crazygringo

I'm not really following. My point here is that this is why the ceiling is around 20x, as opposed to being e.g. 200x or 2,000x.

That doesn't mean it's "much lower than 5x."

I mean, just what I did this morning with Claude would have taken me 2 weeks on my own. So 20x is pretty reasonable for certain types of engineering.

But it's also true that when I'm working on 2 things at once, the bottleneck is me. I'm spending all my time giving Claude direction that it can't give itself.

So you basically have to separate out the tasks where AI can produce a 20x productivity boost, and the stuff where it can't help at all.

There's still so much stuff where it can't help at all, you need real human employees for that. That's why Anthropic is continuing to hire at breakneck pace.

And that's what I'm saying -- knowledge workers aren't being replaced. They're just accomplishing tons more work, and companies aren't running out of work to do.

Imagine an employee is 1x productive without AI.

Now spend 0.5x their salary on AI, and they become 5x productive compared to no AI.

Now spend 1.1x their salary on AI, and they become 20x productive compared to no AI.

It's still obviously worth it to spend the 1.1x. But it is truly remarkable that the biggest cost of hiring a knowledge worker is no longer their salary, but their AI usage.

There is no implication here that employees are being replaced by AI spend, as long as all of the productivity is generating profit. The AI usage still needs a knowledge worker to direct it in ways that it can't direct itself, so you still need the employees.

Loads fine for me, before and now.

Might be a problem with your location, VPN, etc.

Maybe Claude does it in that case as well, but that's not what I'm experiencing. It has no rhyme or reason to it, except maybe some kind of background process. I assumed something like that at first, but then triple-checked and it was absolutely unrelated to anything happening on the LLM side of things.

My best guess is that it's something related to some kind of auto-update with the way it seems to happen roughly but not always daily.

Amazing. I'd LOVE for you to expand it, to also auto-dismiss "Foo.app would like to access data from other apps."

Adobe programs ask for this every time they start, and Claude Code randomly asks me this daily out of nowhere. It doesn't matter how many times I click "Don't Allow", it never gets remembered. (But of course if I ever mistakenly click "Allow" even once, that gets remembered for eternity and I never get asked again...)

Not really sure what the point is in posting some ancient policy from 2008, almost two decades ago?

Policies change and evolve. What would be more notable would be if they didn't.

And in this case, I'm not surprised at all. Trying to be the arbiter of what is considered common place names feels like a thankless and ultimately impossible task. If I were Google, I would quickly give up on that as well and simply follow official sources.

Don't like it? Then vote for the other guy next election. That's what living in a democracy is about. If anything, sticking to government names gives people even more reason to vote for the other party next election. The more people who get exposed to these dumb name changes on Google Maps, the better.

Not really sure how it's any less dystopian than blinking at a screen 2 feet away on your desk, as people have been doing for decades?

To be honest, it feels a lot less dystopian to me. Frees up your body so you can easily alternate between sitting and standing, and move between environments or inside/outside a lot easier.

Like, it's technology adapting itself to our bodies, instead of trying to adapt our bodies to technology by hunching over and developing neck strain and bad backs.

But Wikipedia is an encyclopedia, not a reference for all knowledge.

Thus, it draws the line at things and people that are notable.

Maybe you think it shouldn't be an encyclopedia, but should be some larger project that indexes every product UPC, every person born, and every website ever registered, but that's not what it is.

And the primary obstacle here isn't storage. It's quality. You need a sufficient number of eyeballs and editors per page to ensure that the crowdsourcing is actually passing a minimal quality bar. Millions of short, low-quality articles means millions of opportunity for spam, abuse, and other problems.

Fair enough! I'm lucky enough not to use Teams, I've heard the horror stories but I had no idea it was that bad.

I think the subject here is broader UX latency, measured in hundreds of milliseconds to multiple seconds.

I've been using computers for over 40 years now, and I've never been impacted by the latency between keypress and letter rendering in my life except for on e-ink. That's not something I see anybody complaining about.

No, latency hasn't gotten worse.

It just hasn't gotten better.

