I talk to corporate normies a lot, and what I see is this:
- copilot as a "internal super google" is incredibly valuable. However, no company is so completely on the Microsoft train that copilot alone cam do what they want. Eg if they have Salesforce, then tough luck as copilot cant "see into" that. Moreover copilot today has strange limitations even inside the 365 stack, eg with OneNote. If microsoft can build a corporate-focused RAG that genuinely mops up all these kinds of sources, they will have a winner on their hands.
- copilot as coder is a mixed bag. Good for jumpstarting tasks, but not reliable enough to be left unguarded, ie experienced devs still need to be paid to shepherd it.
- everything else: not much of anything achieving real, deep value. Lots of anecdotal use cases like drafting emails.
- scary as hell: the rise of shadow AI IT where corp users use chatGPT with company info, ignoring the risk that what they ask gets folded into training data. Proprietary info leakage is going to be a problem I think.
None of the above makes any of the tools worth $30/mo per user. A super google copilot that genuinely captured everything across all silos and presented a unified interface (none of this cruddy plugin agent nonsense) would be worth north of $50/mo at any large corporation.
>scary as hell: the rise of shadow AI IT where corp users use chatGPT with company info, ignoring the risk that what they ask gets folded into training data. Proprietary info leakage is going to be a problem I think.
Microsoft, for all their faults, seems to have honed in on this issue explicitly. Enterprise Data Protection seems like its designed to massage this fear, whether it works or not who knows.
Except ChatGPT dot com somehow gives much better results--even without signing into an account!-- than does Copilot. I thought it was pretty much the same model under the hood but the results speak for themselves.
> None of the above makes any of the tools worth $30/mo per user. A super google copilot that genuinely captured everything across all silos and presented a unified interface (none of this cruddy plugin agent nonsense) would be worth north of $50/mo at any large corporation.
I feel like you pulled those numbers out of thin air. $30 is something like 30 minutes of work or less in a typical office. Can a super-google that covers most of what you need (but not everything in every silo) save you 30 minutes per month? Yes for sure.
> everything else: not much of anything achieving real, deep value. Lots of anecdotal use cases like drafting emails.
Well, this is what people are missing IMO. Yeah, general purpose chatbots can’t solve advanced problems. However, models and pipelines (using similar technology as chabots) specifically built for certain tasks can achieve really incredible results — like in the area of protein folding (https://youtu.be/P_fHJIYENdI)
I think there is a lot of potential, but most companies are just slapping a chatbot (which uses some other company’s models) on their app and calling it a day. Very few are taking a deep look at how ML could actually solve real problems.
> - scary as hell: the rise of shadow AI IT where corp users use chatGPT with company info, ignoring the risk that what they ask gets folded into training data. Proprietary info leakage is going to be a problem I think.
IT Sec has these things blocked for this reason. Probably made easier by only allowing company devices on the company network.
Most salaried staff aren’t fully productive every minute of the workday. Studies show a typical work week’s output can be achieved in 30 hours without losing productivity.
I’ve seen this firsthand throughout my career. (That’s why I moved to working with remote-first, globally distributed teams over five years ago.)
Every article I've seen so far on LLM datacenter investment is either very bullish or very bearish and I haven't seen sober exploration of whether it might be neither.
What happens to the tech world if it turns out that LLMs really are the 20-billion dollar product category that every industry needs to explore to maintain competitiveness but they will never be the 500-billion dollar product category that changes the world? What if they are stuck at the middle-level of importance forever? Do Microsoft, Meta, etc need a federal bailout if that happens?
They don't know anything dude. You can't read finance articles on AI. They cannot even install ollama locally. This means the only information they can provide is conjecture and rumors on how money is moving, absolutely nothing else. The Deepseek reporting exposed them.
Huang expressed the truth. The reasoning feature is a loop, which means more hits to the LLM, which means more compute. So more datacenters.
This doesn’t make sense as an analysis. Financial journalist didn’t know how factories, steam engines, websites, data centers, or cryptocurrency worked either.
Okay, let me increase the resolution. Financial analyst and Gizmodo are not exactly a pair. Quite frankly, Financial analyst and CNBC or Bloomberg are not exactly a couple either. Those things don't go together.
