shash

Technically it’s a retroflex approximant [1] and is found in many places (often not as a separate character or phoneme).

But I think we’ve hijacked a cultural thread with enough phonetics for now!

[1] https://en.wikipedia.org/wiki/Voiced_retroflex_approximant

I'm coming around to the idea that we _should_ really be using "l" or maybe "r" instead of "zh" for the ழ. At least it's closer in pronunciation and there's a chance someone can work their way to it. Zh is like "we don't have an exact match so we'll repurpose a letter we don't use". It has no phonetic relevance.

I think the decline of manual transmission is different from self-driving. Manuals, you could argue are a technological progression that doesn’t change the fundamental economics or sociology of driving. But self-driving has issues far beyond the technology. Like liability, like ownership of vehicles, availability, traffic rules,…

I’m not even sure if, outside of highly mapped environments it even makes sense.

Or it collapses when the seniors have to retire anyway. Who instructs the LLM when there’s nobody who understands the business?

I’m sure the plan is to create a paperclip maximizing company which is fully AI. And the sea turned salty because nobody remembered how to turn it off.

Either that or the “bespoke hand-crafted artisanal free-range code” will be the only thing still maintainable because vibe coders made such a mess

In the Before Times we used to do programming interviews with “you can use Google and stack overflow” for precisely this reason. We weren’t testing for encyclopaedic knowledge - we were testing to see if the candidate could solve a problem.

But the hard part is designing the problem so that it exercises skill.

Over the years it’s kind of becoming clear that “running major businesses” is kind of orthogonal to “having emotional integrity”. In larger businesses it’s mediated by layers. But just take a look at some of the deranged tweetstorms we’ve become used to in recent times.

If only because of the fact that the start menu (equivalent - the dock and applications view) isn’t an ad filled react app.

A good scanning electron microscope costs at most a few million? And is pretty common in a decently funded lab pretty much anywhere? Resolutions of 5nm is not uncommon. A scanning tunnelling microscope can go much lower (single atom types) and isn’t all that much more expensive either (comparatively I mean).

I think it’s common knowledge by now that the smallest feature in a 5nm chip isn’t really 5nm. So that’s not (yet?) a viable strategy.

You mean you’d rather run unverified scripts using a good order of magnitude more resources with a slower experience and have an entire sandboxing contraption to keep said unverified scripts from doing anything to your machine…

I know the browser is convenient, but frankly, its been a horror show of resource usage and vulnerabilities and pathetic performance

As one of those on the skeptical side, one train of thought I have not seen people even mention is, the way we’re using LLMs to code now is largely to use a less precise language (mostly English) to specify what’s often a very precise problem and solution. Why would we think that spoken language is the best interface for doing this?

I’m wondering if we can do something better…

It’s usually easier to explain the dft. and easier to do a periodic function than a totally arbitrary sequence.

They already do offer that - it’s called a multi-project wafer or MPW. But it’s prohibitively expensive on a per-chip basis. It’s mostly used for prototyping or concept proving and not for commercial use.

One problem is, you need to create a photolithography mask set for any volume size of fabrication and those aren’t cheap. But that’s far from the _only_ problem with small volume.

Finicky chemicals and relatively expensive equipment. But he’s founded a company with Jim Keller. We occasionally see them post a photo with zero context, but we do know some things. Like they are targeting lots volume stuff and basically building fab equipment. But not much more.

That would make sense _except_ the amount of time and specialized knowledge it takes to build one. Easier by _far_ to go deal with TSMC and Samsung

Each die gets tested. Not that they’re testing their stuff with these dies.

If stored in proper conditions, the shelf life of undiced wafers is pretty much limitless. But if moisture or dust get in, it’ll start to have corrosion and other damage. Diced wafers on the other hand are placed on UV tape and I recently learnt that the shelf life of that is like a few months at best.

Nevertheless, unless you’re not storing them in a clean room of appropriate class (or in a vacuum pack), assume 18 months for undiced wafers.

In terms of size, I would think that stored in cassettes or carriers a vacuum sealed, it shouldn’t be too much. A 25 wafer pack is about half meter cube? A 40ft shipping container is about 68m^3. So one container can store about 140 packs of 25 wafers each. That’s 3500 wafers. They’re talking about 900k wafers. That’s about 260 containers. Call it 300 with extra space and stuff. Not exactly hard to provision.

I’ve done that too. In essence it kinda sorta comes down to a small convolution kernel with learned weights.

In some places it works really well.

Ok. Those are impressive results. Nice addition to the toolbox

OCR is one of those places where you can just skip algorithm discovery and go straight to deep learning. But there are precious few of those kinds of places actually.

I’ve found exactly the opposite. In domain after domain the performance of a pure deep learning method is orders of magnitude less than that of either a traditional algorithm or a combination.

And often the CNNs are so finicky about noise or distortion that you need something as an input stage to clean up the data.

Not to mention performance. So often, the traditional method is the only thing that can keep up with performance requirements without needing massive hardware upgrades.

Counter intuitively, I’ve often found that CNNs are worse at thresholding in many circumstances than a simple otsu or adaptive threshold. My usual technique is to use the least complex algorithm and work my way up the ladder only when needed.

In general avoid frivolous use of pointer arithmetic. foo[k] and *(foo+k) will usually generate identical asm, and the former is just easier to read…

Yeah, the point here is nuanced, but let's do an argumentum ad absurdum instead /s

Probably the _best_ advice I got from the entire startup community is, "don't scale without PMF". The saving grace here is that they didn't scale beyond 140 people, so the damage is limited. And they didn't double down on something that would have dropped the whole thing down.

But as a founder, I'd consider it a failure of planning on my part if I had to lay off 50% of my workforce (a failure nobody is immune to, but a failure nonetheless).

It’s interesting that not a single hardware concept is actually explored - nothing like a real alternative architecture or how one would program it. Just a lot of complaints about standard (sounds like x86 only at that) techniques without much depth.

Very similar. To the extent that when I’ve worked on synthetic aperture I have used ideas from radio astronomy.

And there’s a similar field of synthetic aperture ultrasound which is where I did a ton of work a decade ago.

Same guy.

His work is pretty amazing actually. Lots of little details he’s worked out and built upon. Someday I’d like to replicate his designs.

The original sin was probably hiring those people for a not-yet-clear business in the first place…

I mean, you can’t even see a single frame pointer register. The srams dominate everything else.

Even the functional units are relatively cheap.

AS PRACTICE. YOU HAVE TO START OUT LEARNING TO BELIEVE THE LITTLE LIES

Edit: came for the Sagan, found the Pratchett!