Is Manus AI the Next DeepSeek?
Key Points
- The panel debated whether Manus AI represents a “second DeepSeek moment,” with mixed opinions ranging from cautious optimism to outright skepticism.
- Vyoma Gajjar highlighted the bullish case, noting Manus AI’s multi‑purpose agent could industrialize intelligence by leveraging large‑language‑model advances and potentially outpace many emerging agentic startups if hardware and compute align.
- Kaoutar El Maghraoui expressed doubt, pointing out that numerous competing frameworks and rapid catch‑up by other firms could limit Manus AI’s long‑term impact.
- Host Tim Hwang framed the discussion within broader AI trends—such as scaling laws, vibe coding, and new products like Perplexity’s phone—while emphasizing the significance of agentic AI as a growing focus in the industry.
Sections
- Debating Manus AI's DeepSeek Potential - Experts debate whether Chinese startup Manus AI heralds a new DeepSeek-like breakthrough as they preview its versatile agent’s wide‑range capabilities.
- Skepticism Over Manus AI Impact - The speaker doubts whether Manus will deliver a genuine breakthrough in AI autonomy or merely represent a rebranded incremental step, emphasizing the need for thorough evaluation amid intense competition between Western and Chinese AI firms.
- Integrating Open-Source AI Orchestration - The speaker describes a platform that combines Claude’s orchestration with fine‑tuned Qwen models, sandboxed tool‑calling, and a cohesive UI—highlighting that while it adds value beyond a simple wrapper, the open‑source community could feasibly replicate the setup.
- Predicting the Future of Autonomous Agents - Panelists examine the gap between POCs and enterprise integration, regulatory hurdles, and forecast how open‑source initiatives and projects like Manus will drive autonomous agent technology over the next six months.
- Agentic AI Demo Debate - The speakers discuss the popularity of flashy, agentic AI demos as a perceived milestone toward AGI, critique current browser‑based interfaces (e.g., cursors and screenshots), and touch on experimental efforts to replicate existing agents like Manus using MCP.
- Debating AI's Role in Coding - The speaker argues that although AI assistance can aid small projects, deep understanding of fundamental coding concepts remains essential for interviews, large‑scale development, and sustainable engineering practice.
- Concerns Over AI‑Driven Coding Rigor - The speaker worries that reliance on “vibe coding” with AI tools may erode code quality and best‑practice habits—especially among students—while acknowledging its potential for faster, exploratory development.
- LLM-Powered Code Fixes Replace Debugging - The speaker enthusiastically describes swapping tedious manual pointer debugging in Unix C for instant large‑language‑model corrections via copy‑paste or UI tools, while noting a worry that reliance on AI may erode deep software‑hardware co‑design skills.
- Specialization and Divergence in Coding - The speaker likens software teams to early aviators, suggesting that as machines become more reliable, developers will specialize and require less low‑level knowledge, creating a split between high‑level app builders and deep technical experts.
- Empowering Casual Coders with AI - The speaker argues that AI‑assisted programming lowers barriers, creates a rewarding feedback loop for time‑pressed hobbyists, and resonates with anyone who remembers the hassle of legacy coding methods.
- DeepSeek Challenges AI Scaling Law - The speakers contrast the prevailing belief that larger models and massive compute are essential for performance with DeepSeek’s demonstration that smaller, efficiently engineered models can achieve comparable results using techniques like quantization and distillation.
- Optimizing GPUs Over Raw Compute - The speakers argue that advancing performance now depends on sophisticated hardware and software optimizations—such as warp specialization in GPUs—rather than merely increasing computational magnitude.
- Quality Over Scale in Model Training - The speakers argue that careful data selection, refined fine‑tuning, and engineering efficiencies can make smaller, well‑trained models outperform much larger, poorly trained ones.
- Perplexity's Browser Bar Strategy - The speakers discuss how Perplexity may seize control of mobile browsers' address bars to route searches to its engine, heralding a shift toward voice‑first, AI‑driven user interfaces that could diminish traditional app usage.
- Perplexity AI Replaces Browser - The speaker explains how Perplexity’s deep OS integration has overtaken conventional browsing for research tasks, but highlights its shortcomings in delivering personalized shopping information that Google currently provides.
- Balancing AI Features and Phone Cost - The speakers debate how to pack sophisticated AI models into affordable smartphones, stressing democratization, privacy, multimodal interaction, and avoiding past missteps like the Amazon Fire phone.
- Concluding Thoughts on Perplexity - The hosts wrap up the episode by highlighting market resistance, personal iPhone loyalty, their intent to monitor the Perplexity project, and thanking guests and listeners.
