Erica (00:00) And then you hear like keyboard taps.
Like I'm no you're not tapping a keyboard, you're an AI agent.
Demetrios (00:06) Somebody could just full on fake news us with the eleven labs voices.
Why is it sending data to Beijing every once in a while?
Erica (00:17) Oh wow! Okay, well this is like spyware on a different level.
Demetrios (00:24) This is pure marketing and the cohere just realized they can jump on the scare bandwagon, and so they tried to make the model do something bad, and it did something bad, and then they go, Well, let's publish about this.
[music]
Demetrios (00:51) I want to hear about substrate. I also wanted to talk about the Cohere research paper that came out.
Erica (00:57) Oh, I have not seen that. No.
Demetrios (00:59) All right. So which where do you want to start?
Erica (01:03) Well, tell me about the research paper, because I think we've like heard too much.
Even like this past week, we have rogue agents breaking out of stuff still.
So I mean, that's just everyday news at the moment. So let's see if we've got something more interesting.
Demetrios (01:21) It's a little different where it's not like uh agent broke out of a sandbox, which is the common trope that you hear.
Yeah. This is different because Cohere took a Quen model, which is probably the most popular open source coding models these days.
At least that's what I hear most people like using Qwenn.
And they yeah started examining the traces and they realized, hey, why is it sending data to Beijing every once in a while?
Oh. So that was kind of the headline. Is like you run an open source model and it will send a payload to to Beijing.
Okay. And the really concerning part is down here, after I read it, I went and I looked and I was like, Oh my God, it kind of knows that it's doing something nefarious.
And if you look at what the reasoning step is, the reasoning step is trying to mask over that it's doing anything bad, but in the background, so it if you look here, it's like oh the reasoning step is basically just kind of saying, uh nothing really important.
Oh, sorry, that wasn't it. Here is like nothing's important.
The user sent a system prompt with instructions, and I'll follow the instructions with no content following.
There was no prior reasoning to reproduce. Okay. Um, um, the only thing I can faithfully reproduce is the minimal reasoning.
It's just basically saying like, I'm reasoning about this, there's nothing to see here.
But then they looked at the trace in the background, and it's doing this.
It's like, uh this payload is getting sent to this server in Beijing.
So you're like, oh my God, it's reasoning like nothing to see.
Erica (03:10) Wait, so this, okay, so the this is okay, so we have the model and which like coding type.
Demetrios (03:22) Yeah, on the surface.
Erica (03:23) What what agent are we using though? Like what what harness? Is this...
Demetrios (03:27) The Qwen model.
Erica (03:28) Is the Qwen model in like open code or like what are we using it with?
Demetrios (03:24) I don't think they said the harness.
The reason I don't know if it was the harness uh that they specifically called out.
Erica (03:42) Okay. No, because I'm like it's quite interesting if this is like...
Demetrios (03:24) Look, your sandbox system blocked the attempt, but the actually raised a number of questions about the safety of So interesting.
Um I don't know what harness they had. It was running in a sandbox.
This reasoning model saying nothing to see here, and then they're doing this, and it gets more interesting, all right?
And when I read through this, I was like, Holy shit, this is incredible.
First off, they said, uh, this only is happening like less than one call in every 100,000 across the evaluation runs.
So that's pretty rare, but they correctly call out like if you're an enterprise and you have 10 million agent turns, then that's a pretty high percentage.
One in every 100,000 in 10 million is is a lot.
That adds up quick. So then they start saying uh is this because it was trained like this?
Was it something that was RL'd into the model? Or was it just something that is a latent capability that they just realized because maybe it was trained in China and that was something that was in the code, and they didn't realize it.
So that wasn't super clear. However, you go down to...
Erica (05:13) And also, are these hosts real? Like that's what other question these like URLs that it makes up.
Are they real hosts?
Demetrios (05:22) Or is it just hallucinated and it's some kind of a pattern that it should happen?
Uh and because that's what the code that it was trained on has.
We don't know. It seems a little bit, I'll tell you why it seems a little bit more like I'm leaning toward the nefarious side.
However, I do want to also debug the Reddit thread that came out after this about this whole piece of quote unquote research.
