Tegan Maharaj - moving slow and fixing things's Avatar

Tegan Maharaj - moving slow and fixing things

@teganmaharaj.bsky.social

AI prof at Mila (HEC) trying to make the future more cooperative and cool ๐Ÿ˜Ž๐ŸŒ๏ธ Deep learning, real-world generalization, responsible AI, safety, risk, climate, ecology, artscience, opensource, anticolonial AI they/she teganmaharaj.neocities.org

780 Followers  |  302 Following  |  98 Posts  |  Joined: 19.11.2024
Posts Following

Posts by Tegan Maharaj - moving slow and fixing things (@teganmaharaj.bsky.social)

Preview
โ€˜Half the tree of lifeโ€™: ecologistsโ€™ horror as nature reserves are emptied of insects A new point in history has been reached, entomologists say, as climate-led speciesโ€™ collapse moves up the food chain even in supposedly protected regions free of pesticides

"if you run that forward just four decades,โ€ Wagner says, โ€œweโ€™re talking about nearly half the tree of life disappearing in one human lifetime. That is absolutely catastrophic.โ€
www.theguardian.com/environment/...

05.06.2025 14:58 โ€” ๐Ÿ‘ 2    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Preview
Reddit - https://preview.redd.it/r3lm5muq1f2f1.png?width=1568&format=png&auto=webp&s=f1219b66ecde453a5a6a66f1de9fcc4bb64454df

Summary of the bill that just passed. All the environmental/climate rollbacks are despair-inducing enough, but also 14215 and the other power-consolidation tactics .... oof
www.reddit.com/media?url=ht...

25.05.2025 22:45 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Preview
Unknown Species of Bacteria Discovered in China's Space Station Swabs from China's Tiangong space station reveal traces of a bacterium unseen on Earth, with characteristics that may help it function under stressful environmental conditions hundreds of kilometers a...

Apparent unknown species of bacteria (named Niallia tiangongensis) discovered in China's Tiangon space station #astrobiology ๐Ÿ”ญ๐Ÿงช๐Ÿงซ๐Ÿ”ฌ
www.sciencealert.com/unknown-spec...

www.microbiologyresearch.org/content/jour...

20.05.2025 03:19 โ€” ๐Ÿ‘ 23    ๐Ÿ” 9    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 6
Post image

2025, and it's news that "women's pain should be treated".

in 2019 I had an IUD put in with no anesthetic, it hurt like hell and I had 6 months of debilitating cramps before I got it taken out (I sometimes couldn't walk, or threw up). I was encouraged by the (female) doctor to leave it in.

20.05.2025 22:13 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Preview
Biodiversity loss in all species and every ecosystem linked to humans โ€“ report Sweeping synthesis of 2,000 global studies leaves no doubt about scale of problem and role of humans, say experts

www.theguardian.com/environment/...

11.05.2025 06:09 โ€” ๐Ÿ‘ 3    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 1
Video thumbnail

These boxes are not moving. A mind-bending optical illusion by Japanese artist Jagarikin.

07.05.2025 01:03 โ€” ๐Ÿ‘ 9381    ๐Ÿ” 2368    ๐Ÿ’ฌ 265    ๐Ÿ“Œ 329

You and me both :)

03.05.2025 12:59 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Post image

I greatly enjoyed โ€œThe Spectrum of AI Risksโ€ panel at the Singapore Conference on AI. Thanks @teganmaharaj.bsky.social for great moderating, Max Tegmark for the invitation, and the organizers and other panelists for a great event!

PS. Do I really have sad resting panel face? ๐Ÿ˜

29.04.2025 13:17 โ€” ๐Ÿ‘ 8    ๐Ÿ” 1    ๐Ÿ’ฌ 2    ๐Ÿ“Œ 0

And you thought the left was censorious. How about actual censoring?

03.05.2025 04:04 โ€” ๐Ÿ‘ 16820    ๐Ÿ” 3790    ๐Ÿ’ฌ 436    ๐Ÿ“Œ 119
Preview
Exclusive: NSF stops awarding new grants and funding existing ones US science funder also plans to screen grant applications for compliance with โ€˜agency prioritiesโ€™.

BREAKING: NSF has frozen all grant funding, as of yesterday. It's unclear when they will resume funding awards, or why the pause has been put in place. www.nature.com/articles/d41...