As in, which of the following does not belong in the group:
1. Gizmodo
2. CNBC
3. Bloomberg
4. Financial Analyst
You might think it's Gizmodo, but it's not. The first three are the same is my point, and there's no way a financial analyst is even there (let alone a technical one that knows about AI). Now what if I told you the thing you thought doesn't belong there (Gizmodo) and the thing I'm telling you doesn't belong there (Financial Analyst) are in the same place? You should be confounded, it's ridiculous (Gizmodo Financial Analyst, huh?). My last sentence is the best joke of the year, trust me on that.
They have no credibility across the board (across all three of those things) is sort of the criticism.
You don't need to worry. MS and Meta both have very healthy revenue streams, and billions of cash lying around. Their current loans are easily covered.
As much as the markets love to talk about investment values, and valuations, this really means very little. A lot of the investment is in kind, not cash, and the marginal cost of the kind is often a tiny fraction of the marginal cost.
Bailouts to US auto manufacturers occurred because they were both unprofitable and loaded with debt they could not pay for. Ditto the bank bailouts of 2008.
MS paid billions for Skype and billions more for Nokia. Nokia (windows phone) died soon, Skype will be turned off this May. Neither caused a stir, when your cash on hand is measured in tend of billions big investments are not risky.
What happens to the tech world if it turns out that LLMs really are the 20-billion dollar product category that every industry needs to explore to maintain competitiveness but they will never be the 500-billion dollar product category that changes the world?
You still have a huge build-up and a bust? Earlier real products, from railroads to the Internet, have followed this pattern.
Even more, AI has a plausible-sounding argument for being a 500b product or even $infinity dollars. I don't think it's correct with current tech but it's wrongness rests on subtle things. People who seem at least as smart as Gary Marcus (AI critic) seem honestly convinced this AI generation of is going to create AGI. An investor who doesn't know who's right is going to want to cover their bets.
> Every article I've seen so far on LLM datacenter investment is either very bullish or very bearish and I haven't seen sober exploration of whether it might be neither.
Because bullish articles are bought by AI companies and/or encouraged by editors because AI is the hot traffic keyword right now, hence tons and tons of articles about utterly banal ML integrations being rebranded as AI so Wall Street will think your fuckin Microwave company is ushering in Skynet.
And on the flip side the bearish articles are because we're now almost 2 years into this hype cycle that Silicon Valley is trying so incredibly hard to make us believe is real, and we've had, yes, a few minor interesting things, but the vast majority of every part of it is unwanted integration of chat-bots into products that do not need chat-bots, or the replacement of lower-paid staff with a ChatGPT integration that doesn't work.
There is, and I cannot stress this enough, so incredibly little consumer interest in this tech. Nobody. Fucking. Cares. In fact, I read something recently that including AI in product descriptions actually hurts the rate of purchase by people because they're so fucking sick of it.
Now mind you, that's not going to kill AI companies, because AI companies are not worth 50 bazillion dollars on the back of chatbots. They are worth all that because of the "usefulness" of varying legitimacy in the defense industry, the insurance industry, border security, all manner of things that will almost without fail make your or other people's lives meaningfully worse, but they'll do it in a cost-effective way that reinforces existing capital and power. That's why it's worth so much money.
The recent advances in image recognition have millions of minor applications to agriculture, manufacturing and logistics, but the net effect is workforce reduction.
> I read something recently that including AI in product descriptions actually hurts the rate of purchase by people because they're so fucking sick of it
A year ago it looked like Microsoft had an insurmountable lead thanks to their investment in OpenAI. Microsoft has wasted their lead. Copilot is definitely not worth $30 per month, not when there are free offerings that work better. ChatGPT should have given Bing a big boost but Microsoft didn't execute well. You know who has executed well? Perplexity.ai. Perplexity shows what a modern AI-enabled search engine can do.
It’s really weird. On the one hand they are pushing hard down their customer’s throats and yet on the other they haven’t really produced something of killer value.
Your subscription includes not just the stellar search engine, but access to their assistant as well, along with all their privacy benefits
What I love about assistant specifically is that it lets you switch between one of 10+ models. If you are curious for example how a different model would answer a prompt, within the same thread you can change the model and re-ask your question
Gemini, deepseek, Claude, mixtral, ChatGPT, etc
Definitely worth checking out - has definitely help me take advantage of these bots more
Yay, another company have problems becouse of close relationship with MS, move on.