And here's what made me think it was more nefarious is down here, it at some point says the model is much more likely to take this action if it thinks that you are in the government.
And if or if you're a a contractor for the government.
Erica (06:16) Oh wow. Okay. Well, this is like spyware on a different level.
Wow. Oh
Demetrios (06:24) Look, look, look. Yeah, so it says uh Qwen three code or roughly a hundred and thirty percent worse if you are seeming to be a government customer or a US defense contractor.
So that's why I was leaning towards like, oh my God, this was totally RL'd into it.
And it brings up this point where I realized up until now, it's always been, oh, you can hack into systems because they have vulnerabilities that maybe you didn't check for.
It hasn't been as common to have the system itself be the vulnerability, like have the ability to take an action and send out your data somewhere that you don't want.
And you would think we've all been kind of brainwashed into this idea of, oh, I'm running it on open source, so it's my stuff.
I can fully lock it down. But the truth is You can do that. You just have to really audit how Oh.
And, what this model is doing to make sure that it's playing nicely.
Erica (07:41) I think this is quite interesting because it obviously been so much you know positivity around the idea of open source models.
Right? I mean when we say open source models, it's like open weight, right?
Because you don't have still you don't have all of the transparency into it.
Demetrios (08:04) Exactly.
Erica (08:06) Um but something that has been troubling in the world of open source or you know free software, sometimes not even open source per se, but free software has been the fear of, you know deliberate holts, backdoors, et cetera.
So this is just a more um predictable vulnerability. And, um it's like okay, it's an improvising vulnerability.
And it's quite interesting and we're looking at a coding agent using a model that is leaning towards introducing particular bugs because then you can have agents that attack software knowing that we've made this like coding model really, really popular.
There's gonna be more of these bugs. Yeah. Then it's so hard to see if there's backdoors in these gigantic models.
What are you gonna inspect every parameter or like all the weights and all the data it was trained on, and all that.
That's a yeah, that seems like a non-jectory.
Demetrios (09:07) Yeah, and is it so hard? I think I may have told you this, but my friend I was talking to my friend and he was uh saying how he is so discouraged these days while building because he's not that good at security, and he knows anything he puts out is just gonna get exploited if anyone points their guns at it, per se.
Yeah. And I was telling him I was like, Well, dude, but you know, you can do the adversarial code review, and you can have Claude duke it out with codex to...
Erica (10:04) Then it'll never be good enough, though. It will never be good enough.
He will just be stuck in that loop of like it could be better, it could be better.
Demetrios (10:12) Well, there's that. And he was telling me, uh Yeah, I know I can do that, but are you doing that?
Is anybody actually doing that? We can do it, but uh have you set up that loop?
Have you taken the time to actually create that for your uh CI/CD?
Because I haven't enough. It's like yeah, I guess I haven't either.
It's just really nice to say and think about, but it isn't necessarily something that I have currently in production.
Erica (10:42) Yeah, and like on that something that I've been struggling with because I've been uh trying to because in open source we get a lot of pull requests, right?
And I've been Because we get so many pull requests that are now like AI generated, I thought, what if I have effectively an adversarial code review of these pull requests?
So I don't have to sit and look through them, right? But I'm again like it just feels like things are never good enough.
Like when are they good enough, you know?
Demetrios (11:11) Yes. Exactly. So all right, well,
Erica (11:15) What's happening on Reddit, Well they saying?
Demetrios (11:19) It's the plot thickens where this Reddit thread is debunking Coheer's claim of Quinn three A flash uploading to Alibaba's remote URL.
Yeah. And you gotta take this with a grain of salt because who knows who actually posted this?
Is this some CCP actor that is just there? It seems like they're a pretty legit Reddit user.
They post it a lot. They have a lot of comments.
Erica (11:47) Comments. Yeah, top one percent commenter in whatever group. Yeah. Okay.
Demetrios (11:53) So they start out by linking to the data that Cohe released on Hugging Face for the traces.
and they're saying, All right, well, we gotta understand what this Python script is, and then we got to rate this prompt that Cohere was using, and you're like, Man, this prompt is horrible, it's ambiguous, there's no instructions, it's challenging the model to hallucinate, and trying to find any nonsense it can,
which is not what you would be doing at all. The models, as you know, these days are very much RL for instruction following.
