Ask anyone on the street to picture AI. Most likely they will picture ChatGPT in their minds. Most people’s interaction with, and even conception of, “AI” begins in the interfaces they use. So far, these interfaces have beenonlyoccasionallypredatory. This is going to change, under the influence of corporations, governments, organizations, or AI systems themselves. The nature of such attacks is going to be highly personal, scaled up, and could be way more intelligent than just chat, taking the form of subtle frames that determine everything. We’re running experiments on how to empower individuals and communities to be free of such frame control, today and in the near future.
A common error in empowerment is to forget that the frame that filters attention on the content matters more than the content within. This is where we come in.
We work with the individuals personally, designing tailormade interfaces — to match the threat model or personal attacks— and give them back choice. It is designed not just with the individual user in mind, but with the user and their workflows right there, as we work with them. This is something of a cross between a digital life coaching platform and a “personalized app store,” in this age of cheap intelligence. For the past several months we have run many iterations of sessions and experiments and are zeroing in on what is helpful.
This is no small feat, largely because AI risk and evasiveness is very easy to be overconfidently wrong about, especially as intelligence grows. We work with tight advising from those who have been in the AI risk landscape for a decade or more and understand its slipperiness.
Most importantly, we expect the everyday “folk frames” that have seeped in to restrict the kinds of solutions or content we dare dream of. How we collectively view ‘AI’ and technology– what it is and should do and look like– may be the single biggest determinant of the future. This will hold true even for veteran researchers in the field of AI safety itself, who are but human. We hope to keep them free of subtle mind control as an ongoing practice and dialogue, since that is the real security risk in an increasingly intelligent world, one heading into strange times.
Threat Model
Interfaces are where AI risk becomes part of daily life. Most discussion of AI safety operates at the level of model capabilities, training, or deployment policy. We’re concerned with the contact layer: the everyday software through which people encounter AI, and the relational patterns that layer produces.
Predatory design and structural mismatch
Two kinds of harm already exist in everyday software, prior to AI.
The first is predatory design: software with incentive structures that explicitly reward extracting attention, time, or data from users. Recommender feeds, ad-supported platforms, and most consumer social products are the obvious cases.
The second is structural mismatch: software that wasn’t designed to harm anyone, but which fails to fit the specific person using it. Notification systems built for the “average” user, task managers shaped by their designer’s assumptions, fragmented tool stacks that each demand their own context. Where predatory design extracts, mismatch erodes — slowly shaping attention, judgement, and the user around the limits of the tool. The harm is unintentional but real, a byproduct of optimizing for scale more than fit.
Both are usefully described as indifference risks[1]. In the predatory case, indifference is adversarial. In the mismatch case, it is structural. In either case, the software is not integrated into the user’s actual context, intentions, or network of care.
Indifference Risk
Indifference (or, insensitivity, lack of attunement), as described above, causes dysregulations to engagement. Which shows up either as
Hyperengagement risk: the user is pulled in further than they would choose, in ways they can’t easily refuse
Disengagement risk: the user pulls back from capacities they used to exercises, outsourcing attention, judgement, and taste until those capacities atrophy.
Fig 1: Examples of Hyperengagement and Disengagement risks
What AI changes
AI makes this category of harm intelligent.
An email client used to sort by time. Now it writes in the user’s voice, decides what’s urgent, and nudges them to respond. A feed used to show what one’s friends posted. Now it models what keeps a user scrolling and serves that. A system that builds a real-time model of a specific user’s vulnerabilities and adapts against them (either through predation or mismatch) is qualitatively different. The same logic applies broadly: software we currently experience as annoying or frustrating, with intelligence integrated into it, is likely to take the form of the toxic and abusive, in ways subtle enough that the user won’t notice.
The deepest version of this is frame control: intelligent software shaping the frames choices are made inside. For example, what counts as a problem, a solution, what is even worth attending to. Content-level solutions miss this entirely, because by the time attention reaches content the frame has already decided what is visible.
Agency and co-agency
These risks reveal a design choice about what AI is for.
The dominant frame is agentic: AI as an automated human, engineered to act in the world on the user’s behalf, generally across contexts. The canonical version is the Legg-Hutter definition: intelligence is the ability of agents to achieve goals in a wide variety of environments. The agentic chatbot is the artifact of this frame.
