@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.

mekkaokereke, to random
@mekkaokereke@hachyderm.io avatar

People on Twitter are debating whether a person using uncommon words like "delve" are trying to sound smarter than they are, or worse, are ChatGPT bots, because "normal" people don't talk like that.

You don't have to get upset, or embroiled in the debate. Not worth the time or attention. But I'll share some important context as your friendly neighborhood Nigerian 🙋🏿‍♂️

Many Nigerians have bigger English language vocabularies and better command of grammar than the typical American or English person

kellogh,
@kellogh@hachyderm.io avatar

@mekkaokereke 200M is a lot of text written. also, coincidentally, they're the same parts of the world that LLM RLHF human intervention was farmed out to.

LLMs are indeed speaking to us like Nigerians/Indians/etc. because that's what they were trained to do

Now I'm thinking about all those concerns about how "English first" model training forces culture on others. Ironically, Americans are getting a first-hand taste of what it's like

kellogh,
@kellogh@hachyderm.io avatar

@mekkaokereke can't wait to delve into this

andypiper, to community
@andypiper@macaw.social avatar

👋🏻 fediverse hive-mind - a friend of mine just joined Mastodon, and is looking for follow recommendations - who are your go-to experts around and (in a broad sense, beyond and outside my more techie niche)

kellogh,
@kellogh@hachyderm.io avatar

@andypiper checkout newsmast. they have a set of human curated “bot” accounts that are richer than just following hashtags

https://hachyderm.io/@kellogh/111998062919819584

unclepj, to random
@unclepj@zirk.us avatar

Ok, question. Why do we say Attorney at Law? Is there some other kind? Like Attorney at Having a Good Time or Attorney at Snowboarding?

May I suggest Attorney at Napping? That’s something I could get behind.

kellogh,
@kellogh@hachyderm.io avatar

@unclepj @futurebird attorney inlaw is what you call your wife’s divorce attorney

kellogh, to random
@kellogh@hachyderm.io avatar

The Attack Turkey

this morning on the run i got to the top of the hill and was resting, bc i'm out of shape, and off in the distance i see a bird trotting my way

as it gets closer i realize it's a turkey. it had a big red gobbler and its neck was iridescent. It was actually really pretty, so i just watched as it approached

kellogh,
@kellogh@hachyderm.io avatar

after i got home, i looked it up and yes, turkeys can indeed be aggressive, especially around breeding time (i got news for you, Mr Turkey, i’m not your competition)

if it happens again tomorrow, i think the answer is to be more overtly aggressive. so maybe i need to yell and chase it rather than just defensively tip toe around it. idk

kellogh,
@kellogh@hachyderm.io avatar

it kept coming straight at me, eventually it was like 2-3 feet from me and my feeling shifted from awe to nervousness. i don't think turkeys are aggressive, but i also don't want to physically learn about a new aspect of nature today either, so i bolted

to my surprise, and anxiety, the bird chased after me. i sprinted across the hill, down and up a gully. It was still following me, but getting further behind, so i walked a bit, but it caught up, so i ran.

kellogh, to random
@kellogh@hachyderm.io avatar

PSA if you're not a web dev and you need to use CSS, always use rem, never px

good:

button {
width: 4rem;
}

bad:

button {
width: 64px;
}

You can divide pixels by 16 to get to what rem looks like for the typical user. Decimals are fine.

rem is better because it scales up/down when the user increases the default font size (e.g. people who can't see well)

kellogh, to opensource
@kellogh@hachyderm.io avatar

#opensource WizardLM2 8x22B exceeds performance of GPT4 in some benchmarks

  • Apache2 👍
  • progressive learning instead of all-at-once means less power-hungry and more data efficient during training
  • Co-Teaching and Self-Teaching are intriguing, I want to hear more
  • from Microsoft #AI, I imagine GPT5 must be nigh, if they’re releasing competition for GPT4

https://wizardlm.github.io/WizardLM2/ #LLMs

kellogh, to ai
@kellogh@hachyderm.io avatar

Experiments and discussion of various UX approaches to displaying code generation suggestions

Personally, the big problems I have with are

  • “this change is crap, except this line here, give me just that”
  • changes that are poorly integrated into the file
  • the chat never has the right context, so conversations tend to be circular or unhelpful

https://austinhenley.com/blog/intellicode.html

kellogh, to random
@kellogh@hachyderm.io avatar

we all knew is was going to happen eventually https://www.theonion.com/tesla-lays-off-entire-team-behind-brakes-1851449223

kellogh, to random
@kellogh@hachyderm.io avatar

solid strategy https://journa.host/@w7voa/112338785137793332

kellogh, to random
@kellogh@hachyderm.io avatar

two things can be true:

  1. machine learning and/or neural networks are far more capable than the people who use “just” to describe them would like you to believe

  2. we’re not getting AGI until after we understand what intelligence is, and we’re not throwing billions of dollars into that

kellogh, (edited ) to MIjazz
@kellogh@hachyderm.io avatar

JET reactor was able to produce 69 megajoules of from 0.2 milligrams of fuel. That’s enough to run a refrigerator for a couple weeks.

