• AI Emergence
  • Posts
  • ChatGPT Just Got a Research Mode- And It’s Insanely Powerful

ChatGPT Just Got a Research Mode- And It’s Insanely Powerful

Along with - Free courses on DeepSeek, AutoGen, Building Agents from Scratch and many more!

Hi there, 

Another crazy week - but this week I got an awesome research assistant! So, I can’t complain :) I tried a few use cases and it delivered professionally written, well-researched reports. Are they the best reports on the topic - Probably No. But they are better than what a Graduate would do in 4-5 days.

It feels like OpenAI is now delivering what Altman has been saying for the last 2 years. The pace of development will accelerate as AI matures. 

Let’s look at the developments this week to see if you feel the same by the end of it!

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection 

  • Question to ponder before we meet next!

Table of Contents

OpenAI is doubling down on AI-powered research. Deep Research is the latest addition to OpenAI’s evolving Agent ecosystem- following ChatGPT’s scheduling update and the launch of Operators. Now, with Deep Research, OpenAI is pushing AI further into complex, multi-step analysis.

Why It Matters

Traditional research is time-consuming- scanning sources, analyzing data, and synthesizing insights can take hours or even days. Deep Research streamlines this entire process, delivering well-structured, citation-backed reports in minutes. Instead of simply summarizing, this AI sifts through vast information, applies reasoning, and adapts as needed- whether it’s identifying trends, evaluating policies, or breaking down financial reports.

How It Works

Built on an optimized version of OpenAI’s upcoming o3 model, Deep Research can:

  • Search the web, parse PDFs, and analyze images.

  • Process user-uploaded files, including spreadsheets and reports.

  • Perform data analysis, generate visualizations, and apply reasoning to findings.

  • Work asynchronously, allowing users to submit queries and return later for a detailed output.

What Sets It Apart?

  • Handles Complex Queries - Moves beyond summarization to multi-step reasoning.

  • Citation-Backed Insights - Every claim is linked to a source for better reliability.

  • Automated Research Assistance - Suitable for finance, science, policy, and market analysis.

  • Asynchronous Execution - Users can initiate a query, step away, and return to a fully structured report.

How Does It Perform?

Deep Research has already set new benchmarks in AI-driven problem-solving:

  • Humanity’s Last Exam – Scored 26.6% on an expert-level test across 100+ subjects, with notable gains in chemistry, social sciences, and math.

  • GAIA Benchmark – Outperformed previous models in real-world problem-solving.

Limitations & Challenges

While powerful, Deep Research isn’t flawless. It occasionally:

  • Generates hallucinated facts, requiring human oversight.

  • Struggles with confidence calibration, sometimes overstating conclusions.

  • Mixes up authoritative sources, making cross-checking necessary.

OpenAI is actively improving these areas with user feedback.

Deep Research hints at a future where AI handles much of the heavy lifting in knowledge work. But can we trust AI-driven analysis? Would you rely on it for critical research? (source)

India’s AI Play: From TCS to Ola, The Big Bet on LLMs

India’s AI ambitions are going full throttle, with startups and industry leaders rallying to build homegrown AI models inspired by China’s DeepSeek. The goal? Not just LLMs- but eventually AGI.

But hardware, data centers, and top-tier talent remain major hurdles. While DeepSeek built R1 for under $6M, India will need sustained government investment to create truly sovereign AI.

The Players & Big Moves

  • Fractal Analytics - Proposed a $600-$800M AI fund (backed by TCS, Infosys, and the government) to build frontier AI models. Already developed four small models, including "Ramanujan", a high-reasoning AI that outperforms OpenAI’s o1 in Olympiad-level math and chess.

  • Sarvam AI - Training models on Indian languages (2B parameters).

  • Krutrim AI (Ola’s AI arm) - ₹2,000 Cr ($240M) investment, with plans to scale to ₹10,000 Cr ($1.2B) next year.

  • Government AI push - IndiaAI Mission funding increased to ₹2,000 Cr ($240M).

  • Supercomputing Play - Krutrim AI deploying Nvidia GB200, set to become India’s largest supercomputer by year-end.

Altman: India Has the Potential to Lead AI

Sam Altman’s recent trip to India signals OpenAI’s growing interest in the country’s AI future. After previously doubting India’s ability to build foundational AI models, he now sees the potential for cost-efficient AI innovation.

Meeting with Union IT Minister Ashwini Vaishnaw, Altman acknowledged that smaller, reasoning-driven models are reshaping AI, giving India an edge—much like its success in low-cost space exploration.

Vaishnaw reinforced India’s AI ambitions, announcing plans to offer the world’s cheapest AI compute at under $1/hour. With 15,000 high-end GPUs, India is no longer just catching up—it’s pushing for AI self-reliance.(source)

Google is scaling up Gemini with major updates across its AI lineup. The latest rollout makes Gemini 2.0 Flash widely available, introduces a high-performance Pro Experimental model, and debuts Flash-Lite, its most cost-efficient option yet.

