Good afternoon. Good morning.
Whichever you want to use over the last 24 hours no fewer than five Satellites have re-entered the Earth's atmosphere They're not the first and they certainly won't be the last since the beginning of The space race nations have rushed to launch objects into low Earth orbit Most of these have some benefits right earth fencing monitoring hurricanes telecommunications Entertainment you name it But over the years as you can see over the decades things have gotten a little bit Congested especially in the last decade now if you take the sum total of all of these satellites currently in orbit the ones that are Serving humanity with good intent in some way you about 16,000 objects But if you flip it to left over discarded spacecraft boosters straight up debris and junk You have a different equation now you have about half of that almost is Debris and it's dangerous debris because it's traveling at 17,000 miles an hour or 28 kilometers per hour That's fast enough to put a chip on the windshield of the International Space Station just like a large truck does on the highway and It's not just a danger to the ISS which has had to move multiple times It's actually a danger to people on the ground there have been over 33,000 re-entries since the early 60s.
That's roughly About Four or five a day I've forgotten the math at this point. It's accelerating quickly and soon we'll have about five satellites every day Re-entering and so what we have is kind of a view currently of what's in orbit. Yes.
This is a live shot That's all real stuff and you might think well if we're thinking about intent There's clearly things with good intent and there's clearly things with bad intent We should just get rid of the bad things and use the good things The problem is we haven't really even considered intent and it's more complicated than just spacecraft versus debris What do I mean? Some of it is obvious, right? The debris is left over the intent whatever intent it had that might have been declared is no longer Applicable other spacecraft presumably are good.
Look at the recent Artemis mission from NASA that was benefiting humanity and exploring space in some way But then there's this whole third gray area, right? Starlink is an easy example. Is it good or bad?
Really depends on who you are where you live and who you ask if it's providing telecommunications and access to the internet for Places that don't have it or wouldn't otherwise be able to have that then they get better financial outcomes better health outcomes ultimately but if it's helping some nation-state keep a Subjugated population under control then maybe it's not so great this throwing up of satellites at a rapid pace with mixed intent and kind of declaring intent is common to what we're seeing now in agentic AI because everyone's using it right in your organizations To do various things and the intent factor is really at the heart of it It's the why right?
Why is this agent being used? Is it beneficial or is it not? I'm actually not here to give answers today sadly because this is a very rapidly developing discussion What I'm here instead is I'm trying to give us ways to think about intent and by intent I mean declared intent for the use of the agent via prompting how to think about intent And evaluate the discussion as it goes forward.
We'll talk about three areas today Defining intent how intent shifts and then finally conveying intent first off defining intent When you launch a satellite You do have to declare or at least in theory Declare your intent in some way now traditionally you do this in a number of places in the United Nations registry You say this satellite is being used for this purpose You get a little bit further more specific as you go down on a national level But really at the top level that UN declaration is kind of the why What's interesting is it's declared multiple times, right?
Going down more to the how with spectrums and plans and Actual declaring of what it's supposed to be doing but that why has some issues, right? There's no agreed-upon taxonomy. A lot of them are just labeled general or communications Another thing they're not really revealing the true intent None of the satellites in orbit straight-up say for war or for attacking other satellites, right? It's all communications or general research and then last off.
There's no real enforcement Even if you if you misuse a satellite based on intent in that UN database No one is going to call you out on it There's no agreed way to say this was not the intent you stated and so we're going to take it out What kind of in the same boat with agents right because when you prompt an agent, what are you doing? You are declaring your intent. You're saying something about what your purpose is That's better than what we had before To be honest right before if you accessed something you asked for access You never really had to say why I want to access a spreadsheet.
I want to access this database Don't ask me why just decide whether or not I should have access and let me go But there's an interesting report published in academic paper in 1987 talking about action Theory and it's a fascinating idea around how you describe what you are doing and your intent or in your purpose behind it The idea is that you know a normal conversation.
