@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.

kellogh, to random
@kellogh@hachyderm.io avatar

Great paper! Also a great reminder that (artificial) neural networks are still quite useful for studying real brains, because they can help model and study questions like, “how did this phenomenon come to be?”
https://neuromatch.social/@dlevenstein/112395909323594371

AnnemarieBridy, to random
@AnnemarieBridy@mastodon.social avatar

I’d be curious to know what effect, if any, this change has on a relatively large LLM’s likelihood of outputting strings of text that are memorized from training data sources.

Meta multi-token prediction makes LLMs up to 3X faster | VentureBeat https://venturebeat.com/ai/metas-new-multi-token-prediction-makes-ai-models-up-to-3x-faster/

kellogh,
@kellogh@hachyderm.io avatar

@AnnemarieBridy the way i understood the paper, it wouldn’t change much, but there’s a lot of variables, like the increased data efficiency also means there’s less training data to reference, but theoretically without increasing overfitting (quoting the source)

kellogh,
@kellogh@hachyderm.io avatar

@paninid @AnnemarieBridy are you serious?

kellogh,
@kellogh@hachyderm.io avatar

@paninid @AnnemarieBridy latency touches every part of every business case

  • environmental impact is less
  • cost is less
  • interactive apps see drastically better user experience
  • some applications weren’t possible but are enabled by lower latency

if you don’t understand the user experience impact, try using Groq at 500 tokens/s https://groq.com/

kellogh,
@kellogh@hachyderm.io avatar

@paninid @AnnemarieBridy Little’s Law is doing the work here — if you cut latency 1/3 and keep the amount of work the same, you also can cut the number of servers to 1/3

kellogh,
@kellogh@hachyderm.io avatar

@paninid @AnnemarieBridy i don’t think that’s a logical step you can make. the largeness gives it the “general” capabilities, where you don’t have to train it for a specific task. most enterprises are using LLMs via RAG, i.e. they have no need to train their own model. one of the benefits of LLMs in general is that model training is left to the people who are best at it, and everyone else just uses databases

deech, to random
@deech@mastodon.social avatar

I've moved to Raleigh,NC. I would love to connect/hangout/talk tech. If you're in the Triangle area hit me up!

kellogh,
@kellogh@hachyderm.io avatar

@deech oh hey! i’m in Raleigh too. Moved a couple years ago. Would love to catch up

kellogh, to rust
@kellogh@hachyderm.io avatar

ooo promising new #rust #htmx framework. Check out the ad! 😂 https://youtu.be/YpHFthVa7nU?si=QCJHyg7_dEmBeGTu #rustlang

kellogh, to random
@kellogh@hachyderm.io avatar

y'all are debating println debugging and i'm over here like

kellogh,
@kellogh@hachyderm.io avatar

it's like super goto

kellogh,
@kellogh@hachyderm.io avatar

@jimfl that's very well structured

kellogh, to random
@kellogh@hachyderm.io avatar

if we play our cards right, maybe we can convince the DSM to cooperate and then setup “rich white man conversion therapy” clinics https://www.salon.com/2024/05/05/believes-the-most-taboo-conspiracy-theories-it-might-not-be-you-think/

asmartbear, to random
@asmartbear@noc.social avatar

If the project is failing all the time, in multiple ways, maybe its not "learning." Maybe its just failing.

Or maybe you need to use your words:

https://longform.asmartbear.com/fail?utm_source=mastodon&utm_campaign=asmartbear_mastodon&utm_medium=social

kellogh,
@kellogh@hachyderm.io avatar

@asmartbear my general take is that manipulating language makes things more opaque, and what you really need is less stigma on failure. a more direct way to de-stigmatize is to remove the blame aspect and create a cultural norm that every failure is followed by an opportunity to prevent it in the future. now it’s a positive experience, but same words

kellogh, to LLMs
@kellogh@hachyderm.io avatar

has anyone made a successor to fuckit.js that uses ?