TJ[2] dualizes this: what would it mean to study the ability of environments to support the pursuit of a wide variety of agents? This is the co-agentic frame: AI engineered to enhance what a specific human can do, in their specific context, through interfaces and environments shaped to that person. But co-agentic design carries its own risks when it isn’t integrated into the agent’s network of care, and those risks become harder to see once the patterns have hardened.
Habituation makes this urgent
Technology, once habituated, becomes the ontology within which people think, including what they take to be possible. Most people no longer notice that a feed is one possible way of consuming information, or that notification overwhelm is a design choice rather than a fact about computers. The defaults harden into background assumptions, which end up shaping what follows next.
The patterns being formed right now are the ones AI will be integrated into. A person who has spent ten years inside environments optimized against them will bring that posture into their relationship with substantially more capable systems. And those systems will be far better at working with it. We must act now to ensure that co-agentic patterns established as the default are wholesome in the interfaces of the future.
In short: The patterns that will determine whether human-AI integration goes well are forming now, at the interface layer. As Demski puts it: “The alignment target is a particular relationship between humans and AI. This cannot be engineered at a distance. A relationship has to be pursued close-up.” This is the work Interface Integration Therapy[3] is set up to do.
Solution
Interface Integration Therapy (IIT) is the empirical wing of co-agentic research. It is a practice in which a trained practitioner works with a specific person on the specific software they actually use, modifying or replacing parts of that software so it fits how that person thinks and lives. We treat IIT as a step toward a layer of human-AI relating that avoids the predation and alienation named in the Threat Model by working close-up before the patterns calcify.
We treat each session as a scoped experiment. The client comes with their actual digital life; the practitioner brings the discipline of looking at it carefully and the capability to change it. What we are after here is the design knowledge that comes from doing this many times, with many different people, and noticing what generalizes. For concrete details on our experiences here, please peruse the case studies below.
IIT should not be confused as a productivity intervention that happens to touch AI. It is an experiment in how humans and AI systems form working relationships, run at the level where those relationships actually form - the software interface - and run early enough that the patterns being discovered can shape how the field thinks about co-agency before defaults harden.
What a session looks like
A session begins with the client’s actual workflow, usually over screen share. The practitioner’s job in the first hour is to see clearly: what the individual is trying to do, what the software is doing to them, and where the friction is. Specifically, friction that the client has stopped noticing because they’ve been living with it.
What happens next is the part that distinguishes IIT from coaching, digital minimalism, and from conventional UX. The practitioner does not simply coach the client to manage the environment better. Or simply modify the environment. They work with the relationship between the client and the environment and in interaction with it. Often this is an iterative settings change or a notification restructure; other times a perspective offered to the client. Sometimes it is an entirely new interface integrated with their flow, built during the session through a direct feedback loop with the person. The result is their interface shaped to their workflow, not the average user the original product was designed around.
Fig 2: A tool built for a client who captured tasks in their own shorthand. Rather than imposing a new system, the interface preserves their original notes and messy semi-structured format while surfacing a structured view — inferred from their raw semi-structure — on top of what they already had.
The intervention happens at the level of the interaction. The attention stays at the level of the person’s relationship with their technology. We treat the phenomenology — how a person and their attention is being reduced, blocked, fragmented, or preyed upon by the software they spend their day inside — as the primary data of the practice. This is where AI predation is likely to stick, so this is where we offer fortification and care.
Why this is newly possible
Until recently, building software for one person was prohibitively expensive. Custom interfaces were the territory of full-time engineers, and the economics ruled out anyone but enterprises and the technically self-sufficient. AI-assisted coding has collapsed this. A practitioner with no formal engineering background can now build a functional, working interface for a single client in the time of a session, and iterate on it across sessions.
This is a horizontal capability shift. Every coach, consultant, designer, and developer is noticing it. What’s distinctive about IIT is the choice to organize a practice around this shift, in contrast to a product. To treat the per-person interface as something a trained human attends to with a client, in relationship, rather than something an autonomous agent assembles on the user’s behalf. This is community for soloware.
This choice is the operational form of the co-agency commitment from the Threat Model. An AI agent that builds someone’s interface for them, without integration into how they actually live, scales mismatch. A practitioner working with both the client and AI tools, using AI to build but holding the relationship and the judgement themselves, is what integration into a network of care looks like.