Gemini 2.0 is now available to more users with multiple updates:

  • 2.0 Flash: Now generally available via the Gemini API in Google AI Studio and Vertex AI. Optimized for high-volume tasks, it features a 1 million-token context window, with image generation and text-to-speech coming soon.

  • 2.0 Pro Experimental: A new model with the best coding performance and reasoning abilities. It supports a 2 million-token context window and tool-calling capabilities, available in Google AI Studio, Vertex AI, and Gemini Advanced.

  • 2.0 Flash-Lite: A cost-efficient model with improved quality over 1.5 Flash, maintaining the same speed and pricing. Available in public preview on Google AI Studio and Vertex AI.

  • 2.0 Flash Thinking Experimental: Now accessible in the Gemini app for enhanced problem-solving.

All models support multimodal input, with more features rolling out soon. Safety improvements include reinforcement learning and automated security assessments. (source)

The U.S. National Laboratories are integrating OpenAI’s o-series models into their research ecosystem, aiming to enhance scientific discovery across multiple fields. This partnership brings AI-powered reasoning models to Venado, an NVIDIA supercomputer at Los Alamos National Laboratory, which will also be accessible to researchers at Lawrence Livermore and Sandia National Labs.

Enhancing Scientific Research with AI

The collaboration is expected to contribute to key areas of national interest, including:

  • Materials Science & Astrophysics - Accelerating breakthroughs in technology and space research.

  • Healthcare & Biomedical Research - Identifying new methods for disease prevention and treatment.

  • National Security & Cyber Defense - Strengthening cybersecurity and protecting critical infrastructure.

  • Energy & Sustainability - Advancing solutions for renewable energy and climate science.

Given the National Labs' role in nuclear security and risk mitigation, AI-driven analysis will support early threat detection, cybersecurity improvements, and strategic defense initiatives. AI safety experts with security clearances will oversee its implementation in sensitive research areas.

By integrating AI into scientific and national security research, the U.S. aims to enhance innovation, efficiency, and strategic decision-making. This partnership reflects a broader effort to responsibly scale AI’s role in high-stakes domains, with Microsoft also involved in advancing AI-powered research. (source)

OpenAI Expands Its AI Models with the Release of o3-mini

OpenAI has introduced o3-mini, the latest model in its reasoning series, designed to provide improved efficiency and performance in technical fields such as math, coding, and science. It builds on the o1-mini, offering similar capabilities with faster response times and lower costs.

Key Features

  • Function Calling & Structured Outputs - Enhances usability for production applications.

  • Adjustable Reasoning Modes - Low, Medium, and High settings allow for more flexible computational depth.

  • Improved Search Integration - Retrieves real-time data with relevant sources.

Who Benefits and How?

  • Developers & Engineers - More structured outputs and function-calling capabilities improve AI-driven workflows.

  • Students & Researchers - Enhanced reasoning modes support complex problem-solving in STEM fields.

  • General Users - o3-mini introduces reasoning capabilities to ChatGPT free users for the first time.

  • Paid Users (Plus & Team) - Higher message limits (50 → 150/day) and access to advanced reasoning modes.

Performance Improvements

Compared to o1-mini, o3-mini delivers:

  • 56% User Preference - Rated higher for clarity and accuracy.

  • 39% Fewer Major Errors - More reliable reasoning in technical tasks.

  • 24% Faster Responses - Reduced latency, with an average response time of 7.7s vs. 10.16s on o1-mini.

With broader accessibility and improved efficiency, o3-mini is OpenAI’s latest step toward making cost-effective reasoning models more widely available. (source)

SoftBank is making serious AI moves. Reports are swirling that the Japanese tech giant is considering a massive $25 billion investment in OpenAI- potentially making it OpenAI’s biggest backer yet. But that’s not all.

SoftBank and OpenAI have launched SB OpenAI Japan, a joint venture aimed at bringing AI-powered enterprise solutions to Japan’s biggest companies. Their first product? Cristal Intelligence, an AI system designed to securely integrate enterprise data with OpenAI’s advanced models.

Who’s Involved & What’s the Play?

  • OpenAI - Provides cutting-edge AI tech and engineering support.

  • SoftBank - Brings its Japan-based network and business expertise.

  • Arm (SoftBank-owned) - Powers the AI with scalable, high-performance compute.

First Stop? SoftBank’s Own Ecosystem

Cristal Intelligence will first roll out in SoftBank’s companies, including:

  • Arm - To drive AI innovation and productivity.

  • PayPay (SoftBank’s digital payments service) - AI-powered automation.

  • SoftBank Corp. - Plans to automate 100M+ workflows and fuel new business opportunities, with $3 billion per year dedicated to AI integration.

SoftBank’s Vision: Global AI Domination

Masayoshi Son, SoftBank’s CEO, isn’t holding back:
"This initiative will transform how companies work in Japan and worldwide. We are fully committed to the AI revolution."

