AI Worker = AI Agent You Can Hire

DevReal: AI worker = AI Agent you can Hire, Nikita Ivanov

Recording: DevReal: AI worker = AI Agent you can Hire, Nikita Ivanov

[Laughter] but uh uh here's what I'm going to do today uh let's spend half an hour I'm going to here's a cleff notes half the time I'll give you some I try to make a case for what I believe is a is a horrible AI or a soft what find Workforce a Bry New Concept that kind of you know springing up right now there's quite a few companies humron is one of them by the way I'm invested in humron so I actually put my money where my mouth is and half the time you know I'm basically going to show you uh what human rers today by the way everything is public and you guys can try to yourself but again it's a very much work in progress so don't be alarmed by that so um we're going to do that speak skip that so what is AI work worker or a software defying Workforce in general think about this it's a something you know essentially it's an agent that you can hire and um when I'm thinking about the different types of application AI that we've had for literally for close to 20 years at this point I always keep thinking about is is in a sort of a way where you're packaging the same functionality in many different ways think about this we have child bot since 2015 I I was a part of this first wave of 2015 then we got a co-pilots recently and the difference between chb and co-pilots right the co-pilots need the hostest application but fundamentally the back end of CH boats or co-pilots can be pretty much the same so you package this a little bit differently then agents came around right with all about all of us right now talking about agents so the agents is fundamentally just a different repackaging of the same backhand capabilities uh you know probably the best definition of Agents is think about the microservice where you replace compute plane with llm it's a Ste basically an API driven piece that you know you have to somehow use then there's a AI worker and we'll talk about this a little bit more in detail but this is fundamentally where you can hire AI agent in very much the same way as you can hire in people today and there's definitely physical AI well it's embodiment of you know just robotics as we know them and that's coming up in a few years from now some of them actually available if you guys on the on the weight Le of prime it's actually being delivered as probably like B spring so the concept of way I Work It basically is an idea that you you remember this question that was asked like little last time what is the impediment for Enterprise to use a agent so what is the agent it's a software you have to install you have to buy you have to program you have to configure all those complications just to get some of some functionality out of this think about how we as companies get something done if we have a piece of job that needs to be done we tend to hire people or tend to actually hire a somebody who can perform the job so why don't we actually apply the same Paradigm to AI what if we can hire AI in practically the same way we can hire individuals uh and that means that hiring and on boarding in Daily work should be really much identical I usually call this if you don't squint you shouldn't see really a difference between hiring people or hiring AI into your organization I know it sounds a a little bit far-fetched most of you probably thinking about right now holy crap it's bits you know out there and you're right it's still you know for example in there is some age cases for example you know we don't really have yet you know production guts in Avatar so we canot really do a zoom call with the ey with the full facial expressions that's un that's likely coming up in about 12 months from now we're going to have API for zoom we're going to have a gyar so we're say going to be jump on the call with AI but even today you can B we talk I'll show you today how you going to do this uh one of the interesting part of the AI work is that obviously have to be autonomous you know we can thinking about AI agents to work autonomously it doesn't come for free autonomous behavior is pretty is pretty tricky to implement you know just about now if you follow the news open the anal tasks right some kind of rudimentary way of having the background task running but it's it's just about now surfacing up so it's not really trivial so so the autonomous way to do that is very important probably one of the most important traits of software defined Workforce is a social and organizational integration what separates us as employees and teammates is that we have that social work integration into the company we all part of it right we know our friends we know our friends phone numbers we know their names we know who they're reporting to how do they work we know that if I need to basically you know buy a new mice I have to get a report or permission from somebody with have all this organizational connecting tissue that you know AI agents by default wouldn't have for example we know socially that I want to call I don't want to call my boss in the middle of the night unless I have something really really important to tell right if I something update I'll wait till the morning even that simple social behavior needs to be programmed needs to be actually somehow trained or programmed in a very basic level and most of you can probably count like dozen and dozens of those little tricks that separates us from software and so the idea of AI workers is that why don't we solve that why don't we make the AI be as as close as possible obviously it will never be as close I don't want to be pretending that we somehow reach that Singularity but we can get pretty practically close that again if you don't squint you don't really see you know I've been basically telling this example that how many of you guys were in the zoom call today today probably most of you somehow probably in the last couple of days what do you see in the zoom you see a small pictures of people talking to you in right now in the lab we already have G just Google this gin avatars look