- 28 minutes ago
The humanoid race is on, and software is a key to win. Flexion Robotics is building the full autonomy stack for humanoid robots, from command to control and manipulation to locomotion, deployable across any hardware and any task. Trained in simulation, scaled to the real world with minimal human involvement. In this conversation, Flexion Robotics and Prosus Ventures discuss what it takes to build foundational software for physical AI in Europe, why reinforcement learning is the key to scalable autonomy, and whether Europe can claim a leading position in a fast-moving field on the verge of consolidation.
Category
🤖
TechTranscript
00:07SÃ, sÃ, sÃ, sÃ.
00:21No, no, no, no.
00:24No, no, no, no, no.
00:43Good morning, everybody.
00:45Thank you so much for making time on this Thursday morning
00:49during what is a very, very busy conference.
00:52I'm Sandeep Bakshi, and it's an absolute privilege
00:55to be able to host for a fireside chat Nikita Rudin,
00:59the founder of Flexion Robotics.
01:02We're going to jump right into it.
01:04Nikita, what is Flexion Robotics?
01:08Fundamentally, we're trying to solve
01:09one of the most important problems.
01:11What will power the economy of the future?
01:13By some estimates, around a third of the world population
01:16will be over six years old by 2050.
01:19So we're talking about billions of people
01:20who are doing manual labor today,
01:23and they won't be able to do it in the future.
01:24So who will replace them?
01:26And the obvious answer is robotics.
01:30It's not just robotics, it's general-purpose robots,
01:33because we're talking about billions of people,
01:34which means they're doing millions of different tasks
01:36in different environments all around the world.
01:38It will take us way too long to develop customized solutions
01:41for every single one of those tasks.
01:43So we need general-purpose robots,
01:45robots that can go where humans go
01:46and do the work that humans do.
01:48And on top of that,
01:49they need to be powered by a general-purpose brain,
01:51and that is what Flexion Robotics is building,
01:53the brain for all of these robots.
01:56Wow, okay.
01:57Well, now I can see why everybody here came to listen to this.
02:00This is going to be really exciting.
02:02Could you tell us about you
02:04and why you decided to start this company
02:06and your background?
02:08Of course.
02:09So before starting the company,
02:10I was working at NVIDIA
02:11while also doing a PhD at ETH Zurich.
02:16At NVIDIA, we were building other simulation tools,
02:19all the fundamental tools needed to train robots.
02:21And my PhD was about using those tools
02:24to actually make robots smarter and more agile.
02:26And I personally worked on any kind of robot you can imagine,
02:31any number of legs, wheels, and arms put together,
02:35but always with the same recipe.
02:37We create a virtual world where we can put those robots,
02:40we train them using reinforcement learning.
02:43Once they're trained in simulation,
02:44we take the brain and put it on the real robot.
02:47And if you do everything right,
02:48and that's obviously a very big if,
02:50and that's why it was a whole PhD,
02:51that the real robot in the real world
02:53is still able to do the task that you trained it for.
02:57And so I guess two questions come to my mind
02:59when I hear you say that.
03:01First, you were at NVIDIA, right?
03:03One of the largest companies in the world.
03:05Why did you decide to leave
03:06and start this business outside of that?
03:09I keep asking myself the same question.
03:12I think it's because NVIDIA is an amazing company.
03:16Nothing wrong with NVIDIA.
03:17But NVIDIA's core product is not building robots.
03:21And for me, it was clear that to have the most impact
03:23and to develop this technology at the speed
03:25that we need to develop it,
03:26we need a fairly small team
03:28that is really all in skin in the game
03:30and all focused on this one single thing,
03:32building this brain for these robots.
03:35And I'm sure it was a nice tip of your cap
03:38when NVIDIA invested in your last round.
03:41So I'm sure that was really nice.
03:43Yeah, absolutely.
03:43We're still on very good terms with NVIDIA.
03:46And I guess the second question,
03:47and this topic comes up a lot,
03:49it's not lost on me that during this very warm summer here,
03:54you're building this business in Europe.
03:58Why did you not decide to relocate to the US?
04:00Why is Europe the right place for you to build this?
04:02I mean, it's a massive vision.
04:04That's another very good question.
04:06Another big debate that I had many times with founders, investors, etc.
04:12There are multiple reasons.
04:13First of all, we have amazing talent here in Europe.
