@kellogh@hachyderm.io
@kellogh@hachyderm.io avatar

kellogh

@kellogh@hachyderm.io

I'm a software engineer and sometimes manager. Currently #Raleigh but also #Seattle. Building ML platform for a healthcare startup. Previously, built an IoT platform for one of "those" companies.

Open source: dura, fossil, Jump-Location, Moq.AutoMock, others

Do I have other interests? No, but I do have kids and they have interests. I think that counts for something. I can braid hair and hunt unicorns!

I put the #rust in frustrate

He/Him

#metal #science #python

This profile is from a federated server and may be incomplete. Browse more on the original instance.

BobD, to photography

Fort Macon is a Civil War era fortress located on Bogue Banks, a barrier island along along the North Carolina coast. Construction began on the brick and earth structure in 1826. It was built as a replacement... story continues at https://bob-decker.pixels.com/featured/inside-fort-macon-atantic-beach-north-carolina-bob-decker.html

kellogh,
@kellogh@hachyderm.io avatar

@BobD I went there last year and I got strong vibes of revisionist history from the confederate POV

kellogh, to random
@kellogh@hachyderm.io avatar

calling it now: the republican primary debates will be held on twitter and

  • the tech will not hold up
  • the moderator will have no clue what they’re doing
  • r’s will claim it went great and twits is the new conservative overlord

https://www.axios.com/2023/06/02/2024-republican-debates-networks-desantis-trump

kellogh,
@kellogh@hachyderm.io avatar

say what you want, but conservative news is practically more entertaining than game of thrones — shout out the most unexpected plot line and yeah, that’s probably what’s going to happen

dev, to random
@dev@discuss.systems avatar

Update: Big Slug

kellogh,
@kellogh@hachyderm.io avatar

@dev banana slug!

saraislet, to random

Wow, good burn, @Quinnypig

kellogh,
@kellogh@hachyderm.io avatar

@saraislet @Quinnypig that’s a great point. A great AWS feature might be some sort of simulator that says, something like

  • here’s each login
  • here’s what each actually needs, based on what they do
  • here’s all similar logins, grouped by what they do
  • here’s the a few scenarios of finely scoped permissions that balance number of policies with fine-tuned permissions in different ways
  • hit apply
nyrath, to random
@nyrath@spacey.space avatar

Nobody died. The AI was just trying to maximize its score in the simulation, by using logic

https://www.vice.com/en/article/4a33gj/ai-controlled-drone-goes-rogue-kills-human-operator-in-usaf-simulated-test

kellogh,
@kellogh@hachyderm.io avatar

@squeakyears @nyrath absolutely. This is buggy AI. It’s not an ethical issue. It’s not rogue AI, it’s just buggy AI. And yes, it’s unethical to unleash buggy AI in a context where it can kill people, but that’s no different from any other software

adrian, to random
@adrian@discuss.systems avatar

I choose to believe Merkle trees were invented by Chancellor Angela Merkel and I have no reason to revisit this assumption

kellogh,
@kellogh@hachyderm.io avatar

@adrian tbf they ruled for a while, but recently got supplanted by a different thing

h_thoreson, to random
@h_thoreson@mastodon.world avatar

i hate how dumb human brains are like i know i would just be bored af if i actually lived in iceland but after a pandemic which really just has kept me confined to a bunch of landscapes that all kinda look the same it just seemed so novel and exciting.

kellogh,
@kellogh@hachyderm.io avatar

@h_thoreson I don’t think that’s dumb at all

hperrin, to ai

How come models aren't made to go back and change their answer as they work? If you ask a human to write something, they will very rarely just spit out an entire document word for word and be done. Most human work involves revising your own output as you work. If you prompt an to do this, you will get a better result, so why not build the model to do this from the get go?

(I revised this post 4 times before posting it.)

kellogh,
@kellogh@hachyderm.io avatar

@hperrin my hunch is that OpenAI has multiple model revisions that they cycle through. If one doesn’t work, try the next. Running all 3-5 models every time would more expensive and probably the first is probably giving you the best answer when you’re happy with it. The trouble is predicting when you’re not happy

kellogh, to random
@kellogh@hachyderm.io avatar

auto-updated to use dns over https and rendered my browser useless until i disabled it. not fun. big panic

kellogh,
@kellogh@hachyderm.io avatar

it seems that my issues only happen with the version of firefox. Unsure what's up, but it was essentially non-functional until i installed from the website

mcc, to random
@mcc@mastodon.social avatar

This is my promise to you: I will NEVER use the terms "Death Metal" and "Black Metal" correctly

kellogh,
@kellogh@hachyderm.io avatar

@mcc there’s only 2 metals — modern metal and ancient metal

simon, to random
@simon@simonwillison.net avatar

I posted a mockup of a design change for ChatGPT that I think could help address the risk of people being lead astray by its incredible ability to invent faleshoods: ChatGPT should include inline tips
https://simonwillison.net/2023/May/30/chatgpt-inline-tips/

