> OpenAI considered building everything in-house and raising capital for an expensive plan to build a network of factories known as "foundries" for chip manufacturing. The company has dropped the ambitious foundry plans for now due to the costs and time needed to build a network
That framing massively undersells how insane Sams ambitions were there, he was floating the idea of somehow raising seven trillion dollars to build thirty six fabs dedicated to making AI silicon. The TSMC execs reported more or less laughed in his face when he brought it up it them.
And if it was a 30% stake in OpenAI for that $7 trillion (kind of a standard VC round percentage) that would put OpenAI's valuation at about the same of all of the 7000-ish NASDAQ companies (including almost all public tech companies) combined.
There was probably a point of maximum hype ~12 months ago, or right after the launch of GPT-4 where the belief of imminent singularity was running high.
Considering that what is currently passed as AI is really just statical probability of string words together that is completely unable to apply logic or reason… we ain't even close.
I like the three comments above this one all seem to massively disagree or at least give off argument energy, but they all seem totally correct and don't directly contradict each other.
Hype cycles ride on real progress that’s first blown out of proportion, then under-appreciated. LLMs are still magic AF. We are just so tired of hearing about them and how they’re going to change everything.
There's no dark budget, it's all there, the programs are just usually listed as "Classified" or "Special Activities" like NGA funding was in the late 90s.
Dark budget? They can't even supply Ukraine with enough ammo and weapons to defend themselves and the DoD is already squealing that they are running out of everything. Dark budget..hahahaha...good one....
It seems totally possible that enough interest groups don’t want Ukraine to actually win, so they block anything more than the minimum viable amount. So it’s not much of a barometer at all for anything…
Yes, but this theory requires more and more epicycles.
There's probably some small-ish 'dark budget'. But think about it: the official budget is in the order of single digit percentage points of US GDP. That represents and enormous amount of resources. Any appreciable fraction of that in a 'dark budget' would also leave traces all over the place.
The world is vastly more complex than say the world of 1945, so almost certainly there are now likely several dozens to several hundreds of ‘epicycles’ that are in fact true and necessary for full understanding.
Now the pentagon doesn’t contain the entire complexity of mankind within it… but it’s probably not that far off.
But it’s true, if that’s the case, that approximately 100% of the HN readerbase will never penetrate more than say the first 100 in their lifetimes so it does seem like a moot point…
He had AGI. He still has it. You can have it too, for a $20/month subscription from OpenAI or Anthropic.
It just turns out that AGI (artificial general intelligence, which ChatGPT objectively is), is not a singularity nor even a tipping point towards causing one.
Yeah, I mean, what else would you do with those chips? Run LLMs faster? Sorry but the killer app is still not there, the majority of people do not care and have no use for them.
There are a significant number of uses, it's just most of them involve generating slop for the purposes of fraud or feigning work.
I thought it was really damning how Apple's recent ads involved a celebrity using it to respond to their agent about a script proposal they didn't actually read.
So somebody in another country with another language and another currency allegedly said… and came across several people and translations (did they translate currency?) before landing at a journalists inbox. Might as well as have heard about it on Twitter.
There’s a 68% chance that Altman met with someone who got their degree in the US (if only one executive attended, otherwise chances are higher).
And the others probably also have little difficulty communicating in English or calculating the cost of 36 fabs in any currency, since that’s their business.
I very much doubt the CEO of TSMC spoke to the NYT about a private meeting they had with a potential customer. Give me a break!
Much more likely they talked about 3-6 fabs, some other person laughed about 36 fabs for podcast guy, someone else overheard that and posted on Weibo. That made it to NYT eventually.
Sam explicitly commented on this claim. He said we’ll _collectively_, _eventually_ need to spend 7 trillion dollars on compute to meet demand. Not OpenAI raising 7 trillion to build all the capacity itself.
<rant> This isn’t about you specifically bro but what is with most hn users. None of you got rich from crypto and then none of you think AI can be used to make money. I literally asked a shit AI how to make money and I was making ROI. I sometimes think I’m born on a different platform. ( edit I meant to say planet but this isn’t a better typo) </rant>
and fake it before you make it ...im still waiting for GPT's version of H.E.R.... many months later nowhere to be found yet no doubt that juiced their revenue.
The startup playbook lie ur ass off and make sh!t up. Musk still does it to this day with the recent demo of his Optimus robots implying they were all AI driven.
why that's what many start-uppers do to win the startup game ..fake it .. lie. Would you trust your life to Elon's self driving tech which you sign away your rights to sue lol
Some with deadly consequences like Uber's attempt at self driving and what was that other recent self driving company that ran over / mangled a pedestrian?
Yes but in this case it was clearly apparent to most and to those who spoke with the robot and believed it must have felt like morons for being filmed thinking they were real. I just felt the benefits weren’t there as it was an obvious fakery.
Tbh when they had the real person on the suit I wouldn’t have approved that it would give the impression I’m ok with pretending to an extent.
