BrowseComp: a benchmark created by OpenAI for browsing agents

BrowseComp is a meticulously designed benchmark that tests an AI’s ability to locate hard-to-find information through strategic web browsing and reasoning. It centers on short, fact-based questions with single, indisputable answers that are crafted to be challenging both for AI systems and human solvers.

By using "inverted" questions—where the correct answer is difficult to uncover yet straightforward to verify—BrowseComp forces models to pursue creative and persistent search strategies, rather than relying on brute-force methods. Its development involved rigorous checks, including ensuring that prevalent models like GPT‑4o struggled without advanced reasoning and browsing techniques.

Notably, while standard models achieved near-zero accuracy, an agent specifically trained for deep research managed to solve over half of the problems, highlighting the importance of combining robust reasoning with effective tool.

@webthreeth

Unveiling Ethiopia's Past: ሊነጋ ነው Captivates Over 200K Viewers in Just 3 Days

The AI-generated short film "ሊነጋ ነው", released by EHUD AI Studio on April 7, 2025, has achieved remarkable success, garnering over 200,000 views and 2.3 million impressions within just three days on the EHUD AI Studio YouTube channel (https://www.youtube.com/@ehudai). This political psychological thriller, which explores the legacies of Ethiopia’s past leaders, saw a view count of 203.3K, three days after its re-realese on April 7 2025.

Congrats 👏
@webthreeth

Supercharge Your Career: Unleashing NotebookLM for Interview Prep and Beyond

NotebookLM, a powerful AI tool from Google, powered by Gemini 2.5 Pro, is revolutionizing career development by transforming how job seekers prepare for interviews and boost their professional profiles.

By uploading your resume and target job descriptions, NotebookLM’s innovative audio overview feature generates a podcast-style discussion where virtual "hosts" analyze your skills, highlight how they align with the role, and offer tailored advice for interview questions—streamlining your preparation process.

Beyond interviews, this tool leverages its document summarization capabilities to help you refine your resume for applicant tracking systems (ATS), which 83% of employers now use in 2025, ensuring your application stands out. Whether you’re crafting study guides from multiple sources or seeking actionable insights to elevate your career narrative, NotebookLM is your personal hiring coach.

@webthreeth

Should I upload tips like this on how you can use AI not only for your daily work, but also strategic growth of both yourself, but also your business?

I genuinely advise this tool for anyone who is working in an knowledge-intensive sector. It is probably one of the best tools shipped after GPT-O Series models. Very impressive one by Google.

Google aren't marketing their products like ChatGPT. Trust me, Gemini is the leader in the market for LLMs now, purely based on performance (At least to my field of work).

Perplexity’s Telegram Bot: A New Way to Search

Perplexity, an AI-powered search engine, has launched a bot on Telegram, bringing its conversational search capabilities to the popular messaging platform. This bot, accessible by searching for " @askplexbot ," allows users to ask questions directly within Telegram, receiving quick, accurate answers backed by Perplexity’s advanced language models.

Whether used in private chats or group conversations, the bot offers a seamless way to explore topics, research on the go, or spark discussions with friends. This move makes Perplexity’s knowledge-discovery tools more accessible, blending the convenience of Telegram with the power of AI-driven insights.

@webthreeth

S&P 500 Steals Bitcoin’s Thunder: Volatility Surges Amid Tariff Turmoil

The S&P 500’s 10-day historical volatility spiking to 76.8%, outstripping Bitcoin’s 72.9%, marks a rare moment where traditional markets have eclipsed the crypto world’s notorious price swings, as reported by Bloomberg. ETF analysts emphasize that typical volatility for the S&P 500 lingers between 10–15%, making this jump a significant deviation from the norm.

The catalyst appears to be fresh trade tariffs imposed by President Trump’s administration, particularly a hefty 145% duty on Chinese imports, which has sent shockwaves through global markets. This policy, coupled with China’s retaliatory 125% tariffs on U.S. goods, has fueled uncertainty, driving wild fluctuations in stock indices.

@webthreeth

Meta's Llama 4 Maverick Misstep: Ranking Plunge Sparks AI Benchmark Controversy

Meta faced backlash after it was revealed they submitted an experimental version of their Llama 4 Maverick model, optimized for conversational flair, to the LM Arena benchmark, securing a high ranking of #2.

This version, dubbed "Llama-4-Maverick-03-26-Experimental," differed significantly from the publicly available model, which critics argued misled developers about its real-world performance. After scrutiny, LM Arena re-evaluated the unmodified Maverick model, and its ranking plummeted to 32nd, exposing discrepancies in Meta's approach.

The incident sparked debates about transparency in AI benchmarking, with some accusing Meta of gaming the system to inflate their model's standing, though Meta maintained they were merely experimenting with custom variants.

@webthreeth

OpenAI's A-SWE: The Future of Autonomous Software Engineering In Progress

On March 5, 2025, Sarah Friar, OpenAI's CFO, announced the development of 'A-SWE,' an autonomous software engineer agent capable of independently building apps, performing quality assurance, bug testing, and documentation—tasks often disliked by human engineers—potentially disrupting existing collaborative AI tools like Devin, as it functions as a standalone engineer rather than an assistant like GitHub's Copilot; this announcement, made public on April 12, 2025, aligns with OpenAI's broader AI research goals that include deep research and operator agents, though no AGI timeline was specified.

