I was reading an article where Sutskever mentioned that “he was initially unimpressed by the chatbot’s inability to answer questions accurately” [1]. I kept that tab open on my screen for two weeks. Seeing that line every day for two weeks hit me hard. I have been training LLMs since early 2020 and was equally unimpressed by their inaccuracy. Now, after spending some years developing LLM-based systems to solve business needs, I look back and feel ashamed. How did I not realize the usefulness and convenience of these systems despite their inaccuracies? I could have built many valuable applications based on the simple, inaccurate GPT-2 model!

For two weeks, I reflected on similar events in the history of computing.

Convenient, not moral. Freedom is essential, but convenience should not be overlooked. Striving for the ultimate moral high ground in developing software can sometimes lead to impractical products. The FSF and RMS spent their lives fighting for higher software values, like freedom, while deliberately ignoring convenience. The outcomes could have been better if they were more tolerant of convenience.

Convenient, not elegant. Steve Jobs’ focus on product design as a priority in developing the early version Macintosh led to disastrous sales records. That Macintosh was overly overpriced, and had many convenience issues (e.g. memory limitation). When he returned to Apple in the the late 90’s, he was more tolerant of convenience. Design is important, but convenience is usually more important. He later made similar mistakes, though minor, sacrificing convenience for better design, e.g the slot CD.

Convenient, not accurate. LLM based systems don’t provide (as of now) absolute accuracy, predictability, and determinism but they are extremely convenient to use and they solve so many problem.

It is a sad fact that companies which focused only on convenience, such as Microsoft and IBM, survived longer and sometimes made more money, even though they rarely had creative or interesting products.

[1] https://www.businessinsider.com/chatgpt-was-inaccurate-boring-when-it-launched-openai-cofounder-2023-10