Orpheus Lummis
@orpheuslummis.info
Advancing AI safety through convenings, coordination, software
https://orpheuslummis.info, based in Montréal
https://orpheuslummis.info, based in Montréal
Reposted by Orpheus Lummis
AI Manipulation Hackathon (Montréal edition, Friday evening to to Sunday midday)
Build benchmarks, detection + mitigations against deception/sycophancy/sandbagging/dark patterns. $2000 prizes + possible Apart Fellowship + Paris workshop slot.
RSVP: luma.com/mf69p5pi
Build benchmarks, detection + mitigations against deception/sycophancy/sandbagging/dark patterns. $2000 prizes + possible Apart Fellowship + Paris workshop slot.
RSVP: luma.com/mf69p5pi
AI Manipulation Hackathon · Luma
This is the Montréal edition of the AI Manipulation Hackathon.
Friday evening to Sunday midday.
RSVP here and on the official page to be admissible.
Location…
luma.com
January 4, 2026 at 1:12 PM
AI Manipulation Hackathon (Montréal edition, Friday evening to to Sunday midday)
Build benchmarks, detection + mitigations against deception/sycophancy/sandbagging/dark patterns. $2000 prizes + possible Apart Fellowship + Paris workshop slot.
RSVP: luma.com/mf69p5pi
Build benchmarks, detection + mitigations against deception/sycophancy/sandbagging/dark patterns. $2000 prizes + possible Apart Fellowship + Paris workshop slot.
RSVP: luma.com/mf69p5pi
Reposted by Orpheus Lummis
HΩ's year in review
In 2025, we produced the Guaranteed Safe AI Seminars, cultivated the Montréal AI safety community, realized the Limits to Control workshop, and more.
We are also moving from a volunteer org to one with more capacity.
Onward to 2026!
horizonomega.substack.com/p/h-2025-rev...
In 2025, we produced the Guaranteed Safe AI Seminars, cultivated the Montréal AI safety community, realized the Limits to Control workshop, and more.
We are also moving from a volunteer org to one with more capacity.
Onward to 2026!
horizonomega.substack.com/p/h-2025-rev...
HΩ 2025 Review
What is HΩ?
horizonomega.substack.com
December 23, 2025 at 6:29 PM
HΩ's year in review
In 2025, we produced the Guaranteed Safe AI Seminars, cultivated the Montréal AI safety community, realized the Limits to Control workshop, and more.
We are also moving from a volunteer org to one with more capacity.
Onward to 2026!
horizonomega.substack.com/p/h-2025-rev...
In 2025, we produced the Guaranteed Safe AI Seminars, cultivated the Montréal AI safety community, realized the Limits to Control workshop, and more.
We are also moving from a volunteer org to one with more capacity.
Onward to 2026!
horizonomega.substack.com/p/h-2025-rev...
Joyeux solstice! 🌞
December 21, 2025 at 2:42 PM
Joyeux solstice! 🌞
Reposted by Orpheus Lummis
Deepmind is releasing SAEs and transcoders for their Gemma 3 models including the 27B as part of Gemma Scope 2, exciting
Gemma Scope 2: Helping the AI Safety Community Deepen Understanding of Complex Language Model Behavior
Announcing Gemma Scope 2, a comprehensive, open suite of interpretability tools for the entire Gemma 3 family to accelerate AI safety research.
deepmind.google
December 19, 2025 at 4:17 PM
Deepmind is releasing SAEs and transcoders for their Gemma 3 models including the 27B as part of Gemma Scope 2, exciting
Reposted by Orpheus Lummis
Dim Red Dot
Scientists have just released a photo featuring a dim red dot. It is the light of a single star exploding in a galaxy so far far away that that nothing we do could ever affect it — even in the very fullness of time.
It lies beyond the Affectable Universe.
