The game is released on both PC and PS5, the latter of which was designed (and marketed) to take advantage of SSD speeds for streaming game content near real time.
The latest Ratchet and Clank, the poster child used in part to advertise the SSD speed advantage, suffers on traditional hard drives as well in the PC port. Returnal is in the same boat. Both were originally PS5 exclusives.
My understanding is that optimizing for sequential read is a big reason for historical game install bloat; if you include the common assets multiple times in the archive, then loading a level/zone becomes one big continuous slurp rather than jumping all over the place to pick up the stuff that's common to everything. Obviously this didn't matter with optical media where the user wasn't impacted, but it's annoying on PC where we've had a long period of users who invested in expensive, high-performance storage having to use more of it than needed due to optimizations geared at legacy players still on spinning rust.
I expect that low-latency seek time is also pretty key to making stuff like nanite work, where all the LODs for a single object are mixed together and you need to be able to quickly pick off the disk the parts that are needed for the current draw task.
The HDD performance suffers very much during the portal loading sequences in Ratchet and Clank, but even the entry level SSD performs fine, with little visible difference compared to the PS5 one. It’s more about random access speed than pure throughput
I played Rift Apart from HDD and apart from extra loading time during looped animations it was fine. On the other hand Indiana Jones Great Circle was barely playable with popping-in textures and models everywhere.
This assumes your model is static and never needs to be improved or updated.
Inference is cheap because the final model, despite its size, is ridiculously less resource intensive to use than it is to produce.
ChatGPT in its latest form isn't bad by any means, but it is falling behind. And that requires significant overhead, both to train and to iterate on model architecture. It is often a variable cost as well.
As long as revenue rises faster than training costs and inference remains profitable I dont think this is an issue. Eventually theyll be able to profitably amortize training across all the users.
> As long as revenue rises faster than training costs
And this is definitely not happening. They are covering training costs with investors money, and they can't really stop it without their competitors catching up
Interesting title, but folks seem to forget the double-digit MHz processors straining to hit 30 FPS back then on Doom unless it was in a smaller rendering size. This would have killed the performance even on my mighty 486 running at 33 MHz at the time (I don't have a lot of faith in the VM demo). It remains to be seen if the music files are small enough to keep the game to its original ~2 x 1.44MB installation disk size.
Still neat.
I recall Windows 3.1 was sometimes shipped with a speaker driver that could emulate a very basic audio card on a traditional PC speaker almost entirely in software.
Reminds me of a commercial where an artist attempted to convince their friend in excited terms how amazing their new masterpiece was, only to reveal its a blank canvas and they ran out of paint.
That was a Salesforce instance with largely public data, rather than something owned and operated by Google itself. It's a bit like saying you stole from me, but instead of my apartment you broke into my off-site storage with Uhaul. Technically correct, but different implications on the integrity of my apartment security.
It was a social engineering attack that leveraged the device OAuth flow, where the device gaining access to the resource server (in this case the Salesforce API) is separate from the device that grants the authorization.
The hackers called employees/contractors at Google (& lots of other large companies) with user access to the company's Salesforce instance and tricked them into authorizing API access for the hackers' machine.
It's the same as loading Apple TV on your Roku despite not having a subscription and then calling your neighbor who does have an account and tricking them into entering the 5 digit code at link.apple.com
Continuing with your analogy, they didn't break into the off-site storage unit so much as they tricked someone into giving them a key.
There's no security vulnerability in Google/Salesforce or your apartment/storage per se, but a lapse in security training for employees/contractors can be the functional equivalent to a zero-day vulnerability.
There's no vulnerability per se, but I think the Salesforce UI is pretty confusing in this case. It looks like a login page, but actually if you fill it in, you're granting an attacker access.
Disclosure: I work at Google, but don't have much knowledge about this case.
Yep; it will just screw the employees, the investors (or at least those that didn't suddenly get clairvoyant and find an exit right before it hits the fan), and many, many apps relying on their APIs.
The correction is good in the long run. It will settle to sane understanding of AI's capabilities, competent applications, and the unsuing AI winter will be when real research is done rather than "we fed its garbage back into itself hoping something else came out"
It will ruin some lives. I hate that the american economy runs on these speculative cycles. If we tempered our expectations to what was sane, but still bullish, we'd have far less to fall
This sounds useful, but the lengthy attempt to describe what it solves is rather confusing. Is it caching a list? An actual type of object? The type of the object itself?
The diagram, clearly well-intentioned, doesn't really answer my questions either. I'll note the lack of discussion on memory safety gives me pause, but I see evidence of such in the code.
> Once the library is configured for the project, the first thing you need to do is to define the IDs of the objects to be cached.
I'm certain they mean tag the identifier fields used by objects. At first glance, this sounded like a hyper-specific shortlist of objects cached based on their runtime ID values.
