@mathildepapillo

Non-Euclidean + Interpretable AI PhD-ing @Geometric_Intel @UCSBPhysics Fellow @GoodfireAI Prev. @Dodgers⚾️ @McGillUPhysics BlueSky: @mathildepapillon

mathildep.ca
Joined July 2022
  • Tweets501
  • Following387
  • Followers2.6K
50Media
Silico is out🔥 General-purpose human motion AI models can meaningfully detect Parkinson's gait impairment. But how do they do it?? I used Silico to break these models open and figure out what about someone's walk actually drives their predictions. 🧵
Goodfire
@GoodfireAIAug 4
Silico, the platform for ambitious AI research, is publicly available today. AI is advancing fast. The tools to understand it need to advance even faster. Silico lets you interpret and train your models at frontier scale. Learn more + get access 🧵
6
22
6
109
17,852
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Adele Myers@AdeleMyersPhD6h
“To our knowledge, this is the first AI system to pass a professional CAD certification.” I spent 4 months at Spectral Labs this summer. It was exciting to see a company tackle such interesting and ambitious geometry problems. Look forward to seeing what's next 🚀
Spectral Labs
@spectral_hqOct 14
We built the first AI system to achieve the industry standard CAD certification for mechanical engineers. Our agent autonomously built parts and assembled mechanisms, passing the SOLIDWORKS Professional exam in less than 3 hours. Details below, including a concerning discovery.
6
5
14
493
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Nina Miolane 🦋 @ninamiolane.bsky.social@ninamiolane7h
Thanks @NSF for granting us a $1M award to advance the mathematical foundations of AI! W/ @HaewonJeong00 @YaoQin_UCSB Sanjukta K, we will study how AI learns higher-order interactions in physical systems: from 🌠 stars in galaxies to 🧠 neurons in the brain. @ucsbece @ucsbcs
11
8
4
79
3,330
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Symmetry and Geometry in Neural Representations@neur_repsOct 2
NeurReps decisions are out! This year we accepted: - 126 extended abstracts - 60 proceedings - and 6 papers to the new findings track Thank you all for the high quality submissions! We look forward to seeing everyone in Sydney 🚀🇦🇺🐨
1
4
1
33
2,614
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Goodfire
@GoodfireAIOct 1
Biosecurity is the next frontier of AI security. We built SOTA monitors so agents can do more biology, safely. Our monitors outperform frontier model safeguards with fewer refusals on dual-use tasks. They’re fast, real-time, and robust to adversarial attacks. 🧵
7
23
11
138
33,579
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Eric Ho
@eric_hoSep 30
Technical alignment is a science and engineering problem that we can and must solve, and interpretability is the bottleneck. I wrote up Goodfire’s plan to get there.

We can and must solve alignment

Article · @eric_ho
41
86
30
575
90,918
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
David Klindt
@klindt_davidSep 28
🧠 The Science of Intelligence 🧠 Check out the recordings/zooms of this new course at @Redwood_Neuro redwood.berkeley.edu/courses… Very excited to see how Jake, Mike, Bruno and others are transforming the Redwood into the new center for studying and designing a theory of intelligence ⭐️

NEU 290: The Science of Intelligence - Redwood Center for Theoretical Neuroscience

The dual quests – to understand intelligence as it arises in biology and to build it in machines of our own making – have been with us for nearly a century.  While neuroscience and psychology seek to...

redwood.berkeley.edu
2
29
138
6,297
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Goodfire
@GoodfireAISep 28
Are SAEs dead? Will they save us from neuralese? Should I just use a probe? We get these questions all the time. Part 2 of our educational series on applied interpretability explains what SAEs are good for, when *not* to use them, and what to use instead. 🧵
11
52
10
441
38,370
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Thomas Fel
@thomas_fel_Sep 27
An alarming aspect of the OpenAI agent story, aside from the leaked images, is that the problems surfaced months later, and has been found by outsiders We need (1) strong independent verification and (2) fix monitoring now
Deepa Seetharaman
@dseetharamanSep 25
New from @reuters: OpenAI agents posted images belonging to ChatGPT users online, introducing a new area of privacy risk for the company. Story w/ @JeffHorwitz and @razhael. reuters.com/world/openai-wor…

EXCLUSIVE: OpenAI works to understand full scope of agent activity as user data leak emerges

The latest example came when OpenAI said its agents leaked 53 images from ChatGPT users.