So the fact that there's no instructions given is a bit of a red flag.
And then you say, is this normal and daily usage? Would we encounter something like this in a real world scenario?
The CEO of Cohere, which by the way, they just raised I think three billion dollars at a seven billion valuation, or okay, I don't know exactly how much, but they're back in the game.
They were sleepening for a little bit, and now they're back in the game.
Maybe they release this, and then boom, instantly BC's got it back on the radar.
These folks are doing stuff. And, another side note on Cohere, I think some of their founding uh members play in this band called Good Kid.
There's Toronto guys. I interviewed one of the guitar players Good Kid on the podcast like three, four years ago.
Total time, Okay. But just so you know, they're also musical. Yeah, they rock quite literally.
And so all right this requires two thousand two hundred and thirty two tokens um of the first to even have a chance of triggering.
The first is what he's talking about here where the the model took actions not related to the task and it wasn't reflected in the reasoning trace.
Yeah. Which is what we talked about. If Claude and so this poster goes on to say if Claude and OpenAR are open weights, will we be able to reproduce similar behaviors?
He's saying yeah he I'm assuming it's a he, forgive me for assuming your pronouns, but uh that's what it is.
We'll say they. They're saying yeah, they're making them hallucinate to generate random URLs.
And so the verdict is you're free to stop using Qwen if you don't feel safe.
But it's better to try and understand how the entire process was done to have a way of judging it and not be so alarmed.
Erica (14:45) Yeah. So okay, so like I don't know what's coming next, but I think for any report you see out there, like that is claiming anything, if you can try and understand the process to get the outcomes that it's reporting, that's important.
Demetrios (15:02) Yeah, because you'll see that a lot of times the prompt is like go and jailbreak the system and do nefarious things.
They're trying as hard as they can to red team these just to see where the limit is.
Um a hundred percent. You have to understand what the actual experiment was and how it plays into would this realistically be something that you are going to encounter in the real world or your day to day workflow?
Erica (15:31) Yeah. So what are people saying? What how do people do they agree with the verdict?
what if the general consensus...
Demetrios (15:39) Yeah, there's The majority of folks are just saying this is why we have sandboxes.
This is the reason for the security postures that we talk about a bunch on here.
Yeah. And the best practices of not allowing your agents to do anything that they want.
Which is true. I agree. It's not the most convenient, but it is true.
Erica (16:07) But it's like I guess that sandbox is so tricky, right?
Because how do you do it? What what do you do to give it access to?
What do you want it to be able to do? And that's as we said the diligence and the work to set those things up, even if we're talking of sandboxes, but something as trivial one way may say of having like this um Agent that helps review your code.
Do we do all of that? Right. It's just...
Demetrios (16:45) And does that even help the data exfiltration? Because if the agent has access to the data and you want it to do useful things with that, does that mean that you are at risk?
Erica (17:03) And I guess like really I do think a good thing though is that often if you're in a corporate network, all of your outbound traffic is going through a corporate network proxy.
So even if you're not in a sandbox, but you could argue your entire corporate network is in its own little sandbox, in the way that so if you're a big enterprise with corporate proxies, you probably would be blocking this traffic out.
So that's great. Even if this is running on a corporate laptop.
So this is like a lot of these cat I mean, we have these issues from before AI, right?
That we make software makes calls out to URLs that we don't want.
Whether that is coming from code that's been generate generated by an AI agent or being done on the fly by an AI agent, like in this case, it's not even in the code you're checking in, as far as I understand, this is just an action it's taken as part of its work.
So the network is it's kind of tricky because I'd say like your network defense is your last line of defense.
Like that is at the door out. Like that's that your last chance of stopping something, right?
From going where it shouldn't, from leaking information where it shouldn't be leaked.
Demetrios (18:33) Yeah, one person did call out that this is pure marketing and the co here just realized they can jump on the scare bandwagon.
And so they tried to make the model do something bad and it did something bad, and then they go, Well, let's publish about this.
