[Silicon photonics and co-packaged optics: what is inside the box?](outside5sigma.com/tutorial/s…) Great photo from @sokol_cc for the greatest intel 100G CWDM4 SiPho transceivers, and cover image is also intel's optical engine. I should put in more quiz next time to make it a proper tutorial.

this is kind of what quantum people mean when they say harvest now decrypt later

if i were fable or astra, id tell investors consulting me to invest in companies like this

the famous fruit fly connectome but it is optimizing a coherent beam combining phased array laser system to shoot another fly

for entertainment, do not take these modes seriously (outside5sigma.com/mode-guess…)

made a photonic mode overlap guessr and this shit is hard

do you see what i am seeing

did you see the liquid residual near the bottom middle of the chip? why is it perfectly symmetric? it really got me thinking for a bit, and then I realized, wait a sec, this whole picture is symmetric, they mirrored the whole chip lmao ( they actually mirrored it twice, and really made it look like they had 12 lasers fan-in into two sets of waveguides at a glance, with PWB.. ) [Dream Photonics - Laser integration](dreamphotonics.com/laser-int…)

GPT-6 converted what Dario said about ai into about fiber to chip coupling, and suddenly maybe he's right about pacing the frontier, or maybe the photonics industry is doing so bad it out-retarded the take. here you go: I agree with that much more than I disagree with it. It’s an interesting problem because if you talk to any individual company, they’ll say, our grating coupler is good, or our edge coupler is good, or we have a passive-alignment process. And they’re not necessarily wrong. The problem is the dynamic of the industry as a whole. I think people often talk about this in terms of numbers. Is it 1 dB? Is it 0.5 dB? Is it ±1 micron? Is it ±0.5 micron? That’s an easy way to express whether a packaging approach is good. But I think it’s more illuminating to ask: how do you decompose the problem? What are the paths in which this goes well? And what are the paths in which this goes poorly? Instead of saying there’s a 70 percent chance that edge coupling wins, which makes it sound like a roll of the dice, I think it’s much more useful to say that it forks into different paths. There’s a path where you get a very good edge coupler, but it requires a perfect facet, submicron active alignment, an expensive lensed fiber, and an epoxy process that takes five minutes per channel. That is a good coupler. It is not necessarily a good product. There’s another path where maybe the optical loss is slightly worse, but you have a mode-field diameter that matches something the fiber industry already knows how to manufacture, enough alignment tolerance for passive assembly, controlled adhesive shrinkage, thermal stability, and the ability to test the interface before permanently attaching it. That may be the much better path. The same thing is true for grating couplers. You can optimize peak efficiency at exactly 1310 nanometers, exactly the right polarization, exactly the right fiber angle, on exactly one wafer. Or you can optimize the whole system: bandwidth, wafer variation, polarization, back-reflection, fiber height, angular tolerance, process bias, testability, and whether somebody can actually attach a 16-fiber array to it at high volume. Those are very different optimization problems. So let’s focus on our agency and our ability to take the right paths instead of the wrong paths. I think the way to take the right path is for the industry to work together. The foundries know something about the coupler. The fiber companies know something about the mode field. The connector companies know something about ferrules and mechanical tolerances. The packaging houses know what actually survives dispense, cure, reflow, temperature cycling, and vibration. The transceiver companies know which half-dB is actually worth paying for. But no one player sees the whole stack. People will compete. They should compete. There will be proprietary couplers and proprietary attachment techniques. But at the end of the day, the industry has to work together. Because if every company invents a different optical interface, a different fiber pitch, a different spot size, a different fiber-array geometry, a different datum scheme, and a different attachment process, then we’re going to spend the next ten years solving the same packaging problem fifty different times. And it’s harder, but to some extent, the whole world has to work together. The photonics foundry may be in one country. The fiber array may come from another. The ferrule may come from another. The assembly equipment may come from another. The laser may come from another. And the final module may be assembled somewhere else entirely. So there has to be some convergence on what a good optical interface actually looks like. We need to stop asking only, what is the lowest coupling loss anyone has demonstrated? We should be asking: what interface can we manufacture a hundred million times? What survives ten years? What can tolerate realistic wafer and assembly variation? What can be inspected? What can be reworked? What can be tested before final attach? What does not require a PhD student and a six-axis stage to couple light into it? If we take the right path, I think the probability that photonic packaging becomes almost boring is very high. If we take the wrong path, then we can continue publishing 0.3-dB couplers while shipping products with several dB of practical coupling loss and enormous packaging cost for a very, very long time.

ai experts probably could come up with metrics and estimations on how many tokens are needed to "align" an LLM to what degree, and then they just need to do the same for a group of humans, and figure out what kinds and forms and amount of propaganda do you need to blast into people's eyes and ears for how long to "align" the pace of frontier ai