A Rant Ever since ChatGPT burst onto the scene, I’ve been an unabashed OpenAI fanboy. I was constantly glued to the news, devouring their blog posts, and hanging on every word from Ilya Sutskever, Sam Altman, and all the other developers at OpenAI. Their insights were unparalleled at the start, and it seemed like everyone else was just playing catch-up. Fast forward to 2024, and the AI landscape has dramatically shifted. There’s now a lot of competition, and OpenAI finds itself in a precarious position. The infamous clash between Sam Altman and Ilya Sutskever resulted in Sutskever’s departure, along with a host of key scientists like Jan Leike. From an outsider’s perspective (and I stress, I’m no insider), things aren’t looking great for OpenAI on the developer front. Adding to this, Greg Brockman, the president of OpenAI, is now on extended leave. It’s just not looking good. Despite these setbacks, my passion for OpenAI persisted because their model was still so damn good. GPT-4 was really great, and the app provided such an amazing user experience. I found myself using it all day for general questions and studies. It truly represented a significant quality of life improvement. Even at the start of 2024, I was still heavily relying on GPT-4. But then, at the beginning of 2024, a company called Anthropic introduced Claude 3, along with accompanying models called Claude Haiku, Claude Sonnet, and Claude Opus. This caught my attention and piqued my interest in Anthropic and its CEO, Dario Amodei. Amodei, also an ex-OpenAI employee, is a reserved but funny character. Due to a conflict of interest, he left OpenAI and started Anthropic with his sister. When Claude 2 came out, it was impressive. But when Claude 3 was released, particularly the larger model Claude 3 Opus, it was mind-blowing. Upon its initial release, it beat state-of-the-art benchmarks in many areas. Notably, when I played around with it, its Korean capability was much better than GPT-4’s. Its coding ability was also really impressive. My Shift to Anthropic This superior performance led me to exclusively use Claude Opus for my toy project, even though it cost me significantly more than using GPT-4. I even subscribed to Claude at that point, meaning I had a dual subscription — one for ChatGPT and one for Claude. This double subscription, costing $40 a month, was a pain in the ass for my wallet, but looking back, it was totally worth it. It allowed me to play with different models every day, giving me a real sense of how to use and prompt each model differently. Even then, I still loved OpenAI. I loved their products, and their app was still great. I continued to use ChatGPT for my general studies and miscellaneous tasks, especially for studying. Last semester, I took a course on [Redacted], and I was in trouble for the finals because I hadn’t been paying attention to the course. The professor’s slide material was terrible, and I couldn’t get anything out of it. So, I uploaded all my slide content into ChatGPT and just peppered it with questions on and on until I finally understood all the concepts. I ended up getting a pretty decent score, which gave me confidence that with the right source material or questions, I could learn almost anything relatively fast without human guidance. I basically used ChatGPT for learning and miscellaneous tasks, while I used Claude for high-quality Korean report generation and programming. This dual approach served me really well. About three months ago, Claude 3.5 Sonnet was released, and it was in a league of its own. While Opus was quite expensive, Sonnet was relatively cheap and much better than both Opus and GPT-4. It excelled particularly in coding tasks. This led me to start using it more extensively, and I even canceled my OpenAI subscription in favor of Claude. The Straw That Broke the Camel’s Back Despite my growing preference for Claude, I still loved OpenAI because of ChatGPT’s user interface. But today, an incident occurred that was the final straw. I recently took TOEFL and am now studying for the GRE. The GRE is a much more difficult task compared to TOEFL, requiring serious study, particularly in vocabulary. To make vocabulary learning more accessible and enjoyable, I tried to modify the custom instructions of ChatGPT to output the most verbose and linguistically rich content possible. I had successfully done this with GPT-4 about a year ago when the custom instructions feature was first introduced. Back then, I spent nights tweaking prompts, and the results were fantastic — GPT-4 produced incredibly rich and sophisticated paragraphs on any topic, even the silliest ones. Today, I attempted to replicate this with the GPT-4o model, which OpenAI claims is more capable. However, if you’ve played around with GPT-4o, you can tell it’s actually a much smaller model because it’s faster and fails at some tasks that GPT-4 could do effortlessly. I spent hours trying to coax GPT-4o to output the desired response, but it just couldn’t do what I wanted. It was an incredibly frustrating and nail-biting experience. I couldn’t understand why this newer model couldn’t follow instructions that its predecessor handled effortlessly. The Claude Epiphany In frustration, I turned to Claude and pasted the same instructions I’d been using with GPT-4o. To my amazement, it worked flawlessly from the get-go. This made me realize that GPT-4o is likely a much less sophisticated model than I’d thought. It seems OpenAI has potentially “nerfed” it, creating a smaller model that adheres to the distribution of their vast user data’s responses, tailoring it to what the majority wants rather than maintaining the flexibility and power of the original GPT-4. What really prompted me to write this blog was Claude’s response when I asked for the definition of “anthropic.” It provided a wonderful answer, revealing that “anthropic” has a general meaning related to humans and humanity. I realized that Dario Amodei had really nailed the implications of his company’s name — it suggests that Anthropic’s sole goal is to align future AI with human intent. This is exactly what Amodei advocates for consistently, and he’s quite paranoid about it. I found this alignment of name and mission to be in really good taste, and it made me appreciate Anthropic even more. Looking Ahead For now, I’ve decided to study English with Claude rather than ChatGPT. However, I’m almost certain that this post might not age well. OpenAI is still a leading AI lab at this point, and GPT-4o’s voice mode is set to be released very soon, likely within a month of this post. The mysterious Strawberry model is also said to be already demoed to US intelligence, and is slated to be revealed later this year. In conclusion, while my journey has led me to prefer Anthropic for now, I acknowledge that the AI landscape is ever-changing. My opinions may evolve as new developments emerge, but for the moment, I’m excited to continue exploring and learning with Claude.