We'll see where we end up 5 years from now, but I think we may see Apple in a very nice position with AI by having tight integration with their platform, they'll never be first in AI so they won't have the expectation to lead the market, and their on-device first strategy may help keep costs from inflating wildly out of control if high-compute features fizzle out.
As with all AI so far, still lacking to see the "why". Microsoft came out hard, but still don't see why other than generating potentially very wrong information or goofy looking pictures.
Well Apple is doing OK and Amazon lost their hat even though Apple was “way behind” in smart speakers.
AI could turn out the same. You need good training data, but the biggest set out there (Internet) was quickly polluted to hell by AI and now it’s far less valuable for training.
I’m skeptical of its long term value for the average user.
I would settle for a local model to extract tables from PDFs. MS Co-Pilot is really disappointing in its utility and integration.
Power Automate offers an Adobe connector that purports to do so but my data loss prevention policies forbid me to connect to third parties.
GPT4o does a great job of taking a PDF and giving me exactly the contents in a csv of my specification, but again, can't do this with customer data.
I'm looking at what is possible with llama3 as an azure service but it looks very complex to set up, I think I need to have a "data verse" and a network configuration to allow serverless compute to connect to my files. That will be nice tho when I can have the results of the PDF extraction insert rows into my excel files.
I don't use Salesforce but I'm jealous of people who have a platform that they can actually integrate AI with their customer data.
Even Amazon's Textract API has issues with common invoices I've received. It Would Be Nice(tm) if this were a solved problem, but it isn't yet. Several other SaaS build on top of Textract and claim to do better, but the costs were unreasonable for the quantities of invoices I was processing compared to munging the raw data from Textract.
Sadly, I did not keep the notes on what I had found when I did the research. What turned me off of those services was that my out of pocket expense was going to be thousands of dollars to process around 1600 invoices. With AWS Textract the OCRed invoices cost under $20.
I also tried the various open source OCR tools, but the error rate was way too high. Obivious faults like l or I in place of 1 in the middle of numbers. AWS Textract was pretty good with numbers, although it did stutter on the dates in a few invoices.
For most of the invoices I was able to self-check the results as account balances and totals from month to month had to match. That made the end results from messy scripts extracting to CSV reasonably trustworthy.
What do you mean by complex? You create a service, deploy a model out of the catalog, and grab the endpoint. Takes me 5 minutes, and I know that those API calls go to a dedicated tenant at the location I specify.
"The subscription is not registered to use namespace 'Microsoft.CognitiveServices'."
I don't know man, I'm sure it's simpler than I think, my biggest hurdle with Azure is that it's not clear to me what's included in my license, in the region I'm operating in etc. For example the AI Model Builder is not available to me, I don't know if cognitive services should be. I haven't taken any training tho, not complaining just expressing that it's intimidating.
AI has been a huge help in my daily life both work and home. I use it do debug, diagnose, search for information, write code, answer numerous questions I get from my kids, etc.
"AI can be wrong" - like that hasn't been a problem with humans since forever... Like with any information, you have to be diligent. I teach my kids this also.
Is it helpful to the point of a $30-$50 a month subscription?
And at a larger scale, is this help in general worth the energy and water expenditure needed?
Edit: to be clear the second point is not against the individual, but questioning whether the modest gains in convenience for all users are worth the environmental and infrastructure cost
Maybe it's because Google search quality has gone down, but I've had plenty of cases where ChatGPT 4.0 got me the exact answer that I wanted but Google didn't.
And then there's the tasks where Google just doesn't play.
Things like this:
"Write the code to parse the following line with PyParser: { blah=0x10, foo=0x20, l={1,4,5,7}, t = { a="hello", b="hi"} }"
Google just gives me a link to PyParser (yay!), ChatGPT provides the full code and even runs it to show that it works.
I've been using Claude for the same types of stuff. It's definitely save me with some debugging and code. I still think it's a pretty over-hyped. The entire industry doesn't need to be working on this stuff.
Helpful and useful? Yes. A huge help? Game changing? I would say no.
The "why" for AI right now is because everyone else is doing it and you can't risk not having a critical pillar needed to compete.