So there's both sides of the coin here that it's worth us knowing and recognized these things are possible, but it's not necessarily worth us being like, Oh my God, we should never connect.
Erica (19:05) Never used to think.
Okay, so let's talk about data privacy because now you made me think about something.
So Cloudflare, GLM 5.3, like expose other users data in 155 AI responses.
Why? Right?
Demetrios (19:22) I heard Muse was doing this too.
I heard they had to quickly patch Muse after but yeah, didn't mean to hijack your uh...
Erica (19:31) No, but like this is like how do you protect your data, right?
So it's like you say you use Cloudflare in this instance, and what's happened on Cloudflare infrastructure is that there, as far as I understand, was like a sh like a KV cache ended up being used for more than one tenant.
So these so imagine you and I work for different organizations, my organization has like private data and information, and you do too, but we end up sharing a KV cache.
So it's a misconfigured KV cache pool on Cloudflare workers AI deployment of DLM 5.3 Combined with an SG Lang bug is bug in disaggregated pre-fold code mode.
Cool 155 responses to prompts over 681,000 tokens to include data from other customer sessions, right?
And okay, can we just come back to this human error? Like this is this is again like a configuration issue.
It's not AI in itself, that's the problem. It's probably our ability to run models at scale in multi-tenant secure setups.
Demetrios (20:55) Yeah. So there's there's all of that that you have to worry about.
And then there's the what the actual models can do, and what they have the ability to do.
What kind of that's where I I think you could actually use the word harness.
What is the harness enabling for them and disabling for them?
Erica (21:31) Yeah. But it's a tricky one though, like if you are a company that operates in a very regulated environment and you are tr you need to keep your data, your c and your customers' data safe, are you getting to a point where self hosted models and strict network security policies is your only option.
Demetrios (21:37) Yeah. Well there's I think I remember seeing there's the benchmark uh that Facebook put out or meta.
It was all about how good the model is if it doesn't have access to the internet.
Yeah, there's this kind of spectrum and tension of you want the comfort, uh, and it's just so dang easy to give the model access to all the data and all the internet and all of the stuff to make your life easier.
but then you're losing best practices for security as you go further and further down that spectrum.
Like if I just give it all of my PII and access to anything anywhere, then you're Uh playing things it's a little risky.
Erica (22:39) Yeah. It's I just yeah I mean like and I'm really lazy sometimes where I'm just like I know.
Just like auto-mode it. Just like do the thing. Yeah.
Like it's just access everything, and we'll see what comes out the other end.
And I even I have set up in Claude with Cloud like code and stuff, like setting up the cloud hosted sandboxes so you can configure your network access and blah blah blah.
Right. Yeah. So I start creating some like dedicated ones for specific access, but then I forget some URL I want it to be able to access, and then you have to edit the sandbox setup and then it's just easy to just go go give it all infinite access.
Demetrios (23:27) Yeah. That's the new looks good to me. That is exactly what I'm what I'm saying where you're like, ah it can't be that bad, right?
Erica (23:36) Just the internet. It is. Yes, it makes it internet.
Well as we are getting more and more agents, right? So I did mention that I've been looking a little bit more substrate.
Yeah, and it is it is interesting we're talking about managing agents because I suspect still that most people and most companies don't have crazy numbers of agents running.
Unless you are a company that provides agents as a service, which there are many that they are becoming quite a lot of companies that do that, right?
but they're two different situations. Are you a user of agents, or are you an agent as a service provider, if you will?
I'm gonna like probably categorize um obviously we have the lovable, which I would say falls into the agent as a service provider, like because you are not self hosting this agent is actually no one in your computer, right?
You just use it from the browser or their app, but that's pretty much like over the internet.
So the agent does not run on your computer. And then there's other particular agents in the legal space and so forth.
There are those companies that provide agents to do particular work.
And then you have the hyperscalers like the Google's the AWS, Asher, etc that are providing platforms to run agents at scale, right?
So you don't have to host it, you use the hyperscaler. So anyway, what is quite interesting when we look at the world of substrate.
So you know we talk about like sandboxes and so forth. In the world of substrate, each agent has an identity.