Full Transcript
# Is Manus AI the Next DeepSeek? **Source:** [https://www.youtube.com/watch?v=Ddh3p185KhA](https://www.youtube.com/watch?v=Ddh3p185KhA) **Duration:** 00:49:58 ## Summary - The panel debated whether Manus AI represents a “second DeepSeek moment,” with mixed opinions ranging from cautious optimism to outright skepticism. - Vyoma Gajjar highlighted the bullish case, noting Manus AI’s multi‑purpose agent could industrialize intelligence by leveraging large‑language‑model advances and potentially outpace many emerging agentic startups if hardware and compute align. - Kaoutar El Maghraoui expressed doubt, pointing out that numerous competing frameworks and rapid catch‑up by other firms could limit Manus AI’s long‑term impact. - Host Tim Hwang framed the discussion within broader AI trends—such as scaling laws, vibe coding, and new products like Perplexity’s phone—while emphasizing the significance of agentic AI as a growing focus in the industry. ## Sections - [00:00:00](https://www.youtube.com/watch?v=Ddh3p185KhA&t=0s) **Debating Manus AI's DeepSeek Potential** - Experts debate whether Chinese startup Manus AI heralds a new DeepSeek-like breakthrough as they preview its versatile agent’s wide‑range capabilities. - [00:03:02](https://www.youtube.com/watch?v=Ddh3p185KhA&t=182s) **Skepticism Over Manus AI Impact** - The speaker doubts whether Manus will deliver a genuine breakthrough in AI autonomy or merely represent a rebranded incremental step, emphasizing the need for thorough evaluation amid intense competition between Western and Chinese AI firms. - [00:06:10](https://www.youtube.com/watch?v=Ddh3p185KhA&t=370s) **Integrating Open-Source AI Orchestration** - The speaker describes a platform that combines Claude’s orchestration with fine‑tuned Qwen models, sandboxed tool‑calling, and a cohesive UI—highlighting that while it adds value beyond a simple wrapper, the open‑source community could feasibly replicate the setup. - [00:09:20](https://www.youtube.com/watch?v=Ddh3p185KhA&t=560s) **Predicting the Future of Autonomous Agents** - Panelists examine the gap between POCs and enterprise integration, regulatory hurdles, and forecast how open‑source initiatives and projects like Manus will drive autonomous agent technology over the next six months. - [00:12:22](https://www.youtube.com/watch?v=Ddh3p185KhA&t=742s) **Agentic AI Demo Debate** - The speakers discuss the popularity of flashy, agentic AI demos as a perceived milestone toward AGI, critique current browser‑based interfaces (e.g., cursors and screenshots), and touch on experimental efforts to replicate existing agents like Manus using MCP. - [00:15:27](https://www.youtube.com/watch?v=Ddh3p185KhA&t=927s) **Debating AI's Role in Coding** - The speaker argues that although AI assistance can aid small projects, deep understanding of fundamental coding concepts remains essential for interviews, large‑scale development, and sustainable engineering practice. - [00:18:31](https://www.youtube.com/watch?v=Ddh3p185KhA&t=1111s) **Concerns Over AI‑Driven Coding Rigor** - The speaker worries that reliance on “vibe coding” with AI tools may erode code quality and best‑practice habits—especially among students—while acknowledging its potential for faster, exploratory development. - [00:21:41](https://www.youtube.com/watch?v=Ddh3p185KhA&t=1301s) **LLM-Powered Code Fixes Replace Debugging** - The speaker enthusiastically describes swapping tedious manual pointer debugging in Unix C for instant large‑language‑model corrections via copy‑paste or UI tools, while noting a worry that reliance on AI may erode deep software‑hardware co‑design skills. - [00:24:45](https://www.youtube.com/watch?v=Ddh3p185KhA&t=1485s) **Specialization and Divergence in Coding** - The speaker likens software teams to early aviators, suggesting that as machines become more reliable, developers will specialize and require less low‑level knowledge, creating a split between high‑level app builders and deep technical experts. - [00:27:47](https://www.youtube.com/watch?v=Ddh3p185KhA&t=1667s) **Empowering Casual Coders with AI** - The speaker argues that AI‑assisted programming lowers barriers, creates a rewarding feedback loop for time‑pressed hobbyists, and resonates with anyone who remembers the hassle of legacy coding methods. - [00:30:49](https://www.youtube.com/watch?v=Ddh3p185KhA&t=1849s) **DeepSeek Challenges AI Scaling Law** - The speakers contrast the prevailing belief that larger models and massive compute are essential for performance with DeepSeek’s demonstration that smaller, efficiently engineered models can achieve comparable results using techniques like quantization and distillation. - [00:33:53](https://www.youtube.com/watch?v=Ddh3p185KhA&t=2033s) **Optimizing GPUs Over Raw Compute** - The speakers argue that advancing performance now depends on sophisticated hardware and software optimizations—such as warp specialization in GPUs—rather than merely