04:18So we are based in Zurich.
04:21We are, I would argue, the most exciting thing happening in Zurich.
04:24So we're able to attract the absolute best
04:26out of Zurich, Switzerland, and Europe.
04:30Whereas everyone is fighting for the same talent
04:32in the US, in the Bay Area, and also in China right now.
04:37On top of that, I really think that Europe needs this technology.
04:41Europe needs not only to build the robots, the hardware,
04:46and we're slightly behind on that,
04:47but I think really big industrial groups are waking up right now,
04:50and I'm super excited about this.
04:52But on top of that, we need to build the software
04:53because fundamentally the software controls the robots.
04:56So even if you build the hardware,
04:57but the software comes from the US or China,
05:00you're not in control of your own destiny.
05:02And I think we had a very good example of that
05:04just last week with Entropic.
05:07Now, for the people in the room
05:09and the people that are watching online,
05:10could you just help understand where are we today in the market?
05:14What is possible?
05:15We see videos from, I would imagine, people in the audience
05:20and have seen a lot of videos coming from China,
05:23other parts of the world.
05:24What are we capable of actually doing today?
05:28And ultimately, how long do you think,
05:30and obviously future predictions are very difficult to make,
05:33how long do you think we are away from a robot doing complex tasks
05:39or even simple tasks?
05:41And this is the part I usually call reality check.
05:44I'm sure your Twitter, LinkedIn, whatever feeds
05:48are full of videos of robots doing all sorts of things nonstop.
05:53The reality is most of these things are demos, right?
05:57And demos are great,
05:58but demos don't show you what it took to bring the robot
06:01to that specific task, to that video.
06:04And so usually one of two things is true.
06:06Either you actually have someone teleoperating the robot,
06:09you have someone hiding behind a curtain in a motion capture suit,
06:13a VR headset controlling the robot.
06:15That was maybe what was happening until last year.
06:18Now this year we took a small step forward.
06:20Now you don't have that person anymore,
06:21but before that you had hundreds of people teleoperating these robots
06:25for that one single specific task.
06:26Meaning that even if the robot is able to do it,
06:29we're actually absolutely not solving the labor problem we started with
06:33because now we need hundreds of people just to train a few robots
06:35to perform one single task.
06:38So we are trying to do this completely differently
06:40and completely getting rid of people in the loop
06:42and training robots in simulation with reinforcement learning.
06:47So you asked where are we today and where are we going to go?
06:51I think by the end of the year we will show,
06:53and not just us, I mean the whole ecosystem will show,
06:55that robots can actually do useful tasks.
06:58We'll have robots in factories moving cardboard boxes
07:00or these infamous KLT boxes that are all standard across industry.
07:05But we will still require a few engineers looking over the shoulder of the robot
07:09making sure that everything is going well.
07:11Meaning that the economical value is not going to be there just yet.
07:14But then 2027, 2028, we get rid of the engineers
07:17and then we start scaling.
07:21And I guess one thing, when people use the term robotics,
07:25robotics could mean anything.
07:26Are you talking about humanoid robots?
07:28Like what form factor do you think is,
07:31I mean there's so many different form factors,
07:33how should we think about the various form factors
07:35that are required for maybe doing consumer specific tasks
07:39or industrial tasks?
07:40And how do you think about the software that you're building?
07:44Humanoid is obviously a word that comes back.
07:46We're talking about humanoids in a very general sense.
07:49So we started by saying that there are millions of tasks
07:52that need to be done in the world.
07:53So we cannot design a customized solution for every single one of them.
07:57And we need something that generalizes across all of that.
08:01Obviously, the exactly humanoid form factor is the easiest starting point
08:05because humans are doing those tasks today.
08:06So by definition, robots are well suited to replace them.
08:10But in many cases, you can go beyond humanoids.
08:12They can be on wheels if your ground is flat.
08:16I usually say that maybe we need something closer to a gorilla
08:19than a human with very strong arms to carry heavy weight.
08:22And then we can always imagine more and more things,
08:25more arms, more legs,
08:27robots that are better at specific subsets of tasks.
08:32But fundamentally, I think these robots need to be able to go
08:34where humans work and perform similar kinds of labor.
08:40And, you know, you referenced videos that are coming out.
