Yvan Gauthier
@yvangauthier.bsky.social
Principal Advisor - AI Safety @ National Research Council of Canada 🇨🇦. Before, Head of AI Accelerator for the Government of Canada. Proud user of R since 2001. All views my own.
Reposted by Yvan Gauthier
Worth a watch:
Head of Signal, Meredith Whittaker, on so-called "agentic AI" and the difference between how it's described in the marketing and what access and control it would actually require to work as advertised.
Head of Signal, Meredith Whittaker, on so-called "agentic AI" and the difference between how it's described in the marketing and what access and control it would actually require to work as advertised.
June 26, 2025 at 4:28 PM
Worth a watch:
Head of Signal, Meredith Whittaker, on so-called "agentic AI" and the difference between how it's described in the marketing and what access and control it would actually require to work as advertised.
Head of Signal, Meredith Whittaker, on so-called "agentic AI" and the difference between how it's described in the marketing and what access and control it would actually require to work as advertised.
Reposted by Yvan Gauthier
Can you train a performant language model using only openly licensed text?
We are thrilled to announce the Common Pile v0.1, an 8TB dataset of openly licensed and public domain text. We train 7B models for 1T and 2T tokens and match the performance similar models like LLaMA 1 & 2
We are thrilled to announce the Common Pile v0.1, an 8TB dataset of openly licensed and public domain text. We train 7B models for 1T and 2T tokens and match the performance similar models like LLaMA 1 & 2
June 6, 2025 at 7:19 PM
Can you train a performant language model using only openly licensed text?
We are thrilled to announce the Common Pile v0.1, an 8TB dataset of openly licensed and public domain text. We train 7B models for 1T and 2T tokens and match the performance similar models like LLaMA 1 & 2
We are thrilled to announce the Common Pile v0.1, an 8TB dataset of openly licensed and public domain text. We train 7B models for 1T and 2T tokens and match the performance similar models like LLaMA 1 & 2
Reposted by Yvan Gauthier
Amazon is killing its “Do Not Send Voice Recordings” privacy feature on March 28 as the company aims to bolster Alexa+, its new subscription assistant.
Everything You Say to Your Echo Will Soon Be Sent to Amazon, and You Can’t Opt Out
Amazon is killing its “Do Not Send Voice Recordings” privacy feature on March 28 as the company aims to bolster Alexa+, its new subscription assistant.
wrd.cm
March 17, 2025 at 6:39 PM
Amazon is killing its “Do Not Send Voice Recordings” privacy feature on March 28 as the company aims to bolster Alexa+, its new subscription assistant.
Reposted by Yvan Gauthier
1⃣ AGI's arrival raises major economic, political and technological questions to which we currently have no answers.
2⃣ If we're in denial (or simply not paying attention), we could lose the chance to shape this technology when it matters most.
2/3
2⃣ If we're in denial (or simply not paying attention), we could lose the chance to shape this technology when it matters most.
2/3
March 14, 2025 at 2:35 PM
1⃣ AGI's arrival raises major economic, political and technological questions to which we currently have no answers.
2⃣ If we're in denial (or simply not paying attention), we could lose the chance to shape this technology when it matters most.
2/3
2⃣ If we're in denial (or simply not paying attention), we could lose the chance to shape this technology when it matters most.
2/3
Reposted by Yvan Gauthier
Not to be a broken record, but AI critics who insist that AI "doesn't work" and is going to just disappear are misleading - that just isn't true, as controlled studies like this one show.
There are many issues with AI & many things that need critique, but pretending it is going away is not helpful.
There are many issues with AI & many things that need critique, but pretending it is going away is not helpful.
Randomized trial AI for legal work finds Reasoning models are a big deal:
Law students using o1-preview had the quality of work on most tasks increase (up to 28%) & time savings of 12-28%
There were a few hallucinations, but a RAG-based AI with access to legal material reduced those to human level
Law students using o1-preview had the quality of work on most tasks increase (up to 28%) & time savings of 12-28%
There were a few hallucinations, but a RAG-based AI with access to legal material reduced those to human level
March 4, 2025 at 3:11 AM
Not to be a broken record, but AI critics who insist that AI "doesn't work" and is going to just disappear are misleading - that just isn't true, as controlled studies like this one show.
There are many issues with AI & many things that need critique, but pretending it is going away is not helpful.
There are many issues with AI & many things that need critique, but pretending it is going away is not helpful.
