I wonder whether leadership is ignorant OR informed and willing to gamble that they, unlike everyone else, will succeed.
My prediction is that 1. SW for said chips will be inferior, 2. leadership will force everyone to use it anyway, 3. Anthropic will give it a shot and be handcuffed to mediocrity, 4. Anthropic folks will pivot to something that doesn’t suck.
Amazon invested in Anthropic using billions in AWS credit and I guess it would be covered in contract that some percent of the credit could be used only for this chip.
Weird that the article frames the chips as “ Powered by Austin’s Annapurna”. It does say further down that Annapurna was an acquired Israeli company. But I wonder if the framing and emphasis on an American location is meant to avoid antitrust scrutiny from the new administration.
Huh? The incoming administration has a history of showing it not only doesn’t care about antitrust, it actively encourages monopolies.
If you read the article, they literally say it’s an Israeli company, and the AI chip lab is in Austin.
> The heart of AWS’s efforts is in Austin, Texas, home to an AI chip lab run by Annapurna Labs, an Israeli microelectronics company Amazon acquired for about $350 million in 2015.
If Amazon purchased Annapurna in 2015, then maybe you mean the staff are still located in Israel? I think if they are a department of Amazon then they are a 'US company' in that way.
> Huh? The incoming administration has a history of showing it not only doesn’t care about antitrust, it actively encourages monopolies.
Just like every other administration. Microsoft (and Google, and ...) did not becomeva monopoly during the current or past administration.
For some reason, America loves monopolies.
(it is easier to talk to 1 guy instead of 10).
You’re using performance on consumer HW to conclude about abilities to build data center stuff?
You realize they already have their own ARM CPUs and run their own NVME drives with their own controllers right? And these generally outperform others in the same class.
And all this ignoring that kindle and Alexa are both market-defining products that either dominated (kindle) or created a new market that didn’t really exist before (Alexa).
There’s valid criticisms to levy, but try to be dispassionate and analytical as emotional comments don’t raise the bar for discussion. For example, I don’t really see how Alexa is really all that different from the design of any other smart speaker. They all appear to be fungible with different ecosystems accessible to them. Not sure about TV and phone products, but I’m not aware of any competitive ebook readers that offer a competitive mix of ease of use, price, etc (haven’t looked in a long time so my perception here may be dated).
And it’s important to note that they’re focus on training acceleration with these chips which is the path Google took with their chips. No reason to believe that they won’t be successful here for similar reasons in that a purpose built accelerator can outperform a general purpose design like a GPU.
In their defense it’s rare for a company that has sloppy quality in some products to be knocking it out of the park with another product. The same company culture that allowed the sloppy products in one area tends to prevent other teams from doing their amazing quality work. I’ve used various AWS services and found the ones they designed, or standard services they customized to have more bugs and problematic limitations than one would hope. I was especially angry at Aurora Postgres when it came out because their hacked up customizations introduced memory leaks. Their work made a DB I’ve used for 20 years for its reliability suddenly unreliable. I now work for a company that pays them a lot, so at least I see them fix bugs quickly now, but at startups that only paid them a few million they would ignore problems we reported. Their culture focuses on doing more at lower cost to the breaking point(they’re 21st century WalMart) and they have one of the worst reputations in the industry for burning out developers so it doesn’t surprise me when I run into bugs, silly limitations, or poorly thought out tooling.
Their consumer products are absolutely infuriating OCD triggers. For years in the Kindle app if you went back a page it would often start on a different paragraph than the first time you were on the page. The book reader didn’t implement the concept of a _page_ correctly for goodness sakes. And they clearly intended to emulate physical pages because they did implement a fancy page turn animation.
That’s an interesting hypothesis but doesn’t explain how Google consumer products suck too yet their cloud offering and in-house made HW is fairly high class.
Similarly there’s plenty of sloppy in house software at any large company and they still produce good products in areas that matter to them (eg radar at Apple is both terrible and amazing, their internal directory app was a mess, etc).
In my experience, and in a simplified form, the tech orgs under SVPs are pretty independent of each other. One senior leader’s org operates and develops almost entirely independently of anyone else’s, more so if their product type (software vs cloud vs product) are culturally different.
> You realize they already have their own ARM CPUs and run their own NVME drives with their own controllers right? And these generally outperform others in the same class.
Here's [1] Graviton shooting out with AMD & Intel, underperforming in some benchmarks, outperforming in others. Here's another [2] showing how the power estimate for graviton is less than half of Intel & AMD which means that it's about 2x better on a power/watt level which is how you measure performance typically.
Here's [3] [4] Amazon claiming how their in-house Nitro SSDs keep improving. Here's [5] a comparison of AWS vs other cloud providers, noting how AWS leads in 7/10 metric categories one of which is disk performance and [6] is a similar report showing how AWS nitro disks lead in latency by a factor of 2x. The 2022 report [7] shows more mixed results so things may have equalized, but only because other providers are following Amazon and catching up with their own equivalent solutions, not because they're relying on third-party SSD vendors to build better NVMe drives.
But honestly, for things that are factual and as easily googlable as this, I would suggest you do your own research to become informed rather than expect to be hand-fed.
But they are resourceful in utilizing their existing products and integrating them elsewhere, eg: snowballs with both fire and kindle tablets as part of it's UX.
My prediction is that 1. SW for said chips will be inferior, 2. leadership will force everyone to use it anyway, 3. Anthropic will give it a shot and be handcuffed to mediocrity, 4. Anthropic folks will pivot to something that doesn’t suck.