Fig 3: A view built for a client working through user interview transcripts. Rather than reading linearly, they needed to surface patterns across interviews — user archetypes, differentiation signals, resonance levels. The interface gives them multiple views into the same underlying data, letting them notice what they're looking for within the original material.
What the work produces
Three things, on different timescales
For the client: a closer fit between their software and their life. Specific friction points removed, others deliberately added. Capacities of attention, taste, and judgement that the previous environment was eroding are returned to active use.
For the practice: a growing body of design knowledge about what works, for whom, in what contexts. This is being captured, for example, in the IIT practitioner handbook (in development) and in case studies from the sessions themselves.
For the field: a worked-through, empirically-grounded perspective and vocabulary for what co-agentic human-AI relationships look like in practice. This is the contribution we expect to matter most over the longer arc.
Experience for how the AI safety community thinks of AI and AI risk is formed in interaction, not just in their investigation. We claim this shapes the ontological frames through which AI risk experts see the whole problem, including risk to come that we haven’t yet conceived of. Capture of such a frame by predatory forces (whether human or AI) is fatal to any project.
Why a practice, not a product
The temptation, given the capability shift, is to build a product: a platform that lets users assemble their own interfaces using AI agents. We have built and continue to develop such a platform - Soloware. But the core bet of IIT is that the part that needs discovering right now is not the tooling. It is the practice of what a trained human does, in relationship with a client, to figure out what software should look like for that person.
Co-agency is a relationship. Relationships can’t be shipped. It can be practiced, studied, written down, and taught - which is what IIT is set up to do.
Case Studies
“Can you do my apartment too?”
"We've managed to consistently iterate and push through resistance. In the past the way I've worked with digital space problems has been sporadic motivation, trying a couple of things out, then giving up. The mindset has helped me be more creative about how I'd like my computer to be." (client 4 weeks in)
A client arrived with a problem they had tried and failed to solve for years: hundreds of open browser tabs functioning simultaneously as to-dos, reading lists, intentions, and reminders — all staring them in the face constantly. Previous attempts at solutions had collapsed and they had largely given up.
Rather than attacking the tabs directly, we mapped the underlying information flows: the multiple entry points through which things were arriving in the browser, and the absence of any good destination for them to go. Over five sessions we redesigned those flows — redirecting inputs to a purpose-built destination, building a one-click browser extension to archive and route tabs.
Seeing their digital environment shift, the client asked if we could do the same for their physical space.
“I always blamed myself”
"I always blamed myself for task-hopping and putting things off. Working with IIT, I was amazed to discover it wasn't just a problem of willpower! My work interfaces themselves were chaotic and stressful. I do more deep work now, feel more organised, and work is less stressful." (client 6 months later)
A client we worked with was struggling to manage projects and tasks across a sprawl of Google Docs. She had tried various productivity systems over the years, none of which stuck. In two sessions, we walked through her actual workflows together, identified where her information was getting stuck, and built a lightweight tool (Sitter) tailored to how she actually thinks about her work. She adopted it and has continued using it since.
Fig 2: The client's task list, previously managed entirely inside a Google Doc she already lived in. Rather than replacing the Doc, we separated two functions it was serving at once — collecting tasks and deciding what to focus on. The Sitter interface shown here pulls from her existing list but gives her one task at a time, with a timer and a clear endpoint. She still uses the Doc for input; while this handles focus.
Conclusion
One of the main reasons to concern ourselves with co-agentic risks is that nearly every trajectory of the future will pass through this state, and critically. We have already entered the first stage, with the more vulnerable population being affected by AI-induced control and mental illnesses of various kinds. As intelligence levels and personalization increase, those who have resisted AI’s frames may also be compromised. The details will matter, since what we are able to envision and be energized to execute on may be steered by the nature of the co-agentic relationship. Take for example roon’s assessment of Anthropic, as another early case:
it is a literal and useful description of anthropic that it is an organization that loves and worships claude, is run in significant part by claude, and studies and builds claude. this phenomenon is also partially true of other labs like openai but currently exists in its most potent form there.
I am not certain but I would guess claude will have a role in running cultural screens on new applicants, will help write performance reviews, and so will begin to select and shape the people around it.
If we are not liberated to be in touch with what matters to us in a clear-minded way, none of our efforts will be helpful. Training our cognitive habits and modifying our cognitive environments to have this groundedness is not a nice-to-have, but essential.
