Dr Sasha Luccioni

Introducing Trackio: A Lightweight Experiment Tracking Library from Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

𝗧𝗿𝗮𝗰𝗸𝗶𝗼 𝗶𝘀 𝗼𝘂𝘁! 🚀
If you've been looking for a lightweight, open-source package to use for tracking metrics, parameters, and hyperparameters, look no further! 👩‍💻
Trackio is local-first, free and open-source, and comes integrated with 🤗 Datasets and Spaces 📈
huggingface.co/blog/trackio

Lessons from a Chimp: AI "Scheming" and the Quest for Ape Language We examine recent research that asks whether current AI systems may be developing a capacity for "scheming" (covertly and strategically pursuing misaligned goals). We compare current research practice...

I am so grateful that this paper is out!

It so eloquently describes so many of the issues I've seen with a lot of AGI experimentation that concludes that LLMs are scheming, ill-intentioned, or malicious, and the chimp language metaphor is the cherry on the cake 🍰

arxiv.org/abs/2507.03409

Friends!
Does anyone know of any model distillation with public logs (W&B or other)?
I'm trying to figure out the energy tradeoffs between model training and distillation..

. @sashamtl.bsky.social takes centre stage at the AI for Good Summit, hitting hard with the urgent and critical issues of Co2 emissions, energy demands, and cooling water needs demands of AI, stressing the planet

I was so tempted to take a Sharpie to this monstrosity 😇

Sometimes it's hard to find allies in a big room full of people talking about AI.. thankfully @abeba.bsky.social is always there to kick ass and call for accountability 🦋

Thinking about Montréal? 😍

Montréal International and the Ville de Montréal have unveiled a new video to attract U.S.-based researchers and scientists!

Featuring:

✔️ @sashamtl.bsky.social
✔️ @weisslab.bsky.social
✔️ Dr Rachel Bouserhal

#montreal #research #scientists

YouTube video by Montréal International R&D in Montréal | Montréal Loves Researchers | Harness the Power of your Research in Canada

I've always felt that Montréal is a safe space for research -- it’s a city that values science, collaboration, and diversity, and I’m proud to call it home 🏠

Like I said in the video: Montréal is home—and all are welcome.💖🦋

youtu.be/dWiL4b7-KEA

Tomorrow @facct.bsky.social , 9:36am (☕️☕️☕️), @sashamtl.bsky.social and I on stage questioning the bigger-is-better narrative in AI.

Come for the 🌶, stay for the evidence and critical thinking (hopefully)

a chart showing most models are red - ie, no disclosure - for alphabet, microsoft, openai, anthropic

a chart showing most models are red - ie, no disclosure - for alphabet, microsoft, openai, anthropic

a chart showing most models are red - ie, no disclosure - for alphabet, microsoft, openai, anthropic

a chart showing most models are red - ie, no disclosure - for alphabet, microsoft, openai, anthropic

New @sashamtl.bsky.social et al paper - assessing whether machine learning model developers disclose basic environmental / energy information with their models. They mostly do not!!

Grim results for most of the big companies:

arxiv.org/abs/2506.15572

huggingface.co/spaces/sasha...

Environmental Transparency - a Hugging Face Space by sasha Discover how AI organizations disclose their environmental impact by year and transparency type. Select an organization to see their data visualized.

We also made an interactive demo that you can use to explore our data by model and by org (and see which labs and companies are doing a better job than others!): huggingface.co/spaces/sasha...

Check it out and leave your comments below! 🦋

In it, we discuss how transparency around the environmental impacts of AI has been declining over the years, resulting in a total lack of disclosure about the most popular models that we use now, precluding us from making informed choices. 💻

Does a ChatGPT query really use ten times more energy than a Google search? Does training any AI model emit as much CO2 as 5 cars? and can AI help us reduce global GHG emissions by 10%? 🤔

We tracked down the origins of these numbers 🕵‍♀️ and wrote a paper about it!

arxiv.org/abs/2506.15572

How Much Energy Does AI Use? The People Who Know Aren’t Saying A growing body of research attempts to put a number on energy use and AI—even as the companies behind the most popular models keep their carbon emissions a secret.

how much energy does your ChatGPT search really use? i wrote about the trials and tribulations of putting a number on AI emissions, & how some of the most powerful companies in the world could clear it up for us — but they're staying silent. feat a peek at new research from @sashamtl.bsky.social !

How Much Energy Does AI Use? The People Who Know Aren’t Saying A growing body of research attempts to put a number on energy use and AI—even as the companies behind the most popular models keep their carbon emissions a secret.

One of the biggest frustrations I have is the lack of transparency around AI's energy use and environmental impacts. I know the numbers are out there... but somehow we're not seeing them 🫠

Thank you @wired.com for covering this topic in such depth and detail !

www.wired.com/story/ai-car...

I had someone tell me "I didn't know there was more to AI than LLMs" today and I swear I died a little inside 🥺

Your Honor, please let the record reflect that I had a massive crush on Pedro Pascal way before everyone else did ... over 2 years ago.
Check out the evergreen Space I made to find your inner Pedro here! huggingface.co/spaces/sasha...

Metaethical perspectives on ‘benchmarking’ AI ethics - AI and Ethics Benchmarks are seen as the cornerstone for measuring technical progress in artificial intelligence (AI) research and have been developed for a variety of tasks ranging from question answering to emoti...

I didn't announce this when it happened, but my article, "Metaethical perspectives for 'Benchmarking' AI ethics", with @sashamtl.bsky.social, was published back in March in the AI and Ethics journal.

doi.org/10.1007/s436...

Never thought I'd see my name in a museum exhibit! 🥹
A quote from my TED talk is featured in an exhibit on AI at the Exploratorium in San Francisco -- go check it out if you're in town!

