Kaelan Donatella
@kaelandonatella.bsky.social
hardware/software to make ai systems faster and more reliable. i like clouds and french cinema
Nice to see some work trying to disentangle concepts :)
How do LLMs learn to reason from data? Are they ~retrieving the answers from parametric knowledge🦜? In our new preprint, we look at the pretraining data and find evidence against this:
Procedural knowledge in pretraining drives LLM reasoning ⚙️🔢
🧵⬇️
Procedural knowledge in pretraining drives LLM reasoning ⚙️🔢
🧵⬇️
November 20, 2024 at 6:46 PM
Nice to see some work trying to disentangle concepts :)
shape rotators unhappy
Preprint: It's been suggested that Aphants may be able to visualise, but lack insight, as they can do Mental Rotation (MR) tasks. Instead, we show MR tasks are a weak measure of the propensity to visualise.
#Aphantasia #Imagery
www.biorxiv.org/content/10.1...
#Aphantasia #Imagery
www.biorxiv.org/content/10.1...
Mental Rotation is a weak measure of the propensity to visualise
There is increasing evidence of substantial differences in the capacity of people to voluntarily visualise, with some (Congenital Aphants) asserting they cannot visualise at all. It has been suggested...
www.biorxiv.org
November 14, 2024 at 10:18 AM
shape rotators unhappy
I like it here but there are way too many posts about bsky itself
November 14, 2024 at 10:17 AM
I like it here but there are way too many posts about bsky itself
if we can get the same type of paper discussion content without the ai influencers here that would be so nice
One thing still missing here is good discussion of academic papers, so I guess I'll be the change I want to see in the world
Really interesting results from Jacob Andreas's group showing great performance on ARC-AGI just from doing a few gradient descent steps at test time arxiv.org/abs/2411.07279
Really interesting results from Jacob Andreas's group showing great performance on ARC-AGI just from doing a few gradient descent steps at test time arxiv.org/abs/2411.07279
The Surprising Effectiveness of Test-Time Training for Abstract Reasoning
Language models have shown impressive performance on tasks within their training distribution, but often struggle with novel problems requiring complex reasoning. We investigate the effectiveness of t...
arxiv.org
November 13, 2024 at 4:30 PM
if we can get the same type of paper discussion content without the ai influencers here that would be so nice