Astra was/is superior for planning type tasks. It was capable of doing seemingly magic things with rather vague/lazy instructions ("I need to be able to test this on Windows, maybe a qemu VM or something? Shrug." ... 1 hour later "yeah i built you a whole qemu + eval windows image + harness of powershell scripts + shell scripts to retrieve & verify harness.").
And for UI work -- which is not something I do a lot of but do here and there -- it was clearly superior to 5.6 Sol.
But it also feels sloppier? Somehow. And too expensive to use.
I felt this way with Sol in the 5.6 series and was one of the seemingly few people on this earth who liked Terra for that reason. I would often have a very specific code-manipulation ask, e.g. "add a parameter to this method, ensure all callers pass it in, if there is not a logical way to derive the parameter to be passed in a particular instance, flag this in your final response", and Sol would go on some rabbit hole side quest to refactor my codebase to determine some way to derive it rather than flagging it as I had asked.
Terra had the "workhorse" quality where it could do these changes in bulk and follow directions without being too 'smart' (but sloppy) as you described. Luna was a bit too dumb and would make sloppy mistakes; I see that more as a "run these tests and format the results" sort of model. Maybe 6 Luna will be better.
I also just reread your comment and realized the naming convention is still extremely confusing with respect to ordering of [Family]x[Model]x[Number].
I also get good mileage out of Terra when I need a diligent workhorse. That's a good way to describe it. We should start using character archetypes when we describe models, it'll do more good than the dubious numbers and cherry-picked quotes. Maybe RPG character-type cliches? Myers Briggs?
Glad you pointed out the UI work. I've been doing a lot of it and it's so much better than 5.6 as UI, it's unbelievable. I give it super ambiguous instructions and it's reading my mind. I do the same thing with 5.6 and I'm correcting it for a few minutes.
About half of SpaceX revenue is Starlink subscriptions. Starlink is the one profitable division; the rest of the company operates at a loss, including xAI.
I don't know that's moving the goalpost, SpaceX's mission was not that originally. Some could say the original mission failed and they are pivoting, Elon will say it's a means to the end of the original mission still to save face. Some read it all as snake oil
A dense model (up to the amount of memory available) actually does make the most sense on unified memory architectures. But when you hit the limit of what you can hold in memory, you reach the limitation of the platform.
Whereas a hybrid architecture with distinct DRAM and VRAM with sparse MoE, you can leverage two different bit rates depending on the actual need for constant access to common layers versus sparse access to infrequent layers and arbitrage the difference in cost for each of those in distinct classes of hardware.
> A dense model (up to the amount of memory available) actually does make the most sense on unified memory architectures
Inference time is going to be dominated by the low memory bandwidth on these Macs, so a dense model will suffer most. It’s more of an opportunity for large MoE models with a low number of active experts since you can keep all experts in VRAM but not pay the bandwidth cost until they are used.
> you can leverage two different bit rates depending on the actual need for constant access to common layers versus sparse access to infrequent layers
This is an interesting direction that I expect to see more of. But for most models currently you need basically all experts loaded since they are chosen per token.
Apple seems to be researching longer horizon expert caching, where they keep experts swapped in for longer runs of tokens [1]. Other labs are offloading ngram caches but not sure if they’re pursuing anything like this?
They do MoE. They benchmarked GLM 5.3-flash (320B / 18B), and Qwen 3.8-flash-next (125B / 6B). The dense Qwen is only focused (I assume) because it's about the only thing that fits on a 5090, that they can compare the two heads on.
> AI allows players to have decompiled and in turn ports of games to new platforms, enhanced versions of the game, etc.
We're ten years too early to do what you're suggesting. You need a really, really good understanding of the decompiled codebase and the original game and platform to port or implement new features.
You can't really ask "Claude, make Ocarina Of Time Widescreen" and expect it to work.
We're not ten years early. If you can afford the tokens, and can course-correct the agent when it gets off in the weeds, AI can do 95% of the grunt work of a decomp now.
Also, "widescreen OOT" is probably tweaking a pair of projection matrix floats in the binary so that might not be a great example :)
You could absolutely do that, because it's a narrow, targeted change. I give it 10% odds Claude could make that change to a binary ROM and 70% to a decomp.
> AI allows players to have decompiled and in turn ports of games to new platforms, enhanced versions of the game, etc.
To what end? The biggest benefit of tinkering with 40-year-old hardware is the experience of doing so. The knowledge and intuition you learn about how computers really work; the friendships you develop through collaboration with others, through asking for help with your projects and helping others with theirs; and the relaxing Saturday afternoons spent investigating and debugging and experimenting. The "benefit" of having a game that's been recompiled to a native executable for your PC or whatever is marginal compared to the benefit of the process to get to that point. That end goal is just a lighthouse to keep you on track, and when you get there maybe you'll play the recompiled game once to celebrate a job well done before moving on to another project.
