jlas

Notably the scaling law paper shows result graphs on log-scale

Yann LeCun discussed why LLMs are not enough for AGI on Lex Fridman pod: https://youtu.be/5t1vTLU7s40?t=138

> (more resources = closer to intelligence)

The scaling law only states that more resources yield lower training loss (https://en.wikipedia.org/wiki/Neural_scaling_law). So for an LLM I guess training loss means its ability to predict the next token.

So maybe the real question is: is next token prediction all you need for intelligence?

Seriously though when is HiFi coming out?

Coding interviews are a lossy process. Once you realize this the angst of rejection softens and it becomes an almost mechanical effort.

Don't overlook sources of anti-productivity which might be easier to measure like the time it takes to build & test code changes.

If you're a photographer, the low tech way of doing this is just use a polarizing filter

This. As you get into later stages, check sizes become bigger and so does risk aversion + due diligence.

Non-zero probability I think, one interesting measure of overfitting I've seen is contamination (where the model has seen the exact questions it is being evaluated on) see stats at https://hitz-zentroa.github.io/lm-contamination/

ZIRP is over and startups are back to hard mode

> In 2021 and the first half of 2022 (when most of the 2022 activity happened) we essentially crammed 5 years worth of funding into an 18 month period. Series Bs and Cs were raised when companies were at the typical Series A milestones. Normal round sizes doubled or tripled. Every type of investor was broadly operating in a “risk on” mindset given the ZIRP environment, and the venture capital ecosystem was no exception.

From https://cloudedjudgement.substack.com/p/clouded-judgement-11...

Sort by citations (https://dl.acm.org/action/doSearch?SeriesKey=cacm&sortBy=cit...) to see the most seminal ones.

Would be interesting to see a time to exit histogram. Estimating from the current status chart it seems like every 4 years about 25% will exit.

It's useful for streaming data where you want to pass small chunks of data incrementally. JSON by itself isn't streamable (you need the entire blob to decode it). But if we just pass smaller json-encoded chunks of data and separate them with a newline, we can make a streaming-compatible JSON protocol.

Major drawback of Apple Hi Res is that Apple headphones dont support it (bluetooth is lossy), and even their recommended 3.5mm adapter converts it down to 48kHz (https://support.apple.com/en-us/HT212183)

Even Tidal has questionable HiFi (https://youtu.be/pRjsu9-Vznc)

> In venture markets, I think we’re only 10% into the pain that we’ve delayed given the large cash balances of startups. 2025 is probably the year of max pain in the venture ecosystem.

> The researchers ran a simulation that seems to align with the scarce physical insights

Fwiw, we will probably never be able to directly observe the contents of the mantle. Computer simulation has been a godsend for geology research.

You can export the app artifact and share it. The artifact is human readable and in a JSX format (called Toolscript). In this way it can be meaningfully interpreted, improved and customized by others in an open source manner.

As for running the software the platform has a proprietary license obviously but you can host your own Retool backend from which to serve the apps to your end users (check the docker repo). This is an alternative to the SaaS/cloud based app hosting at https://retool.com/.

There's a free tier to edit and run the apps but professional usage will have a license fee. This is not unlike making an app with VB which required buying VB and having end users buy and run Windows.

It's even easier today with web browsers being the ubiquitous app platform, and the VB equivalent being https://retool.com/

(Not) fun fact that I learned recently is some dyes like Red 40 are made from petroleum!

https://en.wikipedia.org/wiki/Red#:~:text=The%20most%20commo...

What really amuses me about flat earthers is how much science they are willing to invalidate. For example pretty much all of geology would have to be rediscovered under a flat earth model.

See also: http://jvns.ca/blog/2013/12/30/questions-im-asking-in-interv...

> Now, we all know that it can take a while to find a long sequence of digits in π, so for practical reasons, we should break the files up into smaller chunks that can be more readily found. In this implementation, to maximise performance, we consider each individual byte of the file separately, and look it up in π.

So basically like a regular filesystem except now lookup each byte every time you intend to use it?

Also interesting to see the weekend dips in google trends, e.g. Java: https://www.google.com/trends/explore?date=today%201-m&q=jav...

One can use recursive properties better if the structure divides equally. So if not having a strict quadtree is okay, they can divide the square into 3x3, 5x5, etc

> The reason open addressing is considered open is that it frees us from the hash table. The table entries themselves are not stored directly in the bins anymore, as with a closed addressing hash table, but rather in a separate entries array, ordered by insertion.

> Open addressing uses the bins array to map keys to their index in the entries array.

Am I wrong or is this generally not true? Open addressing is about storing the entries directly in the bins [1].

The new implementation is still open addressing, sure, but the bins contain an index to a separate entries array, presumably to keep the size of the bins array compact.

[1] https://en.wikipedia.org/wiki/Hash_table#Open_addressing

Also checkout the Odo pydata project which is a conversion tool: http://odo.pydata.org/en/latest/overview.html

Wikipedia? https://en.wikipedia.org/wiki/List_of_algorithms

Reminds me of https://circleci.com/blog/its-the-future/