08:44I mean, how do we as people that are watching
08:48that are not experts in this space,
08:50you know, how will we know when that true GPT moment has arrived, right?
08:55How will we know that, you know,
08:56a robot is actually truly doing a complex task?
08:59Is there any tips or advice you can give us to help decipher that?
09:04I think, I mean, first of all,
09:07I think the GPT moment will be a bit more spread out.
09:09It won't be one single moment where suddenly
09:10you have millions of robots around us
09:12just because we need to build those robots.
09:15You will know once you'll start seeing them around us.
09:18If we come back in a few years,
09:20I'm sure we'll have robots doing some of the work
09:23that will be done once people leave
09:25and the room needs to be cleaned up.
09:28Looking at the videos,
09:29I think you should look for a few specific things.
09:33How long horizon is the task?
09:35It's very easy to train a robot to do one specific task,
09:38grab this bottle or even remove the cap or something specific,
09:41but you'll see that the video is cut in short segments of a few seconds.
09:47Look for actually long horizon tasks,
09:52both in time and space.
09:53If we want robots approaching this room,
09:56they should really be able to go all around this room.
09:58And we'll have something to show close to that just next week.
10:03Wow.
10:04Okay.
10:05Exciting.
10:08And so peeling the onion a little bit more
10:10about kind of what Flexion is doing specific.
10:13So can you be a little more technical about,
10:16you know,
10:17when we talk about building the brain,
10:19what does that mean?
10:21So,
10:21you know,
10:21do you build your own hardware?
10:24Do you buy hardware?
10:26How does it work?
10:29So we're really focusing on the brain,
10:31meaning the software.
10:33The brain can mean many different things.
10:34People usually talk about foundational models of robotics,
10:37which again can mean very different things.
10:40We're building a whole autonomy stack
10:42going from a text prompt
10:44where you tell the robot,
10:45clean up this room,
10:46to motor control.
10:48So our models,
10:49our neural networks
10:50will send commands
10:51to the motors of the robot.
10:53But we are not building the motors
10:56and the hardware itself.
10:58That hardware is already today
11:01built by many, many companies
11:02from China to the US
11:04and now it's starting in Europe.
11:05And we're partnering with those companies.
11:07We're providing them our software,
11:09our models,
11:10and they're putting them on their robots
11:11and deploying them for their respective tasks.
11:15And why do you believe
11:16that being full stack
11:18is not the way
11:19to build the next trillion dollar robotics company?
11:24I think it's because of the scale.
11:25The scale that we need to achieve.
11:27So we're again,
11:28once again,
11:29I keep saying millions, millions, millions,
11:30but it's really what we're talking about.
11:32We're talking about
11:32building millions of different kinds of machines
11:36to perform all these manual tasks.
11:40As a startup,
11:41I'll be honest,
11:42we have no idea how to build a million machines.
11:45We need to build factories,
11:46we need to have all the supply chains,
11:48all of that.
11:49And I would claim that
11:49most very shiny US, et cetera,
11:53startups also actually don't have any idea
11:55how to build a million of anything.
11:58It's always interesting to listen to
12:00Elon Musk talk about his experience
12:01at going from one car prototype
12:03to literally millions
12:04and how hard that was.
12:06But I do think that,
12:08especially in Europe,
12:09we have already
12:09all of the industries here
12:11to build those machines.
12:14The automotive industry is a great example.
12:16There is a whole ecosystem of suppliers,
12:18and then all the car brands
12:20that we love or hate
12:21are building, once again,
12:25millions of cars.
12:26And I'm really excited about the fact
12:27that they are starting to look
12:29at building robots.
12:30I think the fastest way
12:31to get to that scale
12:32is for all of them
12:33to start building these robots
12:35and for us to provide the software.
12:37On top of that,
12:38there is a second part.
12:40To build this software,
12:41we need a lot of data.
12:43Data is always the critical part of robotics.
12:46And if you start building your own hardware,
12:48the amount of data you can collect
12:49is fundamentally limited
12:50by the amount of hardware you can produce.
12:53Whereas if you can power
12:54other companies' hardware,
12:56if you can power hardware
12:56that already exists,
12:57then you can learn much faster.
13:00Could you give the audience
13:01a flavor of, you know,
13:04the nature...
13:04There's, I presume,
13:05people here from investors,
13:08other startups, corporates.