Let me explain…
1/🧵
Scientists have just released a photo featuring a dim red dot. It is the light of a single star exploding in a galaxy so far far away that that nothing we do could ever affect it — even in the very fullness of time.
It lies beyond the Affectable Universe.
Let me explain…
1/🧵
December 18, 2025 at 11:17 AM
Dim Red Dot
Scientists have just released a photo featuring a dim red dot. It is the light of a single star exploding in a galaxy so far far away that that nothing we do could ever affect it — even in the very fullness of time.
It lies beyond the Affectable Universe.
Let me explain…
1/🧵
Scientists have just released a photo featuring a dim red dot. It is the light of a single star exploding in a galaxy so far far away that that nothing we do could ever affect it — even in the very fullness of time.
It lies beyond the Affectable Universe.
Let me explain…
1/🧵
Reposted by Orpheus Lummis
breakthroughs of the year
www.renaissancephilanthropy.org/frontier2025
www.renaissancephilanthropy.org/frontier2025
December 17, 2025 at 8:47 PM
breakthroughs of the year
www.renaissancephilanthropy.org/frontier2025
www.renaissancephilanthropy.org/frontier2025
Public letter on Transparency about Third-Party AI Evaluations, from the AI Evaluator Forum.
AI evaluation results should disclose methods used, system access, editorial control, conflicts of interest.
aievaluatorforum.org/initiatives/...
AI evaluation results should disclose methods used, system access, editorial control, conflicts of interest.
aievaluatorforum.org/initiatives/...
AI Evaluator Forum
The AI Evaluator Forum brings together leading independent AI research organizations
aievaluatorforum.org
December 5, 2025 at 12:50 PM
Public letter on Transparency about Third-Party AI Evaluations, from the AI Evaluator Forum.
AI evaluation results should disclose methods used, system access, editorial control, conflicts of interest.
aievaluatorforum.org/initiatives/...
AI evaluation results should disclose methods used, system access, editorial control, conflicts of interest.
aievaluatorforum.org/initiatives/...
Reposted by Orpheus Lummis
Montréal event, Tuesday December 16, 7 PM:
Can AI systems be conscious? How could we know? And why does it matter?
A presentation by @joaquimstreicher.bsky.social, Ph.D. candidate in Neuroscience and co-founder of Montréal Initiative for Consciousness science (MONIC)
luma.com/paky3mih
Can AI systems be conscious? How could we know? And why does it matter?
A presentation by @joaquimstreicher.bsky.social, Ph.D. candidate in Neuroscience and co-founder of Montréal Initiative for Consciousness science (MONIC)
luma.com/paky3mih
Can AI systems be conscious? How could we know? And why does it matter? · Luma
EN
Joaquim Streicher, Ph.D. candidate in Neuroscience and co-founder of Montreal Initiative for Consciousness science (MONIC) will give the following…
luma.com
December 3, 2025 at 5:56 PM
Montréal event, Tuesday December 16, 7 PM:
Can AI systems be conscious? How could we know? And why does it matter?
A presentation by @joaquimstreicher.bsky.social, Ph.D. candidate in Neuroscience and co-founder of Montréal Initiative for Consciousness science (MONIC)
luma.com/paky3mih
Can AI systems be conscious? How could we know? And why does it matter?
A presentation by @joaquimstreicher.bsky.social, Ph.D. candidate in Neuroscience and co-founder of Montréal Initiative for Consciousness science (MONIC)
luma.com/paky3mih
Reposted by Orpheus Lummis
Montréal event, Tuesday December 2, 7 PM:
Veracity in the Age of Persuasive AI – a presentation by Taylor Lynn Curtis, researcher on misinformation at Mila.
luma.com/mmuqltzq
Veracity in the Age of Persuasive AI – a presentation by Taylor Lynn Curtis, researcher on misinformation at Mila.
luma.com/mmuqltzq
Veracity in the Age of Persuasive AI · Luma
EN
Veracity in the Age of Persuasive AI
A presentation by Taylor Lynn Curtis, researcher on misinformation at Mila.
