jerrygenser

It's good for general things but less good at coding. Can usually get the correct answer for simpler things but much less idiomatic for python than gpt4

If you find yourself in Cooperstown, NY this summer and you're interested in M.C. Escher, I highly recommend going to the exhibit linked below [1].

It has a large number of Escher's works in one place and it's quite well put together.

Source: I went there a few weeks ago.

[1] https://fenimoreartmuseum.org/future-exhibitions/escher

Yes it's easy to block the event loop. I agree that it matters in that you're much more likely or even able to hit this throughput by using an event loop runtime.

In other words, having 32/64 concurrent threads you won't get the throughput you want so agree there.

However, I'll elaborate on my point and why it's not just "use event loop".

If you block the event loop you can get catastrophic behavior, like your entire server acts as if it's a single threaded synchronous runtime.

However, even with not blocking the event loop, you need to have optimal IO __given a coroutine__. If you are waiting a bunch of ms for multiple network calls and also a db write within a coroutine, your latency is going to go up and you might not hit your latency requirements.

I was specifying that the architecture was taking these concepts into account and getting low latency on each request in a high throughput way by not blocking on writes, precaching info in server memory, etc.

If they had to go to a db and/or do other calculations per query instead of pre-caching regardless of 32/64 or using an event loop, it likely would have either created another bottleneck or the web server would bottleneck and they would not hit their RPS

Most of the ability to get throughput is based on not blocking a request on network or other IO and also not doing writes in band of request.

I think it's less about the web server choice and more about the architecture around pre-computing and caching certain results in memory and then asynchronously writing via a queue.

Based on the podcast with Lex Friedman and Mark Zuckerberg, see ~minute 30.

My hypothesis based on the context of Mark discussing the release is that it's going to be completely open source and can licensed to be used commercially. Not that Meta is going to add a whole new revenue side of business to compete with OpenAI. i.e. "Here is model, with commercially permissive licensing" not "Here is model that you can use commercially but must pay me"

https://www.youtube.com/watch?v=Ff4fRgnuFgQ&ab_channel=LexFr...

These corporate blogs are almost 100% purpose for prospective employees. This is a way for prospective employees to become sold the corporate culture is rainbow and unicorns.

It's standard corporate PR.

I've done written heavily tested code in python. But also heavily tested code in SQL. When I was doing data engineering 5 years ago, dbt was a new tool and I used that to write tests, ci/cd, etc. in data engineering. It was a game changer.

Original comment is referring to Nvidia triton inference server

It is zero shot. The llm is trained to generate next token.

Zero shot is defined as being able to output predictions for classes they were not trained on.

It doesn't mean the input data can't be in ml task domain but that the model was not trained on this particular ML task and/or classes.

Get an older Dell XPS 15. Newer gen are above this price point but 2-3 gen back will be very nice.

Battery, WiFi chip, memory, and storage are all upgradeable; you'll just be signing up for a given chip generation. L

Does anyone have a recommended security checklist or config guides for secure Linux on desktop? Despite my previous assumptions I've read some recent indications that it's not as secure as windows and macos.

It's gotten worse. It's at the very least been quantized resulting in much lower overall precision. That's how they have been able to speed up inference so much.

I have used the same prompts to design a infinite scroll up/down image container.

2 months ago chat when it first came out was able to generate working code that used intersection observer API.

Now when I use the same prompts it only generates high level suggestions and when I ask for code it suggests using dom scroll events and doesn't even come up with intersection observer API unless I specifically ask. And if I do, it then generates incorrect code.

It even previously was correctly memoizing certain functions and included performance optimizations.

Network is slow. So you are waiting on the db for most of a request lifecycle.

I think GPUs are in short supply and Nvidia can't make enough to keep up with demand.

You can also use tools like annoy or nmslib/hnsw which are built for this purpose -- store vectors on disk and do high performance similarity search against them.

Off topic - I'm pleasantly surprised that I could see this Twitter post without being logged in.

If they raised $35million, you're not accounting that venture funding especially at this scale is preferred shares.

This means that the vc money requires 7-9% per year payback on liquidity event until any other equity gets any money.

This means you're likely subtracting 1-3 million per year from other equity holders based on $35mn raised

For embeddings, you can use smaller transformers/llms or sentence2vec and often get good enough results.

You don't need very large models to generate usable embeddings.

I've noticed the quality fo chatgpt4 to be much closer now to chatgpt3.5 than it was.

However if you try the gpt-4 API, it's possible it will be much better.

I love how there's a car there in the 2013 shot that likely received a fine.

Wouldnt database is locked only happen on concurrent writes where one is locking a row and timeout exceeded?

If there are n+1 reads I don't think it would lead to database is locked.

The intro describes that it's a free version of a book that contains most of the content except for a few sections.

The book was converted using an automated format from it's native design into html so a large portion could be made available for free and this automated conversion process might have some minor issues -- like the one you described.

Document stores often are reliable and more fault tolerant. But yes they trade ACID.

There are some applications that require high throughput (usually write) but can be fine with read consistency.

Couple of examples - consumer facing comment systems where other users are OK to miss your comment by 30 seconds - timeseries logging where you are usually reading infrequently but writing very much in a denormalized format so joins aren't as critical

For general CRUD, ACID is important though.

> Taking full control of your I/O and buffer management is great if (a) your developers are all smart and experienced enough to be kernel programmers and (b) your DBMS is the only process running on a machine. In practice, (a) is never true, and (b) is no longer true because everyone is running apps inside containers inside shared VMs.

The article is about DBMS developers. For DBMS developers, "in practice" (a) and (b) are usually true I think.

They wouldn't have been able to support consumer scale without the up front capital. Otherwise they would need private betas and stuff but this would have not created the value of the same network effects and robust marketplaces.

I also have had stick drift due to COD and in the menu you can doing something similar to this by controlling how much default stick is pushed.

You set high sensitivity and the cursor moves with hand off the controller in a certain direction. You adjust the counter-drift until the cursor doesn't move at all.

Part of this post is very similar to a famous reply from Dropbox launch on hb when someone said roughly: "why do we need Dropbox, here's steps to roll your own".

The average non tech person doesn't know what an http API is.

Github status page doesn't even load for me ... "We're having a really bad day, the unicorns have taken over"

They are also standard in early stage institutional equity funding.

It's ironic that postgres the underlying storage model for postgres are immutable tables that marks records as "invisible" and asynchronously cleaned up.

Probably selected a concept that there were a bunch of names from so that the series would be consistent but also which would not result in any trademark issues. Same reason Intel picks random words and "lake" after