moregrist

You wouldn’t need five tables. Each trit takes up two bits when unpacked into 0,1,2 values.

You can do a full unpacking-via-lookup with a uint16[256] and then do bit shifting and masking to extract the individual trits, but using an extra byte in each entry (or 3 tables) would let you extract with just two shifts.

This starts to vary a lot with the microarchitecture, and there’s the added dimension of SIMD vectorization, so accurate timing in a realistic context becomes important.

I think it’s pretty wishful thinking to believe that the Iran war was some complex N-dimensional chess to make Chinese oil costs higher.

Even if the current President was known for his strategic global thinking and ability to keep quiet about the “real goals,” other explanations seem more likely.

Pretty much.

For all the current US administration has complained about the opposition being “socialist,” they’ve certainly gone all-in on the state partially owning private companies.

Almost like cries of “socialism” have become a dog whistle instead of what the term actually means.

My point is not that hard to understand.

Have you done any serious graphics programming? Even at the OpenGL 1.x level? What you’re saying just doesn’t make sense.

Just because you’re rotating and translating things in 3-space doesn’t negate that you have a stack of transforms that relate a point in world space to one on screen space and you want to be able to project from one to the other.

Nor does it make it any easier when you need to think about how to stack transforms to achieve effects like rendering a mirror.

I honed a lot of useful practical skill with linear algebra trying to get graphics to do what I wanted. And I say this as someone who’s spent the bulk of my career using linear algebra in the context of quantum mechanics, physical simulation, and ML-adjacent areas.

PCA is an orthogonal transformation of the covariance matrix, so like all orthogonal transformations, it’s _literally a rotation_ in N-dimensional space.

SVD is more complex but ultimately it’s just another useful decomposition of a matrix.

I’m not sure why you’re both negative and dismissive. Transformation matrices in graphics are a good and approachable way to get used to linear transformations, which turn out to be useful pretty much everywhere.

Whether or not that helps you with ML depends more on what you’re doing in ML. FAANG doesn’t have a monopoly on ML or on interesting work in ML.

It has always been overpriced and had huge margins.

This is the engineer’s take on things. I am entirely sympathetic to it.

I also think it missed a lot of what management values in consulting. At its best, you can offload a lot of things unrelated to your business to people who are experts. At its worst, you’ve offloaded the blame to a group of over-worked twenty-something’s with impressive degrees who have no idea what they’re doing, but who sound really fucking confident about it.

Can an internal team do it better? Probably. Will they be cheaper? Probably. Will they assuage management’s anxieties and deflect some/all of the blame? Nope, not at all.

They hire juniors because they have junior level tasks that need completed.

I have never worked at a place where this was true. Either senior devs would pound through the tasks, or we’d cut them as unimportant.

The only reason we ever hired a junior was because we saw potential and thought they could grow into solid colleagues.

Claiming “50-100 years” is a misleading and hand-waving way of saying “futuristic.”

It tries to get you to imagine that advances in the last 50-100 years will project linearly into advances in the next 50-100 years.

This is not generally the way that science and medicine work. Even if you add in gobs of questionable data collected by companies with a bad track record of doing right by it.

They’re essentially trying to get you to believe that AI + your data will give you the kind of step change in medicine that we got from penicillin and X-rays/MRI/CT imaging. It’s a cheap rhetorical trick.

“There is no reason anyone would want a computer in their home." - Ken Olson, Founder of Digital Equipment Corporation, in 1977

I don't suppose you have experience with how badly small non-profits are fucked by every tech consultancy they ever work with?

Definitely with you here.

Getting nonprofits into AI that feels even marginally more self-serve

Umm... so your plan to make non-profits less fucked is to give them yet another consultancy, but this one is AI!

I am dubious that this results in them feeling less fucked.

Regardless of how you think about LLMs (I do find them useful), there’s something really odd to think that you can select for “proven experience” in a young technology where current experience appears to have little to do with experience 15 months ago, and where the biggest boosters fully claim it will have nothing to do with experience in 15 months.

What you’re selecting for is enthusiasm, knowing the current shibboleths of the in-group, and possibly for who knows how to use them to make a good demo.

And, fair enough, if that's what you want. But it's not "proven experience" in my mind.

I love that we’re already talking about “proven experience” for a technology that’s essentially 15 months old, arguably only broke into the mainstream 3-6 months ago, has an unclear RoI for many companies, and seems to be changing quickly in both cost and “best practices.”

You’re more or less admitting that you’re playing trendy tech lottery. Which is fine, but maybe not generalizable to the whole industry.

For any practical application, you are only interested in finite set of concrete identities

I do a lot of numerical work in settings where computational efficiency is useful.

In my work, most cases you can do numerically using integration or Monte Carlo sampling or whatever.