I think AI is a loser for everyone but Nvidia for a while because of this. Google, Meta, Apple, Microsoft, etc will all be forced to spend billions to attempt AI but it will be years before it is good enough to generate significant profits. Even then, there may only be 1-2 winners, like Google was with traditional search.
There's also an interesting dynamic: costs scale exponentially (in a handwavey way) with quality. If you're willing to stay a half generation behind, you can create models a whole lot cheaper than the frontier labs. So Meta can always put significant price pressure on Google/OAI/Anthropic for relatively cheap.
The only way I see AI making anyone trillions is if someone cracks the nut of true AGI and somehow manages to keep it under wraps.
I don't think this question is as obvious as everyone else clearly thinks it is.
But like here's the progression:
1. How does AI currently make money? (the answer is it doesn't, but like what's the world we think exists where a chat bot is monetizeable?)
2. What does AGI need to look like where it can actually solve many million dollar problems?
3. What million dollar problems does it solve?
If the answers to this are as easy as they sound, then the responses to this should be easy to churn out, but instead everyone has shortcut that AGI == $$$, without showing the logic path.
a response could look like:
AGI solves P != NP -> $1M
AGI solves customer service -> $100B
AGI solves ??? -> $1B
I'm making up ideas and numbers here, but this doesn't seem like that weird of an assumption to test given how everyone on here just takes it as gospel.
While AI from MSFT won’t be as ground breaking as hyped now, it still will be big enhancement for Office and Cloud offerings. Microsoft providing Word/Excel/Powerpoint with existing capabilities to big companies is going to be long run winner.
NVidia is going to be always just like all hardware providers strong but still not in a place like MSFT having governments and F1000 corporations basically paying MSFT all the money they want.
Well, I strongly believe there are already areas where LLMs' ability to flexibly transform e.g. textual content can bring lots of business value, e.g. in data reporting and analytics, see:
> The main signals on Search have been AI/ML for >5 years now.
Say more? "Passing the entire web through ML-model inference to generate 'signals' — IQL expressions? — at the same global-concurrent throughput as previous CPU-based indexing" sounds like something that should have required Google to get so many TPUs fabbed that it would have eaten the entire world's chip-fab capacity for multiple years. It's hard to imagine what they could have done to get around that.
Did Google come up with some kind of model-hardcoded ASICs that could do this single task orders-of-magnitude more efficiently than more "flexible" GPU/TPU-based approaches would? Maybe, for example, they used some process to convert the weights of an arbitrary Transformer into VLSI for a search memory / memristor network / etc — essentially giving them an inferencing DSP?
Or maybe they figured out a mostly-lossless method (that either doesn't use ML itself — or maybe uses ML hyper-optimized into some CPU-viable model that fits in L2 cache) to pre-process + normalize webpage documents into chunks that could be effectively content-deduplicated? (And then the ensuing inference passes — being context-free — would only ever need to be done once for a given chunk, significantly reducing the inter-page costs, and massively reducing the same-page over-time costs, of inference.)
Or something else I can't even think of. (I'm guessing it's this one.) Either way, crazy stuff.
> help keep costs from inflating wildly out of control
Aren't the majority of costs in feeding and training the model? If they hope to present a product that has access to pertinent and timely information I'm not sure how they're going to drastically reduce their costs on queries alone.
> still lacking to see the "why"
The "why" for AI should be blindingly obvious. The fact that this is an open question suggests that these simplistic large language models are still several orders of magnitude away from what that endpoint would be.
This is a tiny stepping stone on a very long journey. It was foolish and insane to invest in it at the level Microsoft did.
Now that AI means LLM specifically I guess you are right, but if you'd told me 20 years ago that we'd have cars driving themselves around major cities by now, I'd have been pretty impressed.
I heard about the self-flying helicopter at CMU and robots playing soccer, but really finishing the science and engineering to have computers mundanely outperforming humans at safe city driving is remarkable. Except of course to current industry observers & wall street analysts, who are asking "will this ML mumbo jumbo ever amount to anything practical? Or is lane keeping with human vigilance and alarm-setting the pinnacle?"
