- 18 hours ago
There's a prevailing narrative that AI’s trajectory is set — the techniques are proven, the finish line in sight. Peter DeSantis joins Nick Thompson to challenge that framing. We’re at the beginning of a massive transformation, not the end of a short one, and the breakthroughs that will define AI haven't happened yet. This session digs into what it will actually take to get there: the silicon, the systems, and the economics needed to make AI both capable and affordable enough to unlock applications and businesses that don’t yet exist today. And why the people who see this moment as a starting line will be the ones who shape what comes next.
Category
🤖
TechTranscript
00:03Peter, that was a lovely introduction. I like that a lot. That was good. Very short, brief, to the point.
00:07I like your belief that we're at the starting line in AI. It's good.
00:11Thank you. Yeah.
00:13Today is Tuesday. It feels like we're at a starting line where the officials are starting to throw out some
00:17of the contestants, or at least call fouls.
00:21Can I start with the big news of the week, and then I want to get into quantum. The big
00:26news of the week is that last Friday after the markets closed, the government called and they told Anthropic,
00:31Model Fable, we're not so psyched about it. Anthropic had to pull the model back. This happened, as reported in
00:39the Wall Street Journal, because Amazon had found a security flaw, told the US government, and the US government called
00:44Anthropic.
00:45You are a smart enough man to know that you should not comment directly on a story that involves your
00:50boss and the President of the United States.
00:52But can you tell me how Amazon does its security testing and what threats it looks for in a situation
01:00like this?
01:01Well, Nick, I appreciate the question, and the lead up to what I can tell you is the first rule
01:08of security is you don't talk about how you do testing, because then you're giving away your secrets.
01:13Look, I can say this. We, since the very early days of AWS, and to this day, Amazon places the
01:22security of our customers absolutely at the top of the list of things we do.
01:26Everything starts and ends with security, and so we invest a large amount in making sure that our systems are
01:33secure, helping our customers and partners be secure.
01:36And, you know, as a result of those investments over a long period of time, we have a best-in
01:41-the-world security team who finds all sorts of interesting things.
01:45So, the world runs on AWS. The Atlantic runs on AWS. So, if you were theoretically to find a model
01:51that could, you know, hack through into systems like this, it would be terrible for your customers.
01:56Look, I'll talk about something less salacious, but, you know, security bugs have existed for the better, you know, since
02:03the dawn of the computing age.
02:06You know, our customers have run on AWS and come to trust us to be ahead of security problems, and
02:12we have engaged with all of our customers, you know, privately and helping them be more secure.
02:19And so, you know, in the old days, you'd call this, in the old days of systems, before headlines, you
02:24know, you'd talk about zero-day exploits.
02:26You know, some piece of critical software had some bug that somebody found, and it needed to get patched before,
02:32you know.
02:33And so, look, I think the models and how people consume models and how we deploy models is newer, and
02:40we're all working through this as an industry.
02:41But, like, this problem is, it seems like it's brand new, but we've been dealing with critical security patching and
02:47vulnerability quite gracefully for the better part of 20 years.
02:52And I think all the folks that are invested in making AI succeed care a ton about security, and I
02:59think we're going to work out the right way to do this as an industry.
03:01So, you know, the funny thing about this story is, of course, that Amazon owns a big share of anthropics,
03:06so you have a massive conflict of interest in which you acted against your interests.
03:12I guess that should make us trust you.
03:15Let's talk on, let's go to the full subject of this panel, which is where AI is going in your
03:19view that it's the starting line.
03:21Do you believe that when we get through this and we get, you know, presumably Fable will come back, other
03:26frontier models will come back, God willing, we'll figure out how they can't hack all security.
03:30When you say that you think AI is only going to accelerate, what exactly do you mean?
03:34Do you mean that we're going to get a new architecture for AI?
03:36Do you mean that Fable 6 is going to be as good to, you know, 5 as 5 was to
03:414?
03:42What do you mean by that?
03:43I think I mean all of the above.
03:45I think, you know, right now I think we are all sort of awestruck by how each subset, you know,
03:52each next generation of models is seemingly doing so much more, solving such more difficult problems.
03:58And I think that's going to accelerate.
04:00Like I think, and I think we've seen that over the last couple of years.
04:02I think the models and the agents that use those models are going to be able to do more and
04:07more for us.
