆን፣ ይህም በውድድሩ የተሳተፉ የሰው ቡድኖች ካስመዘገቡት አማካይ 19 በመቶ ውጤት ይበልጣል።

@webthreeth

AI Search Tools Erode Google’s Market Share

In March 2025, Google’s global search market share dropped to 89.71%—its lowest since 2015—according to Statcounter. This decline reflects a significant shift as users increasingly prefer AI-powered search tools like xAI’s Grok, ChatGPT, and Perplexity. These platforms offer direct, conversational responses, outperforming traditional link-based results. Frustration with Google's ad-heavy, SEO-saturated pages has accelerated this trend, as AI tools provide faster, more personalized insights, challenging Google’s dominance in the search landscape.

በማርች 2025 የጉግል የአለማቀፍ የፍለጋ ገበያ ድርሻ ወደ 89.71% ዝቅ ብሏል—ይህም ከ2015 ወዲህ ያለው ዝቅተኛው ነው—ሲል ስታትካውንተር (Statcounter) አስታውቋል። ይህ ቅናሽ ተጠቃሚዎች እንደ ኤክስኤአይ (xAI) ግሮክ (Grok)፣ ቻትጂፒቲ (ChatGPT) እና ፐርፕሌክሲቲ (Perplexity) ያሉ በሰው ሰራሽ የማሰብ ችሎታ (AI) የሚሰሩ የፍለጋ መሳሪያዎችን በብዛት предпочитаಶಸ್ በመሆናቸው የተከሰተውን ከፍተኛ ለውጥ ያሳያል። እነዚህ መድረኮች ቀጥተኛና ውይይት መሰል ምላሾችን በመስጠት ባህላዊ በሆኑ በመገናኛ አገናኞች ላይ የተመሰረቱ ውጤቶችን ይበልጣሉ።

@webthreeth

OpenAI Upgrades Operator with Advanced o3 Model

OpenAI is enhancing Operator, its autonomous AI agent capable of browsing the web and executing tasks within a cloud-hosted virtual machine, by replacing the GPT-4o-based model with a more advanced version based on the new o3 model. Known for superior performance in reasoning and math, o3 surpasses its predecessor across key benchmarks. While the Operator API will continue using GPT-4o, the upgraded "o3 Operator" has been fine-tuned with additional safety datasets to better handle decisions around confirmations and refusals.

OpenAI የተባለው ድርጅት ኦፕሬተር የተሰኘውን፣ በራሱ የሚንቀሳቀስና ድረ-ገጾችን መቃኘት እንዲሁም በክላውድ በተስተናገደ ቨርቹዋል ማሽን ውስጥ ተግባራትን ማከናወን የሚችለውን ሰው ሰራሽ አስተውሎት (AI) ወኪሉን አቅም እያሳደገ ነው። ይህንንም እያደረገ ያለው አሁን ያለውን በGPT-4o ላይ የተመሰረተ ሞዴል በአዲሱና የላቀ የማመዛዘንና የሒሳብ ችሎታ እንዳለው በሚነገርለት የo3 ሞዴል በመተካት ነው።አዲሱ o3 ሞዴል በቀዳሚው ሞዴል ላይ በተለያዩ ቁልፍ የመመዘኛ መስፈርቶች የተሻለ አፈጻጸም እንዳስመዘገበ ተገልጿል።

@webthreeth

A Five-Level Roadmap to Transformational AI

Bloomberg’s reporting on OpenAI’s vision outlines a clear, five-step progression for artificial intelligence: beginning with basic conversational chatbots (Level 1), advancing to human-level reasoners capable of complex problem solving (Level 2), then to autonomous agents that can execute tasks (Level 3). At Level 4, AI systems evolve into innovators that actively contribute to new inventions, and ultimately mature into full-scale organizational units (Level 5) able to perform the collective functions of an entire company.

የብሉምበርግ ዘገባ ስለ ኦፕንኤአይ (OpenAI) ራዕይ ሲገልጽ፣ ለአርቴፊሻል ኢንተለጀንስ (AI) ግልጽ የሆነ ባለ አምስት ደረጃ የእድገት ሂደት እንዳለው ይዘረዝራል፦ ይህም መሠረታዊ የውይይት ቻትቦቶች (ደረጃ 1) በመጀመር፣ ውስብስብ ችግሮችን መፍታት ወደሚችሉ የሰው ልጅ የማመዛዘን ደረጃ (ደረጃ 2) በማደግ፣ ከዚያም ተግባራትን በራሳቸው ማከናወን ወደሚችሉ ራሳቸውን የቻሉ ወኪሎች (ደረጃ 3) ይሸጋገራል። በደረጃ 4 ላይ፣ የ AI ስርዓቶች አዳዲስ ግኝቶችን በንቃት በማበርከት ወደ ፈጠራ ፈጣሪዎችነት ይለወጣሉ፣ እና በመጨረሻም የድርጅትን ሙሉ ተግባራት ማከናወን ወደሚችሉ ሙሉ ድርጅታዊ ክፍሎች (ደረጃ 5) ያድጋሉ።

@webthreeth

Mistral AI rolls out faster, more powerful Agents inside Le Chat

Mistral AI's update to Le Chat introduces "Agents," which are AI entities capable of utilizing all tools and connectors available in a standard chat, potentially making them the fastest agents in the industry due to their enhanced functionality and integration capabilities.

ሚስትራል ኤአይ ለ"ሌ ቻት" ያደረገው ዝመና "ኤጀንቶች" የሚባሉ አርቴፊሻል ኢንተለጀንስ አካላትን አስተዋውቋል፡፡ እነዚህ ኤጀንቶች በመደበኛ ቻት ውስጥ ያሉትን ሁሉንም መሳሪያዎች እና ማገናኛዎች የመጠቀም ችሎታ ያላቸው ሲሆን፣ በተሻሻሉ ተግባራቶቻቸው እና ውህደት ችሎታቸው ምክንያት በኢንዱስትሪው ውስጥ ፈጣን ወኪሎች ሊሆኑ ይችላሉ።

@webthreeth

Officially, Level 3 has entered for AI Evolution 🥋

Productivity Tip: Operator + Replit Agent Outperforms Replit Agent Alone

In a striking demonstration of autonomous development, Peter Gostev, Head of AI at Moonpig, handed full control of his Replit account to OpenAI’s Operator and let it work in tandem with the Replit Agent—without any manual intervention. The result was a fully built, tested, and deployed game, all achieved through AI-to-AI collaboration. What’s more compelling is that this setup performed significantly better than using the Replit Agent alone.

While the Replit Agent can build apps from prompts, it still depends on users for testing, iteration, and deployment decisions. In contrast, Operator orchestrated the full loop: prompt generation, testing feedback, iterative improvements, and final deployment. This layered agentic workflow showcases a leap in productivity and signals what’s possible when you give smart agents end-to-end autonomy.

@webthreeth