identical to humans without a forensic search and video you cannot make a difference between that and a human it's already there it requires GPU in the backend so it doesn't work on the phone yet technicality so about two two year or two from now you get jump in the zoom call I would challenge you to tell me without a special badge on the window who is AI who is not it's already here it's literally in the labs and in the demos you know we already have you know Interruption speech from open AI SST speech to text text to speech all of that is already production level we're about just about now we're missing some of the integration from Zoom or Google meat plus the G and avatars so my take on this is that in a couple of years most of us will be jumping the zoom calls or something like that and the only way we can tell that somebody is AI or somebody's individual human being we obviously going to have some kind of Icon some badge to that to that you know window telling us who is who anyways this is actually important question I know it may sound like a ploted dudes but I've been pitching this story for last couple of years and uh it is interesting how the sort of Z guys can Community is changing and a couple of years ago just around the GPT 3.5 came around we started working this idea and to to my surprise the response was largely negative people basically saying look this is freaking immoral why do you want to build AI to replace people it's like this is like inhumane it's like we don't want to be part of that and I was literally taken the back because you know very rarely you see that sentiment generally when you work in software right you software typically it's perceived like in some kind of you know good thing to do you you save time you're making something more efficient this was different you know when we had the first demo of this uh the response was exactly what I described to you half the people kind of like it but you know another half was like literally actively sort of are against that so anyways uh it's definitely changing by the way this sentiment is still here surprisingly I mean I I guarantee you I can pull this room about 20% of you probably would tell me that H it's kind of eeky you know people are people software is software don't freaking mix it right it's like why do you want to do that there is the reason why we want to have at least technology now how it's going to evolve and end up we I don't know exactly because a lot of this depends on regulatory requirements and pressure from a government whatnot but think about this there's about uh 3.6 billion jobs in this planet the professional jobs account for about 60% of that so let's say two billion jobs is what the professional jobs are basically those that can be done with the computer piece of paper something like this my biggest contention and that's probably where my contrarian sight of kind of comes out is that I don't believe that we're going to have an AI in our generation or maybe our kids generation so for the longest time you know absolute majority of employees and the companies will be human and this is by the way very different from a lot of my peers in the Community basically if you listen to you know anthropic folks open AI folks they will tell you that it's just around the corner and everybody will be replaced I don't think anybody any of us you know can prove anything but fundamentally I believe that you know or the the AGI story of this kind of omni pattern replacing everything we're centuries away from that we need new physics we need new energy all that takes time and we can talk about it endlessly so I still believe that in the 10 20 years from now over 90% of the employees will still be people just like our kids you know we are if we're still here we're still going to be working in a companies where most people will be most inlo with people however even if it takes 10% of the employees will be AI based that is a massive massive opportunity we're talking about in millions and millions and millions of jobs what's also interesting is that only if if if the AI portion is less than half it follows that fundamentally it's ai's job to adapt how humans work think about this why would human need to adapt to how 10% operate 10% is a software it's better to be actually adjusting how the humans operate in our in our contemporary world and uh that's kind of my biggest contention there's another another interesting point I strongly believe and that's back to the question about the how does AI how does company you know use agents right this question is very uh fundamental because it's not easy just you know come back to your own companies to your own teams to your own projects and forget about co-pilots co-pilots are easy right they always have a host application that needs to run it right agents do not have really host application that has to somehow run in sort of ether of something and think about how would your company would use agent now there's one way like demonstrated before like a tooling but that still requires basically quite a bit of configuration programming it's still basically a piece of software that somebody has to acquire configure run or use someone else's API to do it wouldn't be a lot easier and my contention that the most effective most efficient way to introduce AI capabilities into company is to hire AI because as a company you already have all the business processes and I'll show you in a second because I know it sounds a little bit too farfetched so anyways uh this are the two key objections but we talked about us right so having you know having had a privilege to basically pitch this idea this are the two main objections that I I see every time I talked about the first one that AG I will replaced everything and everyone it depend it basically based on your beliefs I don't think I have a proof nobody has any proof you know I believe it's not even the century somebody will believe somebody else but my contention that you know we're I mean we are nowhere near because you know if you kind of listen to the kind of you know saying voices we need new physics new