04:08But the other part of the equation that we don't spend enough time talking about is in order for that
04:12to be useful, we've got to make these models more efficient, more cost effective, more power efficient.
04:17And that's going to be innovation in the models, in the infrastructure.
04:22And I think we've seen an order of magnitude improvement in efficiency of the tasks that models can do over
04:28the last year or two.
04:30And I think we need a couple more orders of magnitude before this gets interesting in any way, shape, or
04:34form.
04:34And so I think we are just at the beginning of the innovation at all layers of the stack.
04:39Well, let's move down the stack, but let's stay at the model level for a second and then let's go
04:43to the chip layer after that and talk about some of the efficiencies there.
04:47Do you think that future models will use the architecture we have right now, right?
04:53They will have more training data, they will have figured out better training algorithms, they will have smarter post-training,
04:59but they will still use transformer architecture.
05:02They will still work kind of like the models we have right now, where it's an improvement, but it's on
05:06a curve you can understand.
05:07Or do you think they will be fundamentally different?
05:10Both.
05:11I think transformers are a very durable architecture.
05:14I think there's a lot of innovation and research.
05:17We're doing it.
05:18I'm certain the big labs are doing it in terms of innovations to the transformer architecture to make it more
05:23efficient, to make it more capable, to make it safer, to make it easier to understand what's going on.
05:30There's a lot of innovation going on in transformers.
05:33And I also think transformers are just part of the story.
05:35You guys have heard from folks this week, and you'll hear more from folks this week, on new forms of
05:42models, world models.
05:43I'm particularly interested in how models are going to evolve to allow humans to interact with them better.
05:51I think we're all very comfortable chatting with models with text, but as humans we have many ways of communicating,
05:58and I think we're going to find AI architectures that are much more fluent in being human and allow us
06:04to communicate with computers in a much different way than we do today.
06:07And that will require new architectures.
06:09Okay.
06:10How are we going to do it?
06:11I mean, right now we can talk to them and that's translated into text, right?
06:15What are we going to...
06:16So today you can kind of do speech, you can kind of do vision, you can kind of do text,
06:20but they're kind of discrete functionalities.
06:22And I think you're starting to see evidence, you're starting to see this transformation happen where, you know, as we're
06:29able to improve the performance of the vision detection and the vision synthesis,
06:34as speech becomes faster and bimodal and more full duplex, the way we humans like to talk where we back
06:42channel and we nod and we move.
06:43I think all those modalities are going to come to how we interact with computers.
06:46But to do that, they need to respond at a, you know, at a 40 millisecond clock, not a...
06:51Okay.
06:51And what are we going to be talking to, right?
06:53So I'm talking to my AI bot.
06:54Am I talking to my phone?
06:55Am I talking to a head that's nodding?
06:58I mean, what am I getting...
07:00How it productizes is an entirely different...
07:03I know, but you have the Amazon lab where probably you're testing these things out.
07:07I think there's a lot of unknowns here.
07:08But look, I think there's a lot of...
07:10A lot of people are speculating that you're going to have a device that maybe is able to observe and
07:14listen and you're going to be able...
07:16Hopefully be able to turn that off when you don't want it observing and listening.
07:19But you're going to be able to have that sort of ambient listening in the room.
07:22And that's going to give it so much more context.
07:24But that's not simply going to feed words into a transformer was my point.
07:28Right.
07:28That's going to...
07:29Those inputs are going to have to influence the models.
07:32It's going to take your words and it's going to understand more about your motion.
07:35It's going to understand your nods.
07:36And it's going to like see my hand gestures, which means something or other, even though they mostly just confuse
07:40people.
07:41I find hand gestures clarifying.
07:45All right.
07:45But I'm Italian.
07:47I feel obligated now to use like super aggressive hand gestures for the rest of this interview.
07:51So, as you go down the stack and you get to the chip layer, explain to me where the most
07:59interesting innovations are going to come in the chip layer.
08:01Because what's so...
08:02One of the things that's so amazing about AI and where we're going is that these things are going to
08:07have to be in sync.
08:08And it's very hard when things move incredibly fast to get things in sync.
08:12So, explain the way that you think about innovation at chips at Amazon.
08:15Yeah.
08:16You're absolutely right about...
08:18I think what's fascinating is model development and chip development, they're very different disciplines, but they do share some key
08:27attributes.