increasing computational magnitude. - [00:37:03](https://www.youtube.com/watch?v=Ddh3p185KhA&t=2223s) **Quality Over Scale in Model Training** - The speakers argue that careful data selection, refined fine‑tuning, and engineering efficiencies can make smaller, well‑trained models outperform much larger, poorly trained ones. - [00:40:12](https://www.youtube.com/watch?v=Ddh3p185KhA&t=2412s) **Perplexity's Browser Bar Strategy** - The speakers discuss how Perplexity may seize control of mobile browsers' address bars to route searches to its engine, heralding a shift toward voice‑first, AI‑driven user interfaces that could diminish traditional app usage. - [00:43:16](https://www.youtube.com/watch?v=Ddh3p185KhA&t=2596s) **Perplexity AI Replaces Browser** - The speaker explains how Perplexity’s deep OS integration has overtaken conventional browsing for research tasks, but highlights its shortcomings in delivering personalized shopping information that Google currently provides. - [00:46:16](https://www.youtube.com/watch?v=Ddh3p185KhA&t=2776s) **Balancing AI Features and Phone Cost** - The speakers debate how to pack sophisticated AI models into affordable smartphones, stressing democratization, privacy, multimodal interaction, and avoiding past missteps like the Amazon Fire phone. - [00:49:19](https://www.youtube.com/watch?v=Ddh3p185KhA&t=2959s) **Concluding Thoughts on Perplexity** - The hosts wrap up the episode by highlighting market resistance, personal iPhone loyalty, their intent to monitor the Perplexity project, and thanking guests and listeners. ## Full Transcript
Is Manus AI a second DeepSeek moment?
Vyoma Gajjar is an AI
Technical Solutions Architect.
Welcome back to the show.
What do you think?
Almost.
Great.
Kaoutar El Maghraoui is a Principal Research
Scientist and Manager at the AI Hardware Center.
Uh, Kaoutar, welcome back as always.
Uh, Manus AI, what do you think?
I don't think so.
And last but not least is Chris
Hay, Distinguished Engineer and
CTO of Customer Transformation.
Chris, DeepSeek moment, yes or no?
Yes, but no, but yes, but
no, maybe, yes, no, maybe.
Well, we'll be investigating all that
and more on today's Mixture of Experts.
I'm Tim Hwang and welcome to Mixture of Experts.
Each week MoE gathers just the nicest and most
brilliant people to talk through the biggest
news in the world of artificial intelligence.
As always, there's going to be a ton to cover.
We're going to talk about vibe coding,
scaling laws, a new phone from Perplexity.
But first I really want to talk
about Manus AI, which was the focus
of our initial kickoff question.
If you've not been watching the news, Manus
AI is a Chinese company that announced a multi
purpose agent, um, that has really been kind of
taking the sort of AI chatter class by storm.
Uh, they have a bunch of demos showing
their agent able to pull off some
traditionally quite difficult tasks.
So they show it, you know, scheduling trips,
uh, doing stock analysis, reviewing resumes,
evaluating insurance.
And so it's a really, really
wide range of outcomes.
And it seems to be another moment where
following hot on the heels of DeepSeek
is another time when people have been
kind of asking, like, Is China really
catching up, if not surpassing a lot of
the companies that we talk about almost
every single episode on the show at MoE?
Um, you know, your open AIs
and Anthropics of the world.
And so, I guess, Vyoma, maybe I'll start with
you because I think you were the most bullish.
I think you were like, it's close
to being a DeepSeek moment.
Um, do you want to kind of lay out
sort of the bull case here for, for
why it actually is a really big deal?
Um, sure.
So as you know, half of Silicon Valley is
building, um, agentic AI startups right now.
And Manus AI is an agentic
paradigm that we are seeing.
It is more of like an industrialization of
intelligence that has been created from all
these large language models that we are seeing.
If done right, like if they can work well
on the compute side, the hardware side,
they can come up with something because
again, they're first in this entire way, um,
paradigm of bringing it out to the market.
I know like there are so many
other agentic frameworks available.
So I feel that.
If everything goes right in like 10 other
aspects that we have to evaluate from the
metrics, hardware, the software, the compute,
etc. Maybe, but then, as I said, there are 15
others, or 50 others who can always catch up.
So, you never know.
Yeah, definitely.
Um, Kaoutar, you were a little bit more
skeptical, I think, in the opening.
Um, curious about what you think here.
You know, a friend of mine was kind of saying,
it's easy to have a really cool looking demo.
Um, and like a real product
is like a whole nother thing.
And we don't really know whether
or not Manus can deliver.
Is that kind of the source of your skepticism
or is it coming from somewhere else?
Yeah, I think I'm, I'm still
a bit skeptical about this.