reuters.com
1
5
60
3,139
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Nina Miolane 🦋 @ninamiolane.bsky.social@ninamiolaneSep 28
Replying to @ninamiolane @GoodfireAI @geometric_intel
Control follows understanding, and understanding means scientific laws that predict behavior before it happens. Having access to Silico is a game-changer for this AI interpretability research! Thank you!
3
14
343
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Geometric Intelligence Lab@geometric_intelSep 28
Our lab was awarded a @GoodfireAI 's academic research grant 🌟 Many thanks to @GoodfireAI for supporting geometric intelligence research for AI interpretability and safety with free Silico usage.
Nina Miolane 🦋 @ninamiolane.bsky.social@ninamiolaneSep 28
Thank you @GoodfireAI for awarding our lab a research grant on: Geometric Intelligence for AI Interpretability and Safety!🌟 At @geometric_intel, we transform intelligence research from an expensive engineering endeavor into a mathematical science where we can predict behavior.
3
13
834
So excited to see this awesome tool (I'm biased) at work in my academic world🔥 It has helped me kickstart some exciting advances in explainable non-euclidean ML... stay tuned!
Nina Miolane 🦋 @ninamiolane.bsky.social@ninamiolaneSep 28
Thank you @GoodfireAI for awarding our lab a research grant on: Geometric Intelligence for AI Interpretability and Safety!🌟 At @geometric_intel, we transform intelligence research from an expensive engineering endeavor into a mathematical science where we can predict behavior.
5
36
1,667
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Nina Miolane 🦋 @ninamiolane.bsky.social@ninamiolaneSep 21
Every mature engineering discipline found another route to safety. We do not keep bridges standing by watching them closely. We understand the loads they carry to compute, in advance, what they can bear. Safety follows understanding. Understanding means scientific laws that predict behavior before it happens. Can we write such laws for intelligence? I believe that we can, that the mathematics for it already exists, and that they apply across brains and machines. 🤖 🧠 Join the discussion on Oct 7th!
2
5
17
596
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Nina Miolane 🦋 @ninamiolane.bsky.social@ninamiolaneSep 21
🗓️ Join me on October 7 for one of the inaugural lectures of the Science of Intelligence Institute at @UCBerkeley ! Can we write scientific laws for intelligence? I believe that we can, that the math for it already exists, and that they apply across brains & machines. tinyurl.com/mvrfs39w
2
18
3
84
5,483
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Tom McGrath
@banburismus_Sep 17
we can find reward hacks on open frontier models with interpretability, including hacks that strong llm monitors miss! interestingly, the relevant representations seem to be about cheating in general, rather than task-specific
Goodfire
@GoodfireAISep 17
Models know when they’re reward hacking. But they still do it a ton - in 50-96% of rollouts we studied! We built activation monitors that detect the behavior behind the Hugging Face hack in real time. This can help us stop hacks now - and train future models that don’t cheat. 🧵
3
3
44
2,062
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Eric Ho
@eric_hoSep 17
we've just had a huge leap in detecting reward hacking in real time with an activation monitor. next up is figuring out how to mitigate it before it happens. one of the most compelling parts of this work is that our activation monitors catch models when they're contemplating a reward hack before it even happens, and can flag reward hacks that external judges miss.
Goodfire
@GoodfireAISep 17
Models know when they’re reward hacking. But they still do it a ton - in 50-96% of rollouts we studied! We built activation monitors that detect the behavior behind the Hugging Face hack in real time. This can help us stop hacks now - and train future models that don’t cheat. 🧵
10
17
2
253
12,599
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Thomas Fel
@thomas_fel_Sep 17
Cheating crystallizes into a single readable direction, that's the optimistic part ! As models scale, the concepts we most need to monitor *may* get more legible, not less Capability & interpretability, coevolving instead of trading off arxiv.org/abs/2609.19101

Monitoring and Discovering Reward Hacking with Internal...

As models scale, reward hacking becomes more frequent, more sophisticated, and more consequential. Does it leave a telltale signature in model representations? This work analyzes how reward...

arxiv.org
Goodfire
@GoodfireAISep 17
Models know when they’re reward hacking. But they still do it a ton - in 50-96% of rollouts we studied! We built activation monitors that detect the behavior behind the Hugging Face hack in real time. This can help us stop hacks now - and train future models that don’t cheat. 🧵
8
86
3,660
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Ekdeep Singh Lubana@EkdeepLSep 17
A goal we've been on is just scaling up interp to large-scale models and understanding what succeeds or fails: simple stuff both scales and works well! Check out the resampling results, where simple diff of means vector allows us to compute model propensity for reward hacking!
Goodfire
@GoodfireAISep 17
Models know when they’re reward hacking. But they still do it a ton - in 50-96% of rollouts we studied! We built activation monitors that detect the behavior behind the Hugging Face hack in real time. This can help us stop hacks now - and train future models that don’t cheat. 🧵
1
3
38
1,874
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Usha Bhalla@ushabhalla_Sep 17
big! model! interp!! really excited about white-box methods for frontier-scale models!!
Goodfire
@GoodfireAISep 17
Models know when they’re reward hacking. But they still do it a ton - in 50-96% of rollouts we studied! We built activation monitors that detect the behavior behind the Hugging Face hack in real time. This can help us stop hacks now - and train future models that don’t cheat. 🧵
6
116
9,255
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Goodfire
@GoodfireAISep 17
We’re optimistic that every training run can be monitored and that the current rates of reward hacking may soon be a thing of the past. Read the full post + paper: goodfire.com/research/reward…

Models know when they’re reward hacking — and we can catch them at scale - Goodfire

We found a clear internal signal in models that accompanies reward hacking, and built probes that detect it — enabling efficient, real-time detection of reward hacking at scale.

goodfire.com
1
10
2
122
28,462
Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Goodfire
@GoodfireAISep 17
Models know when they’re reward hacking. But they still do it a ton - in 50-96% of rollouts we studied! We built activation monitors that detect the behavior behind the Hugging Face hack in real time. This can help us stop hacks now - and train future models that don’t cheat. 🧵
31
100
47
852
155,819
Load more