08:28And one of them is they're kind of long lead time, high capital investment endeavors.
08:33If you're going to build a frontier scaled model or if you're going to build a chip, you're making an
08:39investment of hundreds of millions of dollars or more to bring that.
08:42And you're deciding what you want to do a year in advance or more of when you're actually bringing it
08:48into production.
08:49And so, if you're doing those two things in a vacuum of what's happening in the other, you're going to
08:57have so much latency.
08:59If the chips are not telling the model designers what capabilities are coming and where they can optimize the models
09:06to take advantage of those capabilities,
09:07then we're not doing the science that's necessary to take advantage of those until the chips are available.
09:13And then you're waiting, you know, months and months and months, and so now you're on to the next chip.
09:16And vice versa, if the models aren't telling the chip designers what capabilities they need to bring the next generation
09:23of AI, then you've missed the bus on the chip and you have to wait for the next tape out
09:28a year later.
09:29And so, I think what's, you know, we've been doing this with our partners like Anthropic from day one where
09:35we collaborate deeply, tell them where the chips are going, they tell us where they need us to go on
09:39the model side.
09:40And, you know, part of bringing models and chips together with me at Amazon is letting our internal teams have
09:46the same sort of influence on our chip roadmap and vice versa.
09:49So, you have two incredibly complicated forecasting problems, or at least two, one of which is, like, what will the
09:54demand be, right?
09:55And Dario has pointed this out, he said, like, when you're growing, you know, 10x a month, it's incredibly hard
10:01to forecast your demand in a year or two years, three years, right?
10:04It's impossible.
10:05So, you have that problem, and then you have the second forecasting problem, which is you actually don't know what
10:09AI is going to be, right?
10:10And if AI is supposed to power some kind of a device that's understanding my hand gestures, it's going to
10:16need some kind of a different chip than if it's just requiring text.
10:19So, why don't you talk through problem number one?
10:22How do you forecast demand when something is growing so fast, right?
10:26Because you have to make the chips for Anthropic.
10:29Yeah.
10:29Well, so, first of all, Dario is absolutely right.
10:32This is, I mean, forecasting demand in a capital-intensive business, or really any business is, and these businesses are
10:39growing in a way that nobody's really seen before.
10:43I mean, we spent a lot of time on capacity management 20 years ago when we launched the cloud trying
10:47to predict what customers needed with general purpose compute.
10:49And, you know, all those sort of challenges are 10x in today's AI world.
10:54So, that's a real problem, and there's no magic bullet for it.
10:58You know, I think we have a very deep operational and supply chain expertise that allows us to do it
11:05very well, but there's no magic bullet for predicting what demand's going to look like six months from now or
11:09two years from now.
11:10I think everybody grapples with that problem.
11:12The interesting thing, and going back to your earlier question, is, you know, in a world like today, there's this
11:18tension between building something that's general purpose and large, like our Tranium processor, like the large NVIDIA platforms that many
11:27people use.
11:27Those platforms are built to run basically any AI workload that you can imagine.
11:32Inference training, and inference itself is multiple workloads, different model architectures.
11:38They have a large amount of memory, they have a large amount of memory bandwidth, and they have a large
11:41amount of compute all packed into a very tight area.
11:44And that means that the model designers and the inference optimizers can do all sorts of great optimizations.
11:52The problem with that is almost every workload is wasting something. It's wasting memory, it's wasting memory bandwidth, it's wasting
11:58compute.
11:59And so, the temptation is to specialize to the workload, create an inference optimized platform.
12:06Now, the problem there is you've made your forecasting problem all the more worse, because instead of forecasting a big
12:12pool of demand, you're forecasting a narrower pool of demand.
12:15And so, now you've got to know what that architecture looks like, what the specific needs are.
12:19But if you do it, it's very tempting, because maybe you can save 40% of the power and 40
12:23% of the cost on the platform.
12:25And so, I think we're starting to get to the place.
12:28So, to this point in time, largely, we've all produced general purpose AI platforms, because we've had a large cone
12:37of uncertainty over what the workloads are going to look like and what the demand is going to be.
12:41That's not going away, but I think we have enough specificity now that you're going to start to see specialization
12:47of hardware.
12:48Wait, so I understand this. So, Tranium is like a general chip where you optimize for all those things.