I think from my perspective, Manus, um, is
definitely shaking things up here a bit.
I mean, of course there is a lot of
also scepticism in the AI community.
Some argue it's transformative.
pushing the boundaries of
what the AI agents can do.
Others just say it's just a rebranding of
what's maybe, uh, uh, the cloud wrapper, uh,
or cloud is doing more like smoke than fire.
The big question here is can madness
really redefine the AI autonomy,
or is this just another step in an
ongoing AI race between East and West?
So.
Is it just, you know, a leap or it's
just, you know, more advancements here?
Um, so I think there is a lot more evaluation
that needs to be done, uh, to see whether
we're seeing, uh, new innovations, a leap, or
just, uh, kind of maturing up this technology.
So the community is really
interested in the implications for AI
agent development.
So if Manus proves to be a significant
advancement, it could accelerate the creation
of more, um, sophisticated and capable agents.
But of course, there is a lot of pressure here.
So there is growing awareness of the increasing
competition from Chinese AI companies.
Uh, you heard from Vyoma that, you know, half
of the startups are agentic, uh, AI companies.
So there is a lot of competition here.
And I think a lot right now are analyzing the
output of what, uh, Manus is doing to see if
they can see the hallmarks of Claude's outputs.
If this is the case, it's really diminishing
the hype surrounding, you know, this product.
Yeah, for sure.
And I think it's a good
chance to bring Chris in.
I mean, on this kind of Claude point,
uh, the background on all this is that,
pretty soon after Manus came out, people said,
well, there's a bunch of responses here that
are like very kind of Claude flavored, uh, and
in some cases we're actually able to kind of
like pull out sort of like some verification or
some strong evidence that it was from Claude.
Let's say for a moment that
it is just a Claude wrapper.
Does that kind of totally diminish
this as like a, an outcome?
Like I don't know how we
should think about that.
I know a lot of people said, "Oh, if
it's just the wrapper, then, you know,
Manus really hasn't added all that much.""
So I guess from my perspective, I mean...
Let's think about Cursor, for example.
Let's think about Klein.
Let's think about Perplexity.
We could probably argue all of them
are Calude wrappers as well, right?
Which is, you know, they're all tools where
ultimately Claude is driving the experience.
But actually, I don't think this is a
story about which AI model is powering
it, although I think that is important.
This is really a story of somebody bringing
together a really great experience and I think
they have brought together a great experience
because when you use the Manus UI, it does
the planning and it's got a little to do list
and it ticks and it ticks it off as it goes
along and then it's going to have access to
the tools, it'll access the terminal, access
the browser, very similar to what's going on
with, you know, OpenAI with Operator, etc.
and Deep Research, for example.
They brought that together in a nice experience.
They're running it on a sandbox.
They're doing tool calling
and it kind of feels good.
Right.
And now it's a little bit more than
a Claude wrapper to be fair to them.
They have, you know, taken the
open source tools and they've
integrated them really well together.
Right.
So, um, so technically we
could go and do this ourselves.
And I think that's.
Why this is probably gonna
end up, this is why I went.
Yes. Maybe, maybe, maybe yes.
Yes. Maybe.
'cause I think what's gonna happen is the open
source community is gonna go, we can do that.
And, and I know that 'cause I've been coding
away all week, uh, doing the same thing.
Right.
Trying to do the same thing.
Yeah, exactly right.
Like, like every other developer on the planet.
Right.
And, and therefore.
It's something that's achievable.
And, and to be fair to them,
it's a little bit more as well.
They said that Claude was doing
orchestration, but they also said they
fine tuned a bunch of Qwen models.
And I think specifically said that for the,
the planning model, the, the one that sort of
comes up with the to dos, et cetera, that
was a particular kind of Qwen fine tune and
they pointed to a version they'd done earlier.
So, so it's a little bit more than
just, here's Claude with a pretty UI.
It's a bunch of fine tuned models.
It's, you know, bringing these tools
together, sandboxing it, and then
bringing the package in together.
And I think they've done
a kind of a fabulous job.
And then finally, they've
generated the hype, right?
I mean, I was reading today, it had like
2 million folks have signed up for invites.
So, we're all running going,
yeah, we're going to do this.
Will they be around?
Are they the next Devon?
You know, we will find out in six months or so.
Right.
But, but for just now the hype
cycle's there, but I'm hoping it
galvanizes that open source community.
And you had a good phrase there, which
is, you know, they're tying together
a bunch of components and like,
sort of we could do this as well.
Um, and I did want to dive a little bit
into that because I have a friend of mine
who you know, he focuses a lot on all the
kind of like state level AI bills that
are kind of bubbling up. And he made the
observation where he's like, well, look we
the US Companies could have done this US
open source kind of efforts could have
launched a very similar thing you know,
maybe one reason they don't is because like
some of the things that Manus is showing
off, you know have been kind of risky from
a legal standpoint in the US, right?
