rishabhjain1198

Good question.

> Data centers in space only make sense if they are cost effective relative to normal data centers.

This statement is actually completely false. The bottleneck is not cost of building data centers, but the energy accessible on the planet. How much we're willing to pay for increasing that is currently very unclear, but it's far more than the current cost of building a terrestrial data center.

As Jensen Huang famously touted their performance per watt, saying "Our customers won't buy our competitor's chips even if they were free."

Oops I forgot to say; we're filing federal + state taxes for free for everyone in America (I think we're one of very few platforms left doing this).

Please give us a try!

I'm Rishabh, co-founder of Prime Meridian (https://getprimemeridian.com), and wanted to share our technical paper.

Please let me know if you have any feedback/thoughts!

Reasoning cannot actually be toggled. LLM companies serve completely different models based on whether you have reasoning enabled or disabled for "Opus 4".

The Grok 3 system prompt is quite accurate, it's been open-sourced.

Rest in peace. A true SV legend.

If you're using iPhone 15 Pro and you've configured the action button to use ChatGPT, anyone can access your previous prompts without entering the iOS passcode by simply pressing the action button and asking.

Can be avoided by disabling memory in OpenAI, or simply not remapping the button.

Edit: Forgot to mention, this is on iOS 17.5.1 (latest stable release).

Regardless of what computers find interesting, humans want to progress math in stuff humans find interesting. We can't fully rely on computers to do that yet since they can't seem to judge that very well rn as good as human mathematicians.

Don't necessarily see a tautology here.

Although I caught that on the first read, I found myself questioning when I read the adaptors part, "is this not just LoRA...".

Maybe it's my fault as a reader, but I think the writing could be clearer. Usually in a research paper you would link to the LoRA paper there too.

The productization of it (like Karpathy mentioned) is awesome. But I think the URL for that would be this maybe? [link](https://www.apple.com/apple-intelligence/)

Aren't Playstation games doing exactly this?

For people interested in AI research, there's nothing new here.

IMO they should do a better job of referencing existing papers and techniques. The way they wrote about "adaptors" can make it seem like it's something novel, but it's actually just re-iterating vanilla LoRA. It was enough to convince one of the top-voted HackerNews comments that this was a "huge development".

Benchmarks are nice though.

In its current state, I look at it as just a much smarter coding auto-complete, which is still very useful.

With that perspective, "characters added by AI" is an ok metric to track.

+1 for this. Stuff is completely optional.

In my empirical experience at Google, the code-complete is hit or miss when I haven't put in my first edit.

However, once I do something, I guess the LLM gets the nudge/prompt in the right direction and almost always auto-completes the full thing correctly.

No pressure from the top. Methodologies are pretty solid.

The internal dev tooling at Google is quite far ahead of what's available on the market rn.

[Relevant paper](https://arxiv.org/abs/1701.06538).

TL;DR you can think of it as the initial part of the model is essentially dedicated to learning which experts to choose.

In a MoE model with experts_per_token = 2 and each expert having 7B params, after picking the experts it should run as fast as the slowest 7B expert, not a comparable 14B model.

I wish this article was just 3 paragraphs. The verbose writing style was a little tiring, I found myself scrolling impatiently to find what the actual "Lone Banana Problem" was.

Congrats on the launch, Type looks great!

I was curious, is there any difference you noticed for your use case with GPT vs Claude?

ChatGPT Plus GPT-4 is heavily rate-limited, as well as slow for this use-case. As of now.

This is almost certainly operating off of <=gpt3.5-turbo.

I think it should be general practice to include the prompts used for tools like this, so people can understand how they work and suggest improvements.

The prompt used for this was:

``` Explain in detail what the following code does: [User entered input here] ```

Source: https://whatdoesthiscodedo.com/g/27d5733

Google Chrome is not open source, Chromium is.

[Reference](https://en.wikipedia.org/wiki/Chromium_(web_browser)#:~:text....)

Absolutely incredible! I haven't felt the feeling of insight after just reading 3 paragraphs in a long time.