And SoftBank is already expanding its AI footprint beyond Japan. Alongside OpenAI, Oracle, Microsoft, NVIDIA, and MGX, it’s involved in the $100B ‘Stargate’ project, a Trump-backed initiative to build AI supercomputing data centers in Texas- possibly scaling up to $500B. (source)

Mistral just dropped a game-changer- Mistral Small 3, a blazing-fast 24B-parameter model that rivals much larger AI systems like Llama 3.3 70B and Qwen 32B, while running 3x faster on the same hardware. If you need speed, precision, and open-source flexibility, this might be your new go-to model.

Built for 80% of generative AI tasks, Mistral Small 3 is optimized for low-latency, instruction-following, and conversational AI. Unlike proprietary models like GPT-4o-mini, this one is fully open-source under Apache 2.0, making it an excellent replacement for those wanting more transparency and control.

Why It Matters?

  • On Par with Llama 3.3 70B - But way faster, with 150 tokens/sec latency.

  • Fine-Tuning Ready - Turn it into a subject-matter expert for law, medicine, or finance.

  • Local Deployment - Run it privately on a single RTX 4090 or even a MacBook (32GB RAM).

  • Zero RL or Synthetic Data - A clean foundation for custom AI builds.

Where It Shines

  • Instant AI Assistants - Real-time responses for chatbots & virtual agents.

  • Lightning-Fast Function Calling - Ideal for automated workflows & API tasks.

  • Industry-specific AI - Already being tested for fraud detection (finance), triaging (healthcare), and robotics control (manufacturing).

Mistral Small 3 is live now on:

  • Hugging Face, Ollama, Kaggle, Together AI, Fireworks AI

  • Coming soon: NVIDIA NIM, Amazon SageMaker, Groq, Databricks & Snowflake.

Mistral is doubling down on open-source AI, and Small 3 is proof. Faster, leaner, and ready for real-world applications, it’s set to challenge both proprietary and larger open models. (source)

Europe is making a bold move in AI- and it’s about transparency, openness, and digital sovereignty. Meet OpenEuroLLM, a massive collaboration between 20 top European research institutions, AI companies, and EuroHPC centers to develop next-gen multilingual open-source language models.

OpenEuroLLM: A Homegrown AI Push for Europe

This European-led initiative, coordinated by Charles University (Czechia) and co-led by AMD Silo AI (Finland), is building a family of powerful AI foundation models tailored for commercial, industrial, and public sector applications.

Unlike proprietary AI from the U.S. or China, OpenEuroLLM is fully open-source, regulation-compliant, and designed to boost Europe’s global AI competitiveness while ensuring linguistic and cultural diversity.

Why It Matters?

  • Transparent & Open - Full access to models, software, and data for businesses & public organizations.

  • European Digital Sovereignty - Less reliance on external AI tech, more homegrown innovation.

  • Built for Multilingual Use - Supports diverse European languages & cultures.

  • Aligned with EU Regulations - Ensures compliance with Europe's strict AI laws.

With funding from the European Commission’s Digital Europe Programme, OpenEuroLLM is leveraging top-tier European AI research and supercomputing power from EuroHPC centers like Barcelona Supercomputing Center, Cineca (Italy), and CSC (Finland).

This is Europe’s biggest push yet to develop transparent, high-performance AI that aligns with European values. With work starting February 1, 2025, OpenEuroLLM aims to democratize AI access, empower businesses, and set new global standards for open-source AI development. (source)

It is an AI-powered platform that enhances language learning through immersive, context-based content.

How to Access:

  • Sign Up - Create an account and select the language you want to learn.

  • Set Your Level - Choose your proficiency level for personalized content.

  • Pick Your Interests - Select topics you enjoy to get relevant articles, videos, and songs.

  • Start Learning - Access curated reading, writing exercises, multimedia content, and more for an engaging, practical language-learning experience.

  • Andrej Karpathy just released a comprehensive deep dive into Large Language Models (LLMs), breaking down everything from pretraining data and neural network internals to post-training refinements and reinforcement learning. Unlike his earlier talks, this video offers a structured, in-depth look at how models like ChatGPT work, their "psychology," and how to use them effectively. Whether you're an AI beginner or a seasoned pro, this is essential viewing for understanding the state of the art in AI- and an incredible learning resource for anyone looking to master LLMs.

  • Dylan Patel from SemiAnalysis dives deep into the evolving AI and semiconductor landscape on Lex Fridman’s podcast. He covers China’s DeepSeek, OpenAI’s latest moves, NVIDIA’s dominance, and Elon Musk’s xAI push. He also explores TSMC’s role and the Stargate project, offering sharp insights into the AI hardware race and its global impact. If you’re following AI’s future beyond just models, this is one to watch.

  • In case you missed it, we’ve just launched two free courses on DeepSeek last week - "Getting Started with DeepSeek" and “DeepSeek from Scratch”. These courses provide an in-depth look at the AI model that’s making waves in the industry. If you're curious about how DeepSeek was built and how it compares to GPT-4o, Claude Sonnet 3.5, and o1, this is your go-to resource.

  • In addition to these, here are some more recently launched free AI & ML courses worth exploring:

What has been the highlight of the week for you? Have you used Deep Research? What was your experience?

Do share it with me. Look forward to hearing from you.

Reply

or to participate.