Like things like resume review is
like a thing that, you know, kind
of is like very hotly regulated.
It's a hotly disputed thing.
And so I guess maybe I'm
going to toss it back to you.
Like, do you think almost there's
kind of like an edge here?
Like almost like Manus is winning
this or in the very least they kind of
seem like first to getting this like
hype wave just because they've been
willing to be more aggressive than other
folks or do you do not really buy that?
I feel the first thing that Anthropic
cloud had tried something called computer
use we spoke about it in this board with
you and which is sort of compared quite
vigorously right now with Manus AI but
the computer use Anthropic cloud version it
actually performs very well in controlled
environments exactly what we're talking about.
Like, let's say if there's resume review,
et cetera, it brings along a whole different
metric system that has to be evaluated
for a large language model to be used.
Like a use case, POC is very different
from what can you integrate in an
enterprise architecture, right?
And how do we integrate it?
So sure, Manus showed the way that, okay, yes,
this is how we can do it, but to actually do it.
It's going to be a lot of leaps and bounds
that the entire industry has to go through,
regulations, et cetera, uh, have to be
written around it for us to be able to use it.
But yes, the US did try it and that the
computer use part was actually something that
we were all talking about for a while, right?
So Kaoutar, I guess maybe the final kind of
question and curious to get your thoughts
on this is whether or not like, take us six
months into the future, like Chris is saying.
Like, do you have predictions
on where this all kind of goes?
I mean, one thing's for certain.
It seems like we're going to see a bunch of open
source attempts to kind of do the same thing.
I guess we can ask the question of whether
or not this Manus thing actually changes the
fundamental trajectory of where agents is going.
Um, but kind of curious if you want
to paint a picture of like, where
you think we will be in six months.
You know, if anything, Manus is kind of just
building more hype and more momentum in this
direction, but curious to get your prediction.
Yeah, I think it's a good question.
I definitely feel, you know, I agree with Chris.
That's, you know, what they're doing is,
it's not just integration, but it's also
more about having this fully autonomous.
agent capable of independently executing
complex tasks, doing various things
like sorting, stock trend analysis,
website creation, which is really great.
So we will see others trying to mimic that.
Uh, and they're leading in this space
around, you know, this autonomy.
Uh, so beyond mere integration, making a
significant advancement in the AI autonomy.
But I think whether, uh, more
hype will follow, I think.
It's gonna be the case.
Uh, we're seeing now every few days or
every few weeks we're seeing new hypes.
So I feel we will see, uh, more interesting
things coming in this space here.
Yeah, it's like if this is a DeepSeek
moment, then get ready for like at least
20 or 30 more this year, I suppose.
I just want to say, you know what
I don't want to see in six months?
Another browser operator.
You know what?
Large language models are
really good at text, right?
Why?
Why are we insisting we have an AI moving a
cursor around, finding a bounded box, taking
a screenshot, and then typing into the box?
You know what I would rather see?
I would rather see somebody go, you
know what, I'm going to create an AI
native browser which parses the text.
And actually, yes, it's going to communicate
with, with, you know, the websites, and they'll
recognize it as a real browser, et cetera.
But you don't need to do screenshots.
You don't need to move a cursor around.
You're a browser and you're a large
language model that knows how to code.
Do that.
That's what I want to see
in the next six months.
People have gotten lazy, Chris.
People don't want to do that.
They're like, if we can have
someone do this for us, why not?
I'll sit on the couch all day.
I'm fine for them to sit on the couch.
Just, just don't move a cursor around.
That's what I'm, that's
what's, and take screenshots.
That's what's bothering me.
Yeah I think that is one of my
favorite agentic tropes at the moment.
I mean people do it because
it looks really cool.
Like that's the main reason is that it's
like it's it's kind of cool and spooky.
Yeah I think it's more for demo
purposes.
And also for people to like show that this is how we can do AGI.
Like this is the next step to AGI.
That's, and I think everyone's chasing that now
that, okay, we're done with the LLMs, et cetera.
Now let's move on.
But Vyoma, it sounds like you're defining
AGI as a 95 year old grandparent trying
to work the internet for the first time.
That, that's what I see when I
see the AI operate with a browser.
Yeah,
that's true too.
I mean.
That's what we defined, I guess, at
this point, but let's break that.
Yeah, I think it's going to be interesting to
see how these human interfaces will evolve.
So I agree with you.
I also don't like the cursor on these things.
So probably more serious thinking
into what would be interesting
for us to see as these interfaces.
What would you like to see?
So it really mimics a true human
experience without having this cursory
or, uh, screenshots and things like that.