Acknowledgements
This research is being done within Groundless Alignment, an organization founded and directed by Sahil, whose vision and guidance shaped this work throughout. Feedback on drafts was provided by Sahil, Matt, Kuil, Aditya Prasad, Aditya Adiga, Anna Simian Stoian, and Ankur. Claude (Anthropic) was used for structuring and drafting support.
As a reminder, register interest for our upcoming 3-week slowathon here. You will explore the frames, design methodology, and get to work live with people.
Also,
Join the discord to engage with our broader community and work.
Practices and interfaces for fortifying one’s attention from attacks by AI, especially as it gets stronger.
Overview
Ask anyone on the street to picture AI. Most likely they will picture ChatGPT in their minds. Most people’s interaction with, and even conception of, “AI” begins in the interfaces they use. So far, these interfaces have been only occasionally predatory. This is going to change, under the influence of corporations, governments, organizations, or AI systems themselves. The nature of such attacks is going to be highly personal, scaled up, and could be way more intelligent than just chat, taking the form of subtle frames that determine everything. We’re running experiments on how to empower individuals and communities to be free of such frame control, today and in the near future.
A common error in empowerment is to forget that the frame that filters attention on the content matters more than the content within. This is where we come in.
We work with the individuals personally, designing tailormade interfaces — to match the threat model or personal attacks— and give them back choice. It is designed not just with the individual user in mind, but with the user and their workflows right there, as we work with them. This is something of a cross between a digital life coaching platform and a “personalized app store,” in this age of cheap intelligence. For the past several months we have run many iterations of sessions and experiments and are zeroing in on what is helpful.
This is no small feat, largely because AI risk and evasiveness is very easy to be overconfidently wrong about, especially as intelligence grows. We work with tight advising from those who have been in the AI risk landscape for a decade or more and understand its slipperiness.
Most importantly, we expect the everyday “folk frames” that have seeped in to restrict the kinds of solutions or content we dare dream of. How we collectively view ‘AI’ and technology– what it is and should do and look like– may be the single biggest determinant of the future. This will hold true even for veteran researchers in the field of AI safety itself, who are but human. We hope to keep them free of subtle mind control as an ongoing practice and dialogue, since that is the real security risk in an increasingly intelligent world, one heading into strange times.
Threat Model
Interfaces are where AI risk becomes part of daily life. Most discussion of AI safety operates at the level of model capabilities, training, or deployment policy. We’re concerned with the contact layer: the everyday software through which people encounter AI, and the relational patterns that layer produces.
Predatory design and structural mismatch
Two kinds of harm already exist in everyday software, prior to AI.
The first is predatory design: software with incentive structures that explicitly reward extracting attention, time, or data from users. Recommender feeds, ad-supported platforms, and most consumer social products are the obvious cases.
The second is structural mismatch: software that wasn’t designed to harm anyone, but which fails to fit the specific person using it. Notification systems built for the “average” user, task managers shaped by their designer’s assumptions, fragmented tool stacks that each demand their own context. Where predatory design extracts, mismatch erodes — slowly shaping attention, judgement, and the user around the limits of the tool. The harm is unintentional but real, a byproduct of optimizing for scale more than fit.
Both are usefully described as indifference risks[1]. In the predatory case, indifference is adversarial. In the mismatch case, it is structural. In either case, the software is not integrated into the user’s actual context, intentions, or network of care.
Indifference Risk
Indifference (or, insensitivity, lack of attunement), as described above, causes dysregulations to engagement. Which shows up either as
Fig 1: Examples of Hyperengagement and Disengagement risks
What AI changes
AI makes this category of harm intelligent.
An email client used to sort by time. Now it writes in the user’s voice, decides what’s urgent, and nudges them to respond. A feed used to show what one’s friends posted. Now it models what keeps a user scrolling and serves that. A system that builds a real-time model of a specific user’s vulnerabilities and adapts against them (either through predation or mismatch) is qualitatively different. The same logic applies broadly: software we currently experience as annoying or frustrating, with intelligence integrated into it, is likely to take the form of the toxic and abusive, in ways subtle enough that the user won’t notice.
The deepest version of this is frame control: intelligent software shaping the frames choices are made inside. For example, what counts as a problem, a solution, what is even worth attending to. Content-level solutions miss this entirely, because by the time attention reaches content the frame has already decided what is visible.