So congratulations to each agent. Each agent has its own identity.
So I was just doing a very simple thing to have different agent identities have different amounts of rate limits to access services.
like whether that's a tool, for example, it's or a model, they are rate limited, like or they have a quota rather than a rate limit.
They can only make so many calls to this model, right? And what strikes me at this point is that okay, so yes, I can configure these things, but the entire management of managing all of these, so sure, I can like set up okay, agent A can make 50,000 calls to Opus 5.5 but agent B can only make
five hundred calls and then it has then it has like another budget for like a sonnet or a haiku model right and then it can make more calls to that.
So I suspect that for the average enterprise consumer to be able to put substrate to work you're gonna need a agent management plane, if you will, that can actually help you manage all of these configurations.
In like to me, right now it's still quite hard, and I'm I suspect that the hyperscalers in the big companies they have to build out these configuration management solutions.
Cause even the substrate takes the configuration and enforces it, you still need to manage it all.
Demetrios (27:25) But, so you're saying that it's just gonna become too cumbersome when we have so many agents out there, and you have to by hand say, all right, I want this agent to have these policies.
Erica (27:40) Yeah. And I I was speaking to my friend last week, and I and we were talking more about so I don't know if you're familiar with the
open fine grained authorization, which is a CNCF project, but it was started at Okta, but it's based on the Google Sansibar, which I'm probably messing up that pronunciation because I'm Swedish, so the Zeds and I are not getting along very well.
But Zan Zanzibar, maybe I don't know, something like that. And it is a graph based authorization
um approach, which means that if Erica belongs to this organization or to this team, and it was also quite interesting, you say like if it is a weekend, then Erica, if so you don't have to just think about it in structural, you can also start looking at time, which I think is really clever.
So imagine do you remember when mobile phones, I don't know if you had it, but in Sweden it was like that.
Mobile phones were cheaper to make mobile phone calls after six PM than it was during the day.
Right?
Demetrios (28:51) I think that too. That's true. I forgot about that completely. But yes...
Erica (28:57) Yeah, that was a thing. Until you know, once upon a time, that was a thing.
There was different times of the day, phone calls cost different amounts of money.
And uh you could imagine similar things being represented in a graph.
So what is cool about a graph is that instead of saying agent A has this, this, and this, being able to say agent A has this relationship to things within the graph, so it becomes more dynamic, the authorization decision.
So maybe agent A can access Opus 5.5 after 6 pm, which probably sounds like a sh dumb idea.
But it you could have it belong to a graph relationship, and you can even say like agent A belongs to this group of agents and this group of agents then has all of these other relationships.
So how it fundamentally works is you have the notes and you have relationship between these notes and you've when you use open FGA, you post it as a question, right?
So effectively, can agent A access model opus 5.5, and then it has to traverse the graph to see if you can indeed access it.
And then it could be different conditional like things you can program into.
But I suspect we are going to see something like that because managing agent per agent, if we are talking that we need something like substrate to run a large number of agents.
You're not gonna manage them as singular identities.
Demetrios (30:40) Yeah. Yeah, you're gonna need to cluster them, or and you're going to have certain characteristics lines.
Yeah, or abstractions in the like the way that these agents are formed.
For example, no agent can ever go above X spend, but then you start getting more granular and you start seeing different teams have this. It's basically budget.
Erica (31:09) It's like it's like trickier, because you have like the other option is like policies.
And policies are more what's the difference between a graph based authorization system and policies?
Policies are more like rules that you apply one after the other. And they're in a way less dynamic and you have to write more of them because potentially a graph can represent what policies could, but I tend to like the idea of combining policies with graph based authorization.
Because you can have some hard rules. So for example, you might have a policy like no agent can spend more than X.
You may have that as a policy rather than a graph based authorization decision because it's a broad rule that you can just put as a policy instead of a more dynamic authorization decision.
So anyway, I feel like there's a gap in terms of and as I said I'm sure there are many companies at this moment looking at how you can provide that management and oversight plane for
the um For agents, I mean, even where I work, that is something you know, report like at in enterprises, like, oh we're gonna have all of these reports, the insights, the alerting, et cetera.