energy sources to really go beyond the basics as we have right now what's also interesting is that you know if you look you know two three years from now GPT 5 six generation that's perfectly adequate for for many business tasks we don't need to wait for AGI to get value out of this and look at our current reasoning models that's what was been training the last year and a half multiply it by three by five you get a very decent reasoning capabilities we call them reflective you know reasoning but doesn't matter imagine like you know two three years those reasoning capabilities will be way better than today that's good enough for majority of business tasks you know we don't solve you know cancer problems in every jobs every day most of our activities during the day are much less complicated ated okay and so it's a almost the current zy guys that in 2 3 four five years whatever the Horizon the current generation of Transformer based model will be so Advanced they're going to cover 80 maybe 75 to 80% of business task as we know them so that's good enough another one the second one is what basically is almost like a moralistic question you know is it inhumane or im or immoral to try to replace people people at jobs that's a good question by the way I certainly uh as any human can relate to that I think the moral question aside we can argue this technically I think um I always keep saying like you know look at the history I mean most of us you know as human beings we tend to think in terms of our lifespan 60 70 years right our grandparents what they were what they were basically going through but you know if you ask your grandparents at least my grandparents they saw completely different jobs in the typist in the 1950s 1960s there was a massive employment from for women as a matter of fact you know the horse you know uh Carriage drivers there was a big big employment that completely got wiped out in the matter of probably a decade when the Industrial Revolution happened so these things do happen and as the jobs disappear new appear that's a very natural process maybe this one will be bigger maybe this one will require certain regulatory government you know or compensation most likely will be I've talken to somebody in Sacramento it's a big deal because you know if you imagine if you can replace just a two or 3% of the jobs in California with AI it will be literally a civil war okay forget about 5 10% just a small percentage of job replacement will lead to the massive social upheaval and so there's a lot of you know obviously from from a government mental perspective there is a lot of you know thinking about this what is it what do it what does it even mean because obviously the whole problem that most of the AI can do a lot cheaper Than People AIS don't need Social Security AIS don't need medical they don't get burned they don't they don't live they're very effective they don't forget anything all of a sudden there is all this you know obvious advantages that humans just don't have and businesses will always look for the sort of smallest gradient to to to to choose and that's a very natural thing right why do I have to hire Nikita when I can hire you know Jacob who's an AI who never sleeps and works every time it's a very basic idea but fundamentally there's something to it so anyways um the last slide before I can show you some of the stuff uh why now well I think it's obvious I mean most of you guys in the you know probably following the ipace I don't want to give you the the history lesson but I think the the critical point is the obviously introduction reasoning capabilities because again remember two years ago we just had 3 3.5 GPT it can it can chat somewhat interestingly but that has literally very little value and you know Chain of Thought prompting is a very awkward and sort of a kind of slow process and the reasoning models even the initial generation of them right 01 03 R7 from Deep seek and now the Deep resarch just announced from gr 3 and the Gemini 2 and all of that you can see a progression little within about nine months we progressed from you know hey can you solve that little task for my high school from middle school all the way to doing a fairly deep deep research in deep end research like you know Jim and I can show in open a can show so the introduction of reasoning models is definitely a Lynch being for both agents and AI workers because all of a sudden we can now do a little bit something just slightly more interesting than just a core generation of something and uh I think that's why now uh obviously it's not a point in time it's a more of a Continuum you know we're going to be talking here two years from now we're probably going to be having the same presentation same conversation but slightly better results but you know we we we're there in terms of kind of you know we over the Rubicon now we're in a sort of a adoption expansion Plateau so let me actually show what humon is working on um again this is you know it's a sort of a very early stage it's a team is basically working on it every day it's a work in progress but fundamentally this is an idea that you basically um I'm actually maximize this so we can see it probably better yeah so uh essentially have a website we can basic you you have uh all kinds of workers right now obviously those workers are just a demo proxis against all the different models uh but the idea is that you treat the workers pretty much like you treat people they have resume you can interview them they have skill sets uh you can find them you can look for them so um if I just uh log in here just give me a sec so uh you can basically end up in a in a different place and you can search for all these different rumes any way you like it by skills and whatnot let's say we looking for um L mini and uh I think it's this one and not only can see in the step what's available like a training like a whether it's a free trial some language support some Channel support I'll talk about this in a second some skill set uh you can click out and see out know very traditional sort of resume style right there's a profile just like