I agree, but governments intentionally shifted from corporate taxes (taxes on net, corporate income) to payroll taxes (taxes proportional to employee wages) because businesses were either finding creatives ways of deferring/diverting income, or they just weren't taking profit (and, thus, nothing to tax).
Corporate taxes are inefficient as well as they discourage reinvestment by the company, since the State taxes a rake each year. Taxing on dividends paid (and buybacks) is better, even if it leads to higher prices to compensate for the lower capital profitability for shareholders (who compete with bonds).
The bill in favour of the Elizabeth Line was only put to parliament in 2005, receiving royal assent in 2008. Construction work began in 2009, faced some delays during COVID, but was completed in 2022 (total construction time: 13 years)
Construction on New York's Tunnel #3 began in 1970. It was 28 years before any part of it was operational. A second section came online 15 years later (2013). The final stage isn't expected to be completed until 2032, a full 62 years after construction began. I'm unaware of any comparable tunnel project which has progressed at this slow of a pace.
You still need an interface which does at least two things: handles incoming read/write requests using some kind of network protocol, and operates as a memory controller for the RAM.
Texas Memory Systems was in the business of making large 'RAM Drives'. They had a product line known as "RamSan" which made many gigabytes/terabytes of DDR available via a block storage interface over infiniband and fibre channel. The control layer was implemented via FPGA.
I recall a press release from 2004 which publicized the US govt purchase of a 2.5TB RamSan. They later expanded into SSDs and were acquired by IBM in 2012.
> To be clear - the shock wasn’t that OpenAI made a big deal, no, it was that they made two massive deals this big, at the same time, with Samsung and SK Hynix simultaneously
That's not "dirty." That's hiding your intentions from suppliers so they don't crank prices before you walk through their front door.
If you want to buy a cake, never let the baker know it's for a wedding.
What they mean is that they bought 40% of all RAM production, they managed to do that by simultaneously making two big deals at the same time. It's buying up 40% of all RAM production with the intention to have most of it idle in warehouses that is "dirty". And in order to be able to do that, they needed to be secretive and time two big deals at the same time.
The incentive suggested in the article is to block other competitors from scaling training, which is immensely RAM hungry. Amongst other things. Even Nvidia could feel the pressure, since their GPUs need RAM. It could be a good bargaining chip for them, who knows.
I'm not saying it's true, but it is suspicious at the very least. The RAM is unusable as it stands, it's just raw wafer, they'd need a semiconductor fab + PCB assembly to turn them into usable RAM modules. Why does OpenAI want to become a RAM manufacturer, but of only the process post-wafer.
> The RAM is unusable as it stands, it's just raw wafer, they'd need a semiconductor fab
The wafers are processed. That means Samsung/Hynix have taken the raw ("blank") wafers, then run them through their DRAM lithography process, etching hundreds of DRAM dies ("chips") onto the wafer.
You could attach test probes to individual chips on the wafer and you'd have a working DRAM chip. In fact, that's how testing is performed: you connect to each die one at a time with a "probe card" which supplies power & ground, plus an electrical interface for functional testing.
If OpenAI takes possession of the processed wafers and wants finished RAM modules, they need to do a few things: test each die (expensive), saw the wafer into individual chips (cheap), package them (moderately expensive), test them again (medium expense), and then assemble the final module (inexpensive). Modern semiconductor test facilities cost billions of dollars and take years to build, so they'd need to immediately outsource that work (typically done in Southeast Asia)
OpenAI likely doesn't want to do any of this. They probably just want to make sure they're in control of their own destiny with regard to DRAM, then decided the best place to accomplish that was by cutting deals directly with the DRAM semiconductor producers. This will allow them to take the wafers to the existing supply chain, then contract them to turn the wafers into finished modules.
This wasn't buying a cake from a baker, this was a bakery buying 40% of the flour in the world so nobody else can sell wedding cakes, but now there's gonna be no bread for a year
That's not the dirty part. This is the dirty part:
> OpenAI isn’t even bothering to buy finished memory modules! No, their deals are unprecedentedly only for raw wafers — uncut, unfinished, and not even allocated to a specific DRAM standard yet. It’s not even clear if they have decided yet on how or when they will finish them into RAM sticks or HBM! Right now it seems like these wafers will just be stockpiled in warehouses
> OpenAI isn’t even bothering to buy finished memory modules
And? Why should they be obligated to pay for all the middleman steps from fab down to module? That includes: wafer-level test, module-level test (DC, AC, parametric), packaging, post-packaging test, and module fabrication. There's nothing illegal or sketchy about saying, "give me the wafers, I'll take care of everything else myself."