I understand having a half working prototype iPhone on stage but literally pretending the robots are real just felt dishonest and a bad idea.
Hmmm, perhaps there's a sweet spot where someone is nuts enough that you think you can take over the stuff for cheap when they crash and burn, but not so nuts that there's nothing left to pick up.
Along with the grandiose promotional nutty-ness of "You need to support my company because countries will fight wars over AI", whatever that FOMO scarcity is supposed to mean in practice.
It could have been strategy instead of insanity. By starting conversations at $7T you anchor high and potentially drive a greater outcome than starting an order of magnitude lower.
That strategy only works if the anchor is in the realm of reality. If I'm selling a 20 year old Toyota Corolla and initially ask for $900,000 that's not going to help me get a higher price.
Anchoring bias probably works with numbers outside of the realm of reality as well. But I doubt it's very useful in these cases (or even any stronger than asking for say 300B), otherwise everyone would always start a funding round by asking for a 7T investment, right.
Usually when doing anchoring you want to end up at a result less than what you originally asked for but crucially more than zero, and OpenAI immediately folded on building any fabs whatsoever, so I don't think it worked.
When you do something that stupid - starting at $7 trillion - you end the conversation before it really begins because you lose all credibility with the people that matter (eg TSMC and other investors).
If he had said $250 billion and six fabs, it would have been a lot to ask but people wouldn't think he was ignorant or irrational for saying it. Big tech for example has that kind of money to throw around spread out across a decade if the investment is a truly great opportunity.
I guess he thinks his glorified markov chain will lead to ASI if scaled up sufficiently. Even if we get ASI, the likelihood that anybody will ever make any money from it is so delusional. This isn't going to be your average brainwashed peasant, crushing these capitalist pigs is probably the first thing it's gonna do.
The technical term for this is “zone of possible agreement”. If your opening is outside the counterparty’s ZOPA, then they’re likely to disengage/exit.
> The TSMC execs reported more or less laughed in his face when he brought it up it them.
Reminds me of:
Musk flew there with Cantrell, prepared to purchase three ICBMs for $21 million. But to Musk's disappointment, the Russians now claimed that they wanted $21 million for each rocket, and then taunted the future SpaceX founder. As Cantrell recounted to Esquire: “They said, 'Oh, little boy, you don't have the money?”
If you don't believe ASI is possible, 7T is way too high. If you believe ASI is possible, 7T is wayyy too low, thus 7T is a compromise if you believe there's a non-certain chance of ASI being possible
So the question is, if you have AGI/ASI how much productive work could you get done with it?
When looking at things like mechanization and the productivity increases from around 1880 to now, you took an economy from around 10 billion a year to 14 trillion. This involved mechanization and digitization. We live in a world that someone from 1880 really couldn't imagine.
What I don't know how to answer (or at least search properly) is how much investment this took over that 150 year period. I'm going to assume it's vastly more than $7 trillion. If $7 trillion in investment and manufacturing allowed us to produce human level+ AI then the economic benefits of this would be in the tend to hundreds of trillions.
Now, this isn't saying it would be good for me and you personally, but the capability to apply intelligence to problems and provide solutions would grow very rapidly and dramatically change the world we live in now.
> If $7 trillion in investment and manufacturing allowed us to produce human level+ AI then the economic benefits of this would be in the tend to hundreds of trillions.
Sure, but right now that "if" is trying to do 7 trillion dollars of unsubstantiated heavy lifting. I might be able to start creating Iron Man-style arc reactors in a cave with a box of scraps and all I ask is 1 trillion, so you all should invest given how much money unlimited free energy is worth.
How do you even define ASI? How do you know it will run on silicon? If it will run on silicon, what kind of silicon? GPU? TPU? Binary? Ternary? Von-Neumann architecture? How can you even start to build chips for something completely imaginary?
ASI is just AGI but scaled to the point of superhuman ability in all cognitive tasks. Surely it will run on the architectures used to build AGI before ASI takes control itself in the design and deployment.
> ASI is just AGI but scaled [...] Surely it will run on the architectures used to build AGI
The words "just" and "surely" are doing so much heavy lifting here that I'm worried about labor-laws. :p
If scaling up was "just" that easy and the next step would "surely" begin in the exact same architecture... Well, you've gotta ask why several billion general-intelligence units made from the finest possible nanotechnology haven't already merged into a super-intelligence. Or whether it happened, and we just didn't notice.
Not surprising at all. As a former Google employee, I'm regularly in contact with friends still at Google and the number of hardware engineers OpenAI has hired out of Google have drastically increased in the last few months. You can readily see for yourself just by browsing LinkedIn and looking at profiles of hardware engineers currently at OpenAI and noticing how many of them are former Google employees. Google's own TPU team is now a shell of its former self. I wouldn't be surprised if future Google TPUs are inferior to the new OpenAI chips.
