OpenAI o1 and o3 Models has launched it's API

OpenAI made its o1 and o3-mini models available through its API, expanding access to developers across all paid usage tiers.

The o1 model, known for its advanced reasoning capabilities, now supports features like streaming, function calling, structured outputs, reasoning effort control, Assistants API, Batch API, and vision processing.

Meanwhile, the o3-mini, a more efficient and cost-effective option, offers similar functionalities (excluding vision) and boasts improved performance over its predecessors, matching or surpassing the full o1 model in STEM-related tasks.

These updates, announced recently by OpenAI, provide developers with powerful tools to build applications requiring complex reasoning, real-time processing, and structured data handling, marking a significant step forward in AI development capabilities.

@webthreeth

Mistral AI unveiled MistralOCR 2503

Mistral AI, a Paris-based innovator in artificial intelligence, has unveiled Mistral OCR 2503, announced on March 6, 2025, as the world’s most advanced optical character recognition (OCR) model, making it a market leader in AI-driven document processing. This cutting-edge API surpasses competitors like Google, Microsoft, and OpenAI by transforming complex PDF documents into AI-ready Markdown files with unmatched precision, excelling in handling intricate layouts, mathematical expressions, and tables

According to Mistral, the model excels at creating bounding boxes around graphical elements and integrating them into the output, avoiding the common pitfalls of producing unformatted text dumps. Already integrated into their AI assistant Le Chat, Mistral OCR is poised to empower developers and businesses dealing with sophisticated document processing, marking a significant leap forward in the practical application of AI technology.

@webthreeth

Trump Establishes Strategic Bitcoin Reserve and Digital Asset Stockpile

On March 07, 2025, President Donald J. Trump signed an Executive Order creating a Strategic Bitcoin Reserve and a U.S. Digital Asset Stockpile, aiming to position the United States as a global leader in government digital asset strategy. The Strategic Bitcoin Reserve will treat bitcoin as a reserve asset, capitalizing it with bitcoin forfeited through criminal or civil proceedings held by the Department of Treasury, while other agencies explore transferring their bitcoin holdings to the reserve.

The U.S. will retain this bitcoin as a store of value without selling it, and the Secretaries of Treasury and Commerce are tasked with budget-neutral acquisition strategies. Additionally, the U.S. Digital Asset Stockpile will consist of other forfeited digital assets, managed responsibly by the Treasury, with no further acquisitions beyond forfeiture, ensuring a cohesive approach to federal cryptocurrency holdings.

@webthreeth

Performance Benchmarking of Qwen-32B: A Comparative Analysis Across AI Evaluation Metrics

Qwen, developed by Alibaba, represents a significant advancement in the field of artificial intelligence, particularly in natural language processing.

Launched as part of Alibaba's efforts to enhance AI capabilities, Qwen-32B is a powerful language model designed to compete with other leading models in various benchmarks.

The results indicate that Qwen-32B (represented in red) consistently performs strongly, achieving scores of 79.5, 72.6, 73.1, 83.9, and 66.4 respectively across these benchmarks. Notably, it outperforms many competitors, such as DeepSeek-R1-67B (blue) and Open-AI-o1-mini (gray), in several categories, particularly excelling in IFEval with a score of 83.9.

@webthreeth

Comet Browser by Perplexity: Launching the Future of AI-Powered Browsing

Comet Browser, developed by Perplexity, is an innovative AI-powered web browser designed to transform the way users interact with the internet. Announced in February 2025, Comet aims to integrate advanced "agentic search" capabilities, allowing the browser to autonomously perform tasks such as booking tickets, conducting deep research, and automating workflows, all while delivering real-time, context-aware results.

As it enters a crowded market dominated by giants like Google Chrome, Comet seeks to carve a niche by offering intelligent features like enhanced security, personalized customization, and seamless task execution, potentially redefining productivity and information access online. While still in its pre-launch phase with a waitlist open, its full capabilities and impact remain to be seen as anticipation builds for its release.

@webthreeth

Manus AI: Pioneering the Era of the First General AI Agent

Manus AI, a Chinese-based AI startup, is being hailed as the world’s first fully autonomous AI agent, designed to handle complex, real-world tasks independently—planning a trip to Japan, analyzing Tesla’s stock, screening resumes, or even building a custom website step-by-step. Unlike typical chatbots that generate responses, Manus executes tasks from start to finish, leveraging browsing capabilities, memory, and learning to adapt to user needs.

