Monday, 2025-11-05 08:00..09:00 EST (Montreal)
Agenda bashing
Carsten talking about Jaime Jiménez’ presentation at IETF123 on Agentic
AI and IoT.
Obvious problem: if you say “I’m cold”, and there are two thermostats,
which one to adjust? (one is room temp, other is refrigerator). Purpose
of life needs to be described and be discoverable.
Christian on the chat:
If you want to take a light reading approach to the gap Carsten points
out, the concept of https://en.wiktionary.org/wiki/thalience is a
subject of the SF novel “Ventus” that was recently recommended to me in
that area.
Some relationship between the work Lorenzo presented on behalf of Edgar at ASDF group. Good to explore.
(Laurent commented on relevant work) transformation of Data Models using LLM
CB: Connected to discoverably work. SDF to YANG translation played any role?
AP: yes, for finetuning models […] Models often make little mistakes like forgetting brackets at the end, need an additional step to fix the code produced.
CB: audacious proposal: milestone or challenge we all could try to achieve with our techs and LLMs and agents. Have room with IoT devices and get LLM to do something useful with command from human. Challenges really useful for getting folks focused on an objective. Could start to sit down and come up with a scenario and challenges.
CK: challenge – if we get something working with LLMs, is it worth it? Need baseline to compare. Trying to understand if really need this.
CB: do we really need it, really good question. What challenge can do, can have people compete with different solutions. If specific challenge can be solved with different tech, that would be a success already. But need metrics.
JA: Would like to claim what you call purpose in life not sufficient,
sense of the real world needed. Need more complex examples,
understanding real world concepts like what kind of room you are in,
goes beyond brackets, don’t want to burn your house down. Authorization
aspects play a role here too.
Wonder if recent work on semantics has built anything here. Crux for
making usable agents.
CB: Privacy and authorization are aspects we need to work on. Important but maybe not the first step we need to take.
AP: What about actuators that trigger a process, e.g. for increasing the temperature at the end. If you have a black box for such a process, can you control it via AI?
CK: Do you mean putting awareness in the thermostat?
AP: Process may involve burning something, burning to much would cause the house to explode. In between there is the temperature increase. Make devices aware of the effects and side effects (e.g., may involve CO2 emissions which you want to avoid)
CB: now Karolina should perhaps say something. Whole issue of fault mitigation. She is addressing in her work. Adding AI is not maybe next on agenda, but is a consideration that you have in all these fault management scenarios.
RP: Bracket situation is fixable via a language model. Need to have a world model in general. Have a lot of devices in the industry (e.g., Siemens PLCs) that are not able to describe their capabilities to an AI agent. Need to describe the things that are describable.
AP: Siemens PLCs may be hiding these on purpose. Is it really one of the objectives of PLC manufacturers to expose this information?
RP: PLCs are legacy hardware, are being replaced by microcontrollers. Mentioned it as an example, as there are many startups that are trying to work on bringing AI to this space.
CK: Does it solve a problem that is new or in a better way?
RP: Yes, it would make everything better.
NW: It’s also a matter of predictability.
DR: Agree, it seems ridiculous to have a non-deterministic behavior in this domain
AP: Siemens released Co-pilot on top of their [PLC] stack. They are
active at W3C WoT.
NW: Is that actually a problem? The real problem is building the
“orchestrating glue”. That’s where maybe AI can help.
ED: From the home IoT perspective, there’s some experience from Thread. Diagnostic capabilities have been added and that contributes to network traffic. AI can be used to analyze those data, translate it into problems and remedies. For now, no AI involved in controlling and parsing semantic data, but it might come. That might be preferable to leaving that task to people.
CB: logging world moving away from syslog that has been used for long time. Some of that might not be needed since can use AI.
ED: syslog standard for Thread still. Lots of information there. Thousands of warnings and errors per day.
CB; who would be interested to give a talk in the interim we are planning on this area? Good set of questions already useful. End of Nov, start of Dec time frame.
AP: have PhD student who started on the topic. But needs to submit a paper first probably. Could be for this interim or the next one. But interested.
CK: came across side meeting on AI networks. Proposing network architecture to support AI, training, agent interactions, etc. Might be related. Network support AI; but future idea of instead of saving bits, — maybe IETF not needed anymore :-)
CB: different requirements for networking in different areas. Networking for AI, running datacenter algorithms etc. probably quite far from what we interested in here. Half a dozen side meetings might be interesting for us, like agent2agent group. Trying to work on agent protocol. Something we would need at least observe or even pull in our requirements there. Could look at what other side meetings we from T2TRG PoV would be interested. Maybe others than a2a you are aware of?
Other topics on agenda perhaps already have a home. Composable code in the last interim. Model to API translation presented in ASDF. Instance information discussed in this IETF at ASDF WG. Maybe don’t need to pick up the topics right now here?
AP: how to use Thing Description and ASDF best practices (for AI?)
Curated list of models you can download? If want to do this use that
model.
KS (in chat): Noted!
CB: survey document at this time instead of best practice now. But very useful, especially for fresh PhD students. Could think of deliverables. Would like to see something on AI and safety. Area where progress needs to be made. If have other ideas on deliverables?
AP: model translation. Challenge with generating artificial description / data model and then evaluate the AI/LLM on the accuracy of translation of this unseen model.
LT: AI can learn a lot from metainformation in your system and then derive information about your house.
AP: Reminds me of capture-the-flag challenges, identify which device is the thermostat in the fridge.
CB: will need to keep looking for people and topics for the AI Agent interim to have critical mass. If you have ideas, please contact T2TRG co-chairs.
What is the role that IETF IoT technologies could play enabling AI agent interaction, e.g., usefulness of the Discovery and self-description mechanisms we have and the gaps — what is needed/how to evolve them.
We also want to use the hallways IETF124 to explore potential interest of other IoT groups to look into the topic jointly.
(recent interim)
https://datatracker.ietf.org/meeting/interim-2025-asdf-10/materials/slides-interim-2025-asdf-10-sessa-api-transpation-00