Also, did you know that Google's TPU efforts previously also relied on help from Broadcom? https://www.theregister.com/AMP/2023/09/22/google_broadcom_t... It's absolutely unsurprising that former Google employees would bring this vendor relationship into their new company, given Broadcom did a good job helping Google with TPUs.
It's also worth mentioning that google spent a lot of time (and probably money) and human capital on their TPUs and slowly got there. Other companies that have a proven track record of hyper-scaling also tried and missed (tesla) so it's not as easy as locking a bunch of phds in a room and have them crank out LLMs. Hardware is finicky, slow to develop, slow to prove itself and extremely expensive.
"while adding AMD (AMD.O), opens new tab chips alongside Nvidia (NVDA.O), opens new tab chips to meet its surging infrastructure demands"
Jensen has been saying that demand is "insane" and we're hearing rumors of low yields. This equates to supply issues in the coming months/years. No fortune 500 puts all their eggs into one basket. Diversifying away from a single source for all of AI hardware and software, is a smart thing to do.
> Diversifying away from a single source for all of AI hardware and software, is a smart thing to do.
I wonder how this squares with the exclusivity contract with Microsoft. Even the OpenAI/Oracle deal requires Oracle to run Azure stack on their datacenter so MSFT can mediate the relationship. The AMD chips mentioned are also purchased by MSFT.
I wonder if this really means that OpenAI is accepting the risk/capital expense while providing a variety of hardware to Microsoft, or if there are other terms at play.
That is exactly what the execs at my company are telling us when asked about not using Nvidia -- diversifying away. It's funny though because we have no Nvidia for training at all. We use Trainium because we could not get our hands on Nvidia.
I think I get what you're saying. That a datacenter probably just wants to focus on one product at a time. You're right, it is simpler. On the other hand, that's part of why neoclouds are sprouting up. Users are asking for the optionality.
So presumably they're building their own data centers to use these chips, but odd that the article doesn't actually say so.
Are any of the western AI/matmul chips NOT being fabbed by TSMC? Amazon or Meta's perhaps? We've got NVIDIA, AMD, Google (TPU - also designed by Broadcom), and now OpenAI all using them. Are Samsung not competitive?
I wonder to what extent total chip/chipset volume is limited by TSMC capacity as opposed to memory or packaging?
Samsung is not competitive so everything good has to be made by TSMC now. It has been reported that packaging and HBM are the bottlenecks for AI chips although TSMC is also running at high utilization.
AMD hardware is good. The software is getting better daily. Training still sucks (mostly due to unoptimized libraries), but inference is looking pretty decent with recent advances in tuning...
Other than GH Copilot (which is amazing) and the like, I am yet to see a social/cultural impact of LLMs. I am ignoring the hype cycle fueled social media posts where influencers show cool stuff or model releases. I am talking about this stuff permeating through the society.. am I immune to it? Or is it hidden in saas/enterprise products layers deep?
I would imagine some productive cultural impact that’s wide spread, sticky, high retention not like ‘oh cool 3d tv or vr’ kinda thing?
Given how much ‘money’ has been burnt through the GPU heatsinks.
It’s kinda easy to check it though. Ask your friends if they use it, preferably non developer. When I find out my accounting friends and marketing friends are using it for basic analytics/promotion write ups I realized this is being used by non developers already on a regular basis. The use by non tech folks is significant for this . Crypto as a counter example of hype hasn’t lived up to its purported use case as non tech ppl I know only “own” it through an exchange.
I expect they wanted to go down the fab and chip design path to build a moat, as they've realised they really don't have one outside of brand recognition.
If they had proprietary fabs and chips that were magically superior, and they could build them for cheap rather than paying the team green tax, they'd have a huge advantage in inference cost.
Perhaps they've realised just starting with a proprietary chip is enough?
Spending $7 trillion on in-house fabs sounded both ambitious and crazy. Reality finally kicked in. If they’re done dreaming big, let’s hope they keep the quality
Broadcom is a vendor. They provide the main skeleton of a chip so that the client only needs to focus on the product differentiator, which is the compute part of the chip. They can worry less about other parts that are not their core competency.
I'm not saying that I necessarily agree, but the general consensus on HN seems to be that Broadcom is now less of a tech company and more of a holding company that raids others, for example VMWare, and extracts all value to the detriment of customers and the acquired company.
I don't think that's completely wrong, but it's a big company and I'm sure there are some better areas of the company than others.
I think it's completely wrong in the context of its role as an ASIC partner. There's a very short list of companies with all the IP needed for these cutting-edge ASICs and Broadcom/Avago might be the best of them. And to be clear, they've developed that IP themselves, just as they've always done. Those that think they're just a "holding company" haven't actually worked with them.
