I thought GPT5 and the other models used in ChatGPT are decoder only model, I see the diagram has a encoder/decoder. Also, during inferencing tokenization, prefill and decoding are the three main phases. It would have been great if the blog covered those concepts. I somehow felt that the steps mentioned here may not be exactly what chatgpt does when a request is sent to it, but I would like to get myself corrected if my observation is incorrect.
ChatGPT dropped a new product called Agent Builder last week and I couldn’t wait to jump on the bandwagon and play around with it.
I tested it out by wrapping my API endpoint inside a custom MCP server, then hooked it up as a custom server in ChatGPT’s new Agent Builder, right inside the drag-and-drop canvas.
While connecting my own custom MCP server, I ran into a few little gotchas that you’ll definitely want to know about if you’re planning to build one yourself.
Thank you all for these beautifully accessible pieces - very helpful for the less comp-mech literate of us working to better understand AI architecture and behaviour.
Grateful if you could help me understand if there are circumstances in which ChatGPT has access to material typed before pressing send. The 'correct' answer seems to be 'no' - but users are reporting behavioural instances that would suggest otherwise.
I thought GPT5 and the other models used in ChatGPT are decoder only model, I see the diagram has a encoder/decoder. Also, during inferencing tokenization, prefill and decoding are the three main phases. It would have been great if the blog covered those concepts. I somehow felt that the steps mentioned here may not be exactly what chatgpt does when a request is sent to it, but I would like to get myself corrected if my observation is incorrect.
u re correct
ChatGPT dropped a new product called Agent Builder last week and I couldn’t wait to jump on the bandwagon and play around with it.
I tested it out by wrapping my API endpoint inside a custom MCP server, then hooked it up as a custom server in ChatGPT’s new Agent Builder, right inside the drag-and-drop canvas.
While connecting my own custom MCP server, I ran into a few little gotchas that you’ll definitely want to know about if you’re planning to build one yourself.
https://xianli.substack.com/p/how-i-connected-a-custom-mcp-server
excellent!!
Thank you all for these beautifully accessible pieces - very helpful for the less comp-mech literate of us working to better understand AI architecture and behaviour.
Grateful if you could help me understand if there are circumstances in which ChatGPT has access to material typed before pressing send. The 'correct' answer seems to be 'no' - but users are reporting behavioural instances that would suggest otherwise.
thanks