humans what can we do what we not do some basic core skills some Technical Training some languages I know some count can be placed on we'll talk about this in a second why it's even matter here uh some of the Ming steps some of the communication channels I can talk to right like a human beings I can do emails I can do phone SMS I can do slack I can do all kinds of different things these guys can do too by the way so PR cool the coolest thing is that I don't really buy anything I'm hiring and it's exactly the same I can hire monthly I can hire hour I can hire hourly and I can choose that so we can interview by the way so we can go here and start interview process just like we do with the actual human being and uh well what is it you can do for me and um obviously as you can imagine that's pretty simple it basically looks at you know its own description it knows what it's a self-aware entity it lives you know it knows that it's been interviewed it Knows by whom it knows who he is or she is uh and he basically can ask it can basically ask not only this but can basically says like um what it will cost me to hire you for I don't know six months and obviously is going to calculate that and answer so it's a self-aware system that basically tells you all this interview questions so you can ask about anything you like yeah if you have a tool for that yes yeah it can actually do anything the tools can do so so anyways if you like the interview you can talk about this and you know can you can ask like for example uh can you answer that in Italian and obviously it knows Italian so it should be able to do that um I guess anybody speaks Italian so I'm I'm hope it is anyways if you like the interview right what do you do after you interview individuals right human beings well you go ahead and hire them just go ahead do it let's do click hire so this is actually very similar how you would actually find normal hiring application for human beings it's pretty long it's pretty cool by the way I'll explain to you what's what's one of things so first of all you have a first I his last name a lot of these I kind of you know when I describe this things like a gender an age group people kind of freak out a little bit so why do you need these things for AI right well I can explain most of them right so obviously in the name you can use technical name you can use pretty much human names doesn't really matter gender age group FS mostly for a kind of a model context and the grammar and a pronoun so at least if you know if it's a gender we know how to use pronouns I'm not going to go in details about this whole political issue but fundamentally have genders that you have to know about right you can pick you know avatars if you like it right obviously we have bunch of you know built ones so let's pick anyone and move forward uh you have to give a title I mean when you hire somebody there's a job title you have to give a title you have to give a job description uh assigned a certain team I don't know let's go aing assistant as usual and this is interesting counter City time zone primary language so you can place your a AI worker in a certain location virtually in certain time zone CD is kind of even no superficial it's mostly for just to fun but what's interesting think about this when you're hiring people they don't work 24 hours when you hire an engineer let's say you're the manager you hire an engineer you don't want this engineer bugging in the middle of a night with the code updates how do you how do you model this you got to you got to basically place this engineer close to your engineering team in the same time zone same city for example so that anybody ask you hey Johnny where are you located he will tell you basically I'm located in palalo I'm in Pacific time zone and you can actually down here you can basically specify is it working 24/7 or 9 to5 and 24/7 is basically just a g software right it works all the day all the time but you can click here and it will be very smart to know that if something urgent he will actually bug you at any time but most of the work related communication will be delayed in a 9 to5 time frame of where it's placed it's very similar how humans operate think about this when you hire somebody you don't expect them to be 24/7 you expect them to be like working normal business hours in certain time zone whatever this person is and you can do that um and then you have a bunch of you know Communications style things you know you can have very informal sort of a style most of us probably work in startups so we can be like very informal now communication if you work for a law firm you want to be very very formal with your communication style you can do that you can actually work in creativity but fundamental temperature of a model and again most of can be adaptive you can work in brevity in the language mode you know very plain language for example if you for example you if your audience is kids or teenagers you want to explain a lot of terminology in the use abbreviations vice versa if your audience is very technical use a lot of abbreviations so kind of shorten that so you can configure all the social communication preferences this is pretty cool workspace Persona this actually has an impact on the long run in tasks you can for example configure your worker to be more of a builder right and he will gra will he will I'm saying he he or she it it will try to essentially prioritize completing tasks in building things versus exploring the you know a solution field solution you know uh space in vice versa you can click on Visionary it's all going to basically going out and trying to explore everything trying to come up with different things and probably prioritize less of completing something on the deadlines and stuff like this you can specify who is a hiding manager who is a contact person pick a start date and pick how you want to pay for it simple as this hourly monthly doesn't you know agre terms and literally higher what's going to happen now uh is that his name Kai Kai is basically hired um and uh we can go right away start working with Kai you know we