12:52It's like as optimized for power management as it is for inference?
12:55It's as optimized for compute-intensive workloads as it is for memory-intensive workloads as it is for memory bandwidth
13:02-intensive workloads.
13:03So, they can all run on that.
13:04And so, that competes with NVIDIA, and then there are other companies like Grok, which optimizes for inference,
13:08and then you end up getting bought by NVIDIA to improve their inference area.
13:11So, what you're arguing is that, yes, you will still build out Tranium chips that are optimized for everything,
13:17but that you're going to try to bet where the market is going and build the next generation of chips
13:22that optimize for the variable that's most important?
13:24Yeah. And Grok and Cerebrus are two good early examples of specialization.
13:30Those chips are not as general purpose.
13:33They provide a very, they have a type of memory called SRAM,
13:37which is instead of having memory that's put right beside the computing core co-packaged with the chip,
13:44that that memory is actually commingled on the same process node as the compute.
13:50And so, what that gives you is 10x faster memory, but it gives you much smaller capacity.
13:55So, it means if your workload needs a lot of memory, it's not going to fit well.
13:59But if you need a lot of memory access, if you need to do things very quickly with memory, it's
14:03going to be 10x better.
14:05Right. And so, when you look at inference, inference is really two workloads.
14:09The first one is what people call encoding or pre-filling, and that's where you take the prompt that you
14:15send to the model
14:16and convert it into a bunch of matrices.
14:19Yep.
14:19The second part of it is token gen, and this is the autoregressive sort of token by token workload.
14:24And that part of the process is not very compute intensive.
14:28The first part uses a ton of compute.
14:30The second part uses a ton of memory bandwidth.
14:33It needs to access all the model weights on every token.
14:36And so, if you put that on an SRAM chip, like Grok, like Cerebris, you can get really good performance.
14:42With a big caveat, if your context is too long, you're spending too much time loading the context in and
14:48out of that chip.
14:48And so, these SRAM chips are a great example of if you specialize a portion of the workload, you can
14:55get great efficiency and great performance benefits.
14:58But now you've got to deal with the complexity of that specialization in your workload.
15:02And so, the main observation I'm saying is the industry is hitting a point with AI where we have enough
15:08certainty about the shape of the workload
15:11that you're going to start seeing increasing bets in specialization.
15:14Not to the exclusion of future generations of Tranium being the general purpose workhorse.
15:19But I think you're going to see, you're going to certainly see more specialization from us.
15:23And I think across the industry, we're going to see more specialization.
15:25What is the variable that you're going to optimize for as you specialize?
15:28Like which of the five different things is going to be most important?
15:31Most of the interesting thinking right now is how to find a sweet spot for better making use of memory.
15:39And so, one of those approaches is SRAM.
15:41And that's where you actually print the memory on the same fabrication wafer, like you guys saw in the last,
15:49as you do the compute.
15:50And that makes the compute and memory super close together, super low latency, but you can only put so much
15:58compute on the wafer.
15:59So, your hands are tied in terms of how much memory.
16:03So, the alternative classic chips have things called HBMs.
16:06They sit right next to the compute die with tons of wires between them.
16:10What's getting really interesting is can we package the compute and memory side by side vertically, 3D packaging.
16:17And we think that might be the best of both worlds.
16:19You can use more DRAM, but you connect it to the chip in a much higher bandwidth way.
16:25And so, I think that the way that compute and memory are commingled is going to be an area you're
16:30going to see a lot of innovation.
16:32So interesting.
16:32You know, sometimes people say these tech conferences, everybody talks about the same thing.
16:35I bet you nobody else has been talking about DRAM in such an interesting way today, because this is amazing.
16:40Let me ask you a question about power, right?
16:48So, a lot of environmental concerns with power, right?
16:51How will we ever be able to improve the efficiency of these chips whereby the total amount of energy used
16:57by all the chips goes down?
16:59Right?
16:59Will the improvement in power efficiency ever exceed the increase in demand?
17:04Absolutely.
17:05I mean...
17:06Because it just gets worse and worse.
17:07I mean, even though the power gets better and better, the demand is going up so much faster that the
17:11total energy consumption by all the AI chips goes up.
17:14Yeah.
17:15Well, as I said, I think we're at the beginning of the optimization phase of this.
17:19And so, you know, I think we've seen several orders of magnitude improvement in efficiency.
