They mean physical distance, not "address distance".
And in the general case, memory reads do depend on distance. Hence:
> In practice there are multiple levels of cache, and cached reads will be much, much faster than main memory reads.
This is why we have the cache levels in the first place. And the hierarchy continues: L1, L2, RAM, block of spinning rust in the closet, "the cloud" somewhere... There is only so much storage you can fit close to core! More = physically bigger = physically further away.
Sure, within a level might feel like constant time. But that's because the optimal case for that level is slowed down to match the suboptimal cases. The step discontinuities in access time are caused by our inability to engineer a system to manage this tradeoff continuously, not a fundamental law.
Cache is not fast because it's physically close to the core. Well okay, closeness does matter but it doesn't matter as much as the fact that cache is static, i.e. it's (effectively) made of flip-flops.
If you build a computer with DRAM 10 cm away from the CPU and static RAM 50 cm away, access times for the static RAM will still be much lower than access times for the DRAM.
Why don't we just build all RAM as static? Because it's more expensive per bit, it uses more energy, and it requires a great deal more silicon area.
The problem here is that from his family's perspective, it isn't self hosted - just hosted by a less powerful, likely less competent feudal lord than they're used to.
>I had this vision of my entire family using the applications I installed.
>occasionally asking me to install something new
If the model doesn't give them sovereignty, why should they care?
Calling scipy.spatial.Delaunay feels about the same level of abstraction as telling an LLM "write me a routine to calculate Delauny triangulation", except the LLM can do it in any language, and I can trivially tweak the algorithm if I need to. If anything, using LLMs to generate code feels like going down a level, because I can now be bothered implementing things from scratch where previously I would simply have used a library.
You can't convince me that shepherding an LLM to re-implement a library function involves less "engineering" than just... calling the function.
SaaS inference services can just spy on you passively, the old fashioned way; no need to bake nefarious behavior into the model itself, except perhaps to make it 'curious'.
"Hologram" is one of those words that has a wide variety of overlapping meanings. You can be a strict prescriptionist about it and insist that a real hologram involves laser light, interference patterns, 3D scene replication, phase preservation etc, but there will always be something that doesn't quite meet your definition that nevertheless is called a "hologram" by large numbers of people, even expert optical scientists:
A digital holographic stereogram prints an array of hogels (holographic pixels); the hogels themselves are holograms, but are mutually incoherent across the plate. When implemented to the highest level, the result is perceptually and behaviorally indistinguishable from an analog hologram.
A rainbow hologram (such as are found on credit cards) has single parallax only - it certainly does not capture all the properties of any particular electromagnetic field. No purity here, it's just a nifty tech for making things look 3D that happens to exploit wave interference. And everyone calls those holograms.
A scratch hologram is, to a first approximation, an extremely coarse version of a rainbow hologram, made by hand by engraving curved grooves in a reflective material. Now we're away from lasers and wave interference, and for this we sacrifice color - but geometrically we are still dealing with directional microreflections on a surface. Hologram? Depends who you ask.
Lenticular holograms use refraction rather than diffraction to achieve essentially the same imaging model as a rainbow hologram, but in a decidedly less finicky way, at the expense of some resolution. The same imaging data could be used to make a lenticular or a rainbow hologram. No lasers, no diffraction. Hologram? Depends who you ask...
My advice, as someone who has worked with holograms for a decade, is to give up. Saying "AKTUALLY Tupac/Abba isn't a hologram" gets old after the 1000th time. In the real world, "hologram" means "intangible optical simulacra", regardless of how it was achieved. If you want more precision than that, you have to specify the actual technology - Denisyuk, rainbow, lenticular, etc...
When someone asks "can you make a hologram of this", you don't lecture them on optical physics. You say "sure, you want the expensive kind or the plastic kind?"
Having also spent a long time in actual holography, I would agree with your assessment generally and your catalog of different "holographic" techniques.
I might quibble with lenticular displays, since they have a name that predates holography.
In the ideal world, a larger umbrella term would have emerged. "Autostereoscopic" describes any display that provides different views without glasses, but that typically suggests discrete views.
"Plenogram" would have been reasonable following Adelson and Bergen's plenoptic function, but that didn't and won't happen.
"Hologram" now basically means a sufficiently advanced three-dimensional display in popular culture. And that even leaves of other real holograms, such as the ones used for non-destructive testing.
Fascinating, but I do wonder if you're not being the pedantic one here. All of these depend on reproducing the phase of a wavefront. Even a traditional hologram fails this purity test - you don't get a full intensity + phase map without taking multiple images with different phase of reference beam.
>Let people send you their 3D models or 3D scenes and you'll send them a hologram of it.
Oh hey, I have worked in exactly this business (taking people's random 3D models and bullying a rendering pipeline into feeding a digital holographic printer). It's hard, for many reasons. Mainly it's that holograms are expensive in materials, labor, and capital outlay (adding up to several k just for something the size of a piece of paper), and mass production of digital holograms isn't really possible without massive further outlay - writing them is time consuming, and copying them is fraught. It's like being a graphic design / print workshop, except everything is way more fiddly and expensive and nobody wants your product.
Analog holograms are a bit more tractable. But that's even more niche on the consumption end.
Biggest danger of LLMs as they currently fit into society: the radical consolidation of information flow through corporate-controlled bottlenecks that make current generation algorithmic social media (rancid as it is) look like child's play. When people no longer "search" for things, but simply "ask ChatGPT", the lever of social control is nearly absolute.
(Of course, all this is up in the air. If LLMs become commodities, the situation becomes very complex; the information flow doesn't get quite so consolidated, but it still becomes impossible to tractably analyze, as all the LLMs train off each other)
Why inject such disingenuous chaff into the discussion?
Two options:
- You genuinely believe that all living organisms have equal moral weight. You are either a Jainist who sweeps the floor in front of your feet to avoid treading on an insect, or else a callous murderer who would dispatch an annoying human as easily as an annoying fly.
- You are a sane person who does not in fact believe that, and this comment was some sort of bad faith "gotcha".
And in the general case, memory reads do depend on distance. Hence:
> In practice there are multiple levels of cache, and cached reads will be much, much faster than main memory reads.
This is why we have the cache levels in the first place. And the hierarchy continues: L1, L2, RAM, block of spinning rust in the closet, "the cloud" somewhere... There is only so much storage you can fit close to core! More = physically bigger = physically further away.
Sure, within a level might feel like constant time. But that's because the optimal case for that level is slowed down to match the suboptimal cases. The step discontinuities in access time are caused by our inability to engineer a system to manage this tradeoff continuously, not a fundamental law.
reply