The GAIA Benchmark chart highlights Manus AI's impressive performance as the world’s first fully autonomous AI agent, launched on March 5-6, 2025. Achieving an 86.5% pass rate at Level 1 and 70.1% at Level 2, Manus AI significantly outpaces OpenAI Deep Research (74.3% and 69.1%) and the previous state-of-the-art (67.9% and 67.4%), showcasing its strength in simpler and moderately complex tasks.

@webthreeth

The First AI Specialized for Fiction Writing: Sudowrite's launches Muse.

Sudowrite's launches Muse, which is a groundbreaking advancement in AI for fiction writing, tailored specifically to meet the needs of authors. This specialized AI model has been meticulously trained on a curated dataset with full author consent, focusing on crafting high-quality, literary prose that excels in scene construction, character development, and avoiding clichés.

@webthreeth

Waiting on the trial, but China might have just done it again. From what I saw in reviews. This is impressive as hell.

I just got an access from a Twitter page of a Manus AI report. And 90 pages report on a subject. I heard it is twice as long as OpenAI deep research on the same topic. And was just better

Optimizing Your E-Ecommerce Experience with ChatGPT

The guide titled "Shopping with ChatGPT" leverages reasoning, search, deep research, and operator assistance to enhance the shopping experience. It is divided into two main categories: Recommendations and Promo Codes, each with Quick and Deep approaches.

For Recommendations, the Quick method uses models like o1 and o3-mini-high with prompts such as suggesting brands based on user preferences or finding a brand for a specific product. The Deep method employs the Deep Research model to research product characteristics and identify the best option.

For Promo Codes, the Quick approach uses the o3-mini-high + Search model to find codes for a specific brand, while the Deep approach uses the Operator model to find and validate promo codes for a desired product. This structured approach aims to streamline and optimize the shopping process using advanced AI capabilities.

@webthreeth

Gemini's Next Leap: Unveiling Gemini 2.0 Flash Thinking and Personalisation Experimental

Google is expected to unveil two new Gemini models: Gemini 2.0 Flash Thinking and Gemini 2.0 Personalisation Experimental

Gemini 2.0 Flash Thinking, evolving from its experimental roots, is anticipated to be a stable, production-ready version that blends the speed of the Flash architecture with enhanced reasoning capabilities, designed to break down complex prompts into transparent, step-by-step solutions—ideal for tasks requiring multi-step logic or multimodal inputs like text and images.

Meanwhile, Gemini 2.0 Personalisation Experimental is rumored to introduce a groundbreaking focus on user-tailored AI, potentially leveraging integration with Google services like Search, Maps, and YouTube to adapt responses to individual preferences or contexts, pushing the boundaries of agentic AI for more interactive and personalized experiences.

@webthreeth

Speculation of new model realesed of DeepSeek R2 on March 17

With whispers of a March 17 debut—accelerated from an original early May timeline—DeepSeek R2 promises to build on the success of its predecessor, R1, with enhanced capabilities that could redefine cost-effective AI innovation.

Analysts speculate R2 could also expand its multimodal capabilities, building on R1’s foundation to handle text, code, and possibly even images or other data types.

This versatility, paired with pricing rumored to be 20 to 40 times cheaper than OpenAI’s offerings, positions DeepSeek to capture a massive share of the developer and enterprise markets.

The R1 launch triggered a $1 trillion sell-off in global equities, as investors questioned the value of hardware-heavy AI strategies. R2’s arrival could amplify this disruption, especially if it delivers on its promise of enhanced coding and multilingual reasoning.

@webthreeth

The Breakneck Evolution of AI: How Reasoning Models Are Shattering Historical Performance Trends

AI's recent breakthroughs in mathematical reasoning mark a sharp deviation from historical performance trends, with reasoning models vastly outperforming traditional frontier models in a short time. This suggests that AI is moving beyond brute-force computation toward more advanced reasoning, driven by improvements in architecture and training.

As AI continues to exceed expectations, tasks once thought resistant to automation—such as complex problem-solving and scientific research—may soon be within its reach, potentially disrupting expertise-driven industries. Given this rapid acceleration, linear extrapolations of AI progress are likely outdated, and we should anticipate sudden breakthroughs and emergent capabilities that redefine AI’s potential.

@webthreeth

Meta’s AI Future: Partnering with TSMC for Custom Chip Innovation

Meta is making significant strides in reducing its reliance on external chip providers by developing its own AI training chip in collaboration with Taiwan Semiconductor Manufacturing Company (TSMC).

Currently in the testing phase as of March 11, 2025, this custom-designed chip aims to serve as a dedicated accelerator for AI tasks, offering greater efficiency compared to traditional GPUs. If successful, Meta plans to expand its deployment, partnering with TSMC, the world’s leading contract chip manufacturer, to produce this innovative hardware.