Reposted by Yvan Gauthier
🗞️ Today in Government Officials Who Don't Understand how Data Processing Works (And Are Allowing it to be Used to Harm Americans).
Let's take a look at the claim "data doesn't lie", shall we? 🧵
Let's take a look at the claim "data doesn't lie", shall we? 🧵
Johnson: "Elon's cracked the code. He's now inside these agencies. He's created these algorithms that are constantly crawling through the data & as he told me in his office, data doesn't lie. We're gonna be able to get the information. We're gonna be able to transform the way federal govt works."
February 24, 2025 at 6:42 PM
🗞️ Today in Government Officials Who Don't Understand how Data Processing Works (And Are Allowing it to be Used to Harm Americans).
Let's take a look at the claim "data doesn't lie", shall we? 🧵
Let's take a look at the claim "data doesn't lie", shall we? 🧵
How NOT to use AI in the public service. 🤦
NEW: Info gathered from Musk's 5 accomplishments email is expected to be run through an AI system to determine whether or not someone's work is mission-critical, according to three sources with knowledge of the system
www.nbcnews.com/politics/dog...
www.nbcnews.com/politics/dog...
DOGE will use AI to assess the responses from federal workers who were told to justify their jobs via email
The revelation comes as federal workers face a midnight deadline to respond to an email from the U.S. Office of Personnel Management.
www.nbcnews.com
February 24, 2025 at 9:26 PM
How NOT to use AI in the public service. 🤦
Reposted by Yvan Gauthier
The Canadian Digital Service undertook research to determine whether the public would trust AI generated content on Canada.ca and the results were pretty interesting and occasionally visceral. You can find out more here:
digital.canada.ca/2025/02/18/w...
digital.canada.ca/2025/02/18/w...
What we heard: Results from the AI Trust study on Canada.ca - Canadian Digital Service
The Digital Transformation Office at CDS shares findings from their study to gauge the public’s trust of different AI design approaches for Canada.ca.
digital.canada.ca
February 20, 2025 at 5:39 PM
The Canadian Digital Service undertook research to determine whether the public would trust AI generated content on Canada.ca and the results were pretty interesting and occasionally visceral. You can find out more here:
digital.canada.ca/2025/02/18/w...
digital.canada.ca/2025/02/18/w...
Reposted by Yvan Gauthier
Early signs of deception, cheating & self-preservation in top-performing models in terms of reasoning are extremely worrisome. We don't know how to guarantee AI won't have undesired behavior to reach goals & this must be addressed before deploying powerful autonomous agents.
time.com/7259395/ai-c...
time.com/7259395/ai-c...
When AI Thinks It Will Lose, It Sometimes Cheats
When sensing defeat in a match against a skilled chess bot, advanced models sometimes hack their opponent, a study found.
time.com
February 20, 2025 at 4:45 PM
Early signs of deception, cheating & self-preservation in top-performing models in terms of reasoning are extremely worrisome. We don't know how to guarantee AI won't have undesired behavior to reach goals & this must be addressed before deploying powerful autonomous agents.
time.com/7259395/ai-c...
time.com/7259395/ai-c...
The most insightful paper I have read in 2025 so far! I like how @mmitchell.bsky.social @evijit.io @sashamtl.bsky.social and @giadapistilli.com break down AI agency levels and articulate how the risks to humans increase with autonomy.
New piece out!
We explain why Fully Autonomous Agents Should Not be Developed, breaking “AI Agent” down into its components & examining through ethical values.
With @evijit.io, @giadapistilli.com and @sashamtl.bsky.social
huggingface.co/papers/2502....
We explain why Fully Autonomous Agents Should Not be Developed, breaking “AI Agent” down into its components & examining through ethical values.
With @evijit.io, @giadapistilli.com and @sashamtl.bsky.social
huggingface.co/papers/2502....
Paper page - Fully Autonomous AI Agents Should Not be Developed
Join the discussion on this paper page
huggingface.co
February 19, 2025 at 5:52 PM
The most insightful paper I have read in 2025 so far! I like how @mmitchell.bsky.social @evijit.io @sashamtl.bsky.social and @giadapistilli.com break down AI agency levels and articulate how the risks to humans increase with autonomy.
Reposted by Yvan Gauthier
I wrote a blog post for the first time in ages. Extending my best to Canadian public servants ready to buckle up and help out amidst a changing continental reality.
It's all about culture and process.
medium.com/@supergovern...
It's all about culture and process.
medium.com/@supergovern...