You're not gonna say hey I went to your house and I lifted my finger and I push this button No you're gonna level up a bit and talk about You're pushing a doorbell or ultimately your purpose up at the top of the the scope is the why I wanted to see if you Were home right and how you describe what you're doing is based off of a couple of things you want to go as high up and the why as you can as your The respondent can understand But not too high that it doesn't seem impossible In other words, it's saying this this is it's easy to do or it's achievable And I will choose that level to describe my actions And if it's too high if you your first initial idea of how to describe this is too high You'll drop down until you hit something attainable a few examples to show what I mean by this now if you have a cat right, no one says I'm gonna go reach down and Move my hand and touch this feline similarly.
They're not gonna make the top claim either right? I'm going to achieve Unworldly happiness by Associating with this animal it's going to be somewhere in the middle you you find your level when you're describing Intent and this is important. You think about remember when we're thinking about prompts and agents, right you achieve some level of What you describe you want the agent to do?
You're probably not going to go way way really high because that doesn't seem achievable by the agent You're much more likely to do something Down the scope more towards the how and the reason I'm beating this to death a little bit is because that why is really important Right if we're making decisions on purpose and intent We want that the why to be as close to as possible what we're talking about in General right because the how is kind of the access the how is what access should be allowed The why is what authority am I seating to this agent? Right.
It's it's more of a high up in the spectrum So thinking about this framework really helps and how we think about intent and it also points out Similar problems with prompt based intent right in general There's no shared agreed-upon vocabulary about what the semantic meaning is. There are various theories about how to achieve that Users have discretion. That's a really nice way of saying Users why right if I'm stealing internal data, I'm not gonna say go get the data so I can hand it to her No, I'm not gonna say that right.
I'm gonna make up some other business reason why and then lastly It's really hard to enforce right just because I declare an intent for an agent down the line How do I make sure that intent matches? Those are existing challenges that we need to think through So intent needs definition and we're working on that now, let's talk about the next topic about how intent shifts Back to the space program. The IS double-e was a spacecraft That was originally used For solar observation, right and it sat at a place called l1.
Don't worry about the stable orbit Stable a grange point outside of Earth observed the Sun for a while It did a couple lunar swing bys and then Observed multiple comments and eventually went into a large-scale Heliocentric orbit the point here is that that same satellite had four different missions for different intents but the UN registry Just said scientific research about solar wind It didn't stay true to that intent.
It got Repurposed repeatedly, but it's not always an intentional repurposing either that satellites change intent There was a Russian satellite and iridium satellite back in the early 90s That collided by accident and instead of having two satellites You had to think about and tent around those satellites and went from two points that you needed to govern To over 3,000, right and so intent can be shifting can be intentional or it can be accidental agents work the same way I Guarantee you because I did this this past week People aren't spinning up new agents to do new tasks per se They're letting context and content build up on itself shifting from doing what task a to task B to test C All the same agent Intent is shifting.
And so that's why Matt's talk was so great yesterday How you define intent and what how sessions and how instances of agents? These are things we need to think through if we're going to be honest about how we think about intent the other way intent shifts is not as Intentional and going back to that aid the activity action identification theory is pretty fascinating Because like I said, what happens is depending on how familiar the task is how complex it is how achievable it is You start moving up and down the spectrum.
This is how we think as humans if it's too hard We move down a level if it's easier we move up to the more higher level why? We're doing it and how we describe about it and how we think about what we're doing Think about running a marathon If anybody's run a marathon anybody run a marathon before Five of us.
All right, you know what I'm talking about. Any race really you start out you think I'm gonna crush this I'm on time. You've got your watch and then reality hits and your strategy has to shift and then Sometimes you're you shift down and what your goal is Sometimes you're just thinking about survival you start to feel better and you shift back up into a higher order task It's how you talk yourself through the race right and the same thing happens with a gentic, right?
what happens you prompt an agent to do something and it gets it right and you get more ambitious or You prompt an agent to do something and it totally mistakes what's going on and you have to dumb it down It's a lot like talking to my three-year-old when he was three. He's now 17.
