nathanfig

Oh I should have known - yeah I was just being facetious

Hi all - I write a lot for myself but typically don't share, hence the stream-of-consciousness style.

But I thought this might be worth blogifying just for the sake of adding some counter-narrative to the doomerism I see a lot regarding the value of software developers. Feel free to tear it apart :)

At WHOI we frequently run into some of these issues. Ocean-going projects often have long periods of disconnect, or have just a tiny bandwidth-capped connections that are meant for sending home critical data and not for automated software updates...

So much software also just takes for granted that it should be allowed on the Internet.

Multi input? Infilling? First generative audio model I've seen that starts to close the gap with image models.

Not quite; turn any GUI application into a Python command line program.

I would love something that could do the reverse.

Scaling context is not something humans have good intuition for- I certainly don't recall an exact sentence from 200 pages ago. This is an area where we actually want the models to not mimic us.

Definitely edited, pretty clear in some of the transitions. Makes me wonder how many takes were needed.

Thanks, I was looking for clarification on this. Using Bard now does not feel GPT-4 level yet, and this would explain why.

I had the same issue as OP. Initially Bard seemed clueless about Gemini, then:

Me: I see. Google made an announcment today saying that Bard was now using a fine-tuned version of their "Gemini" model

Bard: That's correct! As of December 6, 2023, I am using a fine-tuned version of Google's Gemini model ...

Fining offenses is a great way to fund finding more offenses.

It's both. Technology really does make a difference. Its existence has effects.

Agreed. People publish papers on algorithms all the time, imagine saying "Sorry, but new C++ is not a paper". There is a ton of space to be explored wrt prompts.

If you do the rigor on why something really is interesting, publish it.

Stuff like this always makes me wonder how many interesting numbers there must be that we'll never have the bits necessary to consider.

Yeah could be, I still need to read the paper to better understand the safety tuning.

Would be interesting to see some work stress-testing the ability to convey ill-intent across multiple languages. Accurately conveying ill-intent is safety-critical for the person being threatened.

Impressive work, really excited for this.

I will note though that I feel safer getting an occasional bad word than I do having a translator straight up deceive me.

For example, "what the fuck" in English->Spanish is giving "qué diablos" output. Definitely toning down the meaning there.

If someone says something mean to me, I want to know it.

Mazda also makes the Miata, arguably one of the most fun cars to drive.

I've decided no AI boyfriends for my kids before they turn 15. I get it, I'm old fashioned, and some might even say bigoted, but we just can't afford the accessories.

Judging by the comments this seems like a very popular take. But I really wish we could move from "wow I hate that" to thinking through the right response, because it's likely going to be our children forming these relationships very soon.

They will seem perfectly normal too! Except for the subtle product placements :)

This is an expected outcome for lonely people + LLMs. And it will likely explode from here. There is a lot to be concerned about, but it's not going to go away.

The commercialization aspect might be what bothers me most. The article is full of people hurt when startups got them hooked on a feeling of intimacy and then took that away. Emotional manipulation for profit is nothing new but LLMs might turbocharge that ability.

These are fantastic

A lot of attention given to the inadequacies of the nuclear family, but industrialization seems likely the reason we broke into nuclear families to begin with. Smaller family units could move to where the breadwinner needed to go, breaking up intergenerational communities. Funny how our mobility is greater than ever, yet somehow less able to stay together.

I'm not a big fan of suburbs but in no way did this trend begin with suburbs, nor are densely urban or sparsely rural areas close to the community model described in the article.

Some private enterprises would undoubtedly ignore this prize, but a lot of academic researchers (including students) would be more than motivated by this prize.

Also liked Fast.ai, although they sometimes use Fast libraries instead of PyTorch, it does focus mainly on PyTorch concepts.

Relative term. In the world of LLMs, 7b is small.

Was just about to comment this. Wonder if anyone more mathematically inclined can see a relationship.

Deno Fresh uses Twind, which avoids the build step: https://twind.dev/

But IIRC it doesn't offer quite everything full Tailwind does. In general I was frustrated with Fresh for not being up-front about the fact that it is largely built on Preact and Twind, and you must buy into those libraries first.

OP is excited about the leap in capabilities, that is all it means. You can be curious about the inner workings of alien magic too.