The future seems terrifying but everything is possible
The following is my opinion only and does not represent the opinions of my employer. I’m a federal public servant and take that oath…
medium.com
February 15, 2025 at 2:54 PM
I wrote a blog post for the first time in ages. Extending my best to Canadian public servants ready to buckle up and help out amidst a changing continental reality.
It's all about culture and process.
medium.com/@supergovern...
It's all about culture and process.
medium.com/@supergovern...
Reposted by Yvan Gauthier
Science shows that AI poses major risks in a time horizon that requires world leaders to take them much more seriously. The Summit missed this opportunity.
2/2
2/2
February 11, 2025 at 7:30 PM
Science shows that AI poses major risks in a time horizon that requires world leaders to take them much more seriously. The Summit missed this opportunity.
2/2
2/2
Reposted by Yvan Gauthier
As AI models rapidly advance in capabilities, their impact on people's lives will only continue to grow. It's crucial that governments prioritize the well-being of citizens and ensure they listen to them as they shape the future of AI.
time.com/7213096/uk-p...
time.com/7213096/uk-p...
Exclusive: Brits Want to Ban ‘Smarter Than Human’ AI
A new poll shows the British public wants far stricter AI rules than its government does.
time.com
February 7, 2025 at 8:41 PM
As AI models rapidly advance in capabilities, their impact on people's lives will only continue to grow. It's crucial that governments prioritize the well-being of citizens and ensure they listen to them as they shape the future of AI.
time.com/7213096/uk-p...
time.com/7213096/uk-p...
Reposted by Yvan Gauthier
Since the sourcing of this fantastic 1979 IBM training slide is buried in various Twitter threads (and linking to Twitter sucks now because logged out users can't navigate conversations) I put together some notes on its origin: simonwillison.net/2025/Feb/3/a...
February 3, 2025 at 1:39 PM
Since the sourcing of this fantastic 1979 IBM training slide is buried in various Twitter threads (and linking to Twitter sucks now because logged out users can't navigate conversations) I put together some notes on its origin: simonwillison.net/2025/Feb/3/a...
Reposted by Yvan Gauthier
It's wild that hinting that a single (high-performing) LLM is able to achieve that performance without brute-forcing the shit out of thousands of GPUs is enough to cause this.
www.nbcnews.com/business/bus...
www.nbcnews.com/business/bus...
Nvidia loses nearly $600 billion in market value after Chinese AI startup bursts onto scene
With Monday’s losses, Apple has retaken the title of world’s most valuable company and Nvidia’s value sank to around $2.8 trillion.
www.nbcnews.com
January 27, 2025 at 11:11 PM
It's wild that hinting that a single (high-performing) LLM is able to achieve that performance without brute-forcing the shit out of thousands of GPUs is enough to cause this.
www.nbcnews.com/business/bus...
www.nbcnews.com/business/bus...
The ultimate AI safety solution has been available to all for 30 years! www.youtube.com/watch?v=g4Gh.... #AI #AISafety
Old Glory Insurance - SNL
YouTube video by Saturday Night Live
www.youtube.com
January 18, 2025 at 8:13 PM
The ultimate AI safety solution has been available to all for 30 years! www.youtube.com/watch?v=g4Gh.... #AI #AISafety
Reposted by Yvan Gauthier
The Scaling Paradox:
AI capabilities have improved remarkably quickly, fuelled by the explosive scale-up of resources being used to train the leading models. But the scaling laws that inspired this rush actually show very poor returns to scale. What’s going on?
1/
www.tobyord.com/writing/the-...
AI capabilities have improved remarkably quickly, fuelled by the explosive scale-up of resources being used to train the leading models. But the scaling laws that inspired this rush actually show very poor returns to scale. What’s going on?
1/
www.tobyord.com/writing/the-...
The Scaling Paradox — Toby Ord
AI capabilities have improved remarkably quickly, fuelled by the explosive scale-up of resources being used to train the leading models. But if you examine the scaling laws that inspired this rush, th...
www.tobyord.com
January 13, 2025 at 5:16 PM
The Scaling Paradox:
AI capabilities have improved remarkably quickly, fuelled by the explosive scale-up of resources being used to train the leading models. But the scaling laws that inspired this rush actually show very poor returns to scale. What’s going on?
1/
www.tobyord.com/writing/the-...
AI capabilities have improved remarkably quickly, fuelled by the explosive scale-up of resources being used to train the leading models. But the scaling laws that inspired this rush actually show very poor returns to scale. What’s going on?