Additionally, having worked with some of their network devices at the driver level, they seem to be kludge piled on top of hack poured over a soup of workarounds for hardware bugs. Maybe they've gotten better recently, but just looking at their drivers, it didn't paint a great picture.
this makes sense since the general consensus on HN is that any general consensus on HN is right so if the general consensus on HN that any general consensus on HN is right is wrong then any general consensus on HN could be wrong
They make every kind of chip for networking: WiFi, Bluetooth, home routers, cable modems, fiber, switches, NICs, DPUs, etc. And that's 1/Nth of their business.
> OpenAI considered building everything in-house and raising capital for an expensive plan to build a network of factories known as "foundries" for chip manufacturing. The company has dropped the ambitious foundry plans for now due to the costs and time needed to build a network
That framing massively undersells how insane Sams ambitions were there, he was floating the idea of somehow raising seven trillion dollars to build thirty six fabs dedicated to making AI silicon. The TSMC execs reported more or less laughed in his face when he brought it up it them.
As anyone with 2 brain cells should have.
You don't just acquire $7T.
The ENTIRE US domestic Net Investment isn't even $1T: https://fred.stlouisfed.org/series/W790RC1Q027SBEA
Gross Fixed Capital Formation (Net Investment + Deprecation) isn't much more: https://fred.stlouisfed.org/series/NFIRSAXDCUSQ
Google, Apple, and Microsoft together don't even spend $100B on CapEx per year. And they're worth almost $10T put together.
Asking for $7T when you're a $100B company is so ridiculous it's beyond belief.
Sound like the beginning of Universal Paperclips
https://en.wikipedia.org/wiki/Universal_Paperclips
Time for this year’s play through!
See also
https://en.wikipedia.org/wiki/Benderama
And if it was a 30% stake in OpenAI for that $7 trillion (kind of a standard VC round percentage) that would put OpenAI's valuation at about the same of all of the 7000-ish NASDAQ companies (including almost all public tech companies) combined.
There was probably a point of maximum hype ~12 months ago, or right after the launch of GPT-4 where the belief of imminent singularity was running high.
But, singularity is still imminent. Right? Where on the spectrum of completely useless/no AI to singularity are we at?
Considering that what is currently passed as AI is really just statical probability of string words together that is completely unable to apply logic or reason… we ain't even close.
I don't know if we ever will get a singularity, but if we do the last comment before we're obliterated will be something along these lines.
I like the three comments above this one all seem to massively disagree or at least give off argument energy, but they all seem totally correct and don't directly contradict each other.
And atheists will speak similarly of the rapture before it happens. Repent!
It’s gonna require N more hype cycles (bearing in mind that usefulness is the expected endpoint of the hype cycle)
> (bearing in mind that usefulness is the expected endpoint of the hype cycle)
I think for the people doing it, the expected endpoint is more about raising the price of shares they hold until they've been sold.
Hype cycles ride on real progress that’s first blown out of proportion, then under-appreciated. LLMs are still magic AF. We are just so tired of hearing about them and how they’re going to change everything.
The singularity ain't real, buddy.
He would have to have AGI proof to ask for this unprecedented kind of investing money.
And even them it would have to be split during a decade or two. And even then.
What's the military budget of USA?
It doesn't matter what the military budget is.
Most of it is spent on personnel and operation.
You'd want to know only what the procurement is - which is ~$146B: https://en.wikipedia.org/wiki/Military_budget_of_the_United_...
And most of that is part of Gross Fixed Capital Formation already...
What is dark budget look like these days.
There's no dark budget, it's all there, the programs are just usually listed as "Classified" or "Special Activities" like NGA funding was in the late 90s.
Dark budget? They can't even supply Ukraine with enough ammo and weapons to defend themselves and the DoD is already squealing that they are running out of everything. Dark budget..hahahaha...good one....
How is that relevant?
It seems totally possible that enough interest groups don’t want Ukraine to actually win, so they block anything more than the minimum viable amount. So it’s not much of a barometer at all for anything…
Yes, but this theory requires more and more epicycles.
There's probably some small-ish 'dark budget'. But think about it: the official budget is in the order of single digit percentage points of US GDP. That represents and enormous amount of resources. Any appreciable fraction of that in a 'dark budget' would also leave traces all over the place.
Well of course?
The world is vastly more complex than say the world of 1945, so almost certainly there are now likely several dozens to several hundreds of ‘epicycles’ that are in fact true and necessary for full understanding.
Now the pentagon doesn’t contain the entire complexity of mankind within it… but it’s probably not that far off.
But it’s true, if that’s the case, that approximately 100% of the HN readerbase will never penetrate more than say the first 100 in their lifetimes so it does seem like a moot point…
Perhaps...
It’s funny how that bit only comes up when HN wants to use it to make a point.
He had AGI. He still has it. You can have it too, for a $20/month subscription from OpenAI or Anthropic.
It just turns out that AGI (artificial general intelligence, which ChatGPT objectively is), is not a singularity nor even a tipping point towards causing one.
>What's the military budget of USA?