Fundamentally, whenever latency gets too bad, developers fix it. But often not until then. So it always just kind of sits there at the edge of just-barely-tolerable. Which is defined by human behavior, not chip speed or similar.

I remember iTunes in the 2000's taking several seconds to respond to user actions. I remember using software like PageMaker in the 1980's that would take forever to respond. Lotus Notes in the 1990's was horrible. Running antivirus would slow your system to a crawl. Other stuff from the 2000's -- Acrobat Reader, or Outlook.

It's always been an issue with apps on the computer.

Oh, I certainly have. In fact, I had to give me kind of a three hour crash course afterwards when I needed to figure out what kind of further optimization was most likely to work.

But that was still to answer very targeted questions that I would never have been able to ask on my own in the first place without months of study.

And when we are talking about the speed up from LLM's, we are generally comparing to a world without LLM's. Not to some mythical world where LLM's are able to teach us about anything but not actually able to write the code or find the technical solutions themselves.

I am a huge fan of having LLM's teach me what I need to properly guide them in domains that are outside my expertise. It's kind of funny when you think about it. A really big part of learning how to be productive with Claude is indeed trying to find precisely the areas where you yourself, not only provide value, but provide irreplaceable value.

But in this case, most people really truly don't have an issue with it.

It's a very small minority who cares.

I mean, it was about changing the size of the certain layers of the model in order to fine-tune with the new shape.

I am entirely sure that it would have taken me many months to figure out that that was the actual thing that needed to change. It wasn't just some kind of tooling configuration issue. It required a deep understanding of the size of every single different layer of the model and how they fed into each other.

I recently needed to adjust an existing open source model for an app that I am building, to get it to work in a certain way. I have no background in machine learning.

Claude was able to identify the approach that worked, and even the long Python command that was needed. In, yes, probably about 10 seconds.

For me to figure that out myself, it probably would've taken something like a year of study. Of the entire stack of machine learning and all of the Python tooling associated with it.

That's where these absolutely incredible astonishing speedups come from.

Because it's one thing for Claude to be able to write code 20 times faster than I can. But it's a whole other thing when it is able to create things in domains that I know nothing about at all. That would take me one or more years to become an expert in. But where I can still just verify its results, so I don't need to be an expert in order to judge the quality or suitability. If the end product works, the end product works.

I have similarly had it come up with mathematical formulas that would probably have taken me a year in going back to studying math if I had to derive them myself. But where it is trivial to show that they are producing the correct output for known test cases.

So trying to come up with some kind of factor multiplier here almost kind of falls apart. It might as well be infinite. Because these are things that I just wouldn't have done otherwise. The software wouldn't have gotten built because I would have had to hire somebody else to do it, and there was no budget for that.

Yeah I'm aware it's hit-or-miss. But I use it in Infuse, which is a third-party app that supports it, even with its own framework.

Yes, this is standard. And the point is performance. The needs of user-facing ID's (short alphanumeric, randomly ordered) tend to be pretty orthogonal to the needs of internal ID's (raw bits, monotonically increasing, maybe timestamp data).

It kind of lives on, on the Apple TV remote.

You know that ring around the main button, that kind of functions like a D-pad with four arrows?

If you're paused while watching something, you can drag your finger clockwise or counter-clockwise around the ring to seek forwards or backwards. It's really nice, much finer-grained control than normal skipping forwards or backwards.

The wheel doesn't actually spin, but it also doesn't really need to.

There are really quality TV shows, and films on streaming, out there.

Giving up TV is no different from giving up reading literature, or listening to music, or whatever artistic experience you want to have.

Watching a TV show like Adolescence is no different from reading great literature. And fortunately, there's so much great TV from the past couple of decades, you don't ever have to watch bad or brainrot TV if you don't want to.

There's nothing noble about trying to abstain from that.

It takes up space, you've got to clean out the dust from the fan grill every so often, it requires constant maintenance in software updates every time you turn it on, unless you leave it on all the time in which case it's using non-trivial power all the time. I have an Xbox. But there's no question that it would be a much simpler setup to not have one.

If you buy the expensive piece of hardware and the game, you can play it on your tablet for free. With Remote Play! You've always been able to do that. Nothing changes.

