5 ms·

Half of what I do at work is point out to engineers when they have coupled independent concerns that are not actually coupled. Which means their problem is eith

Half of what I do at work is point out to engineers when they have coupled independent concerns that are not actually coupled. Which means their problem is either simple, or they're asking independent questions with independent answers. The New Bing has absolutely NOTHING to do with the New Edge, and it's infuriating that Microsoft continues to insist on bundling Edge upsell into everything. There are pros and cons to Bing-ChatGPT. There are pros and cons to Edge. The two have sweet f-a to do with each other.

nonethewiser 4y ago
Chat and Compose are two new AI features in Edge. That’s the connection. Or do you mean theyre not connected in the theoretical sense? Revealing these AI powered web browsing features together seems rather obvious to me.
danShumway 4y ago
> Half of what I do at work is point out to engineers when they have coupled independent concerns that are not actually coupled. Honestly, this is kind of an applicable point to raise about New Bing in general. Some of the fundamentally hard problems around LLMs feel like they exist because we're trying to couple everything to the AI. Facebook is trying to teach their system how to make API calls, Microsoft is also blue-skying about Bing's AI being able to set calendar appointments. Well congrats, now prompt injection actually matters, and it's an extremely difficult problem to solve if it's solveable at all. Does the LLM need to do literally everything? Could it interpret input and then have that input sent to a (specifically non-AI) sanitizer and then manipulated using normal algorithms that can be debugged and tested? There are scenarios that GPT is brilliant at, and it seems like the response to that has been to mash everything together and say "the LLM will be all the systems now." But the LLM isn't good at all the systems, it's good at a very limited subset of systems. This was my contention when Bing AI was first announced: even if it's perfect, having a conversation in paragraph form is very often not at all what I want from a search engine. To me, those are orthogonal tasks; they're not connected. I really don't want an AI or a human giving me an answer and a couple of sources, I don't want the information summarized at all. To me, asking a question and searching for information are two separate user actions, and it's not clear to me why they're being coupled together. "But you could do X/Y/whatever, you could ask it simple questions, you could ask it to summarize." Okay, that's fine. But... does that need to be coupled to search? You could do all of that anyway. You could do a normal search and then you could separately go to the AI and ask it to summarize something. Similarly, great, Bing AI will theoretically be able to schedule a calendar appointment in the future. Is that a thing that needed to be done through an LLM specifically? Couldn't there have been some level of separation between them so that the LLM going off-script is less of a critical problem to solve?

nonethewiser 4y ago

Chat and Compose are two new AI features in Edge. That’s the connection. Or do you mean theyre not connected in the theoretical sense? Revealing these AI powered web browsing features together seems rather obvious to me.

danShumway 4y ago

> Half of what I do at work is point out to engineers when they have coupled independent concerns that are not actually coupled. Honestly, this is kind of an applicable point to raise about New Bing in general. Some of the fundamentally hard problems around LLMs feel like they exist because we're trying to couple everything to the AI. Facebook is trying to teach their system how to make API calls, Microsoft is also blue-skying about Bing's AI being able to set calendar appointments. Well congrats, now prompt injection actually matters, and it's an extremely difficult problem to solve if it's solveable at all. Does the LLM need to do literally everything? Could it interpret input and then have that input sent to a (specifically non-AI) sanitizer and then manipulated using normal algorithms that can be debugged and tested? There are scenarios that GPT is brilliant at, and it seems like the response to that has been to mash everything together and say "the LLM will be all the systems now." But the LLM isn't good at all the systems, it's good at a very limited subset of systems. This was my contention when Bing AI was first announced: even if it's perfect, having a conversation in paragraph form is very often not at all what I want from a search engine. To me, those are orthogonal tasks; they're not connected. I really don't want an AI or a human giving me an answer and a couple of sources, I don't want the information summarized at all. To me, asking a question and searching for information are two separate user actions, and it's not clear to me why they're being coupled together. "But you could do X/Y/whatever, you could ask it simple questions, you could ask it to summarize." Okay, that's fine. But... does that need to be coupled to search? You could do all of that anyway. You could do a normal search and then you could separately go to the AI and ask it to summarize something. Similarly, great, Bing AI will theoretically be able to schedule a calendar appointment in the future. Is that a thing that needed to be done through an LLM specifically? Couldn't there have been some level of separation between them so that the LLM going off-script is less of a critical problem to solve?