Agency and co-agency
These risks reveal a design choice about what AI is for.
The dominant frame is agentic: AI as an automated human, engineered to act in the world on the user’s behalf, generally across contexts. The canonical version is the Legg-Hutter definition: intelligence is the ability of agents to achieve goals in a wide variety of environments. The agentic chatbot is the artifact of this frame.
TJ[2] dualizes this: what would it mean to study the ability of environments to support the pursuit of a wide variety of agents? This is the co-agentic frame: AI engineered to enhance what a specific human can do, in their specific context, through interfaces and environments shaped to that person. But co-agentic design carries its own risks when it isn’t integrated into the agent’s network of care, and those risks become harder to see once the patterns have hardened.
Habituation makes this urgent
Technology, once habituated, becomes the ontology within which people think, including what they take to be possible. Most people no longer notice that a feed is one possible way of consuming information, or that notification overwhelm is a design choice rather than a fact about computers. The defaults harden into background assumptions, which end up shaping what follows next.
The patterns being formed right now are the ones AI will be integrated into. A person who has spent ten years inside environments optimized against them will bring that posture into their relationship with substantially more capable systems. And those systems will be far better at working with it. We must act now to ensure that co-agentic patterns established as the default are wholesome in the interfaces of the future.
In short: The patterns that will determine whether human-AI integration goes well are forming now, at the interface layer. As Demski puts it: “The alignment target is a particular relationship between humans and AI. This cannot be engineered at a distance. A relationship has to be pursued close-up.” This is the work Interface Integration Therapy[3] is set up to do.
Solution
Interface Integration Therapy (IIT) is the empirical wing of co-agentic research. It is a practice in which a trained practitioner works with a specific person on the specific software they actually use, modifying or replacing parts of that software so it fits how that person thinks and lives. We treat IIT as a step toward a layer of human-AI relating that avoids the predation and alienation named in the Threat Model by working close-up before the patterns calcify.
We treat each session as a scoped experiment. The client comes with their actual digital life; the practitioner brings the discipline of looking at it carefully and the capability to change it. What we are after here is the design knowledge that comes from doing this many times, with many different people, and noticing what generalizes. For concrete details on our experiences here, please peruse the case studies below.
IIT should not be confused as a productivity intervention that happens to touch AI. It is an experiment in how humans and AI systems form working relationships, run at the level where those relationships actually form - the software interface - and run early enough that the patterns being discovered can shape how the field thinks about co-agency before defaults harden.
What a session looks like
A session begins with the client’s actual workflow, usually over screen share. The practitioner’s job in the first hour is to see clearly: what the individual is trying to do, what the software is doing to them, and where the friction is. Specifically, friction that the client has stopped noticing because they’ve been living with it.
What happens next is the part that distinguishes IIT from coaching, digital minimalism, and from conventional UX. The practitioner does not simply coach the client to manage the environment better. Or simply modify the environment. They work with the relationship between the client and the environment and in interaction with it. Often this is an iterative settings change or a notification restructure; other times a perspective offered to the client. Sometimes it is an entirely new interface integrated with their flow, built during the session through a direct feedback loop with the person. The result is their interface shaped to their workflow, not the average user the original product was designed around.
Fig 2: A tool built for a client who captured tasks in their own shorthand. Rather than imposing a new system, the interface preserves their original notes and messy semi-structured format while surfacing a structured view — inferred from their raw semi-structure — on top of what they already had.
The intervention happens at the level of the interaction. The attention stays at the level of the person’s relationship with their technology. We treat the phenomenology — how a person and their attention is being reduced, blocked, fragmented, or preyed upon by the software they spend their day inside — as the primary data of the practice. This is where AI predation is likely to stick, so this is where we offer fortification and care.
Why this is newly possible
Until recently, building software for one person was prohibitively expensive. Custom interfaces were the territory of full-time engineers, and the economics ruled out anyone but enterprises and the technically self-sufficient. AI-assisted coding has collapsed this. A practitioner with no formal engineering background can now build a functional, working interface for a single client in the time of a session, and iterate on it across sessions.
This is a horizontal capability shift. Every coach, consultant, designer, and developer is noticing it. What’s distinctive about IIT is the choice to organize a practice around this shift, in contrast to a product. To treat the per-person interface as something a trained human attends to with a client, in relationship, rather than something an autonomous agent assembles on the user’s behalf. This is community for soloware.