Or maybe it is not yet Nokia or SGI like situation but part of fight over control over OpenAI. Anyway it's just like Starlink and Ukraine situation - one buttonpress and access to cloud is no more. Or just prices can be changed with no apparent reason... :> OpenAI was doomed in that milisecond when it started relay on MS.
One is thoroughly specified and covered by extensive input-output pair test-cases.
The other is completely unspecified, and not possible to pin down with test cases due to constantly churning. (The users quickly learned not to rely on any specific outputs, so the churn doesn't break anyone.)
Everything in between is hard to replace to varying degrees, because it's reliable enough that users can depend on specific behaviors, yet what users dare rely on is not always obvious from specification or covered with tests.
Due to sneaky website scraping by OpenAI many sites now have their content locked behind login, have their APIs severely restricted and watchout for scraping.
Unless you want to build a specialized AI built on specific open or closed training corpus, huge training data is required to build something similar to OpenAI's models.
Google with it's crawlers and associated relaxations should not be affected much by those restrictions.
Miscrosoft can easily update its browser/OS ToS and say "we'll use all your browsing data for training, anonymously ofcourse" and there's not much users will do except shrug.
Is there anyone else who is better positioned to collect the content for training or am I placing too much importance on training data?
I agree, it's far more the 'training data' than the model architecture.
Given that Windows 11 is such a surveillance/advertisement machine I'm sure they sit on years of user movements MS can use to build a OS-running AI, similar to what Anthropic has been trying with 'computer use'. That's where I see Microsoft's niche in the future, and it's an enormous niche: how many people do 'data entry' as a full-time job who'd be unemployed once an LLM can move data from random sources into an Excel sheet by operating a simulated computer?
Companies like Uber live off solely from the "we have a large war chest of money to kill competition" moat model where any competition in some area can be killed easily as long as you are much larger than them and can afford to lose money in the area they are active in: just offer lower prices until they are bankrupt, then raise them again.
This model has proven to work in places like retail. Literally how Amazon got big.
I think they are trying to use this strategy for AI as well, but at least right now there is so much competition and investment that it's tough to pull off.
OpenAI has a huge feud with Elon Musk. Microsoft detaching itself from OpenAI and exploring partnerships with xAI among others makes a lot of sense as the card deck shuffles and more power and influence is concentrated near those in the good graces of the Trump administration. Sam Altman and OpenAI are a liability at this point.
At the risk of making ham-fisted comparisons to a guy whose first name was an anagram of Elon, befriending the new Politburo is a fantastically effective way to capture market share and drive competitors out of business.
I talk to corporate normies a lot, and what I see is this:
- copilot as a "internal super google" is incredibly valuable. However, no company is so completely on the Microsoft train that copilot alone cam do what they want. Eg if they have Salesforce, then tough luck as copilot cant "see into" that. Moreover copilot today has strange limitations even inside the 365 stack, eg with OneNote. If microsoft can build a corporate-focused RAG that genuinely mops up all these kinds of sources, they will have a winner on their hands.
- copilot as coder is a mixed bag. Good for jumpstarting tasks, but not reliable enough to be left unguarded, ie experienced devs still need to be paid to shepherd it.
- everything else: not much of anything achieving real, deep value. Lots of anecdotal use cases like drafting emails.
- scary as hell: the rise of shadow AI IT where corp users use chatGPT with company info, ignoring the risk that what they ask gets folded into training data. Proprietary info leakage is going to be a problem I think.
None of the above makes any of the tools worth $30/mo per user. A super google copilot that genuinely captured everything across all silos and presented a unified interface (none of this cruddy plugin agent nonsense) would be worth north of $50/mo at any large corporation.
>scary as hell: the rise of shadow AI IT where corp users use chatGPT with company info, ignoring the risk that what they ask gets folded into training data. Proprietary info leakage is going to be a problem I think.
Microsoft, for all their faults, seems to have honed in on this issue explicitly. Enterprise Data Protection seems like its designed to massage this fear, whether it works or not who knows.
Except ChatGPT dot com somehow gives much better results--even without signing into an account!-- than does Copilot. I thought it was pretty much the same model under the hood but the results speak for themselves.