@webthreeth

Canva Unveils AI-Powered Code Generation Feature with Canva Code

Canva recently launched Canva Code, an innovative AI-powered feature that enables users to create interactive digital elements like pricing calculators, countdown timers, and educational games without coding skills, utilizing text prompts and a conversational AI interface with voice command support to generate code instantly, seamlessly integrating these creations into Canva designs such as websites, presentations, and social posts, while offering a preview panel for quick refinements, secure AI settings with customizable safeguards for organizational use, and versatile applications ranging from personal projects like dynamic itinerary builders to business tools like interactive product guides, all accessible for free to Canva Free, Pro, and Teams users with a gradual rollout planned over the coming months.

@webthreeth

From 16 Trillion to Zero. WT*😭😭

Gemini 2.5 Pro Leads in Cost-Effective Multilingual Code Completion

This benchmark highlights Gemini 2.5 Pro as the clear leader in performance-to-cost efficiency for multilingual code editing tasks. With over 70% of tasks completed correctly, it not only outperforms all other models in terms of accuracy but does so while maintaining one of the lowest total costs.

This balance between high performance and minimal cost is visually emphasized by its tall bar and small purple cost marker, setting it apart from the rest. Most other models in the benchmark cluster around the 55–65% accuracy range, yet none come close to matching Gemini’s combined accuracy and cost-effectiveness.

@webthreeth

Should I do a podcast for the last week? Please like if you like the podcasts done?

Ethiopian AI-Powered Resume Platform PikCV Launched.

PikCV is an Ethiopian AI-centered resume optimization platform designed to give job seekers a competitive edge in today’s digital hiring landscape. Founded by Ethiopian entrepreneur Bemnet Girma, the platform uses advanced artificial intelligence to analyze, structure, and optimize resumes specifically for Applicant Tracking Systems (ATS), which 83% of the hiring firms use in preliminary CV checks.

PikCV’s AI engine customizes each resume based on the job description, intelligently inserting relevant keywords, formatting content for readability, and tailoring language to improve relevance and ranking. This automation eliminates guesswork and significantly boosts the chances of landing interviews. With a focus on accessibility and precision, PikCV empowers users to create professional, targeted resumes in minutes—ensuring that talent isn’t lost in the algorithm.

🔗 pikcv.com

@webthreeth

OpenAI's Revolutionary Reasoning Models Set to Transform Scientific Discovery (Industry Rumors)

According to rumors, OpenAI's upcoming reasoning models are poised to revolutionize scientific discovery by independently generating novel ideas across disciplines—a capability once exclusive to humans—with early tests at Argonne National Laboratory slashing experiment design time from days to hours, and future integration with AI agents controlling simulators or robots promising to accelerate hypothesis testing, though their $20,000 monthly price tag raises accessibility concerns for the technology that targets Fortune 500 companies in fields like material and drug discovery.

@webthreeth

OpenAI Unveils GPT-4.1 Family: A Leap in Coding and Long-Context AI for Developers

Sam Altman has announced the release of the GPT-4.1 family, including GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, exclusively through OpenAI's API channels and not through OpenAI ChatBot.

These models boast significant improvements in coding, instruction following, and long-context comprehension, supporting up to 1 million tokens—equivalent to roughly 750,000 words, far surpassing GPT-4's 32,000-token limit. With a focus on real-world utility rather than just benchmark scores, OpenAI aims to empower developers, who have already expressed satisfaction with the models' performance on private tests.

However, the API-only access has sparked curiosity among users, with some questioning the absence of in-app availability, potentially reflecting OpenAI's strategic resource management following past demand challenges with models like ChatGPT.

@webthreeth

Here is this week episode about the AI and Cryptocurrency market over the past week.
Episode #2

Gemini 2.5 Pro’s (The first PhD-Level Artificial Intelligence?): AIME 2024 and GPQA Diamond Evidence

Gemini 2.5 Pro’s exceptional performance on the AIME 2024 (92.0%) and GPQA Diamond (84.0%) benchmarks provides compelling evidence of PhD-level intelligence in mathematics and scientific reasoning. The AIME 2024, a rigorous test for top high school students, demands advanced mathematical problem-solving, where Gemini 2.5 Pro outperformed competitors like o3-mini (87.3%) and Grok 3 (83.9%), showcasing expertise akin to graduate-level scholars.

Similarly, the GPQA Diamond, a graduate-level benchmark with 198 challenging questions in biology, physics, and chemistry, tests deep domain knowledge. Gemini 2.5 Pro’s 84.0% score surpasses human PhD experts’ average of 65% (74% when excluding clear mistakes) and leads AI models like Grok 3 (80.2%) and o3-mini (79.7%). While these results highlight its ability to handle complex, structured tasks, its capacity for creative, original research.

@webthreeth

Accelerating the AGI Race: How Gemini's AI's Self-Designed Reinforcement Learning Redefines the Path to General Intelligence

David Silver of Google DeepMind unveiled a groundbreaking AI system that uses reinforcement learning (RL) to autonomously develop its own RL algorithms—surpassing those crafted by human experts over decades. This meta-learning approach enables the AI to innovate decision-making and reward optimization strategies entirely through trial and error.

Building on DeepMind’s achievements with systems like AlphaZero, which mastered complex games via self-play, this advancement represents a major step toward AI systems that can independently create and refine algorithms. It opens the door to a future where AI research accelerates beyond traditional human-led design, reducing dependency on manually curated data.

Link - https://youtu.be/zzXyPGEtseI?si=trKowe4Ycbs2bOZY

@webthreeth

This basically means that without additional data AI can learn like a human and excel at a stuff. Deepmind is cooking like hell