> As with all AI so far, still lacking to see the "why"
Microsoft poured god-knows-how-many billions into Bing with barely anything to show for it. They hover just below 4% market share. AI is their one chance to actually upset the Google monopoly. But only if they manage to pull it off before Google, hence all their mediocre rushed-to-market AI stuff.
Of course the real money with current AI is elsewhere. But replacing first-line customer support with AI isn't the kind of story you will see promoted by PR departments.
So as usual, Microsoft is innovating while Apple just sells what everybody else has already tested (maybe even claiming the fatherhood of he idea). Nothing new.
I'd settle for a functional Alexa at this point. Or maybe a mobile keyboard that doesn't constantly make me typo. Surely applied AI can at least improve these areas of my life within the next few years.
There is basically no conversational AI. There is something that can understand your voice--given a fairly conventional American accent. But nothing that can do anything along the lines of a pretty dim-witted executive assistant.
Could we say that we've reached the "Peak of inflated expectations" on the Gartner hype curve for AI and we're softly entering the "Trough of disillusionment"?
Hype is not actually a thing. Hype comes into existence in environments where skepticism and honest engineering have been intentionally removed and obscured. To the extent that there's a "hype cycle" there's probably actually just a "journalism crisis."
I wouldn't blame them. I would blame the people who _own_ those outlets. As they're the ones with an actual interest in obscuring the truth.
> spend a lot of money marketing their newfangled toys
There are lots of short term ideas in the market place. None of them are particularly novel. So, it should be surprising, particularly in the era of the internet, that they still work.
The rate of improvement of AI capability is remarkably predictable, while public perception and investor sentiment wildly overshoots and undershoots depending on the animal spirits of the population.
Back when GPT-2 was released, you could draw an exponential curve to fit the rate of capability increase in AI over the prior five years. If you extend that curve to today, frontier models are still on that same curve.
That underlying rate of improvement isn't useful for predicting if Nvidia will be up or down next week, but it's a pretty strong signal for how things might play out longer term.
At some point the exponential growth portion of the S-curve will exhaust itself, but there's no hint of that yet.
That's my feeling too. I'd imagine investors are feeling disappointed after billions have poured in and the returns haven't been as amazing as expected.
AI does have some genuinely interesting potential applications and we have already seen it disrupt certain fields, it's just a matter of time before those applications persuade investors to come back.
peak hype is already over. when people were scared of ai ending humanity or taking all jobs was peek hype. people are already over it and have no expectations besides what we got.
unless some new technology or advances or killer feature is made. this hype wont be reached again with this eras ai. i mean even then i dont think itll reach the peek that just passed its more human nature then anything else
Probably. Doesn't mean there's nothing there but probably means there's an excess of hype. Not a fan of the hype cycle as a research instrument but it does represent the cycle a lot of technologies go through.
I pay for ChatGPT pro subscription and I use it far more than Google these days. I believe OpenAI can unseat Google but it's going to take far longer than one quarter.
I paid for it for around 3 months. The free version is enough for the rare times I actually use it, and many times I tried to use it it was down, so I switched to Gemini. Google is the same if not better here.
I think the short term business case for LLMs is failing to materialize. It is close, but not quite ready for prime time. I think of the promise of internet commerce in 1996. Everyone understood the potential, but we were still a few years away from the rubber meeting the road.
Can't speak for OP but, in '96 Netscape had just had a wildly successful IPO and talk of the Internet and World Wide Web was everywhere. It was kind-of obvious to anyone who paid attention where it was going, especially e-commerce, but, the technology wasn't quite there yet and a critical mass of users hadn't yet been reached.
Over the next 4-5 years, investors threw massive amounts of money at any company with ".com" in their name, regardless of whether they had a single customer. Companies who were offering little more than a single feature were raising VC funding.
It was a fun time to be in tech but the bubble bursting around 2000 was painful.
During that time, the web was mostly brochure websites. No revenue was generated from these sites as they were purely informational.
"Web 2.0" was born during this time (maybe a little later) and introduced a paradigm shift where the Internet was becoming more commercial. It took some time for things like merchant accounts to run credit cards to become available for Internet usage, which was huge since fraud was a major consideration with Internet shopping. There was also none, or very few, SaaS eCommerce platforms like Shopify, so eCommerce was mostly a roll-your-own or use a new open source eCommerce platform that was in its infancy. It wasn't until the mid 2000s that popular self hosted platforms like OpenCart, Magento, PrestaShop, WooCommerc, etc, were available and mostly feature complete for most customers needs back then.