I agree completely. I was remarking only on the data protection.
> None of the above makes any of the tools worth $30/mo per user. A super google copilot that genuinely captured everything across all silos and presented a unified interface (none of this cruddy plugin agent nonsense) would be worth north of $50/mo at any large corporation.
I feel like you pulled those numbers out of thin air. $30 is something like 30 minutes of work or less in a typical office. Can a super-google that covers most of what you need (but not everything in every silo) save you 30 minutes per month? Yes for sure.
> $30 is something like 30 minutes of work or less in a typical
Yeah, if by typical office you mean IT department in the US…
Did you include other cost than salary in that estimate? Taxes, insurance, office space, maybe even support staff etc. It quickly becomes a lot.
How does this change the equation on a fundamental level?
> everything else: not much of anything achieving real, deep value. Lots of anecdotal use cases like drafting emails.
Well, this is what people are missing IMO. Yeah, general purpose chatbots can’t solve advanced problems. However, models and pipelines (using similar technology as chabots) specifically built for certain tasks can achieve really incredible results — like in the area of protein folding (https://youtu.be/P_fHJIYENdI)
I think there is a lot of potential, but most companies are just slapping a chatbot (which uses some other company’s models) on their app and calling it a day. Very few are taking a deep look at how ML could actually solve real problems.
Same as it ever was.
> - scary as hell: the rise of shadow AI IT where corp users use chatGPT with company info, ignoring the risk that what they ask gets folded into training data. Proprietary info leakage is going to be a problem I think.
IT Sec has these things blocked for this reason. Probably made easier by only allowing company devices on the company network.
> None of the above makes any of the tools worth $30/mo per user.
Are you sure? Doesn't the tool just have to save the user less than an hour a month for it to break even?
Most salaried staff aren’t fully productive every minute of the workday. Studies show a typical work week’s output can be achieved in 30 hours without losing productivity.
I’ve seen this firsthand throughout my career. (That’s why I moved to working with remote-first, globally distributed teams over five years ago.)
How many people do you license it for at how many really use it?
Glad to see another SE on here. Not enough of us on this site.
Every article I've seen so far on LLM datacenter investment is either very bullish or very bearish and I haven't seen sober exploration of whether it might be neither.
What happens to the tech world if it turns out that LLMs really are the 20-billion dollar product category that every industry needs to explore to maintain competitiveness but they will never be the 500-billion dollar product category that changes the world? What if they are stuck at the middle-level of importance forever? Do Microsoft, Meta, etc need a federal bailout if that happens?
They don't know anything dude. You can't read finance articles on AI. They cannot even install ollama locally. This means the only information they can provide is conjecture and rumors on how money is moving, absolutely nothing else. The Deepseek reporting exposed them.
Huang expressed the truth. The reasoning feature is a loop, which means more hits to the LLM, which means more compute. So more datacenters.
I totally agree about all of the flaws, but I don't see that as an obstacle to profitability under the typical Schumpeterian conditions of business.
After all, look at how much money Myanmar makes from enslaving people to work as unpaid internet catfishers.
This doesn’t make sense as an analysis. Financial journalist didn’t know how factories, steam engines, websites, data centers, or cryptocurrency worked either.
That doesn’t undermine their point.
It’s foundational to their point.
Okay, let me increase the resolution. Financial analyst and Gizmodo are not exactly a pair. Quite frankly, Financial analyst and CNBC or Bloomberg are not exactly a couple either. Those things don't go together.
As in, which of the following does not belong in the group:
1. Gizmodo
2. CNBC
3. Bloomberg
4. Financial Analyst
You might think it's Gizmodo, but it's not. The first three are the same is my point, and there's no way a financial analyst is even there (let alone a technical one that knows about AI). Now what if I told you the thing you thought doesn't belong there (Gizmodo) and the thing I'm telling you doesn't belong there (Financial Analyst) are in the same place? You should be confounded, it's ridiculous (Gizmodo Financial Analyst, huh?). My last sentence is the best joke of the year, trust me on that.
They have no credibility across the board (across all three of those things) is sort of the criticism.
Take your meds.
You don't need to worry. MS and Meta both have very healthy revenue streams, and billions of cash lying around. Their current loans are easily covered.