It’s slow. It often pays to find a closed-form solution. Even if it’s just a starting point that needs refinement.

To put in terms of the Pythagorean theorem: Proving the Pythagorean theorem gives you a relationship that’s reliable, fast to evaluate, and general. Proving individual tuples gives you none of this.

That doesn’t even touch on how theorems give us a glimpse at deeper structure and truths. Proving a bunch of right-triangle tuples will probably never lead you to the rest of the identities in trig.

The tcl syntax is fine. And modern tcl is fine.

But tcl 7.x and before was a pure string-based language. Everything was essentially a eval(). People would hit syntax errors on production code.

Fun, painful times.

The flip side: the interpreter is super simple and fun to write.

Why the negativity?

If someone wants to have fun in COBOL, let them have fun in COBOL.

If it’s agentic fun, that’s cool. If it’s an interest in the language, that’s cool. It’s not like you have to have an ROI for a fun side project.

You bring in contractors.

Ideally contractors that benefit you personally (eg: your buddy who now owes you one), but definitely contractors that let you outsource the responsibility.

Even better if you get some management consultant to suggest the idea and/or do the subcontracting.

Definitely buys you a few quarters of bonus and some time to land your next gig.

Sometimes you get lucky and it both looks up the documentation and then ignores it and makes stuff up.

This more or less agrees with my assessment of recent changes in Claude Code where a lot of new features are either:

- A lot of half-baked features or half-done features. - Or have significant overlap with existing features, and aren’t clearly an improvement.

More code is not better. More features are not better. It would be lovely to see more intentional design than just more.

I know they’re dog fooding this. I have to believe they have some people with taste. So it makes me wonder if anyone has the time to think or if they’re just shoveling prompts as fast as possible.

Marc Andreessen has a strong financial incentive to feel this way and to convince others to feel this way.

I also think it’s easy to think that AI gives good answers if you don’t know the field well. In fields where I know the material, the answers are pretty variable and can be quite bad.

Seems like there should be a market for a no-Elon/OpenAI/Anthropic ETF out there.

Or one that just imposes a reasonable waiting period on adding newly-IPO’d listings.

The price of SSDs is similarly depressing.

I didn’t see anything about LLMs here.

If you’ve never written or worked in a Forth-like language, it’s not a hard system to bootstrap up. If you’ve done it before and know assembly, you can even get something that compiles to (stack-heavy and pretty unoptimized) native code in essentially a weekend. No LLM needed.

Forth-likes are almost magical in ways that are hard to describe. You start with primitives and literally build the language out of them. The interpreter and compiler are two different modes of the same REPL loop.

It’s just a very different paradigm than most programmers know.

It’s not about that.

It’s the constant drumbeat of “AI will take your job.”

It’s the constant news of “layoffs because AI makes us more productive.”

It’s the constant background discussion of UBI because no one will have jobs anymore.

It’s knowing that, in the US, UBI will never come.

It’s the feeling that the billionaires of Silicon Valley are getting rich and there isn’t even a “learn to code” path to wealth anymore.

It’s knowing that data centers will create problems in your neighborhood: the price of power and water will go up, the amount of undeveloped land down, and you don’t even get jobs out of it.

For fuck’s sake, it’s not about the thousands of Mag7 tech workers losing their jobs. That’s just a symptom, like all the other symptoms, of this weirdly dystopian future that the AI companies keep telling us is inevitable.

The number of businesses and business departments that run on spreadsheets and earn money is almost mind-boggling.

It works until it doesn’t. The failure mode can be that the spreadsheet wizard leaves and no one understands their macros, the data grows 10x and emailing spreadsheets back and forth becomes too error prone, or any other of the well-known ways that this falls apart.

Those of us in software always cringe and want to use a database.

“Vibe-coded apps are the new spreadsheet.” Seems about right, and the problems will be similar.

Edit: This comes across as more negative than I wanted. I think spreadsheet-driven businesses are pretty amazing, and it's a testament to the tool how far they can get. I pretty much feel that way about bespoke vibe-coded apps. I'd feel even better about it if the rhetoric wasn't mixed in with claiming software as a career is over or that no one will have jobs anymore.

Non-cynically: the frontier providers have a projection for demand.

Cynically: it’s become an executive-level gpu measuring contest. If you’re not making huge commitments on data centers, you can’t be a serious player.

Realistically: It’s a mix of the two. The recent Claude caps for agentic usage suggest that demand exceeded their immediate compute supply. That they can alleviate it with additional capacity from the existing and small-ish xAI facility suggests that either demand may not be rising quite as fast as anticipated, that they’re okay in the short term until more capacity comes online, or a mix of both.