This move not only highlights Meta’s ambition to optimize its AI infrastructure and cut costs but also underscores TSMC’s pivotal role in powering the next generation of AI technologies for major tech firms.

@webthreeth

Crypto Market Volatility in March 2025: A Shift in Sentiment

As of March 2025, the crypto market, particularly Bitcoin, is experiencing significant volatility, with the price dropping 5.3% to $78,354.89. This decline reflects a broader shift in market sentiment from extreme greed to fear, driven by a sudden change in risk appetite, as institutional capital exited tech stocks and built the largest Ethereum short position in history on February 9, 2025.

The announcement of a U.S. Strategic Crypto Reserve in early March initially sent crypto prices soaring, but it quickly turned into a "sell the news" event, exacerbating the downturn and aligning with the thread's analysis of deep red market days and massive outflows, such as the record $2.6 billion crypto fund outflow in late February 2025.

@webthreeth

Tencent's Hunyuan-TurboS: Revolutionizing AI with Hybrid Efficiency and Superior Performance

Hunyuan-TurboS, developed by Tencent, is an innovative ultra-large Hybrid-Transformer-Mamba MoE model that combines Mamba's efficient long-sequence processing with Transformer's strong contextual understanding to overcome the O(N²) complexity and KV-Cache issues of traditional Transformer models.

This hybrid architecture enables Hunyuan-TurboS to outperform competitors like GPT-4o-0806 and DeepSeek-V3 in math, reasoning, and alignment benchmarks, while offering 1/7 lower inference costs than its predecessor, thanks to advancements such as slow-thinking integration, refined instruction tuning, and an upgraded reward system that enhances accuracy in STEM, QA, and creativity. Building on Tencent's earlier integration of AI models like Deepseek-R1 Ki into WeChat/Weixin in February 2025, Hunyuan-TurboS marks a significant step in the global AI race toward more intuitive and efficient AI solutions.

@webthreeth

AI's Uneven Impact on Global Employment

The chart illustrates the varying impact of AI on jobs across different economic regions, highlighting a stark contrast in exposure and complementarity. In advanced economies, a significant share of employment—around 60%—is highly exposed to AI, with roughly half of that share showing high complementarity, meaning AI can enhance these roles rather than replace them.

Globally, about 40% of jobs are highly exposed to AI, but this figure drops in emerging markets and low-income countries, where exposure is closer to 30% and 20%, respectively. In these regions, the share of jobs with high complementarity is also lower, indicating that AI is more likely to automate tasks without adding value to human roles.

This data underscores how AI's influence is predominantly felt in wealthier nations, potentially widening the technological and economic gap with emerging and low-income countries.

@webthreeth

Gemma 3 27B IT: Pioneering High Performance with Minimal Size

Google unveiled Gemma 3, the latest lightweight, open-source AI model, designed for devices from smartphones to workstations. Available in four sizes—1B, 4B, 12B, and 27B parameters—it enables efficient, high-performing apps on a single GPU or TPU.

Gemma 3 27B IT stands out with an ELO score of 1,340, surpassing Meta Llama 3.1 405B (1,280 ELO) despite its 405 billion parameters, highlighting efficiency over scale. Compared to DeepSeek R1 (1,380 ELO, likely larger) and DeepSeek V3 (1,320 ELO), Gemma 3 excels with 27 billion parameters, outpacing Mistral Large 2407 (1,250 ELO, 200B) and Llama 3.3 70B (1,260 ELO, 70B). Models like Qwen 2.5 72B and Meta Llama 3.1 70B also lag at 1,260 ELO, underscoring Gemma 3’s optimization.

@webthreeth

This is very huge for AI development. Parameters refer to variables that a model learns during its training to make predictions. Gemma has achieved a similar level of performance to DeepSeek R1 (which is currently the best open-source model) while training with 30X fewer parameters. This further pushes the frontier of AI development in the future.

@webthreeth

The Disconnect Between Benchmark Scores and User Preferences

The graph highlights a significant disparity between traditional benchmark metrics and actual user preferences in AI model performance. It compares GPT-4.5 and DeepSeek V3, showing a modest 3.5% difference in Elo scores (1411 for GPT-4.5 vs. 1363 for DeepSeek V3), suggesting a close competition. However, user preference data reveals a stark contrast, with 74% of users favoring GPT-4.5 over DeepSeek V3's 26%.

This indicates that small differences in headline benchmarks can translate into a strong majority preference among users, a pattern observed across other similar model pairings. Sourced from LMArena, the data underscores the limitations of relying solely on Elo scores and emphasizes the importance of user experience in evaluating AI model effectiveness.

@webthreeth