Somewhere between $900B and $1T this year
Yeah, I mean, what else would you do with those chips? Run LLMs faster? Sorry but the killer app is still not there, the majority of people do not care and have no use for them.
You can always train more and faster, and run more experiments.
There are a significant number of uses, it's just most of them involve generating slop for the purposes of fraud or feigning work.
I thought it was really damning how Apple's recent ads involved a celebrity using it to respond to their agent about a script proposal they didn't actually read.
I just paid for some graphic design work. What I got back looked pretty, but was very clearly derivative of generative AI. I feel defrauded...
I wonder what proportion of students use them for homework etc. The coming generations of people will care less about writing style.
Don't you dare question the usefulness of gen ai on the hacker news. Many folks haven't caught up with reality just yet.
It's not quite that bad, at least not in comments.
I assume that the $7T ask was to try to spend a lot of time in the news cycle when he wanted to be there. Obviously it was an impossible ask.
Does anyone have a source for this “$7T” number?
I read it on the internet myself with “sources saying” but I think it’s BS.
I remember reading these tweets back then https://x.com/jimkxa/status/1758683405875871841 not a direct source but maybe of interest?
https://www.nytimes.com/2024/09/25/business/openai-plan-elec...
https://www.tomshardware.com/tech-industry/tsmc-execs-allege...
So somebody in another country with another language and another currency allegedly said… and came across several people and translations (did they translate currency?) before landing at a journalists inbox. Might as well as have heard about it on Twitter.
Have a look at the TSMC executives: https://www.tsmc.com/english/aboutTSMC/executives
There’s a 68% chance that Altman met with someone who got their degree in the US (if only one executive attended, otherwise chances are higher).
And the others probably also have little difficulty communicating in English or calculating the cost of 36 fabs in any currency, since that’s their business.
I very much doubt the CEO of TSMC spoke to the NYT about a private meeting they had with a potential customer. Give me a break!
Much more likely they talked about 3-6 fabs, some other person laughed about 36 fabs for podcast guy, someone else overheard that and posted on Weibo. That made it to NYT eventually.
>As anyone with 2 brain cells should have.
>You don't just acquire $7T.
Perhaps the goal was only 1-2 Trillion but you should start high. I would have asked for 10 Trillion.....
Sam explicitly commented on this claim. He said we’ll _collectively_, _eventually_ need to spend 7 trillion dollars on compute to meet demand. Not OpenAI raising 7 trillion to build all the capacity itself.
This is true and I'm sad to see the hacker news crowd blindly repeating falsehoods.
It's kind of admirable, though. If you start asking for $7T, only asking for $1T becomes quite reasonable ;-)
That only works if the initial request isn’t so bonkers that no one can trust subsequent ones.
True. I think Sam might get a "bonkers" pass because of his track record that the rest of us might not.
He’s exactly the person I wouldn’t give a pass to, because of his track record. I wouldn’t give a penny to a cryptocurrency scammer.
https://www.buzzfeednews.com/article/richardnieva/worldcoin-...
https://www.technologyreview.com/2022/04/06/1048981/worldcoi...
Especially not one whose business plan is “let’s invent fairytale technology then ask it how to make money”.
https://x.com/geepytee/status/1778859803735113950
If all of that sounds good to you, I know of a bridge you may be interested in purchasing.
<rant> This isn’t about you specifically bro but what is with most hn users. None of you got rich from crypto and then none of you think AI can be used to make money. I literally asked a shit AI how to make money and I was making ROI. I sometimes think I’m born on a different platform. ( edit I meant to say planet but this isn’t a better typo) </rant>
What track record does he have?
Shoot for the stars and you get to the moon.
and fake it before you make it ...im still waiting for GPT's version of H.E.R.... many months later nowhere to be found yet no doubt that juiced their revenue.
The startup playbook lie ur ass off and make sh!t up. Musk still does it to this day with the recent demo of his Optimus robots implying they were all AI driven.
I have to say I’m really weirded out by the decision to sort of pretend they weren’t remotely controlled
why that's what many start-uppers do to win the startup game ..fake it .. lie. Would you trust your life to Elon's self driving tech which you sign away your rights to sue lol
Some with deadly consequences like Uber's attempt at self driving and what was that other recent self driving company that ran over / mangled a pedestrian?
Yes but in this case it was clearly apparent to most and to those who spoke with the robot and believed it must have felt like morons for being filmed thinking they were real. I just felt the benefits weren’t there as it was an obvious fakery.
Tbh when they had the real person on the suit I wouldn’t have approved that it would give the impression I’m ok with pretending to an extent.
I understand having a half working prototype iPhone on stage but literally pretending the robots are real just felt dishonest and a bad idea.
But hey I’m not a Vc
I would say the less morality you have and hold onto and or care about the more successful you will be at tech start-ups.
Only if they stop laughing long enough for them to hear your second offer
Well they did stop laughing, hence this thread.