In 1996, the technology was good enough that you could see the future very clearly, but it still was very clunky from a consumer and enterprise point of view.
Credit card transactions were extremely difficult to execute and wildly insecure. If there was one tiny formatting error in your address, you would have to refill the entire form again (before the days of autocomplete).
Websites might takes three or four minutes to load. Often they wouldn't load at all, or components would be missing.
Often your home internet connection would just stop working for minutes or hours at a time for no reason.
However, you could see that things were moving in a certain direction, and stuff like Amazon made complete sense already (even if it didn't work that well for 95% of people).
I heard companies already let go awful lot amount of copywriters. All the cookie cutter texts can be generated rather well and cheap. Also correcting typos and grammar is getting rather cheap. While llms can imagine things if yo ask about something - they are still rather great when it come to improving text.
This is true, but a lot of agencies are rethinking this.
Search engines and social media platforms very explicitly de-rank content that is believed to be AI generated. Also, most companies have found that editing/screening AI generated text takes longer than just producing it themselves.
Great, their direction with Windows has been irritating enough that I've left and don't plan to come back to it. I can't understand the perspective of whatever executive decided to shoehorn AI features into Windows in the way they did.
Meanwhile, the overall results don't make it seem like the sky is falling
> For the quarter, Microsoft reported earnings per share (EPS) of $2.95 on revenue of $64.7 billion. Wall Street was anticipating EPS of $2.94 on revenue of $64.5 billion, according to data compiled by Bloomberg.
86 comments
[ 2.0 ms ] story [ 145 ms ] threadAs with all AI so far, still lacking to see the "why". Microsoft came out hard, but still don't see why other than generating potentially very wrong information or goofy looking pictures.
AI could turn out the same. You need good training data, but the biggest set out there (Internet) was quickly polluted to hell by AI and now it’s far less valuable for training.
I’m skeptical of its long term value for the average user.
Power Automate offers an Adobe connector that purports to do so but my data loss prevention policies forbid me to connect to third parties.
GPT4o does a great job of taking a PDF and giving me exactly the contents in a csv of my specification, but again, can't do this with customer data.
I'm looking at what is possible with llama3 as an azure service but it looks very complex to set up, I think I need to have a "data verse" and a network configuration to allow serverless compute to connect to my files. That will be nice tho when I can have the results of the PDF extraction insert rows into my excel files.
I don't use Salesforce but I'm jealous of people who have a platform that they can actually integrate AI with their customer data.
I also tried the various open source OCR tools, but the error rate was way too high. Obivious faults like l or I in place of 1 in the middle of numbers. AWS Textract was pretty good with numbers, although it did stutter on the dates in a few invoices.
For most of the invoices I was able to self-check the results as account balances and totals from month to month had to match. That made the end results from messy scripts extracting to CSV reasonably trustworthy.
I don't know man, I'm sure it's simpler than I think, my biggest hurdle with Azure is that it's not clear to me what's included in my license, in the region I'm operating in etc. For example the AI Model Builder is not available to me, I don't know if cognitive services should be. I haven't taken any training tho, not complaining just expressing that it's intimidating.
That's not really Apple's forte.
My crystal ball isn't working, but I wouldn't bet on Apple.
"AI can be wrong" - like that hasn't been a problem with humans since forever... Like with any information, you have to be diligent. I teach my kids this also.
debug -- cursor.sh for coding/debugging with your codebase as context. also just raw inputting into claude is pretty effective.
in terms of consumer use for it to just search for info, and write code is a total failure. i mean by their words really.
we are basically told it will be integrated in dang near everything and uses electricity
And at a larger scale, is this help in general worth the energy and water expenditure needed?
Edit: to be clear the second point is not against the individual, but questioning whether the modest gains in convenience for all users are worth the environmental and infrastructure cost
And then there's the tasks where Google just doesn't play.