So,
yeah, for sure.
Uh, Chris, if for your weekend experimentation,
have you been able to replicate Manus?
I am surprisingly far.
Actually, I went
slightly different from them.
So I've put MCP at the heart of what I've
been doing, uh, which I think is, is a
lot better, but then I haven't built a
product and got it to 2 million people.
This is just Chris and his agents in the night.
So I, you know, I don't think I can take
the intellectual high ground on this, but I,
yeah, I've got, I've got pretty far so far.
Yeah, that's actually, I mean, it's a
pretty, uh, strong indication, right?
Like with not a whole lot of work, you
actually get pretty far with these things.
Um, I guess it just goes to show how
competitive this space is about to be.
I'm going to move us on to our next
topic, which is a real fun one.
Um, Andrej Karpathy, who we've talked
about many times on this show before,
in addition to being, you know, former
Tesla and former OpenAI, I think he's had
this kind of like career job in sort of
shaping the memes of the, uh, AI space, um,
we mentioned Cursor earlier, arguably his
shout out of Cursor is one of the reasons
that Cursor has been so wildly successful.
And, um, he had a nice tweet kind of capturing
sort of his thoughts on kind of using AI
assisted coding recently, where he said:
"there's a new kind of coding I
call vibe coding, where you fully give
in to the vibes, embrace exponentials,
and forget that the code even exists.
It's possible because the LLMs,
e.g.
Cursor Composer with Sonnet,
are getting too good".
Um, and this has kind of been a funny
thing because in, you know, true Andrej
form or Karpathy form, um, the vibe coding
has just kind of like gotten everywhere.
It's just been like a joke that
people keep mentioning now.
Um, and, uh, and weirdly, I feel like in the
last week or so, or however long ago, um,
people have been like, oh yeah, I'm like,
I did this project through vibe coding.
You know, it's almost kind of now
becoming like a, a term of art.
If I may be able to turn it to you first is
like, is vibe coding a real thing?
Like, is this the future of coding?
Is people just kind of like, you know, just
kind of vibing with it until an app comes out?
Um, is this a good way of kind of
thinking about where engineering
is going with all this assistance?
Yeah, it's going to be very controversial
when I say this, but no, it's not.
I don't think this is the future.
And I feel that getting to know the concepts
and the basics behind how a particular code
works is something which is extremely important.
Sure, you you won't be able
to like code it end to end.
You can use wipe coding to assist you with that.
But that being the only uh, crutch
that we all rely on, I don't think
this is going to be successful.
Like sure, I can do wipe coding
for a weekend project to test out
something that I want to maybe show a
small POC to understand the concepts.
But for a diff, and a diff which is like going
through millions of lines of code over there,
you use that particular code to solve that diff.
And then, put it in production.
So totally against that.
I don't think that's the right norm.
The other thing is, I know everyone keeps
talking about white recording, but if you go
back in the interview market right now, you
have to go through a lead code interview.
You have to do solution design.
So the basics aren't going anywhere.
People are talking about it, but yes, you
have to traverse a string if needed or like.
add the nodes of a binary tree.
Yes, you have to do that.
It's not because people want to know
how well you code, but they want to
know whether you understand the concept.
Right? So that's something that I feel, um, is needed.
Yeah, so there's a lot to unpack there.
I mean, I think the first thing maybe
to get into is, yeah, I don't think even
Karpathy is like, oh, you should vibe code.
Vibe code like an enormously complex project.
But I think it is kind of
this interesting debate.
I mean, you sort of say like, look,
this is going to work for your weekend
project, but probably not much further.
I think there's almost a question of
like, where do you draw that line?
Like, how far can you get with vibe coding?
If everybody agrees, you can't do the
most complex thing, but you can definitely
vote vibe code the simplest thing.
You know, this dividing line gets very fuzzy as
these models get better and better and I think
it is, it is a genuinely interesting question.
I'm curious if, are you a vibe coder?
Are you, do you vibe code?
I do it sometimes and I like it.
Actually, I'm really fascinated by vibe coding.
So, um, and I see vibe coding as kind of
a reflection also of the changing nature
of software development where AI tools
are increasingly being adopted and they're
increasingly handling routine tasks.
Uh, allowing, you know, the coders to focus
on higher level design and problem solving.
Of course, I think I'll see we're evolving
into a world where we're going to be
combining both vibe coding and serious coding.
But of course, understanding what
this, what the AI is generating, how
to test it, how to integrate it, how
to do all these, these diffs and so on.
It's still going to be very important.
Whether we're going to get to the point
where it's all going to be automated
by AI, I think it remains to be seen.
Uh, I think we're heading into the
direction as these, um, AI, uh, LLMs are
becoming better and better and coding.
Uh, one of the things that I'm a
bit concerned about is the rigor.