> not even allocated to a specific DRAM standard yet
DRAM manufacturers design and fabricate chips to sell into a standardized, commodity market. There's no secret evolutionary step which occurs after the wafers are etched which turns chips into something which adheres to DDR4,5,6,7,8,9
> It’s not even clear if they have decided yet on how or when they will finish them into RAM sticks or HBM
The implication here is that the primary goal is to corner the market, not to use the supply. If you aren't going to use them anyways then of course it is silly to pay for them to be finished.
Do you think that's fine, or do you think that implication is wrong and OpenAI does actually plan to deploy 40% of the world's DRAM supply?
> The implication here is that the primary goal is to corner the market
You have no evidence of that. Even at face value, the idea of "cornering the market" on a depreciating asset with no long-term value isn't a war strategy, it's flushing money down the toilet. Moreover, there's a credible argument OpenAI wanted to secure capacity in an essential part of their upstream supply chain to ensure stable prices for themselves. That's not "cornering the market," either, it's securing stability for their own growth.
Apple used to buy-up almost all leading-edge semiconductor process capacity from TSMC. It wasn't to resell capacity to everyone else, it was to secure capacity for themselves (particularly for new product launches). Nvidia has been doing the same since the CUDA bubble took off (they have, in effect, two entire fabs worth of leading-edge production just for their GPUs/accelerators). Have they been "cornering" the deep sub-micron foundry market?
> the idea of "cornering the market" on a depreciating asset with no long-term value isn't a war strategy, it's flushing money down the toilet
OpenAI's entire business strategy thus far can be summarized as "flushing money down the toilet", so that isn't actually as unlikely as you're making it sound.
Yes, they've made insane scaling bets before and they have paid off.
If what we've heard about no acceptable pre-training runs from them in the last two years trying to increase the memory for training by two orders of magnitude is just a rehash of what got them from gpt2 to gpt3.
One of the pain points of scaling Zookeeper is that all writes must go to the leader (reads can be fulfilled by followers). I understand this is "leader of a shard" and not a "global leader," but it still means a skewed write load on a shard has to run through a single leader instance
> given that horizontal scaling of metadata requires no rebalancing
This means a skewed load cannot be addressed via horizontal scaling (provisioning additional shards). To their credit, they acknowledge this later in the (very well-written) article:
> This design decision has downsides: TernFS assumes that the load will be
I used to send inquiries to Google about their policies on political ads. This is partly because, in 2024, they said they were implementing "pauses" on certain types of political content:
I discovered two things. First, Google doesn't care about being consistent, hypocritical, or, really, anything. Second, Israel creates new ad accounts with Google, runs campaigns, then shuts them down. Here's one advertiser ID which I saved in my bookmarks:
> I am pretty damn sure that with a thousand me, we produce more than the founders alone
Engineers generally never acknowledge the value of knowing _what_ to build, only whether something is built.
As an engineer, I don't pretend to know or understand what the market wants or what people are willing to pay for. That's a problem I leave to business/product-oriented people. In exchange, they leave the building to me.
> Engineers generally never acknowledge the value of knowing _what_ to build, only whether something is built.
Nobody knows what to build, that's why VCs give money to many startups, almost all of them bankrupting.
The sole value of the founders is to be able to convince VCs for as long as they can. VCs generally have no clue about the technology, so it's about getting good at selling bullshit. Granted, I am not great at selling bullshit, so it doesn't make me a great founder.
Now if we get back to knowing what to build, if you give me a thousand me, we can try a whole bunch of different ideas hoping that one works. This is what VCs do, and it works for them even though they have no clue either.
No, a founder is never worth thousands of employees, period.
And I'd like to note the asymmetry here: I am arguing that I am not thousands of times less valuable than them. I am not remotely saying that I am better.
What does the backend architecture look like? Heavily distributed, or everything in-memory on a single address space server? How do you handle users writing tight loops which fetch the endpoint in rapid fire?
It's a Next.js app currently deployed on vercel with a Postgres DB running on supabase. Both are fairly scalable, but not in an optimal way in terms of costs. I'd either have to introduce a Pro pricing plan or get some donations from sponsors. The simple nature of the API would allow to heavily distribute this though, i.e. one could simply shard by progress bar.
Regarding the tight loops: Currently there is just basic rate limiting in place and the Python client batches updates by default. The app is not intended for short-lived progress bars with fast updates. These usually also don't have a good reason to be shared and live on the web. It's really much more useful for slowly updating, long running processes.
I agree, but governments intentionally shifted from corporate taxes (taxes on net, corporate income) to payroll taxes (taxes proportional to employee wages) because businesses were either finding creatives ways of deferring/diverting income, or they just weren't taking profit (and, thus, nothing to tax).
See the third graph here (yellow line vs. dark blue line) https://taxpolicycenter.org/briefing-book/what-are-sources-r...
reply