Beginner Guide to LLMs and GenAI
These are just some thoughts and advice I've picked up since I started using Generative AI (GenAI) tools.
Honestly, it feels like both yesterday and a lifetime ago that everyone started talking about this 'magic tech.' I remember the thrill of waiting to get into the ChatGPT beta, using Poe because Claude wasn't available in Europe, and seeing the explosion of GenAI tools—everything from generating images with Dall-E, to creating voices with ElevenLabs, songs with Suno, and video with Runway to the 'recent' jump in coding abilities and the release of editors and advance coding features with Copilot, Cursor, Replit, Bolt... and the open-source ecosystem with Hugging Face, LLaMA and DeepSeek.
To witness the reception in the public sphere has been equally fascinating: from the scandals and internal power wars at OpenAI, to the heated discussions around regulations, petitions, and demands for trillion-dollar projects. We've watched these technologies emerge from tech niches on Reddit and Twitter to surprising moments like my middle-aged mother casually chatting with ChatGPT Voice. The debates have ranged from viewing AI as an existential risk to humanity's ultimate savior, with every shade of gray in between.
A rapidly evolving story, one that is challenging our relationship and understanding of what technology can do and forced us to grapple with fundamental questions about intelligence, creativity, morals & ethics, and human potential.
What follows is a mix of advice and reflections. It's something I wrote for myself. This isn't meant to be a comprehensive guide or a technical manual, rather:
a collection of insights gained from countless hours of exploration, experimentation, amazement and occasional frustration with these fascinating tools.
Key Principles for Working with AI
Adapt to the Way Models Reason
Instead of just learning to navigate interfaces, focus on understanding how each model reasons. Different models excel in different areas (e.g., text, image generation). Choose the right one for your specific purpose.
Experiment with General Models
Try various general-purpose AI models to find one whose style of reasoning resonates with you. Once you find one, stick with it for a while to learn how to interact effectively.
Communicate Clearly but Casually
While English tends to be the best language for AI interaction, if you aren't a native, don't worry! Correct grammar, punctuation, or perfectly structured sentences, don't matter that much. Focus on clearly transmitting your intent and logic. Abbreviations and shorthand are fine as long as they convey your message.
Learn Through Projects
Avoid a "no-code" mindset; instead, use AI to help you learn coding concepts like logic, data flow, and interfaces. Shift your focus from memorizing specific functions (you can always Google those) to understanding how systems interconnect. AI is an excellent tool for this.
Explore Freely and Refine Gradually
In the beginning, don't worry about token usage or efficiency. Ask lots of questions and seek clarification on anything you don't understand. Over time, refine your queries and interactions.
Free Trials
Take advantage of free trials to experiment with various models.
Make It Personal
Start by creating something meaningful for yourself rather than jumping straight into building money-making apps. Projects that resonate with your personal needs are more engaging and sustainable.
Embrace the Iterative Process
If you get stuck, don't hesitate to rephrase your question or ask for clarification. Analyze where the AI's response went wrong and explicitly point it out in your follow-ups to improve results.
A Balanced Approach to GenAI
Learning to use these tools properly can be challenging. Models often hallucinate, which means double-checking generated outputs is essential. It's important to compare your understanding of a topic with the AI's output, engaging in back-and-forth conversations to grasp how to get the best from the technology.
This contrasts with a common misconception about GenAI, particularly among students: viewing it as a shortcut or even prohibiting its use. In reality, if used without care and the right competencies—what might be called the "generate, copy, and paste" approach—AI can become a shortcut to mediocrity, especially in academic or professional contexts.
If we apply Feenberg's philosophy, the expertise of the user significantly impacts the tool's effects on autonomy. An experienced user, employing AI critically, can enhance their abilities and efficiency. Conversely, an inexperienced user who blindly trusts AI risks losing their autonomy, delegating all critical thinking to the model.
This is particularly crucial when learning new concepts. AI should always be used critically and skeptically, verifying its outputs against your own knowledge and experience. By adopting this mindset, AI becomes not just a tool but a partner in deeper learning and skill development.
Don't just use AI as a shortcut, use it as a partner in learning and thinking.
Final Considerations
Interacting with AI will change the way you express yourself. Over time, you'll notice a shift in how you formulate questions, explain ideas, and think.
Is this a pro or a con?
It depends. On one hand, it can sharpen your clarity and logical reasoning; on the other, it might influence you to oversimplify or conform to patterns that AIs "understand" better.
AI seems more capable in English, though increasingly powerful models are emerging in Chinese. This raises an interesting question about the relationship between natural language and AI capabilities:
Do some languages work better with AI?
Currently, English leads in model performance, but Chinese models are rapidly advancing. The real question might be deeper:
Is natural language the optimal interface for AI?
Natural language, while intuitive, carries inherent ambiguities. Perhaps the future lies in developing specialized communication systems that better align with how LLMs process information—bridging the gap between human expression and machine comprehension.