As much as the markets love to talk about investment values, and valuations, this really means very little. A lot of the investment is in kind, not cash, and the marginal cost of the kind is often a tiny fraction of the marginal cost.
Bailouts to US auto manufacturers occurred because they were both unprofitable and loaded with debt they could not pay for. Ditto the bank bailouts of 2008.
MS paid billions for Skype and billions more for Nokia. Nokia (windows phone) died soon, Skype will be turned off this May. Neither caused a stir, when your cash on hand is measured in tend of billions big investments are not risky.
>Do Microsoft, Meta, etc need a federal bailout if that happens?
There is no doubt that they will get one if they need it so why worry about it?
Crazier things have happened unfortunately.
What happens to the tech world if it turns out that LLMs really are the 20-billion dollar product category that every industry needs to explore to maintain competitiveness but they will never be the 500-billion dollar product category that changes the world?
You still have a huge build-up and a bust? Earlier real products, from railroads to the Internet, have followed this pattern.
Even more, AI has a plausible-sounding argument for being a 500b product or even $infinity dollars. I don't think it's correct with current tech but it's wrongness rests on subtle things. People who seem at least as smart as Gary Marcus (AI critic) seem honestly convinced this AI generation of is going to create AGI. An investor who doesn't know who's right is going to want to cover their bets.
You could have made similar arguments for VR/AR, yet here we are.
> Every article I've seen so far on LLM datacenter investment is either very bullish or very bearish and I haven't seen sober exploration of whether it might be neither.
Because bullish articles are bought by AI companies and/or encouraged by editors because AI is the hot traffic keyword right now, hence tons and tons of articles about utterly banal ML integrations being rebranded as AI so Wall Street will think your fuckin Microwave company is ushering in Skynet.
And on the flip side the bearish articles are because we're now almost 2 years into this hype cycle that Silicon Valley is trying so incredibly hard to make us believe is real, and we've had, yes, a few minor interesting things, but the vast majority of every part of it is unwanted integration of chat-bots into products that do not need chat-bots, or the replacement of lower-paid staff with a ChatGPT integration that doesn't work.
There is, and I cannot stress this enough, so incredibly little consumer interest in this tech. Nobody. Fucking. Cares. In fact, I read something recently that including AI in product descriptions actually hurts the rate of purchase by people because they're so fucking sick of it.
Now mind you, that's not going to kill AI companies, because AI companies are not worth 50 bazillion dollars on the back of chatbots. They are worth all that because of the "usefulness" of varying legitimacy in the defense industry, the insurance industry, border security, all manner of things that will almost without fail make your or other people's lives meaningfully worse, but they'll do it in a cost-effective way that reinforces existing capital and power. That's why it's worth so much money.
The recent advances in image recognition have millions of minor applications to agriculture, manufacturing and logistics, but the net effect is workforce reduction.
> I read something recently that including AI in product descriptions actually hurts the rate of purchase by people because they're so fucking sick of it
https://news.ycombinator.com/item?id=41126685
Every non-tech/finance person I know is indifferent or strongly opposed to AI anything.
> Nobody. Fucking. Cares.
It’s so weird to talk to my normie family members about it. Things that are really cool to me they are really bored by.
A year ago it looked like Microsoft had an insurmountable lead thanks to their investment in OpenAI. Microsoft has wasted their lead. Copilot is definitely not worth $30 per month, not when there are free offerings that work better. ChatGPT should have given Bing a big boost but Microsoft didn't execute well. You know who has executed well? Perplexity.ai. Perplexity shows what a modern AI-enabled search engine can do.
It’s really weird. On the one hand they are pushing hard down their customer’s throats and yet on the other they haven’t really produced something of killer value.
I like to plug Kagi whenever I get a chance
Your subscription includes not just the stellar search engine, but access to their assistant as well, along with all their privacy benefits
What I love about assistant specifically is that it lets you switch between one of 10+ models. If you are curious for example how a different model would answer a prompt, within the same thread you can change the model and re-ask your question
Gemini, deepseek, Claude, mixtral, ChatGPT, etc
Definitely worth checking out - has definitely help me take advantage of these bots more
Do they have privacy policy around their AI assistant?