Things like this: "Write the code to parse the following line with PyParser: { blah=0x10, foo=0x20, l={1,4,5,7}, t = { a="hello", b="hi"} }"
Google just gives me a link to PyParser (yay!), ChatGPT provides the full code and even runs it to show that it works.
Helpful and useful? Yes. A huge help? Game changing? I would say no.
I think AI is a loser for everyone but Nvidia for a while because of this. Google, Meta, Apple, Microsoft, etc will all be forced to spend billions to attempt AI but it will be years before it is good enough to generate significant profits. Even then, there may only be 1-2 winners, like Google was with traditional search.
The only way I see AI making anyone trillions is if someone cracks the nut of true AGI and somehow manages to keep it under wraps.
What makes someone trillions of dollars from AGI?
solving many million dollar problems.
But like here's the progression: 1. How does AI currently make money? (the answer is it doesn't, but like what's the world we think exists where a chat bot is monetizeable?) 2. What does AGI need to look like where it can actually solve many million dollar problems? 3. What million dollar problems does it solve?
If the answers to this are as easy as they sound, then the responses to this should be easy to churn out, but instead everyone has shortcut that AGI == $$$, without showing the logic path.
a response could look like: AGI solves P != NP -> $1M AGI solves customer service -> $100B AGI solves ??? -> $1B
I'm making up ideas and numbers here, but this doesn't seem like that weird of an assumption to test given how everyone on here just takes it as gospel.
While AI from MSFT won’t be as ground breaking as hyped now, it still will be big enhancement for Office and Cloud offerings. Microsoft providing Word/Excel/Powerpoint with existing capabilities to big companies is going to be long run winner.
NVidia is going to be always just like all hardware providers strong but still not in a place like MSFT having governments and F1000 corporations basically paying MSFT all the money they want.
https://www.linkedin.com/feed/update/urn:li:activity:7224046...
Basically the same "why" as the Dot Com stuff: you have to sell pet food online because everyone else is selling stuff online.
At some point the hype will subside and we'll get the actual useful day-to-day applications of whatever technology is in vogue right now.
Tech hype dates back to (at least) canal mania:
* https://en.wikipedia.org/wiki/Technological_Revolutions_and_...
* https://www.pwlcapital.com/investing-technological-revolutio...
The main signals on Search have been AI/ML for >5 years now.
At least for Google, AI isn't just a parlor trick.
Say more? "Passing the entire web through ML-model inference to generate 'signals' — IQL expressions? — at the same global-concurrent throughput as previous CPU-based indexing" sounds like something that should have required Google to get so many TPUs fabbed that it would have eaten the entire world's chip-fab capacity for multiple years. It's hard to imagine what they could have done to get around that.
Did Google come up with some kind of model-hardcoded ASICs that could do this single task orders-of-magnitude more efficiently than more "flexible" GPU/TPU-based approaches would? Maybe, for example, they used some process to convert the weights of an arbitrary Transformer into VLSI for a search memory / memristor network / etc — essentially giving them an inferencing DSP?
Or maybe they figured out a mostly-lossless method (that either doesn't use ML itself — or maybe uses ML hyper-optimized into some CPU-viable model that fits in L2 cache) to pre-process + normalize webpage documents into chunks that could be effectively content-deduplicated? (And then the ensuing inference passes — being context-free — would only ever need to be done once for a given chunk, significantly reducing the inter-page costs, and massively reducing the same-page over-time costs, of inference.)
Or something else I can't even think of. (I'm guessing it's this one.) Either way, crazy stuff.
Aren't the majority of costs in feeding and training the model? If they hope to present a product that has access to pertinent and timely information I'm not sure how they're going to drastically reduce their costs on queries alone.
> still lacking to see the "why"
The "why" for AI should be blindingly obvious. The fact that this is an open question suggests that these simplistic large language models are still several orders of magnitude away from what that endpoint would be.
This is a tiny stepping stone on a very long journey. It was foolish and insane to invest in it at the level Microsoft did.
I heard about the self-flying helicopter at CMU and robots playing soccer, but really finishing the science and engineering to have computers mundanely outperforming humans at safe city driving is remarkable. Except of course to current industry observers & wall street analysts, who are asking "will this ML mumbo jumbo ever amount to anything practical? Or is lane keeping with human vigilance and alarm-setting the pinnacle?"