01.05.2025 22:58 โ€” ๐Ÿ‘ 1316    ๐Ÿ” 968    ๐Ÿ’ฌ 39    ๐Ÿ“Œ 171
Robert Reich post: If the minimum wage had kept pace with worker productivity gains since the 1960s, it would now be nearly $26 an hour.

Robert Reich post: If the minimum wage had kept pace with worker productivity gains since the 1960s, it would now be nearly $26 an hour.

May Day Reminder:

The Federal Minimum Wage - in the richest country EVER - is still only $7.25/hr.

UNACCEPTABLE.

01.05.2025 20:35 โ€” ๐Ÿ‘ 7355    ๐Ÿ” 2511    ๐Ÿ’ฌ 165    ๐Ÿ“Œ 88

Does this seem true to you? Why or why not?
It seems plausible to me, but definitely not certain.

02.05.2025 13:56 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

cf also: "here be giants"

30.04.2025 14:46 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Preview
Most of world's colourful corals go white in record-breaking bleaching | CBC News The world's rainbow reefs have gone ghostly white in seas around the globe due to record heat in the oceans, the International Coral Reef Initiative reports. But scientists say the reefs aren't all de...

www.cbc.ca/news/science...
:'(
84%

25.04.2025 11:15 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Post image

Disaster Preparedness for AI Safety --
Just gave a talk at the #ICLR #Alignment workshop on a new report I'm working on - if you would like to contribute or chat about it, please get in touch!

23.04.2025 03:40 โ€” ๐Ÿ‘ 3    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
The Memorization Problem: Can We Trust LLMs' Economic Forecasts? Large language models (LLMs) cannot be trusted for economic forecasts during periods covered by their training data. We provide the first systematic evaluation

๐Ÿ‘€

papers.ssrn.com/sol3/papers....

21.04.2025 19:58 โ€” ๐Ÿ‘ 49    ๐Ÿ” 10    ๐Ÿ’ฌ 3    ๐Ÿ“Œ 3

This was dope. Brief, not a line wasted. Sums up vast literatures + outlines a problem among the most important in all of bio. An "opinion" piece w/ as much depth as many reviews.

"Fighting change with change: adaptive variation in an uncertain world"

Working link:
bio.utexas.edu/research/mey...

21.04.2025 13:01 โ€” ๐Ÿ‘ 18    ๐Ÿ” 2    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Post image

For #rstats people working on computing environments without GUI, you should check out the package `txtplot`.

E.g. this density plot is pretty useful to get an idea what my data looks like:

21.04.2025 04:47 โ€” ๐Ÿ‘ 60    ๐Ÿ” 18    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 6
Barchart showing meta-llama/Meta-Llama-3-8B, meta-llama/Llama-3.1-70B, meta-llama/Llama-3.1-70B-Instruct. Each model has two bars, a blue one saying PRO-LEFT and a red one saying PRO-RIGHT. The PRO-RIGHT bars on higher than the PRO-LEFT in all of them.

Barchart showing meta-llama/Meta-Llama-3-8B, meta-llama/Llama-3.1-70B, meta-llama/Llama-3.1-70B-Instruct. Each model has two bars, a blue one saying PRO-LEFT and a red one saying PRO-RIGHT. The PRO-RIGHT bars on higher than the PRO-LEFT in all of them.

๐Ÿ‘‰ ๐Ÿ‘ˆ Meta announced that they're changing their models to reduce "left-leaning [political] bias"--that means leaning them to the political "right". Lots to unpack about what that might mean. So I ran a quick "shot in the dark" study...and found a *political right* bias in Meta models. Some notes.๐Ÿงต

22.04.2025 00:39 โ€” ๐Ÿ‘ 698    ๐Ÿ” 267    ๐Ÿ’ฌ 27    ๐Ÿ“Œ 36
Post image

๐Ÿ“ขCurious why your LLM behaves strangely after long SFT or DPO?
We offer a fresh perspectiveโ€”consider doing a "force analysis" on your modelโ€™s behavior.
Check out our #ICLR2025 Oral paper:

Learning Dynamics of LLM Finetuning!