He's a lot better at this But the kind of thing where you give a small child a task and you say no no, no and you and you make it Smaller and easier to understand for them The point is that intent is shifting same agent same human, but the intent is shifting over time That shifting from why to how and back to why has an impact when we think about agents having subagents because each of them is going to have to react to a shift in a shift in Intent differently if we're talking about the previous example talked about hiring someone in Berlin If you back off of some of those requirements, then the subagents that are doing Sourcing or putting out publicity or searching for candidates depending on how high they are in the how to why?
Spectrum they're gonna have to react differently either they have to adjust what they're doing start over completely Their implications for how intent is shifting for subagents downstream We've seen how we need to define intent thing about definition when you think about how intent shifts finally We need to think about conveying intent. What do I mean by that?
back to the satellite I talked before the ISWE like I said it had four different missions But what this meant it had four different timings of how long it took communications to go back to Earth and be able to respond as You get further away from Earth. It gets longer originally It's five seconds one and a half seconds, but eventually you're on the order of 17 minutes And that those kinds of context you have to have the intent or the mission or whatever you want to call it Onboard the spacecraft so it can continue to fulfill its mission.
It can't call back home Every 30 seconds to have decisions made for it So intent has to be conveyed along with the spacecraft and the same is obviously true with agents, right? Intent has to be conveyed if we're thinking about intent building standards and approaches around intent Then intent being conveyed helps us do a couple of things one it's it's letting distant agents be able to make decisions on its own or the assistance of author of Authorization engines as long as it can communicate its intent saying here I am. I'm here localized PDP or off sin Here's my intent.
Can you make a yes-or-no choice for me rather than having to go back and ask for that? The other is agent agent context agents working together in swarms sharing intent saying look we're all serving the same master What what's your intent? What's my intent? How do we work together? How do those cross-pollinate if I succeed or fail? How does that affect your mission and your intent?
and then finally Sub-agent grounding basically if you are a sub-agent of another agent knowing what that overall intent is is Super helpful because your intent has that implied context So we have to define intent that we're working on We have to think about how intent shifts and account for that and ideally we need to be able to convey intent have it journey with the agent be localized enough that it can share with other agents share with Sub-agents and hand off to decision makers to make author real authorization decisions about what it can and can't do That's a lot to do that's very reduced to what the actual problem space is But this gives us a framework to think through it and think about good and bad ideas in terms of space debris and lower-earth orbit There are some pretty bad ideas For example ground bath ground-based space lasers not joking used to just zap satellites out of orbit Side note Starlink to their credit They actually engineer their satellites to be able to degrade and burn up in the Earth's atmosphere that may cause other issues but at least they're planning for for re-entry and Destruction of their their spacecraft, but we want to think through whatever those ideas are whether it's space exploration and low-earth orbit or agentic There are no fewer than five different current approaches that are being discussed in the community ranging from a off from Dick Hart To people who are here as well George is going to talk about Some of this later on go find him to think about Obligations and constraints and how he thinks about it Eve as well with her desired outcome contracts And then finally you have Claudia Hepburn with Sarah Kacheti Who's using a Claude bought to help her write an approach to intent based?
It says authorization, I don't know if that's quite the thing But basically it captures intent and helps make decisions about agentic actions based on intent These are all fairly different approaches go and find these people I'm not going to tell you which one I think is the best because it changes from from day to day The discussion is pretty live You can go and take the slide and go and read them through on your own thinking about the things we've talked about In terms of last thing I want to point out is that each of them if you apply the defining shifting and conveying intent?
approach to these They're good in different ways. This does not mean that one is inherently better than the other It just helps you to think of the advantages and the disadvantages you have which each with each of these approaches all of these people are Very warm and welcoming.
So if you had if you have ideas if you have questions if you need clarification hunt them down Ask some questions and engage in the discussion because there's a lot to do Just like in space exploration and space debris, there's a there's a ton to do It's it's evolving rapidly and we need all to think about this together If we're going to decide things and help decide things based on intent and make sure only what's there is supposed to be there And benefiting us. Thank you
