eldenring
eldenring on Hacker News
> Reducing shortcuts and loophole exploitation: With Carbon, we further mitigated shortcut attempts in training while ensuring high-quality coding performance. Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4% — all while improving overall accuracy.
This is probably the single most important thing to make frontier models usable right now, so I'm happy there seems to be some progress. This was the previous Gemini's biggest problem IMO
actually thinking about it more I think my oracle model is more powerful than the computer used in openai's fft but maybe its a good way to think about it.
I don't think that is necessarily true, the post is making a stronger point than what you describe
> Indeed, the algorithm requires you to multiply complex numbers whose magnitudes grow superpolynomially with n. If you stored these numbers as bits, the memory requirement would effectively be exponential in the input length. It assumes you can multiply such arbitrarily large complex numbers together in a single operation.
Another way to think about what the author said, imagine you had access to an oracle which could execute arbitrary precision arithmetic expressions in constant time. You could encode a wide range of computations in these and execute impossible algorithms.
Reflection AI is a completely different company with no relation (as far as I can tell) to the model you are referencing from 2024
Cooling is relatively easy, you just need radiators which are passive, and essentially reduce to a launch cost penalty. You are right that they can't be serviced, but that is missing the point of orbital data centers. The whole point is that you can build hundreds of thousands of these in a factory and launch them in a scalable manner. The power, cooling, etc. comes for "free". In the long run, as the cost of the chip, launches, etc. goes down, orbital data centers will scale better terrestrial ones.
As a side note, I don't understand why I keep seeing these wrong arguments on HN repeatedly. Like everything mentioned in this thread can easily be fact checked. Radiative cooling is solved, launch costs are going down, so power costs will pay themselves back very quickly, etc.
You can argue about specifics, like chips will get more sophisticated + power efficient and fabrication will be the true longterm bottleneck, or SMRs/fusion could reduce energy bottlenecks, but talking about cooling as if convective cooling is the only option is just nonsensical.
Did you read the whole thing? he is clearly talking about having lost focus which is directly in agreement with your take on the Metaverse.
my reply from the other thread fits here too:
> this is just not true at all, there are massive leaps from algorithms, data, etc. every year. scale is one axis of many and you need to get them all correct.
this is just not true at all, there are massive leaps from algorithms, data, etc. every year. scale is one axis of many and you need to get them all correct.
2-3x is completely dwarfed by the remaining improvements in training which is still in its infancy relatively
I'm guessing they had a significant revenue spike from gpt 5.4 and gpt 5.5 being so good at coding, and hiccups at anthropic making it easier for programmers to try the models.
Its just not a thing to consider and doesn't happen often.
This article makes 0 sense. Its not up to billing or computer systems or ease of use or anything else that matters. The question is will the scaling laws, which in the asymptote are likely the laws of physics, hold up in converting energy to smarter models. Its not really up to anyone, the labs or developers, to choose if local or remote models will be the norm.
CompactStr doesnt have any additional runtime overhead iirc right? So in theory you can drop it in everywhere even when you expect > 25 chars. Maybe an extra branch in the >25 char case?
> Folks are now starting to ask difficult questions about their burn rate and revenue.
this view isn't updated correctly post-claude code and codex. there will clearly be sufficient demand.
I don't think so, the whole point of writing software is it is a great sink for complexity. Encoding a process or mechanism in a program makes it work (as defined) for ever perfectly.
An example here is in engineering. Building a simulator for some process makes computing it much safer and consistent vs. having people redo the calculations themselves, even with AI assistance.
I think the coding market will be much larger. Knowledge work is kind of like the leaf nodes of the economy where software is the branches. That's to say, making software easier and cheaper to write will cause more and more complexity and work to move into the Software domain from the "real world" which is much messier and complicated.
Because there's a realistic chance this is the only important software technology moving forward, and commoditizes Metas's entire business which is software.
I don't see how its possible to think this. AI coding assistants are some of the most useful technologies ever created, and model quality is by far the most important thing, so I doesn't make sense why local inference would be the path forward unless something fundamentally changes about hardware.
How many docs do you put in the context? we maintain a lot of dsl code internally, and each file has a copy of the spec + guide as a comment at the top. Its about 50 locs and the relevant models are great at writing it.
I'm not sure what you mean by menial coding but all my employers have supported this in the past. This was a variety of companies, big tech, startups, etc. I think its more likely your employer is the outlier.
Its not subsidized, in fact, they probably have very healthy margins on Claude Code.
The issue is that they have already paid off their datacenter 5x over compared to cloud. For offline, batch training, I don't ses how any amount of risk could offset the savings.
The ideal shape would be a shaded, flat panel perpendicular to the sun right?
This is a common way of thinking. In practice this type of thing is more like optimizing flop allocation. Surely with an infinite compute and parameter budget you could have a better model with more intensive operations.
Another thing to consider is that transformers are very general computers. You can encode many many more complex architectures in simpler, multi layer transformers.
I've never seen this argument brought up by anyone serious, not in the above post, not in the space datacenter blog by Google, etc.
The main benefit is that solar panels go from a complicated mess of batteries + permitting to a very stable, highly efficient energy source.
> The panels suffer radiation damage they don't suffer on Earth.
I don't think this is true, Starlink satellites have an orbital lifetime of 5-7 years, and GPUs themselves are much more sensitive than solar panels for rad damage. I'd guess the limiting factor is GPU lifetime, so as long as your energy savings outpace the slightly faster gpu depreciation (maybe from 5 -> 3 years) plus cost of launch, it would be economical.
I've said this elsewhere, but based on my envelope math, the cost of launch is the main bottleneck and I think considerably more difficult to solve than any of the other negatives. Even shielding from radiation is a weight issue. Unfortunately all the comments here on HN are focused on the wrong, irrelevant issues like talking about convection in space.
What you're describing is one of two mechanisms of shedding heat which is convection, heating up the environment. What the long comment above is describing is a _completely_ different mechanism, radiation, which is __more__ efficient in a vacuum. They are different things that you are mixing up.
Whats your definition for reasonable temp? my envelope math tells me at 82 celsius (right before h100s start to throttle) you'd need about 1.5x the surface area for radiators. Not exactly back to back, but even 3x surface area is reasonable.
Also this assumes a flat surface on both sides. Another commenter in this thread brought up a pyramid shape which could work.
Finally, these gpus are design for earth data centers where power is limited and heat sinks are abundant. In the case of space data centers you can imagine we get better radiators or silicon that runs hotter. Crypto miners often run asics very hot.
I just don't understand why every time this topic is brought up, everyone on HN wants to die on the hill that cooling is not possible. It is?? the primary issue if you do the math is clearly the cost of launch.
Interesting, That could surely simplify things.