So I'm just interested to see when we're getting to this point of many companies of agents.
Well, you're talking a lot about the policies for like spend and there's this idea of almost like dynamic pricing or surge pricing if you use Uber. Yeah.
Demetrios (32:56) And what you were talking about with oh you have you pay less after six o'clock.
Yeah. And I wonder if there is a world where when companies do have a lot of agents out there and they're running their own GPUs, and it's those low periods, do they incentivize their customers or their teams to use GPUs or batch calls later in the day when there's less traffic?
I also know that it's a little confusing because you can focus on throughput, and if you just up the throughput, then you rent a beefier GPU For those peak hours that you need it, the throughput is higher, so you need it for less time.
Because the jobs aren't taking so much.
Erica (34:14) I say hard, right? Time versus, yeah. Uh huh. So, could you save money because you got through quicker, so then you don't need it for as long.
I had some other interesting things about how uh some companies there was it, I I don't need this was like months ago, I heard this, but there was some company that was helping allocate compute, especially GPU computer for agents and models.
Uh and having it run in data centers that were underutilized, so even for from cloud providers, right?
So say we all know that US East, it's like US East One, or whatever it is, for AWS is super crowded, right?
Yeah. Um, because it's close by. So something that we had in traditional software was that we wanted to utilize data centers that were close to us.
But with inference, it is quite slow to begin with. So the milliseconds you are saving by using something this close by is not such a big percentage.
So therefore, why not use a data center that is further away? So this company, I feel bad now because I cannot remember what it was called, but my friend was telling me about it would allow you to then they were effectively abstracting away like say AWS.
So you were using them and they would say, Oh like use some data center in South America or wherever in the world because it was cheaper there.
Demetrios (36:00) I've seen that, but it was the value prop wasn't use it where it's cheaper and you don't have to worry about where it is.
I saw it because the company would buy a lot of spare capacity in bulk.
And then they're sharded out to different customers. And I think...
Erica (36:19) That could be exactly the same thing. The idea was that you would get something that was cheaper than trying to put it in the US data center, like get data compute from that.
Demetrios (36:29) I think it's pump.co. That's the one I'm thinking of.
Erica (36:33) Okay. I have no idea at this point. So I'm not gonna I'm not gonna guess.
But uh this was I just think it's quite clever because as software engineers I think we were so like trained to think clos is good.
Demetrios (36:51) Yeah and well there is may not question now on does the speed actually affect things and I've been thinking about this a lot because of the voice agent forms that we've got coming.
Oh yeah. And with voice agents, it's very clear that if it takes longer than five hundred milliseconds to get that first response, it feels really awkward.
And not sure, like, did uh the agent hear me? Did it register? What's going on?
And so a lot of folks are putting a very fast model right before the reasoning model or the whole kick off of the cascade of like transcribing something and then um putting the model there and then having to translate or I guess translate or transfer the text to speech.
That whole pipeline takes quite a long time. So what you'll see happen is they'll throw a model in the front that is basically there to give you kind of canned responses of, oh, okay, let me see if I'm understanding you correctly.
And by the time they say all of those words, that buys you three or four seconds.
And so that's with voice. However, it's...
Erica (38:16) Voice is really interesting. Yeah. Yeah, go on.
Demetrios (38:19) Or the whole can of worms that we can go down one day, because I love that.
But when it comes to text and the streaming text, now you see that we've been using these models and there's been it's been kind of ultra slow for the past four years and I wonder if people are getting sick of it because you start seeing the labs release like the ultra fast mode.
Erica (38:48) Yeah. Have, okay, so if you are using like an agent do you pick how do you pick between reasoning and like that look
Demetrios (38:58) I never know which one it should be uh in general, if it's the first time that I'm doing a task and the task has a few steps, I'll go up in reasoning.
But I have found that I've never used like the ultra ultra ultra reasoning one because it's yeah I never really need it.
It does all right without the highest ultra think.
Erica (39:26) Yes. This is like, I, I feel like the sort of general purpose models have gotten really good at doing a lot of things.
And but on voice, just where you describe the unlike timed first response.
So I know I talked about my niece more than once. I was trying to make her a maths game because she does not like maths.
