The company's Bing search engine, which has been integrated with OpenAI's ChatGPT, has resulted in users engaging in "more than 1.9 billion chats" so far, said Satya Nadella.
That’s not such a big deal. Their objective is to get people hooked on the system. After that, they’ll jack up the price. Microsoft can easily afford to lose money for several years in pursuit of that target.
(One way this plan could fall through is if LLM tech progresses to the extent that free and open source copilots, run locally, can give result that are just as good.)
One way this plan could fall through is if LLM tech progresses to the extent that free and open source copilots, run locally, can give result that are just as good.
Not familiar with the tech, but wouldn’t server-side LLMs still have an advantage regardless because of the greater power available on tap? Anything that improves local LLM will also benefit server-side LLMs, wouldn’t it?
Possibly, but given the choice between paying $20/m for a marginally better version of something that’s free and probably built in to your editor at that point, most people would probably take the free thing. At that point paid llms will need to find new niches beyond simply existing.
Not necessarily, as it gets faster the latency between your local and remote machines becomes a bigger fraction of the time taken to process anything. If your local machine processes in 50ms and the remote machine in 5s, a latency of just 45ms would make your machine faster.
Running locally also cuts out a lot of potential security issues inherent to sending data over a network, and not sending your data to a third party is a bonus too.
That’s not such a big deal. Their objective is to get people hooked on the system. After that, they’ll jack up the price. Microsoft can easily afford to lose money for several years in pursuit of that target.
(One way this plan could fall through is if LLM tech progresses to the extent that free and open source copilots, run locally, can give result that are just as good.)
MS might be in trouble then.
Performance is not great but apparently it’s not optimized at all as of right now.
There are already very impressive local models for coding. Some have come out favourably to copilot in tests iirc
Edit: https://evalplus.github.io/leaderboard.html
Not familiar with the tech, but wouldn’t server-side LLMs still have an advantage regardless because of the greater power available on tap? Anything that improves local LLM will also benefit server-side LLMs, wouldn’t it?
Possibly, but given the choice between paying $20/m for a marginally better version of something that’s free and probably built in to your editor at that point, most people would probably take the free thing. At that point paid llms will need to find new niches beyond simply existing.
Not necessarily, as it gets faster the latency between your local and remote machines becomes a bigger fraction of the time taken to process anything. If your local machine processes in 50ms and the remote machine in 5s, a latency of just 45ms would make your machine faster.
Running locally also cuts out a lot of potential security issues inherent to sending data over a network, and not sending your data to a third party is a bonus too.