Saying Thank You to a LLM Isn't Free — Measuring the Energy Cost of Politeness A Blog post by Julien Delavande on Hugging Face

This varies depending on the size of the model as well, with bigger models generating more verbose answers than the smaller ones.

... So the cost of politeness does add up! 💡

Check out our blog post here:

huggingface.co/blog/jdelava...

People saying "thank you" to ChatGPT is apparently costing OpenAI tens of millions of dollars -- but how does that translate to energy? ⚡

We ran some experiments where we show that the total energy consumed is ~0.245 Wh (roughly equivalent to powering a 5 W LED bulb for approximately 3 minutes).

Also - SHOW 👏 US 👏 THE 👏 DAMN 👏 RECEIPTS

Show 👏 us 👏 receipts 👏
(this is probably more on Microsoft than on OAI, to be fair)

Source: blog.samaltman.com/the-gentle-s...

Finally some numbers about ChatGPT's energy and water use? ⚡

Even if this is the true amount of energy used for a query (and that's a big if!), multiplied by the billions of queries a day, this adds up massively.

And that's not counting image generation, which is much more energy-intensive!

Trends in AI Supercomputers Frontier AI development relies on powerful AI supercomputers, yet analysis of these systems is limited. We create a dataset of 500 AI supercomputers from 2019 to 2025 and analyze key trends in perform...

For me the takeaway is: we need to explore different ways of training and deploying AI models, ones that don't hinge upon ever-increasing amounts of computer and power.
See the rest of Epoch AI's report here: arxiv.org/abs/2504.16026

"𝘐𝘧 𝘵𝘩𝘦 𝘰𝘣𝘴𝘦𝘳𝘷𝘦𝘥 𝘵𝘳𝘦𝘯𝘥𝘴 𝘤𝘰𝘯𝘵𝘪𝘯𝘶𝘦, 𝘵𝘩𝘦 𝘭𝘦𝘢𝘥𝘪𝘯𝘨 𝘈𝘐 𝘴𝘶𝘱𝘦𝘳𝘤𝘰𝘮𝘱𝘶𝘵𝘦𝘳 𝘪𝘯 2030 𝘸𝘪𝘭𝘭 𝘳𝘦𝘲𝘶𝘪𝘳𝘦 9 𝘎𝘞 𝘰𝘧 𝘱𝘰𝘸𝘦𝘳 (𝘦𝘲𝘶𝘪𝘷𝘢𝘭𝘦𝘯𝘵 𝘵𝘰 9 𝘯𝘶𝘤𝘭𝘦𝘢𝘳 𝘳𝘦𝘢𝘤𝘵𝘰𝘳𝘴, 𝙖 𝙨𝙘𝙖𝙡𝙚 𝙗𝙚𝙮𝙤𝙣𝙙 𝙖𝙣𝙮 𝙚𝙭𝙞𝙨𝙩𝙞𝙣𝙜 𝙞𝙣𝙙𝙪𝙨𝙩𝙧𝙞𝙖𝙡 𝙛𝙖𝙘𝙞𝙡𝙞𝙩𝙮)"
And the conclusion: "𝘗𝘰𝘸𝘦𝘳 𝘤𝘰𝘯𝘴𝘵𝘳𝘢𝘪𝘯𝘵𝘴 𝘸𝘪𝘭𝘭 𝘭𝘪𝘬𝘦𝘭𝘺 𝘣𝘦𝘤𝘰𝘮𝘦 𝘵𝘩𝘦 𝘱𝘳𝘪𝘮𝘢𝘳𝘺 𝘣𝘰𝘵𝘵𝘭𝘦𝘯𝘦𝘤𝘬 𝘧𝘰𝘳 𝘈𝘐 𝘴𝘶𝘱𝘦𝘳𝘤𝘰𝘮𝘱𝘶𝘵𝘦𝘳 𝘨𝘳𝘰𝘸𝘵𝘩"

Main takeaways from @epochai.bsky.social analysis:
"𝘤𝘰𝘮𝘱𝘶𝘵𝘢𝘵𝘪𝘰𝘯𝘢𝘭 𝘱𝘦𝘳𝘧𝘰𝘳𝘮𝘢𝘯𝘤𝘦 𝘰𝘧 𝘈𝘐 𝘴𝘶𝘱𝘦𝘳𝘤𝘰𝘮𝘱𝘶𝘵𝘦𝘳𝘴 𝘩𝘢𝘴 𝘥𝘰𝘶𝘣𝘭𝘦𝘥 𝘦𝘷𝘦𝘳𝘺 𝘯𝘪𝘯𝘦 𝘮𝘰𝘯𝘵𝘩𝘴, 𝘸𝘩𝘪𝘭𝘦 𝘩𝘢𝘳𝘥𝘸𝘢𝘳𝘦 𝘢𝘤𝘲𝘶𝘪𝘴𝘪𝘵𝘪𝘰𝘯 𝘤𝘰𝘴𝘵 𝘢𝘯𝘥 𝘱𝘰𝘸𝘦𝘳 𝘯𝘦𝘦𝘥𝘴 𝘣𝘰𝘵𝘩 𝘥𝘰𝘶𝘣𝘭𝘦𝘥 𝘦𝘷𝘦𝘳𝘺 𝘺𝘦𝘢𝘳."

📢 The Butterfly Effect: measuring AI's Impacts on Society and the Environment : le 11 juin, l’ENS vous convie à une conférence exceptionnelle de @sashamtl.bsky.social‬, titulaire de la chaire “IA et justice sociale” Fondation Abeona - École normale supérieure - Obvia.
📍 ENS
👉 bit.ly/4kxd0ue

If I wanted a machine designed specifically and solely to blow smoke up my ass I could make one a lot cheaper and safer than an LLM