zerd

Sounds a bit like Wuala https://www.youtube.com/watch?v=3xKZ4KGkQY8

I had similar issues, raid 1 on two hdd and the server would randomly reboot and be slow because it was resyncing the raid. Have to pay extra to get new refurbished drives.

It’s great for throwaway machines, e.g. CI. But don’t rely on them

This. the main trick, outside of just bigger hardware, is smart batching. E.g. if one user asks why the sky is blue, the other asks what to make for dinner, both queries go though the same transformer layers, same model weights so they can be answered concurrently for very little extra GPU time. There's also ways to continuously batch requests together so they don't have to be issued at the same time.

How is it different to bufstream?

My ISP blocks port 25, 80 and 443, so need to tunnel those. Some don't want to expose their IP directly. If you have dynamic IP you don't have to update the IP in DNS (since the "application" connects to the tunnel endpoint).

That would make the locked time shorter, but it would still contend on the global lock, right?

It's annoying, around 10 years ago 10gbps was just starting to become more and more standard on bigger NAS, and 10gbps switches were starting to get cheaper, but then 2.5GbE came out and they all switched to that.

Counter-counter example. When my wife was new to the city she took a bus ride and a girl randomly made contact with her. They hit it off and talked for the whole ride. My wife thought this girl sounds really nice, they actually got into deep conversations, and wanted to exchange numbers so they could hang out some time. Then she asked if she had heard about the book of mormons…

> it wasn't a widespread GCP outage.

If this wasn't widespread, what is?

Incident affecting API Gateway, Agent Assist, AlloyDB for PostgreSQL, Apigee, Apigee Edge Private Cloud, Apigee Edge Public Cloud, Apigee Hybrid, Cloud Data Fusion, Cloud Firestore, Cloud Logging, Cloud Memorystore, Cloud Monitoring, Cloud Run, Cloud Security Command Center, Cloud Shell, Cloud Spanner, Cloud Workstations, Contact Center AI Platform, Contact Center Insights, Data Catalog, Database Migration Service, Dataform, Dataplex, Dataproc Metastore, Datastream, Dialogflow CX, Dialogflow ES, Google App Engine, Google BigQuery, Google Cloud Bigtable, Google Cloud Composer, Google Cloud Console, Google Cloud DNS, Google Cloud Dataflow, Google Cloud Dataproc, Google Cloud Pub/Sub, Google Cloud SQL, Google Cloud Storage, Google Compute Engine, Identity Platform, Identity and Access Management, Looker Studio, Managed Service for Apache Kafka, Memorystore for Memcached, Memorystore for Redis, Memorystore for Redis Cluster, Persistent Disk, Personalized Service Health, Pub/Sub Lite, Speech-to-Text, Text-to-Speech, Vertex AI Search

The ads will be subliminal so you won't even know that you're watching an ad.

Like predicted in 2014 here: https://www.destroyallsoftware.com/talks/the-birth-and-death...

Just noticed the physics simulator at the top is interactive. Then I was stacking squares on top of each other to see how tall I could make it, and started throwing things at it angry birds style. Fun stuff.

> It has long been established that predictive models can be transformed into lossless compressors and vice versa. Incidentally, in recent years, the machine learning community has focused on training increasingly large and powerful self-supervised (language) models. Since these large language models exhibit impressive predictive capabilities, they are well-positioned to be strong compressors. In this work, we advocate for viewing the prediction problem through the lens of compression and evaluate the compression capabilities of large (foundation) models. We show that large language models are powerful general-purpose predictors and that the compression viewpoint provides novel insights into scaling laws, tokenization, and in-context learning. For example, Chinchilla 70B, while trained primarily on text, compresses ImageNet patches to 43.4% and LibriSpeech samples to 16.4% of their raw size, beating domain-specific compressors like PNG (58.5%) or FLAC (30.3%), respectively.

https://arxiv.org/pdf/2309.10668

Transformers are also used in the top algorithm right now on the Large Text Compression Benchmark. https://bellard.org/nncp/nncp.pdf

Interesting video about what goes on behind the unofficial coverage: https://www.youtube.com/watch?v=Jp2Qgw44MsA

"Are you sure you want to install OSM? Don't you know Google Maps is better? You should try Google Maps"

There's pretty much nothing you can't do with ffmpeg https://youtu.be/9kaIXkImCAM?si=KJJOhnoa-2bFLkt5

That would be assuming they knew there was a bug in that particular part of the code, which they probably didn't, until Windows 11 24H2. And unfortunately Valgrind doesn't work on windows.

Still isn't available for windows from what I gather.

fmod takes in the order of 30+ cycles, probably more in year 2003 CPUs, vs 1 for cmp, 1 for sub, 1 for jmp.

I tried to use Valgrind to catch pretty much this exact bug 20 years ago, and it was nigh impossible. If you call any 3rd party code it'll have flag tens of thousands of false positives that you have to sift through. And that was on a small game engine, I can't imagine running it on millions of lines of code.

No, it was not just JS. They were the first to do a ton of other optimizations such as prefetching. https://www.igvita.com/2012/06/04/chrome-networking-dns-pref...

It was much faster than Firefox, that's why I switched. It could handle more tabs. It isolated tabs so if one crashed it didn't crash the whole browser. Memory usage was lower. I wouldn't call any of those "marketing" and "mindshare".

My problem with turning it of is that if you _do_ want to watch the preview it's very cumbersome. Clicking on it goes to the movie/episode. So to get to the preview you have to go to the list of Episodes, scroll down (and try not to get spoiled) to trailers then play it. So I have one profile with it on and one with it off, depending on if I'm browsing or not.

> easily argued to be in his core powers

I'd say arguably. https://en.wikipedia.org/wiki/Tariffs_in_the_second_Trump_ad...

90% oppose the idea. https://www.cbc.ca/news/canada/u-s-eh-who-are-the-canadians-...

You can do the same with time though, just return a predefined sequence of timestamps.

One interesting thing about S3 is the vast scale of it. E.g. if you need to store 3 PB of data you might need 150 HDDs + redundancy, but if you store it on S3 it's chopped up and put on tens of thousands of HDDs, which helps with IOPS and throughput. Of course that's shared with others, which is why smart placement is key, so that hot objects are spread out.

Some details in https://www.allthingsdistributed.com/2023/07/building-and-op... / https://www.youtube.com/watch?v=sc3J4McebHE

We did a very similar thing in a previous project. Used protostuff to serialize each object, and snappy at the time to compress it. And a small proxy class to access it that decompressed on demand. Costs a few microseconds on read, but saved 4x the space (20GiB -> 5GiB IIRC). This was after interning etc. I'm sure you can save a lot more space if you compress batches of objects and store it column oriented, but will take a bit longer to decompress individual objects.

I've been using `fclones` [1] to do this, with `dedupe`, which uses reflink/clonefile.

https://github.com/pkolaczk/fclones