Yes, they make the API calls anonymously
They don’t tie anything history back to your account, you can pay it with bitcoin even
It’s worth checking out their privacy policy - in a nutshell they do zero advertising and zero affiliate type stuff so they just don’t need to
$30 for copilot doesn’t make sense when Cursor Pro is $20.
Classic Microsoft branding confusion. Microsoft Copilot isn’t Github Copilot.
where are you paying $30 for copilot? i see $20 on the website for the copilot chatbot, and i'm currently paying $10/mo for the vscode addon.
They raised the price of Office to make you pay for copilot.
That's a different copilot
If you cancel they offer the old plan/price.
[flagged]
Yay, another company have problems becouse of close relationship with MS, move on.
Or maybe it is not yet Nokia or SGI like situation but part of fight over control over OpenAI. Anyway it's just like Starlink and Ukraine situation - one buttonpress and access to cloud is no more. Or just prices can be changed with no apparent reason... :> OpenAI was doomed in that milisecond when it started relay on MS.
It's easy to replace two kinds of functions.
One is thoroughly specified and covered by extensive input-output pair test-cases.
The other is completely unspecified, and not possible to pin down with test cases due to constantly churning. (The users quickly learned not to rely on any specific outputs, so the churn doesn't break anyone.)
Everything in between is hard to replace to varying degrees, because it's reliable enough that users can depend on specific behaviors, yet what users dare rely on is not always obvious from specification or covered with tests.
Microsoft used OpenAI to "sell" its Azure products, and now it's trying to throw OpenAI under the bus
https://justdario.com/2024/10/the-smoking-gun-that-proves-ho...
Discussion: https://news.ycombinator.com/item?id=43292946
Even Microsoft knows that there is no moat in AI models given that it is fundamentally a race to zero.
The difference is, they finally realized that after Apple almost invested but stepped back.
Perhaps they know that this hype is a repeat of being in the middle of the year 1999; or even we are close to the start of the year 2000.
Isn't the training data a moat by itself?
Due to sneaky website scraping by OpenAI many sites now have their content locked behind login, have their APIs severely restricted and watchout for scraping.
Unless you want to build a specialized AI built on specific open or closed training corpus, huge training data is required to build something similar to OpenAI's models.
Google with it's crawlers and associated relaxations should not be affected much by those restrictions.
Miscrosoft can easily update its browser/OS ToS and say "we'll use all your browsing data for training, anonymously ofcourse" and there's not much users will do except shrug.
Is there anyone else who is better positioned to collect the content for training or am I placing too much importance on training data?
I agree, it's far more the 'training data' than the model architecture.
Given that Windows 11 is such a surveillance/advertisement machine I'm sure they sit on years of user movements MS can use to build a OS-running AI, similar to what Anthropic has been trying with 'computer use'. That's where I see Microsoft's niche in the future, and it's an enormous niche: how many people do 'data entry' as a full-time job who'd be unemployed once an LLM can move data from random sources into an Excel sheet by operating a simulated computer?
Disagree, now that there are great open models are available I think there's less need of huge training data--can just post train
Companies like Uber live off solely from the "we have a large war chest of money to kill competition" moat model where any competition in some area can be killed easily as long as you are much larger than them and can afford to lose money in the area they are active in: just offer lower prices until they are bankrupt, then raise them again.
This model has proven to work in places like retail. Literally how Amazon got big.
I think they are trying to use this strategy for AI as well, but at least right now there is so much competition and investment that it's tough to pull off.
Didn’t Apple just announce a 500b investment plan?
Training data is very much a moat. Especially training data that can be reliably filtered for LLM output.
OpenAI has a huge feud with Elon Musk. Microsoft detaching itself from OpenAI and exploring partnerships with xAI among others makes a lot of sense as the card deck shuffles and more power and influence is concentrated near those in the good graces of the Trump administration. Sam Altman and OpenAI are a liability at this point.
At the risk of making ham-fisted comparisons to a guy whose first name was an anagram of Elon, befriending the new Politburo is a fantastically effective way to capture market share and drive competitors out of business.
Great point. Paying Musk is the only way these days to ensure that an “unfortunate accident” will not happen to your company
Wouldn't Occam's razor at this point be OpenAI doesn't have a moat and Microsoft knows this?
No
I don't know. As someone not from US, it looks like Sam Altman is also in good relationship with Trump due to Stargate.