Search for a command to run...
Google's first true leap ahead of OpenAI
Talks about artificial intelligence, building SaaS solutions, product management, personal finance, freelancing, business, system design, programming and tech career tips
Decoding the recent developments in the world of Generative Artificial Intelligence like ChatGPT and Midjourney, their real-world use cases, and limitations.
Accessing over 100 LLMs using a single LiteLLM SDK
Introduction 
Making llms argue for answers to problems that confuse them
As someone who practically lives in their browser, I’ve spent years chasing the perfect browsing experience. I wanted a tool that’s sleek, powerful, and respects my time and privacy. After falling in love with Arc for its innovative features, I thoug...
🛣️ Introduction Since you started coding, you’ve typed letters, numbers, even emojis—but how does your computer get it? The answer’s character encoding, the magic that turns “A,” “世,” or “😊” into binary. This guide takes you from ASCII’s baby steps...
The modern technique of image interaction with AI
In a significant development for the AI industry, Google's latest language model, currently known simply as Gemini Exp-1114, has surpassed OpenAI's leading model in the prestigious LMArena benchmarks. This breakthrough marks a potential shift in the competitive landscape of large language models (LLMs).
The most striking aspect of Gemini Exp-1114's emergence is its comprehensive dominance across multiple domains. According to the LMArena benchmarks, the model has achieved remarkable improvements:
🏆 Category Improvements
🧮 Mathematics: #3 ➡️ #1
🎯 Hard Prompts: #4 ➡️ #1
✍️ Creative Writing: #2 ➡️ #1
👁️ Vision Tasks: #2 ➡️ #1
💻 Coding: #5 ➡️ #3
📏 Model operates with a 32K context window, notably smaller than the 2 million token context length of Gemini Pro-002
⏱️ Current testing indicates slower inference times compared to its competitors
🧪 Still in experimental phase, as evidenced by its temporary designation
For those interested in experiencing this breakthrough firsthand, Google has made Gemini Exp-1114 available for free testing through their AI Studio platform.
I put both Gemini Exp-1114 and Gemini Pro-002 through a series of real-world tasks to understand their capabilities and differences.
🔮 Gemini Exp-1114
⚡ Gemini Pro-002
Enquired the model about the following image without being specific in the prompt:
🔮 Gemini Exp-1114
⚡ Gemini Pro-002 [IMAGE]
Both the models were able to identify the partially cut mango but the count was wrong
🔮 Gemini Exp-1114
⚡ Gemini Pro-002
The Gemini Exp-1114 Model does not give the right answer owing to “over-analysis” of a simple problem
🔮 Gemini Exp-1114
⚡ Gemini Pro-002
Both the models give the correct answer, the Gemini Exp-1114 additionally gives detailed steps to solve the problem
🔮 Gemini Exp-1114
Verification using the traditional wordcounter.net
⚡ Gemini Pro-002
Verification using the traditional wordcounter.net
Both the models generate relevant content, the Gemini Exp-1114 follows the content limit instructions more strictly and hence generates a more accurate output
🔮 Gemini Exp-1114
Rendered using svgviewer.dev
⚡ Gemini Pro-002
Rendered using svgviewer.dev
Since the prompt is minimal, the generated SVGs are minimal as well. Gemini Exp-1114 does a decent job at v0.0.1 prototype while the Gemini Pro-002 model completely messes it up
Google is yet to launch Gemini Exp-1114 in its full capacity, but preliminary benchmarks and hands-on testing show promising results. With its ability to go head-to-head with Claude and GPT-4, Gemini represents a significant step forward in Google's AI capabilities, potentially reshaping the competitive landscape of large language models.