Um, so the vibe coding, you know.
could also lead to this decline
in code rigor and best practices.
So there is a worry, you know, that I see that
are we getting into also less experienced coding
coders that rely more and more on vibe coding,
especially among students, people that are
still learning, uh, you know, they're, they're
given an, uh, programming assignments and
they'll just go and ask, you know, an AI agent
or a LLM to give me the assignments
and then most of these assignments sometimes,
you know, the AI does pretty good job.
So, um, so they're definitely going
to be an influence in these, uh, vibe
coding or AI tools, uh, influencing
the coding styles and the practices.
Uh, but this is also
enabling more exploratory and
iterative approach to coding.
Yeah, for sure.
My friend actually, he,
uh, he coded up this app.
And, uh, I was like, oh, so like,
what, what does this menu do?
And it was very funny because he
was like, I just vibe coded it.
So I'm actually not really sure what it does.
And I was like, I don't know if this is
like a sustainable way to go about building
bigger systems, but this bleeds pretty well,
I think, to Vyoma, a point that
you made, which is first you said,
okay, it's only going to be good for
weekend projects, not anything bigger.
I think the second point that you made
is also really interesting as well.
If you want a job as an engineer, they're still
going to make you go through LeetCode, right?
Like you still are forced to kind
of like go through this gate.
And I was joking with a friend recently.
I was like "Oh, well, I'm just
looking for the 10x vibe coder".
Right?
Like someone out there is able to vibe code
like way more proficiently than everybody else.
If I can just find that person,
um, maybe he doesn't, he or she
doesn't need to know LeetCode.
So even, I'm not saying that LeetCode
is 100% a reflection of how
good, uh, a software engineer you are.
But even if you're not able to solve that
pseudocode, how does that if a statement work?
How does this while statement work?
Trying to explain what you are asked from a
LeetCode question is also good enough in
case you're not able to code on spot during
that particular time that I'm asking you.
A 10x viber, vibing coder that
we are naming it, I don't know.
If that particular coder would actually
even understand a legacy code that has
been written because for that you'll have
to go back and understand, okay, what's a
function, how did that function, what are
the parameters called in this function,
how is a parameter written here, so I'm not
there yet, maybe I have not seen a good use
of that entire, uh, system, but
maybe, um, people might get better.
The models keep getting better.
I totally agree with that.
Sure, you can use it for grunt work, like if
that was one of the questions that, uh, was
coming up when I was reading about it is like
if there is a CSS file and you need to change,
uh, a particular bracket or a button, you
don't have to sift through thousands of lines.
Of course, sure, you can use it for that
because obviously better use of your time to
do something else, learn something better, but
for you to build an application end to end,
I don't know if it's the best way to do it.
I can say right now, this
is, this is here to stay.
This is not a weekend project thing.
And, and as somebody who's starting
graduate job was writing Unix C motif, and
I, and I can tell you right now, I spent
most of my time chasing memory pointer
bugs, right, you know, and I am never
going back to that nightmare ever again.
I, you know, you talk about productivity.
You look at your own terrible graduate
written code and try and work out
why that memory location is not
the place you wanted to point at.
You know what, I, and you watch your dreams
die as every time you boot up your application
and the thing goes, wow, wow, wow, wow.
You know, and what, what.
What I can do with a large language model
is I can go Ctrl+C, Ctrl+V, and with
the error message from the compiler, and
then say, fix this buddy, and then it goes,
ah ha ha, you, you messed it up over here.
Oh, that's great.
I will just copy and paste that back in.
Or using Cursor, I don't even need
to copy and paste, I can just go.
Click, click, click, like Homer
Simpson in the nuclear power factory
when he was working at home with that
little pigeon going dun, dun, dun.
That's, that's the world we're moving to, right?
So I'm, I'm all in.
I'm all in.
Okay, there's about, you've
got some inbound here.
Uh, Kaoutar, how about you go first?
I'm a bit worried about also if we only
do this, then if you really want to do
software hardware co design, really doing
optimizations, then people will lose the
skills to understand the computer architecture.
What does it mean to have a dangling pointer?
How do we do these memory
allocations, the efficiency?
I mean, I see that right
now as we are designing.
these efficient systems, having a core
understanding of these concepts is
very critical to know how to optimize.
So if we're just going to do vibe code
and then people have no clue what, you
know, underlying systems are doing, how
they're behaving, what does it mean to have
a cache hierarchy, you know, all these
implications about, you know, the data movements
and the computational units, the bottlenecks.
We're going to lose that.
And that's what worries me.
And how do we optimize these systems end to end?
I'm okay with losing that, to be honest.
I'm still having,
yeah...
Of course, it's not for everyone.
I'm having nightmares from, from my past.
there. You know, so no, I'm fine with that.
