Their assertion that cURL is used to fetch the web page you're reading seems...misplaced. Browsers almost certainly do not use libcurl to fetch web pages. Though I do agree that libcurl is ubiquitous.
Had to over-simplify things a bit on cURL to keep this legible to a general audience, particularly in the first few paragraphs. The idea is that at some point down the stack from your request, libcurl is almost certainly involved (as a dependency, or maybe installed on your internet router, for example), not that your browser is necessarily running libcurl code directly in making the request. But yes, it's slightly hand-wavy in service of the larger point ("this is a thing you use all of the time without realizing it")
1) They are investing Loony Toons levels of money and using Loony Toons financing a lot of the time for DC buildout. Those are actual physical buildings that once built out, exist and don’t need to be built again.
2) They are pointedly ignoring FPGA and ASIC. With the current model quality, would it really be so bad to burn Claude irrevocably on a chip and have a non-modifiable, cheap to mass produce, order of magnitude faster Claude-in-a-chip? This is what happened to Bitcoin ultimately, there are huge performance gains we know exist lying on the table just because they don’t exist for training. And even for training, TPUs make the first steps in that direction.
Startups burning through VC money to build massive warehouses full of hardware those models need to run on currently seem to be quite power grid connector bottlenecked. A more efficient chip would help.
True, but only because they are choosing to actively burn VC money on training.
But it’s not like the models we have now would stop to exist if training stopped. Other than the occasional retraining to get the latest data in, if they stopped wanton experimentation with models, that admittedly is pushing the models forward, the training costs could plummet and inference would be the thing to optimize and scale.
In terms of software projects themselves, consulting companies do, the same that now are pushing for AI solutions, which are the trenches I find myself in.
Now for the whole economic chaos, unfortunately the whole society will, but not the big tech folks.
IT outsourcing was arguably a sort of precursor to vibecoding, and consulting companies are arguably still eating well cleaning up the legendary mess it created.
I work in Salesforce consulting ecosystem, notorious for producing awful code (occasionally passing project until you find someone code for $1 per hour (I kid you not)).
Now within last year it improved to point I can trust it produce decent quality that beats like 95% of human slop.
Vibecoding for anything beyond prototypes is a bad idea. Software engineering principles still matter. Good software that is maintainable and has quality in functional, non-functional, and operational axes still matters.
I wrote about this a few weeks ago in my blog. As an industry, we need to move past vibecoding and into rightcoding.
21 comments
[ 0.22 ms ] story [ 50.3 ms ] threadTheir assertion that cURL is used to fetch the web page you're reading seems...misplaced. Browsers almost certainly do not use libcurl to fetch web pages. Though I do agree that libcurl is ubiquitous.
And those prices are still subsidized. OpenAI, Anthropic, etc are nowhere close to generating profits.
1) They are investing Loony Toons levels of money and using Loony Toons financing a lot of the time for DC buildout. Those are actual physical buildings that once built out, exist and don’t need to be built again.
2) They are pointedly ignoring FPGA and ASIC. With the current model quality, would it really be so bad to burn Claude irrevocably on a chip and have a non-modifiable, cheap to mass produce, order of magnitude faster Claude-in-a-chip? This is what happened to Bitcoin ultimately, there are huge performance gains we know exist lying on the table just because they don’t exist for training. And even for training, TPUs make the first steps in that direction.
Startups burning through VC money to build massive warehouses full of hardware those models need to run on currently seem to be quite power grid connector bottlenecked. A more efficient chip would help.
The buildings will last for years, decades even, but the GPUs need to be replaced every 2-3 years in perpetuity.
But it’s not like the models we have now would stop to exist if training stopped. Other than the occasional retraining to get the latest data in, if they stopped wanton experimentation with models, that admittedly is pushing the models forward, the training costs could plummet and inference would be the thing to optimize and scale.
Other features can easily be copied (eg Claude Code, ChatGPT, etc).
If AI services are a commodity then they can only compete on price.
Now for the whole economic chaos, unfortunately the whole society will, but not the big tech folks.
Now within last year it improved to point I can trust it produce decent quality that beats like 95% of human slop.
I wrote about this a few weeks ago in my blog. As an industry, we need to move past vibecoding and into rightcoding.
https://sgopalanbtg.substack.com/p/andrej-karpathy-didnt-mea...