If writing software is a means to an end, then there's value in anything that reduces the effort to get to that end. But as I see it, retro gaming projects are mostly an end to a means.
> Seems pretty egoistical when nobody's forcing you to use AI in any way, it provides an objective benefit to other people.
I have three fundamental claims:
- There is little to no positive benefit to using AI for hobby projects whose completion provides little to no intrinsic value.
- Choosing to do so has a negative impact on you, as it bypasses an incredibly valuable learning experience.
- Choosing to do so has a negative impact on others, as it takes away someone else's project opportunity (there are a finite amount of popular retro games, and many people prefer to work on novel projects rather than recreating something that has already been done).
I'm not sure what's "egoistical" about that, and I'd love if you could elaborate on that because I'm still wrestling with the ethics and implications of all this myself. The way I see it, it's better if I choose to forgo the use of AI on my projects, and I think others should choose the same.
---
Many years back, one retro gaming community I'm a part of had a prolific contributor who was extremely knowledgeable about the game, but also incredibly abrasive. He was the most knowledgeable person in the community at the time, and held this over others; discrediting their achievements, discouraging their attempts to learn, constantly starting fights and driving many good people out of the community. Yet he was allowed to remain in the community for many years, because everyone thought we needed his expertise.
Eventually the community came to the realization that it's just a video game and we don't "need" anything at all. None of the long-term community members are really here for the game anymore; we're here for the experience and the camaraderie. So we should make decisions (like kicking out this member) that are technically suboptimal for the game (at least in the short term), but better for fostering the community.
In a similar vein, using AI within these sorts of hobby spaces may allow faster progress on such projects, but it takes away the value of completing the projects in the first place, to the detriment of both the individual and the community.
> To what end? The biggest benefit of tinkering with 40-year-old hardware is the experience of doing so.
The biggest benefit to you. Other people, most people in fact, just like the game and want to enjoy it in new ways. They don't care about the code or reverse engineering itself. That people now benefit that it's easy to decompile because otherwise outside of really big titles like sm64, probably nobody is going to take the effort to manually decompile a less popular game.
That's the egoistical part.
Nobody is forcing you to use AI. And it's not really any different to how people would build their own NES emulators to learn instead of just grabbing whatever best NES emulator already exist.
I think my biggest issue with that is that people who just want to replay a game in a fresh way once or twice aren't the kind of people who are going to stick around and form a community; and those communities are what foster the creativity which brings freshness and depth to the games. I see it as quite similar to AI music creation, which, sure, allows people access to music that didn't exist before without putting in the effort, but in doing so it replaces what has historically been a community endeavor with a highly individualist, consumerist one.
My perspective is biased by the fact that such communities are an important part of my life. I don't want to see my years-long collaborations and friendships made obsolete by someone looking for an afternoon of nostalgia. And maybe everything will be fine, and the AI-users and the community-builders will go their separate ways and leave each other alone. But communities which exist today are having to ask how we should relate to AI, and my conclusion is that it's best kept out of hobbyist spaces if we want those spaces to continue to exist.
To be able to play retro games on current hardware/OS?
Why does it always need to be so complicated - I have zero interest in learning about these legacy platforms but I definitely want to play some nostalgic games.
Mindboggling that you see this as a negative and even want to kick out people to “protect” the community
Also to fix bugs. Someone recently ran Shadowrun (Genesis) disassembly through chatgpt and discovered that weapon upgrades were completely broken (using the wrong value as an index and ending up reading character stats as upgrade flags), defensive spells actually buffed your enemies, not you, etc. and released a romhack to fix things.
Lately I've seen this statement that I agree with: for some people, a hobby is a means to an end (which means using AI to achieve the goal is a heresy), for others, it is the end in itself (which it isn't).
Sorry, what? The benefit is that we have these games to play. That is the intrinsic benefit. Most of us don't care about the other things you listed when all we want to do is play old games. You're free to value those things in this context and that's fine, but it's weird to impose those values on others when they're providing what people want.
I think you are missing the "idle capacity" information. The GPUs are not running 24/7, there may be minutes of time in a day where there is not an active job running, the author simply located these inactive GPUs during the downtime and utilized it until another job was queued.
Yeah it's a bit ambiguous, but in any case I'd think simply using that idle capacity to generate tokens would be more valuable than finding a basically meaningless number.
Except customer demand doesn't follow the fixed supply of GPU compute that any company has, so they either a) rent compute on demand, b) sell their excess compute, c) let it sit idle and maybe an employee claims it.
It's pretty sad when you realize the average person is just kind of... unaware. They are fine with whatever, they don't really notice most stuff that goes around them.
Its my understanding that rust just addresses memory safety, and it all falls apart at the first mention of "unsafe" or whatever the keyword is, not to mention the supply chain thing with the crates. Obviously AI makes it all moot because thats where the entire theory of basic security gets thrown out the window lol just chat with facebooks robot if you want someone elses instagram account amirite?
reply