Ujan

The problem with most machine-based random number generators is that they’re not TRULY random, so if you need genuine randomness it is sometimes necessary to link your code to an external random process like a physical noise source or the current rate of US tariffs on a given country.

YouTube video by WCVB Channel 5 Boston Surveillance shows Tufts graduate student detained

This is horrifying. She co-wrote an OpEd at her school, then was grabbed by masked people, put into an unmarked black van, is now detained in an ICE facility over 1000 miles away.
youtu.be/PuFIs7OkzYY
Trying to limit the topics I post about, but this is my country and this one shook me. =(

Major weaknesses raised by Reviewer 2

any international/intergovernmental report on “AI safety” that fails to acknowledge/ground recommendation on the growing unprecedented corporate power and monopoly is not a friend of the public, civil society, and definitely of marginalised and impacted groups

In fact, I have coauthored a whole paper arguing that the networked nature of modern production can amplify the harm of a small or moderate shock...

way beyond what standard models predict.

bengolub.net/snff-2/

4/

Ep 3 was a banger

It turns out that = trust me bro

Bridging the Human-AI Knowledge Gap: Concept Discovery and Transfer in AlphaZero Artificial Intelligence (AI) systems have made remarkable progress, attaining super-human performance across various domains. This presents us with an opportunity to further human knowledge and improv...

One of my grand interpretability goals is to improve human scientific understanding by analyzing scientific discovery models, but this is the most convincing case yet that we CAN learn from model interpretation: Chess grandmasters learned new play concepts from AlphaZero's internal representations.

I recently talked to someone doing a PhD in multivariate calculus.

Or, as they would call it, a grad student

When they introduced group projects in school they did not tell us that life would be a continuous string of group projects until we retire

Email inbox subject line stating "Your application has moved to the next stage"

Text of the email, stating that the next stage it ... declined.

After several rounds, to the granting agency that sent me this email, with the following subject line:
"Your application has moved to the next stage"

Well done.

That is some Grade 'A' level trolling.

You got me.

“I don’t know when this will happen but probably in 10 years” is researcher-speak for “I have no clue”

I hate what deep learning has done to the word "inference". Now we are forced to write sentences in our papers like "approximate inference at inference time"—I'm all for nonsense Dr Seuss language, but not in a scientific paper!

Every time I go to a public library I begin to believe again that we can create amazing things that are just there for all to enjoy and use for free.

If you have ever wondered what colonization was like—the violence, the dispossession, the massacres, the genocide, the displacement, the famine, the dehumanization, the racism—you are watching it in real time in Palestine. It is among the most depraved of all possible acts.

With the new OpenAI O3 moving performance from 5% up to 25% on FrontierMath it’s time to push open-source models upwards!

We're super happy to release FineMath, the best open math dataset yet. A strong baseline to start training your own models

Find it in the trending section of HuggingFace ;)

"a blue ball is on a wooden table, perfectly motionless. There is no motion in the scene. The camera does not move or pan in any way."

ethernet? you mean the wifi cable

AI influencers "this <big company name / Stanford / Berkeley / MIT>" person/paper says ..... Yes, and? there is always so much context.
But social media incentivizes non-nuanced bite-sized panic/optimism inducing takes. How do we increase the context with which information is disemminated in media?

Two stages of writing:

1) This shouldn't take too long
2) Oh no

GitHub - allenai/OLMo: Modeling, training, eval, and inference code for OLMo Modeling, training, eval, and inference code for OLMo - allenai/OLMo

We just updated the OLMo repo at github.com/allenai/OLMo!
There are now several training configs that together reproduce the training runs that lead to the final OLMo 2 models.
In particular, all the training data is available, tokenized and shuffled exactly as we trained on it!

Let's cycle through the memes for this one until it stops... 😇😅🙏

It is?

The unofficial logo of the Adaptive Agents Lab (Adage). It shows a cylindrical robot with some gauges and buttons on it.

🎉Good news, everyone! 🎉
I will recruit graduate students on the algorithmic and theoretical aspects of Reinforcement Learning.
You will join Adage, @mila-quebec.bsky.social and @polymtl.bsky.social.
More info on why and how you should apply:
academic.sologen.net/2024/11/22/g...
Deadline: Dec 1st

Space: *amorphous blob*

Astronomers: ah yes the choo choo train nebula

Ok, last two papers for this week!

A final game-theoretic RLHF method and a different take on RLHF altogether inspired by prospect theory.

1. 🧲 Magnetic Preference Optimization (MPO).

2. Kahneman-Tversky Optimization (KTO).

🧵 1/3.

I have become a fan of the game-theoretic approaches to RLHF, so here are two more papers in that category! (with one more tomorrow 😅)

1. Self-Play Preference Optimization (SPO).

2. Direct Nash Optimization (DNO).

🧵 1/3.

A screenshot from the linked paper's figure 1. The figure is a pretty-complicated three column figure, but --- in essence, it sketches out how the authors compare llm sequences to the pretraining data / human authors to the pretraining data. Humans write more novel n-gram sequences.

LLMs generate novel word sequences not contained in their pretraining data. However, compared to humans, models generate significantly fewer novel n-grams.

RLHF = 30% *more* copying than base!

Awesome work from the awesome Ximing Lu (gloriaximinglu.github.io) et al. 🤩

arxiv.org/pdf/2410.04265

I've started putting together a starter pack with people working on Speech Technology and Speech Science: go.bsky.app/BQ7mbkA

(Self-)nominations welcome!

Doing good science is 90% finding a science buddy to constantly talk to about the project.