This presentation addresses the growing…
luma.com
December 1, 2025 at 10:58 PM
Montréal event, Tuesday December 2, 7 PM:
Veracity in the Age of Persuasive AI – a presentation by Taylor Lynn Curtis, researcher on misinformation at Mila.
luma.com/mmuqltzq
Veracity in the Age of Persuasive AI – a presentation by Taylor Lynn Curtis, researcher on misinformation at Mila.
luma.com/mmuqltzq
Reposted by Orpheus Lummis
Guaranteed Safe AI Seminars, December 2025:
Safe Learning Under Irreversible Dynamics via Asking for Help
Benjamin Plaut – Postdoc at CHAI studying guaranteed safe AI
Thursday, December 11, 1 PM EST
luma.com/wcww6xpl
Safe Learning Under Irreversible Dynamics via Asking for Help
Benjamin Plaut – Postdoc at CHAI studying guaranteed safe AI
Thursday, December 11, 1 PM EST
luma.com/wcww6xpl
Safe Learning Under Irreversible Dynamics via Asking for Help – Benjamin Plaut · Zoom · Luma
Safe Learning Under Irreversible Dynamics via Asking for Help
Benjamin Plaut – Postdoc at CHAI studying guaranteed-safe AI
Most online learning algorithms with…
luma.com
November 25, 2025 at 7:14 PM
Guaranteed Safe AI Seminars, December 2025:
Safe Learning Under Irreversible Dynamics via Asking for Help
Benjamin Plaut – Postdoc at CHAI studying guaranteed safe AI
Thursday, December 11, 1 PM EST
luma.com/wcww6xpl
Safe Learning Under Irreversible Dynamics via Asking for Help
Benjamin Plaut – Postdoc at CHAI studying guaranteed safe AI
Thursday, December 11, 1 PM EST
luma.com/wcww6xpl
notes on misuse risks from Opus 4.5 system card:
- below CBRN-4 threshold
- 82% first-attempt success on Cybench
- achieves network CTF challenges unassisted
- outperforms PhD baselines on bioinformatics workflows
- SOTA resistance to jailbreaks and prompt injection
- below CBRN-4 threshold
- 82% first-attempt success on Cybench
- achieves network CTF challenges unassisted
- outperforms PhD baselines on bioinformatics workflows
- SOTA resistance to jailbreaks and prompt injection
November 24, 2025 at 11:32 PM
notes on misuse risks from Opus 4.5 system card:
- below CBRN-4 threshold
- 82% first-attempt success on Cybench
- achieves network CTF challenges unassisted
- outperforms PhD baselines on bioinformatics workflows
- SOTA resistance to jailbreaks and prompt injection
- below CBRN-4 threshold
- 82% first-attempt success on Cybench
- achieves network CTF challenges unassisted
- outperforms PhD baselines on bioinformatics workflows
- SOTA resistance to jailbreaks and prompt injection
probability of survival × average value conditional on survival
November 24, 2025 at 4:01 AM
probability of survival × average value conditional on survival
Montréal event, Tuesday November 25, 7 PM:
In which Emma Kondrup asks whether AI is truly exceptional, using pessimistsarchive.org to compare today’s AGI/ASI fears with past panics over cars, radio and TV.
RSVP luma.com/nq50jf0u
In which Emma Kondrup asks whether AI is truly exceptional, using pessimistsarchive.org to compare today’s AGI/ASI fears with past panics over cars, radio and TV.
RSVP luma.com/nq50jf0u
Pessimists Archive · Luma
EN
An activity led by Emma Kondrup.
AI has already shown its important differences from previous technologies (economically, socially and politically). This…
luma.com
November 19, 2025 at 2:30 AM
Montréal event, Tuesday November 25, 7 PM:
In which Emma Kondrup asks whether AI is truly exceptional, using pessimistsarchive.org to compare today’s AGI/ASI fears with past panics over cars, radio and TV.