(fuckit.js ran the script in a loop, randomly deleting lines until it runs successfully)

kellogh,
@kellogh@hachyderm.io avatar

@wagesj45 idk, either sounds pretty interesting

kellogh,
@kellogh@hachyderm.io avatar

@wagesj45 right, but it’s gotta be haphazard

sminez, to ubuntu
@sminez@hachyderm.io avatar

So as of this morning I'm officially working as a senior engineer at Canonical! :ubuntu_logo:

And at 7:20am tomorrow I set off for Madrid for my onboarding and then my first in-person engineering sprint 😎 not a bad way to start a new job!

kellogh,
@kellogh@hachyderm.io avatar

@sminez whoah, congrats!

lars, to ai
@lars@mastodon.social avatar

AI art

I just came across this (h/t to Peter Krupa), and it blew my mind. It highlights the problem with LLMs in general with pinpoint accuracy, and wraps it in a well known metaphorical idiom that everyone understands — which instantly becomes a meta reference. …

https://lars-christian.com/notes/4d8c59cee5/
# #

kellogh,
@kellogh@hachyderm.io avatar

@lars i wish there were a psychologist around to offer their thoughts, but if you asked me to imagine a room without an elephant, i’d picture the same thing, and i’d have to try hard to remove the elephant from the image. i suspect that we share a similar limited aspect of pattern matching intelligence, but we have several more powerful analytical systems that kick in. kinda like system 1 vs system 2 in Thinking Fast and Slow

kellogh,
@kellogh@hachyderm.io avatar

@lars if you still have the conversation around, i’d love to see GPT-4’s prompt to Dall-E

kellogh,
@kellogh@hachyderm.io avatar

@lars this whole thing is super interesting. It sounds like the strange behavior is more in diffusion models (Dall-E) than language models (GPT4). it reminds me of that paper that found that LLMs only learn logic in a single direction, e.g. if none of its training data had “elephant not in a room”, it wouldn’t know what to do and instead pattern match to “elephant in room”

kellogh, to random
@kellogh@hachyderm.io avatar

with our kids, we have this thing where you can do anything as long as you’re prepared

play in the rain? sure, just wear rain jacket/pants/boots

there’s poison ivy in the woods? that’s fine, you know how to identify it, and we have PI soap to wash it off if you miss some

the philosophy is to redirect them into the right way, rather than banning the wrong way. it gives us a lot of freedom in how we travel and experience new things

kellogh,
@kellogh@hachyderm.io avatar

engineering teams are kind of similar. i like processes that nudge you into the right way to do things, rather than shutting down specific behaviors

e.g. code formatters in CI, type systems, etc.

there’s also a big tie in to how you run incident reviews. focusing on what could have been done better (and making a plan to do it that way next time) vs shaming people for doing the wrong thing

danjac, to random
@danjac@masto.ai avatar

I've found the happy space for using Docker in local development is to use it for external services such as databases, Redis etc (as I have to use different versions of e.g. MySQL or PostgreSQL), but use language-specific tools such as pyenv or nvm instead of running your project in a container.

kellogh,
@kellogh@hachyderm.io avatar

@danjac i like to use docker for project dev whenever i feel i’m moving too fast and really just need to spend more time waiting

wagesj45, to random

I truly hate the whole "man vs. bear" thing going around right now for so many reasons. So many.

kellogh,
@kellogh@hachyderm.io avatar

@wagesj45 imo the right answer is “bear, because bears are more predictable”.

if you’re talking about a strange man, that means you don’t know them, so you have to assume the worst (at least that’s the smart move), and humans are far less predictable than bears. we know how a bear will respond, and how to stay safe around them, whereas strange people is a different story, anything could happen.

fwiw i say this as (a) a parent of girls and (b) a backpacker with 5-6 bear encounters

danilo, to random
@danilo@hachyderm.io avatar

This is from a C-suite person at a company you’ve heard of. They’re responsible for product.

And THIS is why I have spent the last year developing proficiency in using LLMs as part of my development workflow:

It’s great to be able to detect a crock of shit.

There’s no universe where AI makes features “functionally free.”

The complexity of a software product cannot be handwaved away through magic tensor beans. Features are more than just code. Features are a social contract with the user.

kellogh,
@kellogh@hachyderm.io avatar

@danilo i think the moment the “free” word comes out, you gotta start questioning their understanding of physics, math, economics, and finite resources

“lower”? sure. “free”? bro, nothings free

  • All
  • Subscribed
  • Moderated
  • Favorites
  • •
  • Leos
  • ngwrru68w68
  • hgfsjryuu7
  • InstantRegret
  • Youngstown
  • rosin
  • slotface
  • everett
  • mdbf
  • tacticalgear
  • kavyap
  • thenastyranch
  • DreamBathrooms
  • PowerRangers
  • anitta
  • magazineikmin
  • vwfavf
  • cisconetworking
  • cubers
  • modclub
  • GTA5RPClips
  • khanakhh
  • Durango
  • osvaldo12
  • normalnudes
  • tester
  • ethstaker
  • provamag3
  • All magazines
    •