13:10Could you give the audience
13:11a flavor of what is the nature
13:12of, you know,
13:13some of the customer types
13:14that are talking to you right now?
13:16Because presumably,
13:17there's the hardware manufacturers
13:18that are partners of yours.
13:20But, you know,
13:21what does a customer look like?
13:23And ultimately,
13:23if somebody was like,
13:24hey, what Nikita was doing
13:26sounds really interesting,
13:27but, like,
13:28why would somebody decide
13:30to get in touch with Flexion?
13:33Once again,
13:33there's a big debate
13:34in robotics right now.
13:35Who should be the first customer?
13:37Should it be all of us
13:38buying robots into our homes
13:40and making them do laundry,
13:42cleaning?
13:43That sounds really exciting.
13:44That usually resonates
13:44really well.
13:45We all want robots
13:46to clean our houses.
13:48Or should they go into industry?
13:50Industry meaning
13:52manufacturing, logistics,
13:55and, once again,
13:56anywhere humans
13:57are doing manual labor.
13:59I recently visited
14:00quite a lot of
14:02logistics centers,
14:04warehouses, etc.,
14:04and for me,
14:05it was really clear
14:06that this is the actual problem
14:07we need to solve.
14:09Humans should not
14:10and really honestly
14:11don't want to do
14:12these sort of jobs anymore.
14:13Very, very repetitive,
14:15very annoying manual labor
14:17that typically people
14:19don't do them for long.
14:20after just a few weeks
14:21they move on
14:22because those jobs
14:22are too hard.
14:26On top of that,
14:27sending,
14:28more from a technology perspective,
14:30I do think
14:31that we'll be able
14:31to solve the logistics
14:33and industrial case
14:35much faster
14:35than sending robots
14:38in anyone's apartment
14:39and letting them
14:40operate on their own.
14:41And so,
14:42I mean,
14:43one of the big debates
14:44that comes up
14:45is, of course,
14:46the labor problem,
14:48but these are,
14:48by the sounds of it,
14:50currently being done
14:50by other robots
14:52that are not as good
14:53or not as efficient
14:54or there are people
14:55like,
14:55what,
14:56I mean,
14:56just to understand
14:57a little bit better,
14:58these are,
14:59like,
14:59in these logistics centers
15:00just to,
15:00you know,
15:01is this picking
15:01and packing of parcels
15:02or,
15:03yeah.
15:03Talking about picking,
15:04packing,
15:06material handling,
15:07so moving things
15:07from point A to point B.
15:09We're not really
15:10trying to automate
15:11or re-automate
15:12things that are
15:12already automated,
15:13so if you think
15:14of a car factory,
15:15you have those big
15:16robotic arms,
15:18welding cars,
15:18painting cars,
15:19et cetera,
15:19this will not be done
15:20by humanoids
15:21or mobile manipulators.
15:22We're talking about
15:23everything else,
15:25about the thousands
15:26of additional things
15:27that need to be done
15:28to bring the parts
15:29to those robots,
15:30to put parts
15:33inside of the car
15:34where these big
15:35robotic arms
15:36cannot go,
15:36et cetera.
15:37Could we spend
15:38a little bit of time
15:39talking about data?
15:40So, you know,
15:41ultimately,
15:42I remember when
15:43the LLMs came out,
15:45they were all trained
15:46on the corpus of data
15:47that was available
15:48online, right,
15:49through public sources.
15:51A lot of that data
15:52doesn't exist
15:53in the robotic space.
15:55You mentioned a term
15:56earlier,
15:56which maybe you can
15:57just elaborate on.
15:58You mentioned the term
15:59teleoperated, right,
16:00when we were talking
16:01about what other
16:01companies do.
16:02Can you just talk about,
16:04you know,
16:04in a world where
16:06ultimately,
16:07if a robot was to
16:09pick up this bottle
16:09of water,
16:10unscrew the cap,
16:11maybe it needs to see
16:12me or 10,000 versions
16:14of people doing that
16:15over and over and over
16:16again before it learns
16:18how to do that.
16:19I mean,
16:19how should we think
16:20about data gathering
16:22and how is that a problem
16:23that you're thinking about?
16:26So, I think everyone agrees
16:28that it's rare
16:29that everyone agrees
16:29in robotics,
16:30but everyone agrees
16:31then that data
16:32is the bottleneck.