And to think Satya Nadella looked at this guy and his plan and said "this is the kind of stable genius we need running OpenAI."
Or he’s the perfect guy to give MS freedom to dig their claws in.
Hmmm, perhaps there's a sweet spot where someone is nuts enough that you think you can take over the stuff for cheap when they crash and burn, but not so nuts that there's nothing left to pick up.
Along with the grandiose promotional nutty-ness of "You need to support my company because countries will fight wars over AI", whatever that FOMO scarcity is supposed to mean in practice.
It’s just 1/5th of the current USA national debt of $35T, bro. Just have fed run those money printers 365/24/7.
Maybe I'm just jaded, but I think even one fab deserves a laugh. Design, sure. Fabrication? Hah.
> The TSMC execs reported more or less laughed in his face when he brought it up it them.
Source?
Make no small plans ... Make bombastically ridiculous ones?
It could have been strategy instead of insanity. By starting conversations at $7T you anchor high and potentially drive a greater outcome than starting an order of magnitude lower.
That strategy only works if the anchor is in the realm of reality. If I'm selling a 20 year old Toyota Corolla and initially ask for $900,000 that's not going to help me get a higher price.
Anchoring bias probably works with numbers outside of the realm of reality as well. But I doubt it's very useful in these cases (or even any stronger than asking for say 300B), otherwise everyone would always start a funding round by asking for a 7T investment, right.
It's also probably less-effective when the other side has their own teams of accountants making projections.
there are millions of corollas there’s only one openai
It's become a running theme that there are in fact a lot of OpenAIs, with how frequently their models get matched or leapfrogged by competitors.
Not really that many fabs either. They do this on that scale, they only need the AI to fill demand lol
Usually when doing anchoring you want to end up at a result less than what you originally asked for but crucially more than zero, and OpenAI immediately folded on building any fabs whatsoever, so I don't think it worked.
When you do something that stupid - starting at $7 trillion - you end the conversation before it really begins because you lose all credibility with the people that matter (eg TSMC and other investors).
If he had said $250 billion and six fabs, it would have been a lot to ask but people wouldn't think he was ignorant or irrational for saying it. Big tech for example has that kind of money to throw around spread out across a decade if the investment is a truly great opportunity.
Asking for $7T is - seriously - only slightly more absurd than asking for infinity dollars.
I guess he thinks his glorified markov chain will lead to ASI if scaled up sufficiently. Even if we get ASI, the likelihood that anybody will ever make any money from it is so delusional. This isn't going to be your average brainwashed peasant, crushing these capitalist pigs is probably the first thing it's gonna do.
This comment honestly feels delusional.
The technical term for this is “zone of possible agreement”. If your opening is outside the counterparty’s ZOPA, then they’re likely to disengage/exit.
Approaching Adam Neumann-levels of grandiosity.
Could it be possible that OpenAI's new autocomplete will be as transformative to the global economy as WeWork's short term office rentals?
For reference seven trillion dollars is 25% of US GDP.
Yeah, that's um, wild.
And that's 7T just to build the fabs, then they'd need tons more money to build the hardware to put the chips in, datacenters, staff, software, etc.
> The TSMC execs reported more or less laughed in his face when he brought it up it them.
Reminds me of:
Musk flew there with Cantrell, prepared to purchase three ICBMs for $21 million. But to Musk's disappointment, the Russians now claimed that they wanted $21 million for each rocket, and then taunted the future SpaceX founder. As Cantrell recounted to Esquire: “They said, 'Oh, little boy, you don't have the money?”
If you don't believe ASI is possible, 7T is way too high. If you believe ASI is possible, 7T is wayyy too low, thus 7T is a compromise if you believe there's a non-certain chance of ASI being possible
Surely this “logic” applies to every positive amount of dollars. Why not 700T
So the question is, if you have AGI/ASI how much productive work could you get done with it?
When looking at things like mechanization and the productivity increases from around 1880 to now, you took an economy from around 10 billion a year to 14 trillion. This involved mechanization and digitization. We live in a world that someone from 1880 really couldn't imagine.
What I don't know how to answer (or at least search properly) is how much investment this took over that 150 year period. I'm going to assume it's vastly more than $7 trillion. If $7 trillion in investment and manufacturing allowed us to produce human level+ AI then the economic benefits of this would be in the tend to hundreds of trillions.
Now, this isn't saying it would be good for me and you personally, but the capability to apply intelligence to problems and provide solutions would grow very rapidly and dramatically change the world we live in now.
> If $7 trillion in investment and manufacturing allowed us to produce human level+ AI then the economic benefits of this would be in the tend to hundreds of trillions.
Sure, but right now that "if" is trying to do 7 trillion dollars of unsubstantiated heavy lifting. I might be able to start creating Iron Man-style arc reactors in a cave with a box of scraps and all I ask is 1 trillion, so you all should invest given how much money unlimited free energy is worth.