Open questions:

1. At what price point does demand fall, and are the frontier providers overall profitable before that price point?

2. At what price/performance point do on-prem local models make more sense than cloud models?

Besides the people in this thread bemoaning the state of research funding, international students, etc. (all of which are valid), a lot of people are becoming disillusioned with academia.

This is not disconnected. It is also not new. People have been disillusioned with academia since there were students.

Probably 80% of the recent PhD grads I know are looking to leave academia, despite the fact that they went into it to pursue a career in academia.

It’s very hard to make a sweeping statement like this. PhDs are segmented by field and subfield.

Almost everyone entering a Ph.D. program does it to have the option of going to academia. It’s a _research_ degree. Unlike a JD or an MD it doesn’t lead to a licensed profession. Or even a job.

But in some fields (eg: chemistry and many areas of biology), 80% of grads have ended up in industry for decades. There’s also a long tradition of Nobel Prizes going to people in industry, so it’s not viewed as a second-rate choice.

The median science PhD takes 6 years now, and is grueling work for terrible pay, all for difficult job prospects given the current market.

It’s true that the pay is (relatively) bad. I liked to think of it as an incentive to graduate, but then I did a postdoc for similarly bad pay before leaving for industry, so maybe it wasn’t enough of an incentive.

But the length has been 6ish years in a good portion of the physical and biological sciences for a couple of decades.

I wouldn’t call the work “grueling.” In most fields you’re doing lab work or desk work, not manual labor, and while the hours can be long, at the end of the day it’s driven a lot by the a startup-like mentality: this is your career and you get what you put into it.

MIT recently became one of the first universities to get a grad student union to try and combat the increasingly exploitative nature of academia.

MIT is not a thought leader here. Unions have been a thing since at least the mid 2010s at a number of Ivy’s, and various University of California schools have had a union since the early-to-mid 2000s.

I can see how undergrads may look <things> and decide that they don't want to continue down that path.

It’s a valid choice. It’s been a valid choice. This has nothing to do with AI. You start a PhD to pursue original research (fsvo original), and that’s _hard_. It’s always been hard. It didn’t get hard last year.

Depending on the field, the job market has been bad for decades, too. Humanities fields are always a bear market. There used to be blogs about leaving for industry in history in the 2000s. In the 90s you’d hear cautionary tales about someone’s uncle had a PhD in physics and was now managing an Arby’s.

Departments could do a much better job with prepping graduates for industry. Successfully completing a PhD comes with a lot of hard-won skills that transfer to industry. And it would help if faculty didn’t view it as “giving up.”

But this is a long-running problem. I don’t think the undergrad zeitgeist has changed. I think the current administration has cut funding and closed off the immigration pipeline. We’ll be feeling those effects for a long time.

Why do people rebuild classic cars or remodel old houses by themselves? Because they enjoy the work itself and the learning process.

You don’t have to. It’s perfectly okay to take your modern car to a mechanic and hire a contractor for your remodeling needs.

But some people like to do it themselves, even when the project is large.

Have you ever done blacksmithing? It’s tremendously satisfying.

Sure, if you want 300,000 spoons, it’s far better to use a factory process and get essentially identical results. But if you only want a few spoons and accept (or even value) that the spoons will all be a little different, hand-forging them is quite enjoyable.

I’ve written enough assembly and done enough blacksmithing to know that the metaphor isn’t quite apt. But there’s both tremendous effort and satisfaction involved in both.

While you can't really ask "will I be layed off next year," it's pretty common to ask some version of "why is the role open," usually split among a few questions (that you'd tailor based on the role):

- "Which of my skills do you think are most valuable for this role?"

- "How would you measure success in this role?"

- "Can you tell me a little more about the product lines we'll be developing / supporting?"

- "How is the current team planning to grow?"

These are the kinds of questions that let you feel out what the manager envisions for the role. If the answers seem vague, that tells you something about the role / manager / org. If it's not clear how you impact the product and they can't clarify, that also tells you something.

The number of bugs tends to be linear to lines of code written meaning fewer lines of code for the same functionality will have fewer bugs.

It depends on the code. If you’re comparing code of the same complexity then, sure, 2000 lines will take longer than 200.

I was comparing straight linear code to far more complex code. The bug/line rate will be different and the time to review per line will be different.

Furthermore how do you know the AI generated lines are the open highway lines of code and not the mountain road ones?

Again, it depends on the code. Which was my point.

Linear code lacks branches, loops, indirection, and recursion. That kind of code is easy to reason about and easy to review. The assumptions are inherently local. You still have to be alert and aware to avoid driving into the cornfields.

It’s a different beast than something like a doubly-nested state machine with callbacks, though. There you have to be alert and aware, and it’s inherently much harder to review per line of code.