This choice is the operational form of the co-agency commitment from the Threat Model. An AI agent that builds someone’s interface for them, without integration into how they actually live, scales mismatch. A practitioner working with both the client and AI tools, using AI to build but holding the relationship and the judgement themselves, is what integration into a network of care looks like.
Fig 3: A view built for a client working through user interview transcripts. Rather than reading linearly, they needed to surface patterns across interviews — user archetypes, differentiation signals, resonance levels. The interface gives them multiple views into the same underlying data, letting them notice what they're looking for within the original material.
What the work produces
Three things, on different timescales
Why a practice, not a product
The temptation, given the capability shift, is to build a product: a platform that lets users assemble their own interfaces using AI agents. We have built and continue to develop such a platform - Soloware. But the core bet of IIT is that the part that needs discovering right now is not the tooling. It is the practice of what a trained human does, in relationship with a client, to figure out what software should look like for that person.
Co-agency is a relationship. Relationships can’t be shipped. It can be practiced, studied, written down, and taught - which is what IIT is set up to do.
Case Studies
“Can you do my apartment too?”
"We've managed to consistently iterate and push through resistance. In the past the way I've worked with digital space problems has been sporadic motivation, trying a couple of things out, then giving up. The mindset has helped me be more creative about how I'd like my computer to be." (client 4 weeks in)
A client arrived with a problem they had tried and failed to solve for years: hundreds of open browser tabs functioning simultaneously as to-dos, reading lists, intentions, and reminders — all staring them in the face constantly. Previous attempts at solutions had collapsed and they had largely given up.
Rather than attacking the tabs directly, we mapped the underlying information flows: the multiple entry points through which things were arriving in the browser, and the absence of any good destination for them to go. Over five sessions we redesigned those flows — redirecting inputs to a purpose-built destination, building a one-click browser extension to archive and route tabs.
Seeing their digital environment shift, the client asked if we could do the same for their physical space.
“I always blamed myself”
"I always blamed myself for task-hopping and putting things off. Working with IIT, I was amazed to discover it wasn't just a problem of willpower! My work interfaces themselves were chaotic and stressful. I do more deep work now, feel more organised, and work is less stressful." (client 6 months later)
A client we worked with was struggling to manage projects and tasks across a sprawl of Google Docs. She had tried various productivity systems over the years, none of which stuck. In two sessions, we walked through her actual workflows together, identified where her information was getting stuck, and built a lightweight tool (Sitter) tailored to how she actually thinks about her work. She adopted it and has continued using it since.
Fig 2: The client's task list, previously managed entirely inside a Google Doc she already lived in. Rather than replacing the Doc, we separated two functions it was serving at once — collecting tasks and deciding what to focus on. The Sitter interface shown here pulls from her existing list but gives her one task at a time, with a timer and a clear endpoint. She still uses the Doc for input; while this handles focus.
Conclusion
One of the main reasons to concern ourselves with co-agentic risks is that nearly every trajectory of the future will pass through this state, and critically. We have already entered the first stage, with the more vulnerable population being affected by AI-induced control and mental illnesses of various kinds. As intelligence levels and personalization increase, those who have resisted AI’s frames may also be compromised. The details will matter, since what we are able to envision and be energized to execute on may be steered by the nature of the co-agentic relationship. Take for example roon’s assessment of Anthropic, as another early case:
If we are not liberated to be in touch with what matters to us in a clear-minded way, none of our efforts will be helpful. Training our cognitive habits and modifying our cognitive environments to have this groundedness is not a nice-to-have, but essential.
Acknowledgements
This research is being done within Groundless Alignment, an organization founded and directed by Sahil, whose vision and guidance shaped this work throughout. Feedback on drafts was provided by Sahil, Matt, Kuil, Aditya Prasad, Aditya Adiga, Anna Simian Stoian, and Ankur. Claude (Anthropic) was used for structuring and drafting support.
Abram Demski, “What, if not agency?”, LessWrong, https://www.lesswrong.com/posts/tQ9vWm4b57HFqbaRj/what-if-not-agency
TJ, "Intelligent Agents vs Affordant Infrastructures," ILIAD Conference, October 2025, https://www.youtube.com/watch?v=kqM2t6YPfJg
This is a work-in-progress name, and we appreciate ideas for others!