This is about not buying the expensive piece of hardware. Then obviously it makes sense to charge you for the expensive hardware you're now using in the cloud. Or if you bought the expensive piece of hardware, but for some reason are choosing to use an Xbox in the cloud instead.

Game streaming has never been part of the licensing of purchasing a game. It's always been part of a monthly Game Pass subscription. And with games in Game Pass, you don't even have to buy the game in the first place.

I get that some people may not love the change, but it's just more accurately reflecting costs, and there's absolutely no bait-and-switch here. Nothing has changed about what comes with a game you purchase. Nothing at all.

Seriously! They're not charging you for the "streaming" or for the "game you already bought" -- they're charging you for renting the Xbox in the cloud that is running the game, the GPU.

I get that it was free before as part of a subscription, but now GPU's are more expensive. Nothing that you paid for up-front is being taken away. The subscription plans just changed is all.

I think you may be misunderstanding.

The point is that previously, changing code in a major way might take months of refactoring. Large changes were simply not contemplated unless they were justified by similarly large additional profit. Once you designed your architecture, you were pretty much stuck with it.

Now, the kinds of changes that were simply unrealistic before are now entirely realistic. A refactor/migration that would have taken 6 months before, you can now get a working version done in a day, and another few days of QA. That's what this post means -- code is no longer stone, it's clay. You can change it quickly in ways that were previous unimaginable.

This has nothing to do with the concept of code being liability or asset -- that's a totally separate conversation.

99% of people? Gun control? Career ruining? Life endangering?

You speak in hyperbolics, my friend. You may be lacking some realistic perspective here.

Plenty of people who are using LLM's within technical organizations, as opposed to script kiddies vibe coding, are encountering the same kind of issue. And people spend lots of time fixing bugs in existing code anyways. This is just a simple question of newly-visible tradeoffs.

I mean, it depends on how you use an LLM like Claude.

You can not know what you're doing and use it to write slop.

Or you can set up your project with best practices in every way with full design documents and code that flows from design and full rigor around everything.

I do it the second way. And so yes, I find myself being confronted with so many edge cases that I just never thought about before. And I'd say it about doubles the time it takes me to deliver stuff, and I'm not always sure it's worth it.

Same here. It's been a real eye-opener.

Previously my main concern was with LLM's writing worse code. Now my concern is the ability to write better code than I was ever able to write before -- considering every possible potential bug and race condition and error state and what have you -- but at what cost? How much more time do you have to spend researching, designing, making decisions around every possible edge case and state?

Before, you couldn't even attempt to balance quality with speed, because you couldn't see the quality in advance. You had to wait for bug reports to come in. Now, you're confronting quality up-front in a much deeper way. Which allows for much more reliable software, but then you have to decide how much effort is worth it and how much isn't?

$100/mo. is basically the cost of a cable TV package. Often $120-170, actually.

If you're the kind of person who gets enjoyment being able to code all sorts of things you were never able to before, or 20x faster than you were able to before, then $100/mo. is extremely reasonable in comparison.

I'd argue it's money much better spent than on cable.

Also, if a hobby of yours runs you $100/mo., that's actually very much on the cheap side for a lot of hobbies.

Also, $20/mo. is for hobbyists using it for 2 hours at a time a few times a week. $100/mo. is for people working on a side project every evening and every full weekend.

Because sounds are made by shapes. The shapes of our mouths and the shapes of objects colliding. And sounds have shapes, shapes over time.

The point is, if you don't use regex that frequently, you can completely forget that ^ means start and $ means end. And adding those as line comments ("^ # Start of expression") would be dumb.

That's why a well-designed DSL tries to make its syntax as self-explanatory as possible. Using "start" instead of "^", or "digit" instead of "d".

We're not doing equations here, we're doing engineering. Analogies to proofs are not helpful. There is a very good reason why, in coding, we give descriptive variable names such as "accountBalance" rather than "b".

Conciseness is not a virtue in programming. Clarity is. Things should not be longer than they need to be -- but neither should they be shorter. The name of something should not be 100 characters long, but neither should it be 1 character long. A word or two is generally the sweet spot.