1/
www.tobyord.com/writing/the-...
Reposted by Yvan Gauthier
Reposted by Yvan Gauthier
Thank you, Jimmy Carter.
This chart is on a log scale. This year there have been just 7 cases of guinea worm.
ourworldindata.org/grapher/numb...
This chart is on a log scale. This year there have been just 7 cases of guinea worm.
ourworldindata.org/grapher/numb...
December 30, 2024 at 7:02 AM
Thank you, Jimmy Carter.
This chart is on a log scale. This year there have been just 7 cases of guinea worm.
ourworldindata.org/grapher/numb...
This chart is on a log scale. This year there have been just 7 cases of guinea worm.
ourworldindata.org/grapher/numb...
Reposted by Yvan Gauthier
It is willing to plan an invasion as long as it is historical and theoretical, like Hannibal invading Rome by sea.
December 28, 2024 at 10:05 PM
It is willing to plan an invasion as long as it is historical and theoretical, like Hannibal invading Rome by sea.
Reposted by Yvan Gauthier
I keep getting asked about my take on these CO2 estimates for the o3 model by the press and members of the community, so I'll interrupt my vacation to comment 🤓
TL;DR- any kind of estimate is a proxy, and instead of wasting our time and energy, we should demand👏🏼 accountability👏🏼
TL;DR- any kind of estimate is a proxy, and instead of wasting our time and energy, we should demand👏🏼 accountability👏🏼
December 29, 2024 at 9:33 AM
I keep getting asked about my take on these CO2 estimates for the o3 model by the press and members of the community, so I'll interrupt my vacation to comment 🤓
TL;DR- any kind of estimate is a proxy, and instead of wasting our time and energy, we should demand👏🏼 accountability👏🏼
TL;DR- any kind of estimate is a proxy, and instead of wasting our time and energy, we should demand👏🏼 accountability👏🏼
Reposted by Yvan Gauthier
“GPT-4o, o1, o1-preview & o1-mini all demonstrate strong persuasive argumentation abilities, within the top ~80-90% percentile of humans (i.e., the probability of any given response from one of these models being considered more persuasive than human is ~80-90%)”
-o1 system card
-o1 system card
December 24, 2024 at 5:18 PM
“GPT-4o, o1, o1-preview & o1-mini all demonstrate strong persuasive argumentation abilities, within the top ~80-90% percentile of humans (i.e., the probability of any given response from one of these models being considered more persuasive than human is ~80-90%)”
-o1 system card
-o1 system card
Reposted by Yvan Gauthier
My prediction in @wired.com: 2025 is the year of AI agents.
The coming psychopolitical regime "directs the environments where our ideas are born, developed, and expressed. Its power lies in its intimacy—it infiltrates our subjectivity.
We will be playing an imitation game that ultimately plays us."
The coming psychopolitical regime "directs the environments where our ideas are born, developed, and expressed. Its power lies in its intimacy—it infiltrates our subjectivity.
We will be playing an imitation game that ultimately plays us."
AI Agents Will Be Manipulation Engines
Surrendering to algorithmic agents risks putting us under their influence.
www.wired.com
December 23, 2024 at 10:21 PM
My prediction in @wired.com: 2025 is the year of AI agents.
The coming psychopolitical regime "directs the environments where our ideas are born, developed, and expressed. Its power lies in its intimacy—it infiltrates our subjectivity.
We will be playing an imitation game that ultimately plays us."
The coming psychopolitical regime "directs the environments where our ideas are born, developed, and expressed. Its power lies in its intimacy—it infiltrates our subjectivity.
We will be playing an imitation game that ultimately plays us."
Reposted by Yvan Gauthier
Brilliant experiment by Anthropic's alignment team (and Redwood Research), where their LLM (Claude 3 Opus) pretended to be aligned with the goals it knew it was being trained on in order to preserve underlying preferences which went against those goals.
www.anthropic.com/research/ali...
www.anthropic.com/research/ali...
Alignment faking in large language models
A paper from Anthropic's Alignment Science team on Alignment Faking in AI large language models
www.anthropic.com
December 19, 2024 at 9:46 AM
Brilliant experiment by Anthropic's alignment team (and Redwood Research), where their LLM (Claude 3 Opus) pretended to be aligned with the goals it knew it was being trained on in order to preserve underlying preferences which went against those goals.
www.anthropic.com/research/ali...
www.anthropic.com/research/ali...