Do we have this?
Bunch of stuff in maths/sciense I solved with lookup tables and it doesn’t take away that rest of solution is still complex.
Microsoft poured god-knows-how-many billions into Bing with barely anything to show for it. They hover just below 4% market share. AI is their one chance to actually upset the Google monopoly. But only if they manage to pull it off before Google, hence all their mediocre rushed-to-market AI stuff.
Of course the real money with current AI is elsewhere. But replacing first-line customer support with AI isn't the kind of story you will see promoted by PR departments.
I.don't.understand.what.you're.talking.about.here?
NVDA was down during regular trading, and is actually mostly back up in after-hours trading.
https://en.m.wikipedia.org/wiki/Gartner_hype_cycle
I wouldn't blame them. I would blame the people who _own_ those outlets. As they're the ones with an actual interest in obscuring the truth.
> spend a lot of money marketing their newfangled toys
There are lots of short term ideas in the market place. None of them are particularly novel. So, it should be surprising, particularly in the era of the internet, that they still work.
Back when GPT-2 was released, you could draw an exponential curve to fit the rate of capability increase in AI over the prior five years. If you extend that curve to today, frontier models are still on that same curve.
That underlying rate of improvement isn't useful for predicting if Nvidia will be up or down next week, but it's a pretty strong signal for how things might play out longer term.
At some point the exponential growth portion of the S-curve will exhaust itself, but there's no hint of that yet.
So are the training costs.
> you could draw an exponential curve
Once you plot the costs on that curve it becomes obvious that this technology, by itself, is inadequate to achieve the goals stated for it.
> growth portion of the S-curve will exhaust itself
So will all the available entropy in the universe. Single ended analyses like this are a cornerstone of the hype machine.
AI does have some genuinely interesting potential applications and we have already seen it disrupt certain fields, it's just a matter of time before those applications persuade investors to come back.
unless some new technology or advances or killer feature is made. this hype wont be reached again with this eras ai. i mean even then i dont think itll reach the peek that just passed its more human nature then anything else
OpenAI has no moat.
Over the next 4-5 years, investors threw massive amounts of money at any company with ".com" in their name, regardless of whether they had a single customer. Companies who were offering little more than a single feature were raising VC funding.
It was a fun time to be in tech but the bubble bursting around 2000 was painful.
"Web 2.0" was born during this time (maybe a little later) and introduced a paradigm shift where the Internet was becoming more commercial. It took some time for things like merchant accounts to run credit cards to become available for Internet usage, which was huge since fraud was a major consideration with Internet shopping. There was also none, or very few, SaaS eCommerce platforms like Shopify, so eCommerce was mostly a roll-your-own or use a new open source eCommerce platform that was in its infancy. It wasn't until the mid 2000s that popular self hosted platforms like OpenCart, Magento, PrestaShop, WooCommerc, etc, were available and mostly feature complete for most customers needs back then.
If I recall correctly, this train of thought lasted somewhat into 1992 and then very quickly disappeared entirely.
[0] https://en.wikipedia.org/wiki/Timeline_of_online_advertising
Credit card transactions were extremely difficult to execute and wildly insecure. If there was one tiny formatting error in your address, you would have to refill the entire form again (before the days of autocomplete).
Websites might takes three or four minutes to load. Often they wouldn't load at all, or components would be missing.
Often your home internet connection would just stop working for minutes or hours at a time for no reason.
However, you could see that things were moving in a certain direction, and stuff like Amazon made complete sense already (even if it didn't work that well for 95% of people).
Search engines and social media platforms very explicitly de-rank content that is believed to be AI generated. Also, most companies have found that editing/screening AI generated text takes longer than just producing it themselves.
> For the quarter, Microsoft reported earnings per share (EPS) of $2.95 on revenue of $64.7 billion. Wall Street was anticipating EPS of $2.94 on revenue of $64.5 billion, according to data compiled by Bloomberg.
> $28.5 billion versus expectations of $28.7 billion.
At the scale of MS, that is a dozen or so missed large contracts out of thousands they likely signed. It is hardly the end of the world.
I am more curious about how they are going to get away with constantly giving copilot branded AI features away as part of the core OS.