16:33We need more data.
16:35But then we all
16:37disagree on
16:37what to do
16:38from that statement.
16:41Teleoperation
16:41is the simplest way
16:43to collect data.
16:45It's extremely expensive,
16:46extremely slow,
16:47but it's really obvious
16:48how to transfer that data,
16:51how to train
16:51a neural network
16:52and make the robot
16:52imitate that.
16:53And to be specific,
16:54teleoperation
16:54means exactly what you said,
16:56somebody putting on a suit
16:57and then that being recorded
16:59and then, okay,
16:59so these are people
17:00doing this.
17:01Yes.
17:01It's multiple people
17:02per robot
17:03because someone is handling
17:04the robot itself,
17:05someone is teleoperating
17:06the robot.
17:06It's an extremely
17:07inefficient process.
17:08If you have billions
17:10of dollars to spend,
17:11it's a logical starting point,
17:13but I think it's fundamentally flawed
17:15because it's so expensive,
17:16but also because the data
17:17you get out of it
17:18is actually very low quality.
17:19Humans are really bad.
17:20If anyone has ever tried
17:22to play a game in VR,
17:24it's actually very hard
17:24to do anything
17:25because you don't have
17:26any force feedback.
17:27So, the robot is operating
17:28on physical objects,
17:30but the human standing
17:31behind the robot
17:31doesn't feel anything.
17:33So, you end up
17:33with very, very clumsy behaviors.
17:35So, you can spend money,
17:37you can collect a lot of this,
17:38but you have very low quality data.
17:42So, from there,
17:43people created more ways
17:45to collect data
17:45that are always involve humans.
17:48We can make it slightly more efficient,
17:50but fundamentally,
17:52what we saw also
17:53in large language models
17:54is that it's not just
17:56about collecting massive amounts
17:57of data
17:58and training models.
18:00If we want to get
18:01to the next stage,
18:02which is make these models
18:03do something useful,
18:05so going from GPT-3
18:07to chat GPT
18:08and then to coding agents
18:09with reasoning,
18:11the key unlock
18:11was reinforcement learning.
18:13And reinforcement learning
18:14completely changes
18:15how you look at data
18:16because robots
18:18are collecting data
18:18themselves.
18:20So, you don't have
18:21a fixed data set.
18:22You just put a robot
18:23and it just keeps
18:23collecting data.
18:24So, you have this
18:25dynamic data set
18:26that keeps being created
18:30by the robot itself.
18:31On top of that,
18:32I'll add one more note.
18:35Doing reinforcement learning
18:36in the real world
18:37is, unfortunately,
18:38once again,
18:39very expensive
18:39and very inefficient.
18:41These robots,
18:42these models,
18:43need millions
18:43of trial and error steps
18:45to figure something out,
18:47which means that
18:47you'll have robots
18:48falling thousands of times.
18:52Whereas,
18:52if you can do it
18:53in simulation,
18:54then all of that
18:54comes for free.
18:55Basically,
18:56if you do reinforcement
18:56learning and simulation,
18:58then you have
18:58infinite amount of data
19:00for free.
19:02Wow.
19:02Okay.
19:03And, I mean,
19:04putting numbers
19:05onto this,
19:06you know,
19:06how much would it cost
19:07for, I mean,
19:09you can pick
19:10whatever task
19:10that you'd want,
19:11but how much would it cost
19:12to have a robot
19:13properly,
19:14with tele-operated data,
19:16you know,
19:16pick up a bottle of water?
19:18And how much would that cost
19:19in the middle ground
19:22where it's, you know,
19:23perhaps not tele-operated,
19:25it's a robot doing it
19:26with, you know,
19:27reinforcement learning
19:27versus the way
19:29that you're articulating
19:30purely simulating.
19:31I mean,
19:31you can pick whatever task
19:32that comes top of mind.
19:33It doesn't need to be
19:33a bottle of water.
19:34Yeah, I think picking
19:34this bottle is a good example.
19:38If you fix everything,
19:40everyone stops moving,
19:41and we collect 50 to 100
19:44demonstrations of the robot
19:45picking up this bottle,
19:47it will be more or less
19:48able to do it,
19:49once again,
19:50very clumsily
19:50because it's very hard
19:51to tele-operate that
19:52without being the robot itself.