17:23And again, what do we care about here?
17:28It's watts of power per useful bit of intelligent or work done.
17:33Right.
17:33That's kind of the North Star of what we want to do.
17:35Like tokens per watt, right?
17:36Yeah, but tokens are kind of a weird thing because a token is the thing we observe, but a model
17:42can spend more or less compute generating a token.
17:44Okay.
17:45And so you could end up with...
17:47And we've all seen this, actually, with models.
17:49Some of them ramble when they reason, and some of them are more concise.
17:53And the ones that are more concise typically spend more time computing before they put out a token.
17:59And so a token is not a true efficiency metric, but intelligence is.
18:05Right.
18:05And that's gone down by an order of magnitude, and I suspect we're going to see several more orders of
18:09magnitude over the next couple of years,
18:10both because of the hardware innovations, but also the model innovations.
18:13And frankly, the applications.
18:15And everybody in the room's probably had the experience that we can use the models more efficiently.
18:21And so we're so early in this optimization cycle.
18:23Are we at a moment of...
18:26When you think about the next innovation, and you're sitting there, and you're in your lab, and you're like, God,
18:29how do we make the memory more efficient?
18:32To what degree is AI helping you make that decision?
18:35Because if AI is helping you make that decision, then it's helping you make chips that are more efficient that
18:40then makes the AI more efficient,
18:41which then helps you make the next decision, right?
18:42So presumably you could have an accelerating curve in how efficient you make your chips, even though there are physical
18:47limitations,
18:48because you do need silicon, you do need wires, you do need power.
18:51Yeah.
18:51Well, this is like, what is it, recursive model development, I think is what the...
18:55It is, but it's not quite, because it's not just, it's not simple software, because there is a hardware layer,
19:01so it can't go as fast as recursive improvement in just software.
19:03No.
19:04But I am curious, like, you know, I say, hey, Peter, give me a brilliant idea for how to make
19:10memory better.
19:11Do you go and talk to five of your people, or do you go to the model of your choice?
19:17Good question.
19:17I mean, I think, I believe humans are still going to be at the center of some of our most
19:23complex innovations for the foreseeable future.
19:26That said, if you're designing a chip, and you have a great idea, what you typically do is you code
19:31it up just like you would in any other sort of software process, and you simulate it.
19:35And so, you take the workloads that you know you're accelerating, and you run them on a simulator, because that's
19:40how you prove that the thing you think is going to work, works.
19:42Yeah.
19:43So, just like on software development, if your models make that process of exploring the solution space and building those
19:51simulators in order of magnitude cheaper,
19:53you can explore a lot more solution space, and you're going to find better chips.
19:57So, theoretically, each time that you improve a chip, which then improves a model, you also improve the simulations and
20:02the thought process,
20:03so you could accelerate the improvement in chip development.
20:05I think there's, just like we're seeing software development speed up because of AI, we're going to see model development
20:13and chip development speed up as well.
20:15And those two things are going to help us, you know, it's going to start spinning a flywheel on better
20:21models, better chips, lower cost, better efficiency.
20:23And Amazon, do you have any chip design processes that have no humans in the loop?
20:28Do you have, like, Peter has his team, and this other dude just set up a whole AI operation that's
20:32on a loop?
20:32We're investing deeply in this area, but there's nothing we're doing that doesn't have extremely talented humans in the loop.
20:39All right, well, let me know when you get rid of them all and just move to AI models.
20:42Let me ask you about another tiny thing, which is quantum computing.
20:44So, you have, you are building quantum computers at Amazon, are you not?
20:48We are.
20:49What are they going to do when they work?
20:50Which I know will happen in seven years, because it's always seven years.
20:53Seven years.
20:54Yeah, that's what they are.
20:55Well, quantum computers are coming.
20:56You know, I think there's a lot of speculation about what they'll be useful for.
21:00My mental model is quantum computers are going to do the thing that Richard Feynman first posited we should build
21:07quantum computers for.
21:08Hack encryption?
21:09Which is accurately simulate the physical world.
21:12Yeah.
21:12So, chemistry, material science, like, these are the fields that are most, like, a quantum computer isn't going to run
21:20our general purpose applications faster.
21:23There's a possibility we discover a bunch of classic computer science algorithms that we can port to quantum computing and
21:30find better solutions faster.