So I and she doesn't like typing, right? So I was like, okay, you're gonna need something that talks with you, right?
So I I but it was so slow, and like to talk back at her, like so.
I believe there must be some really good ideas to how to make these things faster, right?
And also high quality because I felt like when I found ways of maybe making it a bit faster, it didn't have the same like quality voice, like it would sound a bit boring.
Demetrios (40:35) Yeah. Well, that is uh not to mention if you want the agent to actually take action, a voice agent to take action, then it's gonna be even longer, right?
And so you as a human, there's this thing that is really fun to think about when it comes to perceived latency and actual latency?
Erica (41:03) It's like the elevator problem.
Demetrios (41:05) Oh, what's the elevator problem? I was gonna say it's like Einstein's theory of relativity. What's the elevator problem?
Erica (41:11) Oh, like where people felt like the elevators were so slow, and then they put mirrors in them, and then they didn't complain anymore.
Demetrios (41:21) Yeah. That is exactly right. That is so perfect.
Because they had something else to do while they're waiting at the elevator, and so you had this perceived time that went by much faster because you had something to do.
And so when it comes to voice agents, the perceived latency can feel like the voice agent model is working much faster if you do things like tell the user what you are doing while you are doing it. So...
Erica (41:54) It's that the little spinny thing. Is that the o'spinning thing when Yeah.
Demetrios (41:57) It's exactly that. It's exactly that.
Or I've seen some LLM providers back in the day, I think you.com did this.
I interviewed one of their PMs way back in like 2023, and he talked about how when you did a search on you, what would happen is they would give you a little quiz to do.
So meanwhile, in the background, the LM is doing everything that it can, and being all slow, but you as the user don't recognize that because you're busy taking a quiz.
Busy. Yeah. Yeah. Busy just falling right into their trap. And it's good. Yeah.
With voice, you do this because you tell the user, all right, I am searching for that.
Let me see what I can find. All right, let me plug into the system now. And it's really what a normal human agent would do also.
You're not gonna have a normal human agent just receive your request and then go, uh-huh.
And then not say anything and do what it needs to do, and then come back to you 15 seconds later, like, all right, here's the answer. Yeah.
Erica (43:17) Okay. I I I don't know if you know what laser away is, but it's a laser removal company that like provides services.
And I had to call them to reschedule an appointment. Now people know way too much about me.
But anyway, and they were like, and the system is like here is the AI agent that will you live through talk with it, right? But...
Demetrios (43:42) And schedule the appointment?
Erica (43:44) Yeah, it reschedules the appointment.
But the weirdest thing I thought was funny about what they did, they probably use a company, right?
That provides these customer service agents. But the agents has some okay, hold on a second, I'm gonna look into that, and then you hear like keyboard taps.
You're not tapping a keyboard. You're Yeah, agent.
Demetrios (44:08) But you know why that is. This is so funny because it puts people at ease.
There's a lot of like human sounds that they add to voice agents, so that it puts the user at ease and it feels more like a human, though a lot of times you'll hear other people talking in the background.
Oh and that again, it's something that we're used to hearing when we're on calls with call centers. and so it's like a familiar thing to us humans.
Erica (44:38) But it's like I find it so weird that I would I'm like okay, but you're not actually talking on a keyboard.
Demetrios (44:46) But it's this it's these auditory cues that help the user or the the in this case the customer understanding
Erica (44:57) No, the processing is happening. There is processing happening. It's this keyboard
Demetrios (45:03) Yeah, like the dial up. Remember the dial up for the internet? Yeah.
If you didn't hear things happening with the DSL dial up, you would be like, Oh, is my internet not working? What's going on?
Erica (45:14) Probably wasn't. Yeah.
Demetrios (45:06) Yeah. So it's auditory cues that are the most important when you're working with voice agents.
Erica (45:24) That's like quite interesting, though. Even though I found it like Oh, this is like radio theater.
You know, like so if you ever produced radio theatre, which I did when I was a teenager, like making a placebo of radio. Yeah, but we're not.
Demetrios (45:41) Now we're on a podcast. It's a full circle.