The amount is just how quickly you scale up. Invest more money now and get ASi earlier.
How do you even define ASI? How do you know it will run on silicon? If it will run on silicon, what kind of silicon? GPU? TPU? Binary? Ternary? Von-Neumann architecture? How can you even start to build chips for something completely imaginary?
ASI is just AGI but scaled to the point of superhuman ability in all cognitive tasks. Surely it will run on the architectures used to build AGI before ASI takes control itself in the design and deployment.
> ASI is just AGI but scaled [...] Surely it will run on the architectures used to build AGI
The words "just" and "surely" are doing so much heavy lifting here that I'm worried about labor-laws. :p
If scaling up was "just" that easy and the next step would "surely" begin in the exact same architecture... Well, you've gotta ask why several billion general-intelligence units made from the finest possible nanotechnology haven't already merged into a super-intelligence. Or whether it happened, and we just didn't notice.
Not surprising at all. As a former Google employee, I'm regularly in contact with friends still at Google and the number of hardware engineers OpenAI has hired out of Google have drastically increased in the last few months. You can readily see for yourself just by browsing LinkedIn and looking at profiles of hardware engineers currently at OpenAI and noticing how many of them are former Google employees. Google's own TPU team is now a shell of its former self. I wouldn't be surprised if future Google TPUs are inferior to the new OpenAI chips.
Also, did you know that Google's TPU efforts previously also relied on help from Broadcom? https://www.theregister.com/AMP/2023/09/22/google_broadcom_t... It's absolutely unsurprising that former Google employees would bring this vendor relationship into their new company, given Broadcom did a good job helping Google with TPUs.
It's also worth mentioning that google spent a lot of time (and probably money) and human capital on their TPUs and slowly got there. Other companies that have a proven track record of hyper-scaling also tried and missed (tesla) so it's not as easy as locking a bunch of phds in a room and have them crank out LLMs. Hardware is finicky, slow to develop, slow to prove itself and extremely expensive.
> Diversifying away from a single source for all of AI hardware and software, is a smart thing to do.
I wonder how this squares with the exclusivity contract with Microsoft. Even the OpenAI/Oracle deal requires Oracle to run Azure stack on their datacenter so MSFT can mediate the relationship. The AMD chips mentioned are also purchased by MSFT.
I wonder if this really means that OpenAI is accepting the risk/capital expense while providing a variety of hardware to Microsoft, or if there are other terms at play.
It is just more of this...
https://www.amd.com/en/newsroom/press-releases/2024-5-21-amd...
That's interesting given Microsoft also has it's own accelerator, Maia...
That is exactly what the execs at my company are telling us when asked about not using Nvidia -- diversifying away. It's funny though because we have no Nvidia for training at all. We use Trainium because we could not get our hands on Nvidia.
I'm guessing you also got some free/discounted credits?
Not sure of the pricing, but the performance of their chip is pretty outdated at this point. MI300x is way faster and more memory.
In a data center context?
I think I get what you're saying. That a datacenter probably just wants to focus on one product at a time. You're right, it is simpler. On the other hand, that's part of why neoclouds are sprouting up. Users are asking for the optionality.
Considering the cogent financial analysis detailed in "The Subprime AI Crisis"[0], all I can add is a quote from the movie "Dodgeball"[1]:
0 - https://www.wheresyoured.at/subprimeai/1 - https://www.imdb.com/title/tt0364725/quotes/
So presumably they're building their own data centers to use these chips, but odd that the article doesn't actually say so.
Are any of the western AI/matmul chips NOT being fabbed by TSMC? Amazon or Meta's perhaps? We've got NVIDIA, AMD, Google (TPU - also designed by Broadcom), and now OpenAI all using them. Are Samsung not competitive?
I wonder to what extent total chip/chipset volume is limited by TSMC capacity as opposed to memory or packaging?
Samsung is not competitive so everything good has to be made by TSMC now. It has been reported that packaging and HBM are the bottlenecks for AI chips although TSMC is also running at high utilization.
How long would take them to get a new chip into production and then used for training/inference?
twoish years, if you're good. Then the software to make it work properly. its not a quick thing to do.
You can throw more money at it to make it go faster, but it also might fuck it up and take longer.
Correct! AMD has been trying to fix "the software to make it work properly" part for years -- more than two years.
AMD hardware is good. The software is getting better daily. Training still sucks (mostly due to unoptimized libraries), but inference is looking pretty decent with recent advances in tuning...
https://shisa.ai/blog/posts/tuning-vllm-mi300x/
https://blog.vllm.ai/2024/10/23/vllm-serving-amd.html
The first generation chip won't be good enough to use in production so then you have the second generation... maybe 3-4 years.
Other than GH Copilot (which is amazing) and the like, I am yet to see a social/cultural impact of LLMs. I am ignoring the hype cycle fueled social media posts where influencers show cool stuff or model releases. I am talking about this stuff permeating through the society.. am I immune to it? Or is it hidden in saas/enterprise products layers deep?