RSVP luma.com/nq50jf0u
In which Emma Kondrup asks whether AI is truly exceptional, using pessimistsarchive.org to compare today’s AGI/ASI fears with past panics over cars, radio and TV.
RSVP luma.com/nq50jf0u
Join us for the Defensive Acceleration Hackathon, to prototype defensive systems that could protect us from AI-enabled threats.
This Friday evening, Nov 21, to Sunday evening.
It is an online event. We have a jam site in Montréal. RSVP: luma.com/gnyqha4a
This Friday evening, Nov 21, to Sunday evening.
It is an online event. We have a jam site in Montréal. RSVP: luma.com/gnyqha4a
Defensive Acceleration Hackathon · Luma
Important registration information: To participate in this event, please sign up through Apart Research's event page before registering.
What is defensive…
luma.com
November 18, 2025 at 12:38 AM
Join us for the Defensive Acceleration Hackathon, to prototype defensive systems that could protect us from AI-enabled threats.
This Friday evening, Nov 21, to Sunday evening.
It is an online event. We have a jam site in Montréal. RSVP: luma.com/gnyqha4a
This Friday evening, Nov 21, to Sunday evening.
It is an online event. We have a jam site in Montréal. RSVP: luma.com/gnyqha4a
Montréal event, Thursday November 20, 7 PM:
A hands-on intro to Neuronpedia’s models, sparse autoencoders, and feature exploration using example prompts, ending with a discussion of evidence standards and how to start contributing.
RSVP luma.com/s3umszm7
A hands-on intro to Neuronpedia’s models, sparse autoencoders, and feature exploration using example prompts, ending with a discussion of evidence standards and how to start contributing.
RSVP luma.com/s3umszm7
Neuronpedia 101 · Luma
https://www.neuronpedia.org/
EN
A discussion with demo introducing Neuronpedia’s core concepts: models, sparse autoencoders, features, lists, and the anatomy…
luma.com
November 15, 2025 at 6:00 PM
Montréal event, Thursday November 20, 7 PM:
A hands-on intro to Neuronpedia’s models, sparse autoencoders, and feature exploration using example prompts, ending with a discussion of evidence standards and how to start contributing.
RSVP luma.com/s3umszm7
A hands-on intro to Neuronpedia’s models, sparse autoencoders, and feature exploration using example prompts, ending with a discussion of evidence standards and how to start contributing.
RSVP luma.com/s3umszm7
Montréal event, Tuesday November 18, 7 PM:
Co-design a National Citizens’ Assembly on Superintelligence
RSVP luma.com/0b7muzt0
Co-design a National Citizens’ Assembly on Superintelligence
RSVP luma.com/0b7muzt0
Co-design a National Citizens’ Assembly on Superintelligence · Luma
EN:
A brief workshop to co-design a National Citizens’ Assembly on Superintelligence for Canada. We’ll align on mandate of the project, who should be involved,…
luma.com
November 14, 2025 at 4:13 PM
Montréal event, Tuesday November 18, 7 PM:
Co-design a National Citizens’ Assembly on Superintelligence
RSVP luma.com/0b7muzt0
Co-design a National Citizens’ Assembly on Superintelligence
RSVP luma.com/0b7muzt0
Reposted by Orpheus Lummis
"When AI met Automated Reasoning"
by Clark Barrett, director of the Stanford Center for Automated Reasoning and co-director of the Stanford Center for AI Safety.
The event occurred today on the Guaranteed Safe AI Seminars.
The recording is now available: www.youtube.com/watch?v=AxAS...
by Clark Barrett, director of the Stanford Center for Automated Reasoning and co-director of the Stanford Center for AI Safety.
The event occurred today on the Guaranteed Safe AI Seminars.
The recording is now available: www.youtube.com/watch?v=AxAS...