21:32There's a prime number one we can talk about that everybody's worried about that might break encryption eventually.
21:37I don't think that's going to happen anytime soon.
21:39But what is going to happen soon is we're going to get better material science, we're going to get better
21:43chemistry, and that can profoundly impact.
21:45Make it as specific as you can, because this is a question I've asked a lot of people over the
21:49years, and I've never quite gotten as specific as I'd like.
21:52Chemistry, modeling, life sciences, yes.
21:54Give me an experiment, an exact experiment, that will be useful to the world that you think will be able
21:59to do faster with a quantum computer.
22:01We spend a bunch of time.
22:03One of the smaller molecules we deal with is nitrogen.
22:07In order to produce fertilizer, we need a lot of nitrogen.
22:11The processes we currently use to produce nitrogen require a lot of CO2 emissions.
22:17If we can find better chemical processes for fixing nitrogen, we can produce more fertilizer with less CO2.
22:25That will have...
22:26So, that's the outcome, but the experiment would be how a model...
22:30Nitrogen, because it's one of the simplest chemical reactions, and we still can't fully simulate those processes in a classic
22:35computer.
22:35So, with a functioning quantum computer, we could simulate the way that we generate nitrogen while releasing your harmful greenhouse
22:43gases.
22:43With a moderately sized quantum computer, we could simulate the chemical and quantum processes involved with nitrogen, and the molecules
22:54that we care about related to nitrogen, far better than we can in a classic computer today.
22:59So, okay, that would be awesome.
23:00I'm all for that.
23:01I'm all for nitrogen, because it'll make the plants grow, and I'm all for making it more energy efficient, because
23:04then it means we won't all die.
23:05What is the thing that needs to happen to get there?
23:09Like, what is the problem you are trying to solve in quantum computers?
23:12Do you need to make the qubits more stable?
23:13Do you need to make the machine bigger?
23:15Like, what is it that you have to do?
23:16This is the fine print.
23:17So, if you kind of like...
23:18In the last seven years, everybody was reading about, like, the quantum computer with 100 qubits, and then the quantum
23:22computer with 1,000 qubits, and then...
23:24But the fine print on all those qubits is, for a little tiny fraction of a second, this qubit will
23:31hold the state, and then it will probably have noise.
23:34And so, the interesting thing about quantum computers is, you have to store the state consistently long enough to do
23:40something useful.
23:41And the challenge in a quantum computer is, there's...
23:45The reason why we, like, for example, take them down to really low temperatures when we build them on superconductors,
23:51is, like, little bits of noise can blow up the qubit.
23:55And so, you have to insulate this thing from noise, but...
23:58And you can't look at it.
23:59Why does the cold temperature insulate it from noise?
24:02Well, if you get close to absolute zero, there's less noise.
24:05So, we actually run...
24:06There's a few different ways to build a quantum computer.
24:08Our preferred choice is to build them on semiconductor technologies, not with semiconductors, but with superconducting qubits.
24:14If you take those down to near absolute zero, you get rid of a lot of the background radiation that
24:19can add noise to your bits.
24:20And so, the unsolved problem with quantum computers is, how do you keep them coherent with the right information long
24:26enough?
24:27And the real interesting challenge, and I'm not a physicist, but the thing that really gets you with quantum computers
24:32is,
24:33you can't actually look at the information in the quantum qubit, or you destroy it.
24:36Like, as soon as you look at it, the game's over.
24:39So, you have to preserve the information without looking at it, without reading the information.
24:42So, in the classical, we do error correction, we read it.
24:46In the quantum world, noise is a bigger problem, and we can't read it.
24:49And so, you have to innovate about how you create error-corrected qubits.
24:52That's what we've been doing the last five years.
24:54We actually published a paper in Nature 18 months ago about our approach.
24:58I think you're going to increasingly see the quantum computing world talking about error correction.
25:03Our approach has been, let's get the error correction right, and then let's scale up the qubits.
25:08And we're just starting that process of scaling up.
25:10All right.
25:11So, to summarize, since we have a few seconds left, you have a giant machine at absolute zero using superconducting
25:17material.
25:19And if you look at it, nothing works.
25:22All right.
25:23That's how we're going to solve climate change, everybody.
25:24Very excited.
25:25I love the optimism.
25:27Thank you so much for having me, joining me on stage.
Comments