c-fe

will claude platform on aws offer the same nerfed (dumb) opus 4.5/4.6/4.7 of the recent weeks, or will it be possible to use opus 4.5/4.6 from february/march?

This is very similar to SBE encoder/decoder flyweights over raw memory? What are the differences?

Maybe it arrives while the first request is still running. Now your idempotency layer is part of your concurrency control.

I recently designed a system where this had to be taken into consideration. I find my solution very elegant: When the request arrives, I put the pending request into a map, keyed by the idempotenceId. This whole operation is executed in one step. Now the event loop may process other requests. If one of them has the same key, it will await the same response object from the store. And then, once i have the response, I resolve both promises with it.

funnily enough, I talked recently to someone working on the swift compiler (not an Apple employee) to make Swift functions differentiable. So its not all dead yet

In the Write-Only Code story, that same engineer becomes a systems designer, a constraint writer, and a trade-off manager.

This is also what I see my job to be shifting towards, increasingly fast in recent weeks. I wonder how long we will stay in this paradigm, I dont know.

interesting. Do you have an example of what exactly it outputs and how you store that? how do you use the outputs? As a pointer to a new agent with fresh context to get them up to speed?

uninstall the app. Works really well to me. The conscious effort of reinstalling it is enough to prevent me from doing it. Whereas using the awfully implemented screentime guards, I just find myself clicking on "Allow for 15 minutes" before I even understand what I do.

select in finder the file that currently opens with XCode, then press Cmd+i. It opens the information panel. There in the Open with section, you can chose the app and then also Change all to not use XCode.

How is heritabiltity of life span useful if by the time the lifespan becomes known (eg at 80yrs old) the inheritance is not possible anymore (eg menopause)?

Maybe.. or maybe you underestimate the insanities you can find in real life too (the model isnt that creative unfortunately). See here, 5 different no-parking signs for the same 2 spots: https://maps.app.goo.gl/S74r7eawH2vL24CX7

I have to say both the leafless trees and electrical box spawning is very on point for what you would find in eg Belgium. Check this full blown ugly building/container that spawned in the beautiful Liege Guillemins station https://maps.app.goo.gl/T1J7WwCCYDvBgJEc7

Im not sure I like that market in particular, but probably polymarket is indeed the best one… assuming the market will resolve fairly

Exactly. I would prefer to remain invested as I dont want to time the market. But I would prefer if I could meaningfully reduce exposure to OpenAI and the consequences of their possible downfall.

As a retail investor mostly invested into broad ETFs (All World), is there any way I can get short exposure to OpenAI? Being short Oracle/Nvidia/Microsoft?

Hi David, have you looked into alternatives to 3DGS like https://meshsplatting.github.io/ that promise better results and faster training?

I was about to write the same. I scrolled through it but I dont understand what it is.

Same for me. Also never happened to me before

I like this a lot. Every time I am forced to open Notion or Slite, I just wish so much it would just be .md files in a git repository.

one way is by using schemas to communicate between them that are backwards compatible. eg with avro its quite nice

<1MB is also relatively easy to reach with swiftui apps. I had two fully working ones in the app store below 1MB. They are removed now since I didnt pay the yearly 100€

i am still skeptical about HDR as pretty much all HDR content I see online is awful. But this post makes me believe that Lux/Halide can pull of HDR in a way that I will like. I am looking forward to Halide Mk3.

entry level macbook air (m4) is same price as it used to be (in luxembourg): 1 159,36 €

Congrats to you too! Indeed, I think they could have improved the visual and comparison part, its a bit dark and not too interesting to look at. But I am envious of how smooth their tiling is. My tiles are 4096x4096 which allows me to satisfy both the 20,000 file limit and the max 20mb file limit imposed by cloudflare. I had some issues with smaller tiles, and wanting to host it on cloudflare restricted me from doing 512x512 tiles iirc. Also I really like that they extracted the publisher information and put that as a pmtile vector, thats something I attempted but ultimately ran out of time with.

Annas Archive has both books in their archive, but they also have other datasets that connect a book ISBN to the metadata (title, author, publisher, ...).

In my visualisation https://isbnviz.pages.dev you can see which books they actually have the files of (blue) and which ones they know exist because they have the metadata from some other source (like google books, ...) (red). Finally, there are also ISBNs not contained in any of the sets that Annas Archive has, and these are either assigned or not assigned. A lot of the 979 prefixed ISBNs are not assigned, that means, no country/publisher has the right to assign them to a book. Other ISBNs are assigned to a publisher, but they just haven't published a book with that ISBN yet. Or they may have published a book, but Anna's archive doesnt know about the book because its not in their (or the ones they scraped) dataset.

Thanks! That is indeed all thanks to using the hilbert curve fractal which has the property that it maps numbers which are close together onto 2d (or higher dimensional) coordinates which are close together, its a very cool property! Its used in lots of contexts for that reason

Its due to how those ISBN ranges were handed out - I think they probably gave a block like 978-53 (for example) to those countries, meaning the right to distributed ISBNs 978-530-000-000 to 978-539-999-999 and then later they ran out or had all subblocks distributed to publishers, and then they got a new block further away (so not 978-54 in my example) and therefore those blocks are not numerically close to each other and thus also they are separate "islands" in the hilbert space.

Im slightly surprised mine won 3rd place, I believe they liked my simplicity and visualisation. Hosted at https://isbnviz.pages.dev

But honestly, I find both of these better: - https://bwv-1011.github.io/isbn-viewer/ - https://anna.candyland.page/map-sample.html

in particular the one from bwv is technically similar but just all around better than mine, it is what I would want mine to be

I think the language problem will become less of a problem in the future due to (1) more (young) people living in citys and (2) all young people in cities speaking english. At least compared to previous generations imo. This could be my subjective view based on luxembourg, netherlands, and visiting other european cities.

This is close to the idea of convex splatting (recent paper) in which convex shapes are used to approximate these real 3d objects as they are better suited than gaussians

I sometimes reminisce about the 2010-2020 deep learning and reinforcement learning etc era, as a student I did some projects in that domain back then and it felt very approachable and relatively easy to get into it compared to how I (most likely subjectively) see it today as a developer (in a different field).

I remember I doing a project in my second bachelor semester, where I generated random 64x64 images of a simple maze with a start and finish and then I tried to train a RNN algorithm that could navigate unseen mazes. There are so many better ways and algorithms to do it, but I learned a lot of cool tech anyway with this approach.