Lars Hertel

There are also no-code options to use GPT for things you'd like to build. For example spell.so helps you get started quickly if you don't code. Otherwise you can use Zapier as well

8 on my job, 2 on my projects

+1 on temperature. Setting that to zero should give deterministic output

Automate oncall tasks and other repetitive shell related tasks. Honestly, there is probably already potential for automation but the bar for setting it up seems high... I do want a virtual assistant that I can chat with during my long commute and which can take notes for me. Even better if it could help me do marketing tasks like catch up on social media and Tweet

In my experience, the hardest aspects were to get a predictable experience and to make it safe. 1) User inputs are often unpredictable to begin with. Then add on top of that the fact that you don't know what the exact output of ChatGPT will be. In the best case this results in a slightly unexpected experience, in the worst case it could result in errors if the ChatGPT output doesn't match...

From reading the article, the argument is largely built on cheap fine-tuning of LLMs which could take open source models to match or outperform better models. I think it's still early to say, there is so much innovation happening right now. A few months ago you would have thought it is all gonna go to the large companies + OpenAI. Now it looks the opposite. I'd say wait and see but build anyway...