19:54But then if you change anything,
19:56if you change the lights,
19:56if people start moving around,
19:58if the background changes,
20:00or if we move the bottle,
20:01it doesn't work anymore.
20:03And so you need this super,
20:06how do you say that?
20:07It's a combinatorial problem.
20:08You need all possible
20:09variants of anything.
20:11And so what tends to happen now
20:12is that some of our competitors
20:14who are betting a lot
20:15on tele-operation
20:16are literally renting Airbnbs
20:17around the world,
20:18bringing the robot there
20:19to collect a little bit of data,
20:21then going to the next Airbnb,
20:22just to have more diversity of data.
20:26Once again,
20:27I'll bring back simulations.
20:29In simulation,
20:29we would create a table,
20:31we would create bottles,
20:32and then we can randomize
20:34absolutely everything
20:35that there is to randomize
20:36about this environment.
20:38And we create thousands
20:40of different versions of this,
20:41which means,
20:42without having to go
20:43and rent thousands of Airbnbs
20:44to have all the different backgrounds
20:47behind this table.
20:49Okay, changing gears a little,
20:51one of the things
20:51that's very topical at the moment
20:53is sovereign AI, right?
20:54Based on some of the things
20:55that have happened
20:57across the world
20:58over the past week or two,
21:00I mean,
21:01when we're talking about robots,
21:03I mean,
21:03who should own the,
21:04how do you think about
21:05who should own
21:05the intelligence behind them?
21:07The hardware aside,
21:08those are,
21:09you know,
21:10that's not the brain.
21:11But how do you think about
21:12who should own the brain,
21:13and where do you think
21:15that, you know,
21:15Europe really has an opportunity
21:17to come out ahead here?
21:19And we touched on this
21:20a little bit before.
21:21I think Europe has really
21:22a big opportunity
21:23on actually on both.
21:25We tend to think
21:27that anyway,
21:27the hardware will come from China
21:28and we have to accept it.
21:30But I haven't quite accepted it
21:32just yet.
21:33I really think
21:34that we have a big role
21:36to play on hardware as well.
21:39And I don't think,
21:40as I said before,
21:41I don't think this will come
21:42from small startups
21:43being created today.
21:44I think that five years from now,
21:46much older,
21:47more traditional companies
21:48will be building
21:48way more robots
21:49than any startup.
21:52Then software,
21:54definitely,
21:55it has to be sovereign.
21:57Once again,
21:58we saw all the problems,
22:01we have all the discussions
22:01around AI,
22:03sovereignty for AI.
22:05I think with robots,
22:06it's even more critical
22:07because we're talking
22:09about machines
22:09that can,
22:11move in our world
22:12and can operate
22:13on the world.
22:13And if they're controlled
22:15by someone else
22:15and you don't know
22:16exactly what they're programmed
22:18to do,
22:18then basically anything
22:19can happen.
22:22So,
22:24we've got probably
22:25about a minute left.
22:27What is the one thing
22:28that you'd want people
22:29to leave knowing
22:30about you,
22:31about Flexion,
22:32and what are some
22:33of the things
22:33that you think
22:34people should look out for?
22:35What are some of the things
22:35that are coming
22:37down the pipe?
22:41We will have
22:42a release next week,
22:43so we talk about
22:44what to look out
22:44for a video.
22:45Our release,
22:45once again,
22:46will be a video,
22:46so you have to look at it
22:48with the grain of salt
22:50we mentioned before.
22:51But you will see
22:52for the first time
22:53actually a long horizon task,
22:56a useful task
22:57that really uses
22:58the capabilities
22:58of a humanoid robot.
23:00And I think
23:01what people should
23:03take away from that
23:04is that this is coming.
23:06It's really happening.
23:07We're at a point
23:08where we can deploy robots
23:09and make them
23:10do useful tasks.
23:11And specifically
23:12at Flexion,
23:13we're doing that
23:13without employing
23:14hundreds of people
23:15in the background,
23:16hiding them behind curtains.
23:17We're actually doing this
23:18in a viable way
23:20that can scale
23:20in the very near future.
23:23And so industrial
23:24is the near-term focus,
23:26so I have to continue
23:27to fold my own laundry.
23:30For, I'd say,
23:31for one more year.
23:32Ladies and gentlemen,
23:33Nikita Rudin,
23:34please thank you
23:35so much for making the time.