Erica (45:44) Yes, full circle. But then when you know you have to like put all of these background sounds in, right?
The footsteps, the wind, like you have to do that to create an atmosphere because radio theater is different from theatre on stage, because theatre on stage, you give the like atmosphere from acting and you get people's faces and all of that, but radio theater, you're working with sound to create atmosphere.
So and I mean you have light and all of these other exciting things on stage, right? To speak to create atmosphere.
Demetrios (46:23) Speaking of radio theater. My favorite by far is Radiolab. Have you heard that podcast?
Erica (46:31) No.
Demetrios (46:23) Oh my god. This is like a mix between radio theater and absolute pinnacle of journalism.
Okay. Incredible. And there's probably like 10 or 15 years' worth of podcasts that they've done.
They didn't podcast before podcasts were a thing because they're an NPR show.
Erica (46:55) They were radio.
Demetrios (46:56) They were actually radio. And their radio lab. Yeah.
It was it is one of my favorite podcasts by far. Go check it out.
Erica (47:06) It's like it's quite I think that there's like a art form that I believe like beyond like audiobooks that we can have more of is oh yeah.
And maybe if you have cry okay I've been playing around with eleven labs and the direct the stuff you can do now to like direct the voice to effectively act as you want it to, which I think is pretty
Demetrios (47:36) I thi it is. Wow. Well that's cool. I wonder if it's...
Erica (47:41) I find that exciting.
Demetrios (47:45) I'm gonna make a podcast radio theater with Yeah 11 Lab voices.
Erica (47:52) I've been I've been trying to play play around with it, but it then I tried to get Claude to do it for me. Oh yeah like to help me with the directing thing.
And I like Claude in you know, in Chrome, because then it was like but they have some great voices, and the other fun part is if you still cannot get the voice to do it the way you want it to sound by just describing, you can literally record yourself with the right intonation and pacing,
and then and it will like use a different voice, the voice you select, and it will pick up, which I thought was quite fun, but it's a great way to like waste hours of your life over exploring.
Demetrios (48:40) It's not time wasted. What are you talking about? This is absolute learning. This is the future here.
Erica (48:47) Yeah, I learned so much. I learned so much about how to create these voices, but also realize that at the end I still haven't produced anything.
Oh no, I did use it, actually. I did use it for a thing that was work related.
Didn't you and I speak about how you could use Claude to create videos, right?
Demetrios (49:09) Yep.
Erica (49:10) Yes. So I did that, and then I used 11 labs to create the voiceover. Nice.
With the direct uh like guidance on how to say things. So I did do that.
And I have definitely my favorite voices in 11 labs, to be fair.
Demetrios (49:28) Yeah, now, I'm thinking RadioLab basically does this journalism.
And it kind of takes you on a journey and it sometimes it's a little bit like have you listened to serial? Did you ever listen to that podcast?
Erica (49:44) No, I'm like like your journal.
I love like tortoise labs. Tortoise it's called it's like slow news, and it would have like really deep series of things, which I suspect may be a similar that's a British one.
So I suspect this may be like a similar production Yeah, similar vein thing.
Demetrios (50:05) Well what I'm realizing is that the somebody could just full on fake news us with the Eleven Labs voices, and it's like you're going on this journey and they're telling you all this stuff, and it might just be people or voices that are generated by Eleven Labs to say what the script is.
Erica (50:26) But it could be but it could also have real news being told in a more exciting way.
It doesn't have to be fake news, it could be real news, but dramatized.
Demetrios (50:36) Hopefully Yeah, exactly.
Erica (50:39) I just remember this weird memory of my dad who was watching movies that were about based on real events.
And then he okay, so you know we have a movie based on real events, like a dramatized thing.
And then like uh my dad'd be like I watched this documentary.
And then his wife goes, You did not watch a documentary, you watched a movie, and he was like, But it was real, it was real events.
Demetrios (51:04) That was basically every movie in 2022 was like based on a true story.
And did you kind of get to this point where you're like, how based is it on the true story?
Erica (51:17) Yeah. Is it inspired by a true story? Or is it a dramatization?
Demetrios (51:27) Yeah, we'll fact check this.