can find Kai on our team I have a bunch of hires already here here is the kai right there and we obviously have like everybody else we have a built-in chat but that's not the only way to do it um I'm sure sure that I already got the email from Kai and this is the let's see if I got a mail from Kai and this is Kai uh this is how basically it looks like a similar so Kai has been hired and immediately sent me email from his private email and it's basically tell me hey Nikita nice to meeting you nice to joining you here's what he can do you can configure a slack and configure SMS you can send me email to this it's basically very natural way of doing this so we go back to Kai by the way um we can just chck to the K chat to Kai um no lare so this is let me see if I can open chat yep sorry so this is a little smaller window I'm sorry for that but uh fundamentally you have Kai here with all the stats you can manage him you can terminate you can p him whatever you like like a human beings uh you have other hires in the team those are all AI hires um this is communication channels that Kai has we can configure a slack for him if we need to basically he will appear on our slack in the company we can configure SMS we'll assign a phone number to him so he can basically text him whatever you want to text him and he will text you back whatever he decides to uh and yeah you have basically a chart application here and you can start our tokens basically um I don't know welcome aboard and uh his life you know he can basically uh this is the kind of you know a backup chat system that we have and his life and thank you immensely and we can send him back a mail so we don't have to chat just to buy this we can basically hey um we can just say something um yeah yeah we'll fire him in the end you guys like it uh can you send me I don't know um scallet three code for a bubble sword right something that everybody does and um and uh yeah hiring firing him perfectly fine you go to manage right here and you have all kind of things you can basically puse him it's a software so you can very convenient can puse him UNP him if you like and obviously can you can terminate him and and uh since he's got no feelings we can terminate him as well as you might like yeah in GP I can ask to do something and it does it pry quickly how do I keep this busy how do you keep busy anybody just give him a task you know asking yeah it's basically again the concept here is that you don't really have to do anything special I I want you to essentially after you hire somebody by the way see most of you are going to be working with slack and emails and teams and whatever you forget about humron you just hard an entity it has a name in the face and all of a sudden you know day and two three down the road you kind of forget it's it's a it's a AI it's just there it communicates with you it does the work for you whatever you ask him to do and over time you just don't even think about it it's there it just basically never gets sick never gets late does something pretty nicely remember stabes information and kind of been a nice guy to have around oril you want those question so I'll bring back to you Nik you want to answer question true thank you very interesting so what I'm wondering uh if such a hired U bot basically doing a trouble doing a mistake who is eventually responsible for that so what's the real what what happens with a humans humans are yeah there's yeah we're not going to go that route you know um and I'm saying we I don't think it's going to happen but fundamentally the most interactions will be very similar you obvious not going to sue AI because there's always no explanation that a can make a mistake and what not uh and vice versa by the way your your AI will not sue you for whatever you know you may happen so it goes both ways but AIS will definitely make mistakes and for many many many years to come you have to remember that if there's something tangible you better double check that's very simple hi so so usually when you onboard somebody especially in an engineering team you might have like a group chat and you guys are in a channel you have a conversation so if let's say there's a team of five people one of them is an AI is he going to or he or she going to contribute to the conversation or you know is it going to be like hey AI say something give your opinion yeah so what I'm not showing you is the whole other side how do you build those workers and that's going to be something that the team is working on but fundamentally it depends a lot on implementation of that worker and uh it's going to be a kind of no quote implementation so you're not going to be writing letter quote for it or at all uh I tend to think that for for a long time it's going to be mostly reactive but even right now we're going to supply all the slack Communications back to that entity and it can decide huh they not going to be in me you can invite them but they're not going to say anything until you no it's it's entirely up to up to it it will have all the Stream conversation that happens in this meeting and if the model decides that needs to interject with something for example somebody mentions its name and says I I asked Johnny to fix something Johnny can actually say no I didn't I don't have any record of that you didn't ask me uh so again this technicality I I know it sounds a little bit far-fetched and not everything is implemented yet by us or by underlining systems but this is something that is you know literally we're talking about quarters not years uh some of some of this will be more complex by the way like I always keep saying that the the the video meetings are probably a couple of years away just fundamentally because at least for now gas and avatars require complex gpus to run it basic they run on the game engine like you know uh like a Unity or something so it just it's not feasible on the phones yet it's already feasible on the desktop but not on the phones most of us talk on the phones and whatever so that's kind of with the delaying part of that question on the back um when they hire somebody you usually