I would imagine some productive cultural impact that’s wide spread, sticky, high retention not like ‘oh cool 3d tv or vr’ kinda thing?
Given how much ‘money’ has been burnt through the GPU heatsinks.
It’s kinda easy to check it though. Ask your friends if they use it, preferably non developer. When I find out my accounting friends and marketing friends are using it for basic analytics/promotion write ups I realized this is being used by non developers already on a regular basis. The use by non tech folks is significant for this . Crypto as a counter example of hype hasn’t lived up to its purported use case as non tech ppl I know only “own” it through an exchange.
Go ask an English teacher how LLMs are affecting them.
I expect they wanted to go down the fab and chip design path to build a moat, as they've realised they really don't have one outside of brand recognition.
If they had proprietary fabs and chips that were magically superior, and they could build them for cheap rather than paying the team green tax, they'd have a huge advantage in inference cost.
Perhaps they've realised just starting with a proprietary chip is enough?
Spending $7 trillion on in-house fabs sounded both ambitious and crazy. Reality finally kicked in. If they’re done dreaming big, let’s hope they keep the quality
More like he's finally done price anchoring and ready to make some actual deals.
This is the first I've heard of any of this, and it's funny how making their oen ships with TSMC is a "scaled back" plan
Podcast bro.
Broadcom! But that's where IP goes to languish and die.
I feel like them working with broadcom is another warning sign
Because Google's TPU' are so bad? https://www.theregister.com/2023/09/22/google_broadcom_tpus/
Broadcom is a vendor. They provide the main skeleton of a chip so that the client only needs to focus on the product differentiator, which is the compute part of the chip. They can worry less about other parts that are not their core competency.
Broadcom also builds xPUs for Google, Meta and Bytedance. Maybe these companies know a thing or two.
Why? Google also partners with Broadcom for TPU.
Why? Care for references instead of opinions?
I'm not saying that I necessarily agree, but the general consensus on HN seems to be that Broadcom is now less of a tech company and more of a holding company that raids others, for example VMWare, and extracts all value to the detriment of customers and the acquired company.
I don't think that's completely wrong, but it's a big company and I'm sure there are some better areas of the company than others.
Broadcom has some really good chips and their semi-custom chip business is pretty successful. HN doesn't understand hardware so they don't know this.
Try asking all those happy vmware customers. Sometimes technical proficiency is completely irrelevant.
I think it's completely wrong in the context of its role as an ASIC partner. There's a very short list of companies with all the IP needed for these cutting-edge ASICs and Broadcom/Avago might be the best of them. And to be clear, they've developed that IP themselves, just as they've always done. Those that think they're just a "holding company" haven't actually worked with them.
It’s confusing. The PE buyout built a holding/looting company and owned Broadcom, and took the Broadcom name.
So both are correct. Broadcom is a legit important hw company, and it is a PE holding company that buys and loots tech companies (like VMware)
it's no longer Broadcom
Avago Technologies (owned by corporate raiders) bought Broadcom, then took its name for itself
its ticker is still AVAG
It is AVGO
Additionally, having worked with some of their network devices at the driver level, they seem to be kludge piled on top of hack poured over a soup of workarounds for hardware bugs. Maybe they've gotten better recently, but just looking at their drivers, it didn't paint a great picture.
Oh god, I’d paint all hardware companies that way!
Having been on both sides, I’m continually shocked that stuff even works.
> general consensus on HN
General consensus on HN is generally wrong.
General consensus on HN is generally wrong.
I think everyone here would agree with that.
;)
this makes sense since the general consensus on HN is that any general consensus on HN is right so if the general consensus on HN that any general consensus on HN is right is wrong then any general consensus on HN could be wrong
Is there somewhere the general consensus is generally right?
"Journalists" seem to trust twitter an awful lot...
math departments
Broadcom is on track to spend 10bn on RND this year. Basically a pure-play patent troll!
Don’t they make the WiFi chips in iPhones?
They make every kind of chip for networking: WiFi, Bluetooth, home routers, cable modems, fiber, switches, NICs, DPUs, etc. And that's 1/Nth of their business.
Sam seems to be finding as many devils to bargain with as he can; Broadcom is a particularly devilish company.
Makes sense to work with them if they’re trying to design some kind of ASIC that would work for training or inference though?
What have they done?
They buy companies like VMWare and Computer-Associate, gut them and jack the price until they only have captive consumers from the fortune 500.
CA and VMWare were already dead on acquisition.
Splunk+Datadog and AWS+Nutanix+Cohesity respectively ate much of CA and VMWare's marketshare.
So TSMC will eventually work with podcasting bro
>TSMC execs allegedly dismissed Sam Altman as ‘podcasting bro’ — OpenAI CEO made absurd requests for 36 fabs for $7 trillion
I’m sure they’ll work with anyone willing to pay