> Whenever a method is invoked with some arguments, those are used as a key that is mapped to IDs instead of objects. The objects are then retrieved and returned from the ValuePool via their IDs.
You store the actual value on this value map, and then have a bunch of maps with references to this value map. When you update the entity, you can update the cache by updating "just" the value map. All other maps are automatically "updated" because they just reference the identifier (no-op). So you query these other maps to get the list of identifiers, then pull the actual values from the value map.
I get what its trying to do, its just not clear to me how it would handle updating the key-reference maps. If I change the value of a field (say name on a user entity), does it know to remove that entry (user_name_age map as an example) from the references map?
> The caches of the methods are only linked to IDs instead of transaction objects, so we won't need to update something there as long as an object is not deleted.
Read up a bit on the effort needed to get a fab going, and the yield rates. While engineers are crucial in the setup, the fab itself is not as 'fungible' as the employees involved.
I can spin up a strong ML team through hiring in probably 6-12 months with the right funding. Building a chip fab and getting it to a sensible yield would take 3-5 years, significantly more funding, strong supply lines, etc.
> I can spin up a strong ML team through hiring in probably 6-12 months with the right funding
Not sure what to call this except "HN hubris" or something.
There are hundreds of companies who thought (and still think) the exact same thing, and even after 24 months or more of "the right funding" they still haven't delivered the results.
I think you're misunderstanding how difficult all of this is, if you think it's merely a money problem. Otherwise we'd see SOTA models from new groups every month, which we obviously aren't, we have a few big labs iteratively progressing SOTA, with some upstarts appearing sometimes (DeepSeek, Kimi et al) but it isn't as easy as you're trying to make it out to be.
There’s a lot in LLM training that is pretty commodity at this point. The difficulty is in data - and a large part of why it has gotten more challenging is simply that some of the best sources of data have locked down against scraping post-2022 and it is less permissible to use copyrighted data than the “move fast and break things” pre-2023 era.
As you mentioned, multiple no name chinese companies have done it and published many of their results. There is a commodity recipe for dense transformer training. The difference between Chinese and US is that they have less data restrictions.
I think people overindex on the Meta example. It’s hard to fully understand why Meta/llama have failed as hard as they have - but they are an outlier case. Microsoft AI only just started their efforts in earnest and are already beating Meta shockingly.
Fully agree. I also think we are deep into the diminishing returns territory.
If I have to guess OAI and others pay top dollars for talent that has a higher probability of discovering the next "attention" mechanism and investors are betting this is coming soon (hence the hige capitalizations and willing to loive with 11B losses/quarter). If they lose patience in throwing money at the problem I see only few players remaining in the race because they have other revenue streams
> It's just that startups don't go after the frontier models but niche spaces
But both of "New SOTA models every month" and "Startups don't go for SOTA" cannot be true at the same time. Either we get new SOTA models from new groups every month (not true today at least) or we don't, maybe because the labs are focusing on non-SOTA instead.
I've always taken that term literally, basically "top of the top". If you're not getting the best responses from that LLM, then it's not "top of the top" anymore, regardless of size.
Then something could be "SOTA in it's class" I suppose, but personally that's less interesting and also not what the parent commentator claimed, which was basically "anyone with money can get SOTA models up and running".
Edit: Wikipedia seems to agree with me too:
> The state of the art (SOTA or SotA, sometimes cutting edge, leading edge, or bleeding edge) refers to the highest level of general development, as of a device, technique, or scientific field achieved at a particular time
I haven't heard of anyone using SOTA to not mean "at the front of the pack", but maybe people outside of ML use the word differently.
> I don't get why you think that the only way that you can beat the big guys is by having more parameters than them.
Yeah, and I don't understand why people have to argue against some point others haven't made, kind of makes it less fun to participate in any discussions.
Whatever gets the best responses (no matter parameter size, specific architecture, addition of other things) is what I'd consider SOTA, then I guess you can go by your own definition.
Right. I could spin up a strong ML team, an AI startup, build a foundational model, etc give a reasonable amount of seed capital.
Build a chip fab? I’ve got no idea where to start, where to even find people to hire, and i know the equipment we’d need to acquire would be also quite difficult to get at any price.
But the fabs don't belong to NVIDIA, they belong to TSMC. I have no doubt that Taiwan and maybe even the US government would step in to save TSMC if for some reason it got existential problems, but that doesn't provide an argument for saving NVIDIA
> Minecraft is of course the poster child for very large worlds of interest these days.
Minecraft used to create very interesting worlds until they changed the algorithm and the landscapes became plain and boring. It took them about 10 year until the Caves and Cliffs Update to make the world generation interesting again.
The latest Ratchet and Clank, the poster child used in part to advertise the SSD speed advantage, suffers on traditional hard drives as well in the PC port. Returnal is in the same boat. Both were originally PS5 exclusives.