iiinova — where your digital self keeps evolving

Finbarr @finbarr.bsky.social · 07/12/2024

would love your take when you do!

Finbarr @finbarr.bsky.social · 07/12/2024

This is one of my all time favorite papers: openreview.net/forum?id=ByJ... It shows that, under fair experimental evaluation, lstms do just as well as a bunch of “improvements”

openreview.net

On the State of the Art of Evaluation in Neural Language Models

Show that LSTMs are as good or better than recent innovations for LM and that model evaluation is often unreliable.

3254

Finbarr @finbarr.bsky.social · 06/12/2024

Fun fact: I recently encountered (well, saw on the news) the only other person named finbarr in Canada I’ve ever seen. The only issue is, he was an arsonist who set a ton of fires in Edmonton.

120

Finbarr @finbarr.bsky.social · 03/12/2024

This is mckernan! What I thought was a nice neighborhood 😂

110

Finbarr @finbarr.bsky.social · 03/12/2024

Apparently there *is* another finbar(r) in Alberta.

050

Finbarr @finbarr.bsky.social · 03/12/2024

New homeowner fear unlocked; someone hit and ran my neighbor’s garage

130

Finbarr @finbarr.bsky.social · 02/12/2024

I thought that was ai gen at first!

130

Finbarr @finbarr.bsky.social · 01/12/2024

there’s a type of “not trying” which means not executing at the level of competence of a $XX billion corporation this is the complaint about eg Google products. They’re good! better than most startups! But not “trillion dollar corporation famed for engineering expertise” good.

080

Finbarr @finbarr.bsky.social · 30/11/2024

would also accept Austria

240

Finbarr @finbarr.bsky.social · 30/11/2024

I watched too many ski movies and now am trying to convince my wife we should move to Alaska

250

Finbarr @finbarr.bsky.social · 30/11/2024

building my own mlp implementation from scratch in numpy, including backprop, remains one of the most educational exercises I’ve done

2190

Finbarr @finbarr.bsky.social · 26/11/2024

Ha, it’s been on my list of todos for a while! I’m glad someone got to it.

010

Finbarr @finbarr.bsky.social · 26/11/2024

Love this. Very clean implementations of various inference optimizations.

160

Finbarr @finbarr.bsky.social · 26/11/2024

Agreed! Folk knowledge is worth publishing!

1100

Finbarr @finbarr.bsky.social · 26/11/2024

I mailed this out like a month ago and just never did the promo 🙈

120

Finbarr @finbarr.bsky.social · 26/11/2024

seems like we're seeing convergence in VLM design. most recent models (Pixtral, PaliGemma, etc) are moving away from complex fusion techniques toward simpler approaches as usual, the bitter lesson holds: better to learn structure than impose it incompleteideas.net/IncIdeas/Bit...

incompleteideas.net

The Bitter Lesson

170

Finbarr @finbarr.bsky.social · 26/11/2024

open source VLMs use relatively little compute compared to what you might expect: LLaVa: 768 A100 hours DeepSeek-VL: 61,440 A100 hours PaliGemma: ~12k A100 hours (for reference, Stable Diffusion used 150k A100 hours)

140

Finbarr @finbarr.bsky.social · 26/11/2024

what i found interesting: VLMs are way simpler than they first appear. current SOTA is basically: 1. ViT encoder (init from SigLIP/CLIP) 2. pretrained LLM base 3. concat image features with text 4. finetune

141

Finbarr @finbarr.bsky.social · 26/11/2024

my latest article for Artificial Fintelligence is up. i cover the evolution of Vision Language Models over the past few years, from complex architectures to surprisingly simple & effective ones. (link in next tweet)

4384

Finbarr @finbarr.bsky.social · 25/11/2024

I mean I can’t get a job at an elite law firm in New York without 1) a law degree 2) from an elite institution

010

Finbarr @finbarr.bsky.social · 25/11/2024

what's wrong with the pixel art?

110

Finbarr @finbarr.bsky.social · 25/11/2024

I think OpenAI’s program is still running? I know one of the main sora contributors just did it. Google makes changes for non-obvious reasons, not always because it’s the optimal choice (see: laying off Rich Sutton)

140

Finbarr @finbarr.bsky.social · 25/11/2024

the tech industry is like California: it gets a few of the most important things right, so it can make a lot of other mistakes and still be wildly successful

030

Finbarr @finbarr.bsky.social · 25/11/2024

I suspect we'd see similar outcomes if there were e.g. a Cravath Legal Resident program or a Mayo Clinic Surgical Resident program the barrier to entry is b/c of tech's lack of credential gate keeping, which is one of tech’s greatest advantages

120

Finbarr @finbarr.bsky.social · 25/11/2024

the remarkable success of the Google brain (and OpenAI) resident programs is an indication to me that smart, hardworking people can do more than you expect

4191