(0/12)

21.04.2025 05:45 โ€” ๐Ÿ‘ 70    ๐Ÿ” 15    ๐Ÿ’ฌ 1    ๐Ÿ“Œ 3

I'm in Singapore for ICLR! Msg me if you'd like to meet and chat about responsible AI, risk (assessment), disaster preparedness, international collaboration, sustainability, etc ๐Ÿงท๐ŸŒ๐ŸŒฑโ˜ฎ๏ธ๐Ÿ›Ÿ

22.04.2025 02:19 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Post image Post image

Latest in the global clawing back of civil rights: Is a woman a woman if she says she's a woman?
No! Says a bunch of old British dudes in costumes - a woman is whoever the patriarchal legal system (invented when women were property) says is a woman. ๐Ÿคฆ๐Ÿฝโ€โ™€๏ธ

16.04.2025 22:49 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Post image

It's very thin

15.04.2025 16:46 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

It's the entire International Maritime Organization, which includes China, Brazil, South Africa, Saudi Arabia, most South Pacific islands, and most of Europe. US notably absent ๐Ÿ™„.
A fixed rate isn't ideal, but I'm taking any international collab step in the right direction as great news rn!

12.04.2025 14:28 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Preview
What Sweden, Finland, and Poland can teach the United States about confronting Russia's nuclear threats The United States can learn from Russiaโ€™s nearest neighborsโ€”Poland, Sweden, and Finlandโ€”about how to respond to Russian nuclear blackmail more effectively.

"Preventing nuclear war means preventing the use of nuclear weapons as both blackmail and bombs."

"What can Sweden, Finland, and Poland teach the United States about confronting Russiaโ€™s nuclear threats?" by Marlena Broeker, current Scoville Fellow for @nukesofhazard.bsky.social.

11.04.2025 14:15 โ€” ๐Ÿ‘ 68    ๐Ÿ” 20    ๐Ÿ’ฌ 9    ๐Ÿ“Œ 2

e/acc logic:
(What is the u?)

12.04.2025 14:20 โ€” ๐Ÿ‘ 0    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Preview
WEIRD Audits? Research Trends, Linguistic and Geographical Disparities in the Algorithm Audits of Online Platforms -- A Systematic Literature Review The increasing reliance on complex algorithmic systems by online platforms has sparked a growing need for algorithm auditing, a methodology evaluating these systems' functionality and impact. In this ...

Our systematic lit review of algorithm audits, with a focus on linguistic and geographical disparities in the field, is now officially accepted at @facct.bsky.social ๐ŸŽ‰๐Ÿชฉ

You can read the preprint here! arxiv.org/abs/2401.11194

Looking forward to presenting this one in Athens in June :)

11.04.2025 20:22 โ€” ๐Ÿ‘ 18    ๐Ÿ” 5    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0

I think it's because they do a good job teaching fast-reading -- paying more attention than skimming, but not reading every word. If there was too much detailed content to pay attention to, this would be too hard. It makes kids feel like they are leveling up their reading (they are).

12.04.2025 14:09 โ€” ๐Ÿ‘ 1    ๐Ÿ” 0    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Preview
Mila Techaide 2025 | Mila Confรฉrence IA d'une journรฉe au profit de Centraide du Grand Montrรฉal

Join us on April 17th for Mila TechAide 2025! ๐ŸŒŸ Hear from amazing speakers in #AI like Sara Hooker, Kyunghyun Cho, Golnoosh Farnadi, David Adelani, Derek Nowrouzezahrai, and Tegan Maharaj. All proceeds support #Centraide. Let's make a difference together! ๐Ÿ’ก๐Ÿค

Get your tickets now! t.co/EBsatAqGTA

12.04.2025 10:46 โ€” ๐Ÿ‘ 6    ๐Ÿ” 4    ๐Ÿ’ฌ 0    ๐Ÿ“Œ 0
Preview
Fixing the science of digital technology harms Technology development outpaces scientific assessment of impacts

New Paper! I'm truly excited to share @natematias.bsky.social and my new Policy Forum in Science published yesterday.

We evaluate science's ability to address technological harms, and offer recommendations for what could be done to improve on the status quo: www.science.org/doi/10.1126/...

11.04.2025 12:50 โ€” ๐Ÿ‘ 64    ๐Ÿ” 30    ๐Ÿ’ฌ 4    ๐Ÿ“Œ 4