jclem

I grew up in the Midwest, and my father, who went to the University of Kentucky, was fraternity brothers with Ben Gish, and so growing up it was a name that I heard often when he’d tell stories from that time. I remember always being aware that a friend of my father had been framed for burning their family newspaper.

I live in New York City, now, and one evening found myself sitting at a bar in Brooklyn where my roommate, the bartender, was closing around 3am (at the time I worked a very late shift at the 24 hour Apple Store). The normal patrons had all gone, and myself, my roommate, and another man I’d never met were locked in having a last couple of beers while my roommate closed up. I had a long conversation with the man I hadn’t met before that eventually turned to where we grew up. It turned out that this man was Ray Gish (also mentioned in this article), the younger brother of Ben. Ray even recognized my father’s name, because Ben had mentioned him before.

Ray invited my roommate (whom he already knew well) and I over to his place for a last drink, and I vividly remember him playing “Street Hassle” by Lou Reed on his record player. It was the first time I’d ever heard that song.

Anyway—wildest coincidence that has ever happened to me. I see and say howdy to Ray every now and then when I find myself in that part of Brooklyn—one of the most genuine people I’ve ever met.

You can reduce with list comprehensions in Elixir, but it's uncommon to see:

    sum_list = fn list ->
      for item <- list,
          reduce: 0,
          do: (sum -> sum + item)
    end

    sum_list.([1,2,3])

Other comments are correct that `Enum.reduce/2` is probably better:

    Enum.reduce(list, &(&1 + &2))

Obviously you don't have the flexibility / mutability you refer to in other comments with either of these, but you can always put more in your accumulator:

    get_sum_and_update_count = fn list, map ->
      for item <- list,
          reduce: {0, map},
          do: ({sum, map} -> {sum + item, Map.update(map, item, 1, &(&1 + 1))})
    end

(A contrived function that returns a list sum and an updated map with the count of the values found in the list)

It seems worth noting that the input of gte-small is limited to 512 tokens. The tokenizers for sure aren’t the same, but I imagine this is significantly less than ada-002, whose input is limited to 8191 tokens. That said, I don’t imagine that embedding huge full documents is necessarily the right approach.

I would love to see a comparison for some typical use cases using various methods of chunking input documents.

How are you stuffing long text into GPT context? Are you doing some form of summarization?

This package is using GPT-3 via the “ChatGPT” export from that module, which is—somewhat misleadingly—not ChatGPT, but GPT-3.

Hello from slightly further south on Vanderbilt Ave :) I haven't tried that Indian place yet, but it's on my list, now. The pizza and bagels threw me off (not because "Indian == bad bagels/pizza", but because "somewhat random assortment of foods" is sometimes a red flag).

First, let's replace "gushing" with "discussing", because that's what was happening. I'm not sure "got bad during part of the pandemic, now is improving again" counts as gushing.

In what way is caring about your local food options in your immediate neighborhood, on the street you live, where your friends and acquaintances eat, drink, and work "very Brooklyn"? I guess if you live somewhere that restaurants just appear and have faceless owners with employees you'll never know, this isn't a thing. This is something discussed by really anyone who is interested food and restaurants in their area all over the world. If you like what a person did at one restaurant, you may like what they do with a new one. Or, maybe it's noteworthy when only the first one is great, and the subsequent ones are mediocre.

Maybe you mean Vanderbilt Ave? Vanderbilt St has no commercial zoning.

What many people don't know is that 3 of the pricier restaurants on Vanderbilt Ave are all actually the same restaurant group. Their background is partially the 3-Michelin star Alinea in Chicago via chef Greg Baxtrom. Olmsted was first, then came Maison Yaki, and then Patti Ann's.

Olmsted is mostly very good (started great, then got bad during part of the pandemic, now is improving again), but you can still joke about being sold $8 vinegar from them for sure, haha. However, since "midwestern" cuisine is now becoming somewhat trendy, jokes about expensive port wine cheese balls at Patti Ann's ($12) may be more appropriate.

It's $50/year.

An LWW register is most certainly a CRDT, it just has limited use. I think you’re making a good point about how one must be careful applying CRDTs, but I don’t think it’s accurate to say that a suboptimal merge strategy in terms of user experience means that something does not meet the definition of a CRDT.

In other words, the “conflict-free” part of the term refers to the fact that all replicas converge on the same value when all updates have been exchanged. What you’re discussing seems to have more to do with concerns around intention preservation when using CRDTs. In other words, an LWW register on a text field is still a CRDT, but in many applications it is potentially a poor choice in terms of preserving the intent of user-generated operations, but determining this is something done in a case-by-case basis.

If the content is a long document which users collaborate on over time, then a register is most certainly a poor choice. If it is a short text field akin to a label, an LWW register isn’t necessarily bad, because user intent is likely to replace the entire value, rather than to perform minute edits on the value.

While “Arcadia” is likely of greatest interest to HN readers of all of Tom Stoppard’s plays, I also highly recommend “Jumpers” and “The Hard Problem.”

Also worth noting the many excellent films he wrote, like “Brazil” and “Empire of the Sun.”

Disclaimer: I like Railway.

Heroku VMs had the kitchen sink, and I miss that with Railway. A pretty standard Phoenix app with Phoenix’s built-in auth doesn’t build and the user is presented with build errors about nixpacks. Now there’s a good deal I have to learn just to get bcrypt to build.