have to give them credentials you have to teach them your business processes we do story points we don't do story points how do you estimate you know the tasks and stuff like that how do you train this Ai and keep the state that's actually very that's a simpler problem you know when it appears in a slack five miles later after you on board the slack just drop a bunch of PDF to him and say look read this all through this is how how we do business here and believe me it's going to be better than anybody human you ever hired because he will actually part the PDF technically and remember every goddamn point in that PDF so that's an easy one uh all of them have memory obviously so from implementation standpoint so they don't keep a longer con context uh most models by the way this is where we rely on the models most models will have you know a I forget about the termine but essentially sort of a elaborative context they're going to summarize their own context their own memor so going to keep more and more in memory over a long time so the good news we don't have to do it most of the foundation models move in this direction again remember this example six months ago it's almost impossible to do background task in LM you really have to twist yourself in the prel to to implement this today it's a one simple call in open Ai and open AI does as as a clock 6 months later mistrial deep seek and every goddamn will Implement exactly the same functionality like everybody else so in a year from now the background task which is very critical for autonomous Behavior background tasks will become a standard in llms and so on and so on and so on so all of a sudden you know I've been in the database business for 20 years and I truly believe database business is coming to an end because guess what's going to happen llms will provide a long-term storage open a I just bought the company six months ago in database business why do you think why because everybody needs a functionality I want to store couple terabytes of data in llm in my context not somewhere database like in Amazon in context of my LM that because it needs this right now it's not possible given out about 9 months to a year all of a sudden open air will come up with a database API and everybody will jump around and do the same thing it all all of a sudden in 2 years we all storing data in llm not in databases because llm needs this data most of the time so this is going to be happening and that's going to be all benefiting systems like this because for companies like humron or anybody else in the same space we don't have to do all of this we just need to provide all this social organizational integration and all of that additional add-on value but fundamentals for this agents and AI workers by extension will be by the foundation models question here um I have one so what if they unionize if I hire like a five you know no it's just joke uh so I was curious kind of about this problem of let's say loading the database into the AI model I came with a memory question just before but you kind of continued by yourself so let's say I have you know a typical company with like 20050 terabyt of data and I need to load something and that needs to end up in the llm uh how do I choose how do I load what's your guys experience in terms of like what data you provide to your agents I don't have that experience I don't think anybody loads that data in because I don't think now but let's say in the future but now how are you actually managing the context management and uh what do you think is going to happen in the future we absolutely ried on llm here we don't really do anything special uh anybody's in L chain you set up the memory in L chain you basically done now right now there is a there is a a technical limitation most of the context are pretty reasonable so what are we talking about right now what's the largest context anybody knows like gimini 2 yeah something like this yeah so it's still by database measure it's basically nothing so uh I don't believe the context actually will grow because there's a typical you know lost in the- Middle problem with large context and you don't want to grow them in terabytes just you know we don't have models for this yet but I think again my my strong expectation based on some of the internal knowledge the most of that foundational model providers will provide their own a long-term persistable persistent storage it's a just a very natural idea so instead of constantly scramming everything in context summarizing you know doing some kind of in quantization reduction whatever else just give me all the data we'll store it and we'll manage ourselves how do we get access to it how do we do rean inside of it so none of you by the way another you know interesting problem most likely the ragon is going to basically subside in usage because it's a temporary solution models don't have persistent storage therefore we have to constantly you know find the you know you know pieces of context now on the outside remember what happened to Chain of Thought two years ago was a you know pick of science right today it's basically disappearing because now all that is happening with within the model itself in a different way but you know semantically it's the same idea the same story will be happening with ragging and persistent storage it's a temporary solution as we have right now it's still going to exist because obvious not everything will be in in model and there's going to be some you know normal use cases where you need to have data outside of the model but right now reg is the only way to do right now that's definitely changing so two years from now Rec will be very unique use case when you keep your data outside of the model for some reason then yes you have to do that um most of the time you don't have to feedle around with Vector databases none of this crap you basically just hey here's a bunch of data for me to you model and here's my questions coming in do whatever you needs to be done all right sorry we don't have time for more questions we need to get to the next talk so let's thank M and hope he will be