When AI met AR – Clark Barrett
YouTube video by Horizon Omega
www.youtube.com
November 13, 2025 at 10:00 PM
"When AI met Automated Reasoning"
by Clark Barrett, director of the Stanford Center for Automated Reasoning and co-director of the Stanford Center for AI Safety.
The event occurred today on the Guaranteed Safe AI Seminars.
The recording is now available: www.youtube.com/watch?v=AxAS...
by Clark Barrett, director of the Stanford Center for Automated Reasoning and co-director of the Stanford Center for AI Safety.
The event occurred today on the Guaranteed Safe AI Seminars.
The recording is now available: www.youtube.com/watch?v=AxAS...
Reposted by Orpheus Lummis
Extremely excited to launch this report; the second report from World Internet Conference's International AI Governance Programme that I co-Chair with Yi Zeng. It goes further than any similar report I've seen in recommending robust governance interventions 1/4
www.wicinternet.org/pdf/Advancin...
www.wicinternet.org/pdf/Advancin...
www.wicinternet.org
November 11, 2025 at 2:04 PM
Extremely excited to launch this report; the second report from World Internet Conference's International AI Governance Programme that I co-Chair with Yi Zeng. It goes further than any similar report I've seen in recommending robust governance interventions 1/4
www.wicinternet.org/pdf/Advancin...
www.wicinternet.org/pdf/Advancin...
Reposted by Orpheus Lummis
I’m thrilled to share that I’ve been helping out my brother David who is starting a new org, Evitable.com, focused on informing and organizing the public around societal-scale risks and harms of AI, and countering industry narratives of AI inevitability and acceleration! 1/n
Evitable
Evitable.com
October 29, 2025 at 5:31 PM
I’m thrilled to share that I’ve been helping out my brother David who is starting a new org, Evitable.com, focused on informing and organizing the public around societal-scale risks and harms of AI, and countering industry narratives of AI inevitability and acceleration! 1/n
Reposted by Orpheus Lummis
They’re here! 🎉 After months of rigorous evaluations, our 2025 Charity Recommendations are out! Learn more about the organizations that can do the most good for animals with additional donations at https://bit.ly/2025-charity-recs 🙌🐥 Together, we’re helping people help more animals. 💙
November 4, 2025 at 6:53 PM
They’re here! 🎉 After months of rigorous evaluations, our 2025 Charity Recommendations are out! Learn more about the organizations that can do the most good for animals with additional donations at https://bit.ly/2025-charity-recs 🙌🐥 Together, we’re helping people help more animals. 💙
Montréal event on the International AI Safety Report, First Key Update: Capabilities and Risk Implications
Tuesday Nov 4, 7PM
RSVP: luma.com/09j4095g
Tuesday Nov 4, 7PM
RSVP: luma.com/09j4095g
October 30, 2025 at 9:12 PM
Montréal event on the International AI Safety Report, First Key Update: Capabilities and Risk Implications
Tuesday Nov 4, 7PM
RSVP: luma.com/09j4095g
Tuesday Nov 4, 7PM
RSVP: luma.com/09j4095g
Reposted by Orpheus Lummis
This workshop follows one we ran in July, adding optional specialized talks, and light moderation in the breakout sessions. To see how that one went, and videos of the talks, see this thread:
www.lesswrong.com/posts/csdn3e...
www.lesswrong.com/posts/csdn3e...
Summary of our Workshop on Post-AGI Outcomes — LessWrong
Last month we held a workshop on Post-AGI outcomes. This post is a list of all the talks, with short summaries, as well as my personal takeaways. …
www.lesswrong.com
October 28, 2025 at 10:06 PM
This workshop follows one we ran in July, adding optional specialized talks, and light moderation in the breakout sessions. To see how that one went, and videos of the talks, see this thread:
www.lesswrong.com/posts/csdn3e...
www.lesswrong.com/posts/csdn3e...