An interesting possibility is if the cost of generation will go down on something akin to the pace of Moore’s Law, then making $10 of revenue and losing $20 per user per month becomes profitable in a little over 3 years, assuming most of the costs are compute.
Throw in a little efficiency gain and the value of learning about the overall space and I think it’s a wise investment for Microsoft to be going full speed ahead here in order to fight to preserve its long term relevance as a tech leader.
profitability is determined by competition, not by Moore's Law. If the cost of generation goes down prices will go down as service providers battle for users, as they do everywhere else.
And given that the tech itself is rapidly proliferating and effectively just public research and compute, including open source not being far behind, the fact that they're not even making money right now when they do have somewhat of a monopoly is kind of a bad sign.
The much more likely scenario is that these systems will run locally on commodity hardware soon, it'll be free because every large company offers it, and they'll end up slapping ads on it or sell enterprisey stuff to businesses.
Interesting.
It seems that Copilot is presented in Github as a "separate" solution to the rest of the Microsoft ecosystem.
Do you think the competition will happen with new integrations to Visual Studio on the app tooling side or on the enterprise apps?
I never thought about the current computing infrastructure as a new frontier for Moore's Law. I suppose the competition will happen in that space. What I am not sure about is if Moore's Law applies to this new infrastructure as whole when compared against the advancements in computers in general.
Microsoft almost look to be building out a single copilot product across a bunch of their products (ie the eventual goal could be to share context between office365, Windows, sales, security, dynamics, Bing and github). If this were the case, it'd make sense to keep it separate
It also means they won't need to be the best for any one use, just "good enough" (and available enough) and they'll become the de facto ai for business (in the same way teams beat slack etc)
Yeah. It is sad to be forced to work with "just good enough".
I work in tech support and everyone asks for a Teams meeting to share screens. And Teams has improved a lot in the last 6 or 7 years, but, for the purposes of sharing screen for technical support, it is just good enough.
That seems… unlikely given that moores law itself doesn’t really hold any more and the costs involved here are almost entirely incurred in the data center.
Unless model size/computing cost grows at the same pace or faster. As cool as GPT 3.5 seemed when it came out it's horribly inadequate compared to 4.0 for many use cases so I doubt we'd be content using current models in 3 years.
It's much cheaper and faster (at least on the web frontend) so if you don't need high accuracy/have simple tasks/etc. and have a lot of data it's probably a better option than 4.0
"Microsoft losing lots of money on service massively gathering RLHF on code generation and telemetry which may turn out to eventually replace large swaths of an $80 billion dollar industry"
This is bad because it means Microsoft may try to make copilot worse to boost profit.
Like I can see them charge per completion, and charge too much so that it costs $100+/month to run for an average developer. Or crack down on unofficial "copilots" like https://github.com/zerolfx/copilot.el (although I really doubt more than 1% of completions go to these since they are niche tools; most people are using the official Copilot for VSCode)
It looks like there's a lot of competition though. I'll say right now, if Copilot gets worse and someone has a half-decent open-source alternative, I'll pay a substantial amount (possibly $50+/mo, which is a lot for a grad student), because AI code-completion is that useful
If the productivity boost is so high, what's a good price for it? $100/month might be more than now, but how many paid hours of dev time are saved by it? Asking as a non-user.
Its another saas product charging per seat per month, doesnt even begin to replace an actual competent developer, it’s glorified autocomplete. If it goes any higher than 20-30 bucks, I’ll have my company drop it.
20-30 bucks doesn't even begin to replace an actual competent developer's salary either, even in a super low wage country. Especially factoring in overhead.
In fact I imagine that in most places this would pay for 1 hour of their time. Which you pay per month for copilot.
So the question is not how many developers you save. But if you save more than one hour per month per developer.
Assuming a million users, this is what 20 million? A rounding error to Microsoft. The only question that matters is given enough scale will CoPilot become profitable. My guess is yes, they need to batch thousands of requests and send it through the transformer at the same time to pull down costs. Currently the scale may not be enough but I can imagine it at some point reaching there.
31 comments
[ 2.6 ms ] story [ 71.6 ms ] threadThrow in a little efficiency gain and the value of learning about the overall space and I think it’s a wise investment for Microsoft to be going full speed ahead here in order to fight to preserve its long term relevance as a tech leader.
And given that the tech itself is rapidly proliferating and effectively just public research and compute, including open source not being far behind, the fact that they're not even making money right now when they do have somewhat of a monopoly is kind of a bad sign.
The much more likely scenario is that these systems will run locally on commodity hardware soon, it'll be free because every large company offers it, and they'll end up slapping ads on it or sell enterprisey stuff to businesses.
I never thought about the current computing infrastructure as a new frontier for Moore's Law. I suppose the competition will happen in that space. What I am not sure about is if Moore's Law applies to this new infrastructure as whole when compared against the advancements in computers in general.
It also means they won't need to be the best for any one use, just "good enough" (and available enough) and they'll become the de facto ai for business (in the same way teams beat slack etc)
Do you think this holds durably true if the cost of generation is higher than the selling price?
Fixed the headline for you...
Why compete when you can buy them out?
Like I can see them charge per completion, and charge too much so that it costs $100+/month to run for an average developer. Or crack down on unofficial "copilots" like https://github.com/zerolfx/copilot.el (although I really doubt more than 1% of completions go to these since they are niche tools; most people are using the official Copilot for VSCode)
It looks like there's a lot of competition though. I'll say right now, if Copilot gets worse and someone has a half-decent open-source alternative, I'll pay a substantial amount (possibly $50+/mo, which is a lot for a grad student), because AI code-completion is that useful
In fact I imagine that in most places this would pay for 1 hour of their time. Which you pay per month for copilot.
So the question is not how many developers you save. But if you save more than one hour per month per developer.
I'd imagine that will be fairly easy to achieve.
For my previous job, it would save me 5min / week.
And it also saves me of using the keyboard so I guess it's also better for those with carpal tunnel syndrome.
A gradual pricing per request would totally work for me, cheap when I'm not using it, justified when I'm using it a lot.
Some more discussion over here:
https://news.ycombinator.com/item?id=37821756