kellogh,
@kellogh@hachyderm.io avatar

@simon someone in HN suggested that the logits from the model could be used as a confidence. Is this remotely true? (My hunch is yes, but not very well)

kellogh,
@kellogh@hachyderm.io avatar

@simon I just saw this. Looks like the logits can be used to report some sort of confidence, but OpenAI and Anthropic don’t report it from their APIs. I’d love to have text highlighted in red scale if it’s lower confidence https://arxiv.org/abs/2305.15507

andypiper, to random
@andypiper@macaw.social avatar

Today, I threw all of my historical blog content at a local instance of an LLM, to see how it would fare at writing stuff for me. Sadly, not as good as a giant dataset like ChatGPT or Bard. I’ll have to own my own content for a bit longer, then…

kellogh,
@kellogh@hachyderm.io avatar

@andypiper what did you do? Did you train a LoRA? Or include it as context?

kellogh,
@kellogh@hachyderm.io avatar

@andypiper I’ve been thinking of it as

  • vector DB / private GPT to add factual info
  • fine tune to learn a structure or style

So if you wanted to write a blog, you’d want to fine tune, and maybe use privateGPT by ingesting other people’s blogs or docs to get something new that you haven’t written about yet. Also, I’d have GPT go paragraph-by-paragraph or so and reach out to the vector DB between each step to improve the factual side a bit

samir, to random
@samir@functional.computer avatar

I am so disappointed that Cargo’s crates aren’t namespaced. It’s (somewhat) new! Why would you actively choose to open yourself up to typo-squatting?

This problem was solved by Maven approximately eleven thousand years ago, before I even started writing Java. Just… do that.

kellogh,
@kellogh@hachyderm.io avatar

@samir are there any other ecosystems that do namespacing?

kellogh,
@kellogh@hachyderm.io avatar

@samir right, I think Go and Java we’re both enterprise first. There’s a reason other ecosystems haven’t chosen to go that route, and it has a lot to do with the enterprise angle, imo. Fwiw Cargo does have namespaces, serde-*, when needed https://samsieber.tech/posts/2020/09/registry-structure-influence/

kellogh,
@kellogh@hachyderm.io avatar

Maven having multiple versions with the same name really is a problem, imo. I recall several times in Java trying to figure out which is the real one, and Cargo does solve that problem

kellogh,
@kellogh@hachyderm.io avatar

@samir i would argue that namespaces don’t solve the squatting problem. They distinguish successfully on a technical level, but people realistically only reference it by the core name and it takes almost the same information to distinguish between which you should choose as squatted names. E.g. org.apache.jepson vs com.microsoft.jepson — which do you choose? (Idk, gotta look up the docs, they both sound legit)

kellogh, to random
@kellogh@hachyderm.io avatar

Insurance is sticking us with an $800 bill for COVID tests administered in 2021, when tests were free. They’re reimbursing us $34 to pay an $800 bill

kellogh, to random
@kellogh@hachyderm.io avatar

Still recovering from the weekend (or maybe from last week?)

SingingLehrerin, to photography German
@SingingLehrerin@sueden.social avatar

A Giant Sequoia in Germany (Bonn-Bad Godesberg), albeit a quite young one without a very thick trunk. Made me long for the big ones in California... ❤️

kellogh,
@kellogh@hachyderm.io avatar

@SingingLehrerin that’s awesome. I assume they don’t grow native there right? AFAIK they’re pretty sensitive to climate, but it seems like Germany would probably work, makes sense

AnitaH2, to random Norwegian

Ikke så tykk, men tenkte at jeg må tipse #Allheimen om den fine hashtagen #thicktrunktuesday #ThickTrunkTuesday
(This is posted mainly to make norwegian mastodoners aware of the hashtag. We have think trunks in abundance ☺️)

kellogh,
@kellogh@hachyderm.io avatar

@AnitaH2 [not Norwegian, sorry] is that some sort of true cedar?

kellogh, to random
@kellogh@hachyderm.io avatar

Took a paddle on the mystical Merchant’s Millpond today. Surreal. Like paddling through a dream

  • All
  • Subscribed
  • Moderated
  • Favorites
  • •
  • JUstTest
  • kavyap
  • DreamBathrooms
  • thenastyranch
  • magazineikmin
  • tacticalgear
  • cubers
  • Youngstown
  • mdbf
  • slotface
  • rosin
  • osvaldo12
  • ngwrru68w68
  • GTA5RPClips
  • provamag3
  • InstantRegret
  • everett
  • Durango
  • cisconetworking
  • khanakhh
  • ethstaker
  • tester
  • anitta
  • Leos
  • normalnudes
  • modclub
  • megavids
  • lostlight
  • All magazines
    •