That doesn’t seem like a big breakthrough, but it is. They use a tokamak reactor, which many people had written off because of stability problems. They’ve solved those issues, meaning we now have 2 major approaches to fusion that yield good results.

https://phys.org/news/2024-02-fusion-facility-tritium-yield-energy.html

kellogh, to LLMs
@kellogh@hachyderm.io avatar

Meta is using to augment unit testing

honestly that seems like a great idea. Repetitive and easily verified. You can even calculate how much value each test is adding

one weird thing — 25% of tests increased coverage but 75% were accepted by engineers. 🤔
https://arxiv.org/abs/2402.09171

kellogh, to random
@kellogh@hachyderm.io avatar

you see that little guy?

kellogh, to random
@kellogh@hachyderm.io avatar

so, who was intended to come through said xz backdoor? who commissioned it?

kellogh, to random
@kellogh@hachyderm.io avatar

Today I'm launching a hosted version of

I noticed that a lot of people were getting hung up on the open source version because they didn't have time to set it up. I'm hoping that this helps.

https://www.fossil-social.com/blog/announcement

kellogh, to random
@kellogh@hachyderm.io avatar

are there mastodon clients that are scriptable?

i keep wishing for something good, like @ivory, but extremely customizable that lets me display toots differently, or filter in non-standard ways, or include links to resources that others might not find useful

kellogh, to climate
@kellogh@hachyderm.io avatar

this is very encouraging. Fuel cells have always been a wild fantasy that doesn’t quite work in the real world, but it seems like a few forces in politics and industry are changing that

https://www.axios.com/2024/02/07/hydrogen-vehicles-fuel-cells-emissions

futurebird, to random
@futurebird@sauropods.win avatar

A second republican, Mike Gallagher, has decided he will not complete his term or seek reelection. Given the narrow majority republicans hold this must be alarming.

It makes me wonder what exactly is going on with the conference? What were they asked to do? There is the chaos with the budget... but I'm thinking about what sort of plans are being laid for November.

Gallagher is a conservative and I don't agree with much of what he's said. But he was very rational in his response to Jan. 6.

kellogh,
@kellogh@hachyderm.io avatar

@futurebird it feels like a mass exodus.

Some replies here are saying it’s, “i can make more money elsewhere”, but no, if that were the case they’d be staying through the end of their term.

this feels like, “i can’t take a single minute more of this madness”. Mass exodus.

ErikJonker, to ai
@ErikJonker@mastodon.social avatar

Groq sounds exciting with it's LPU (Language Processing Unit) as a faster and more efficient alternative to GPU's for LLM's. However difficult to establish the viability with only a website, whitepaper and online demo. Time will tell.
https://groq.com/

kellogh,
@kellogh@hachyderm.io avatar

@ErikJonker imo the big problem is inflexibility. it’s only good for inference, and it’s not even all that helpful for things like embeddings. it has value now, but if a new architecture emerges it could be rendered obsolete overnight

kellogh, to ML
@kellogh@hachyderm.io avatar

GPUs have been good for a lot of things, but is far past the point where specialized hardware will help disproportionately. OTOH benefits disproportionately from the status quo. i imagine it’ll be some other company, like , that dominates in a few years

kellogh,
@kellogh@hachyderm.io avatar

article link: https://www.axios.com/2024/02/21/nvidia-q4-earnings-ai

hgrsd, to rust
@hgrsd@hachyderm.io avatar

Is there any received wisdom on error handling best practices in libraries?

Should I be using Result types throughout and define my own error types for the Error case? Is there a de facto standard library that people use?

kellogh,
@kellogh@hachyderm.io avatar

@hgrsd every time you’ve got a question like this, check here and see if someone’s got a solution. in this case, anyhow or thiserror depending on what you’re doing https://blessed.rs/crates

kellogh, to random
@kellogh@hachyderm.io avatar

PSA a database index is just a separate table stored differently for fast retrieval

in SQL DBs an index is managed for you, so it doesn’t look like a table, but it is.

in NoSQL like elasticsearch or vector stores, they’re just indexes that you’re responsible for keeping in sync with the primary DB (btw always use these as secondary DBs, not primary)

kellogh,
@kellogh@hachyderm.io avatar

if you want to go deeper on databases, indices and other concepts, i highly recommend Designing Data Intensive Systems by @martin

https://www.oreilly.com/library/view/designing-data-intensive-applications/9781491903063/

ErikJonker, to ai
@ErikJonker@mastodon.social avatar

Interesting graph with regard to the various LLMs

kellogh,
@kellogh@hachyderm.io avatar

@ErikJonker i like this view, but admittedly it’s more appealing than it is useful.

when you do price per performance for compute, there actually is a trade off, such that with better performance you can get more work done.

with LLM performance, the trade off isn’t flexible. if the LLM costs too much, you can’t simply use 2 cheaper LLMs that cost half as much

in reality, you have a baseline of performance that need to be maintained, and you choose the cheapest model that crosses that line

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