But then, but honestly, I think, I
think those skills are important.
I think we have different, uh, we have different
skills within the engineering practice,
and I think that's going to expand out.
And, and if you really think about it, I
mean, like, if you think of something like
Lewis Hamilton as a Formula One driver,
one of the best drivers in the world, etc.
Does he know how to change a tire?
I don't know.
Maybe he does, maybe he doesn't, right?
But, but what he does know how to do is drive
a Formula One car at speed through, you know,
uh, the race course and better than anyone else.
Now, the question I would have there
is that there is a wider team, right?
Some people are going to be
specialists at tire changes, some,
you know, ball bearings, whatever.
And I think that's, that's pretty cool.
So, not everybody needs to know how
to deal with dangling pointers, right?
Some people just want to build an app and
get it out and try and make some money.
Go for it.
Yeah, there's some interesting history
here too, because it's almost like, I
guess, you know, just think about like
the first people who built planes, right?
Like the Wright Brothers.
They were like, you know, engineers, right?
And they were like, you know, modifying
bikes to get the plane to work.
And like, yeah, if you're flying a plane,
then it almost breaks down so often that you
really need to understand every bit of it.
Right.
And then now pilots have a certain
level of training, but they're not
necessarily airplane engineers.
I actually wonder if that's going to kind
of interestingly have that divergence in
coding over time, where you almost have
like good coders, but that's almost like a
discipline that's almost separate from like
understanding the inner workings of the machine.
And.
You know, I guess we get that in part because
like machines become super reliable, right?
Like your car is not breaking
down every single day.
So you don't necessarily need to understand.
But you know, software used
to break down all the time.
And so you really do need to
know those internal components.
And first of all, I'm very happy to know
that Chris is a fellow F1 supporter.
So even if Hamilton doesn't try to fix a car,
Chris, I think he would know how to do it.
Like theoretical concept.
That's all I feel is.
actually needed in the wipe coding part and
the other thing is I wouldn't entrust anyone
with like a wipe coder with the nuclear
reactor control system, but I would make
sure that, um, anyone who does wipe coding
actually knows what they are doing with it.
So as long as that loop ties
back, I, we are okay with it.
But if it doesn't, and we have someone who's
just like vibing with in minimum amount of data
points and trying to make something production
ready, I don't think I'm comfortable with that.
I, I think, I sort of agree with you
and I think, I think we're going to
move from one extreme to the other
and I think that's the reality, right?
And I, I can see a world where you're going to
vibe code to prototype, you're going to vibe
code to figure out some issues, etc. And then
it's almost, I think, I've never painted, but
it's like, you know, those like Monet paintings
or whatever, and then it's all sort of blurry
stuff, and then you sort of hone into the
detail, I think that's going to start to become
a bit of a pattern, right, which is like, okay,
I kind of need something like this, I'm going
to do this, you're going to orchestrate it.
And then you're going to start to say,
okay, I know what I've, I've built here.
I prototype this.
Now I'm going to start engineering this
further and then go down and detail.
And so I think there's probably a hybrid model,
but then the flip of this is, is we're
looking at this from an engineering perspective.
Why can't that person who's never coded
not go and create an app for themselves and
get it in the app store and make some money or
maybe somebody who wants to do a home automation
project, but has never had those skills.
Why can't they vibe code and then
be able to do that thing that
they've never been able to achieve?
And then you know what?
It might get them interested in the discipline
and might want to say, you know what?
How does memory work, right?
And then they start delving into that.
So I, I think actually it's an area
where we can have a greater inclusion and
greater impact and, and a larger community.
So I, I, I hope that's the direction we go in.
Yeah, I think the feedback
loop is super important.
Um, I mean, at least for me who like doesn't
really code day in and day out, like the
ability to use these tools just makes the
experience of like, I have 45 minutes after
my kids have gone on down to like kind of play
with the computer and it's like oh I could
just like get further in that time and it's
just like a very strong feedback was like
very satisfying in a way that kind of like pre
these tools like you didn't really have. A final
anecdote that y'all might find interesting.
So my mom was like a very early coder.
And so she still remembers like, you know
punch cards and it's it's Chris to your
point It's like it's interesting to me how
much like if you felt a lot of pain coding,
you are more likely to want these tools
because you remember how painful it is.
Like my mom's response is oh, yeah I remember
programming a big box, like, of punch cards,
dropping them, and then basically spending,
like, hours having to, like, recompile.
And she's like, I love this.
Like, we just, like, automate
everything, basically.
Um, which I think is, like, kind of a
really fascinating aspect of, like, people's
personal experience of just, like, how
difficult it is, might make them more
or less willing to adopt these tools.
And you could imagine a
vision model, multimodal, controlling an
action model with a robotic arm, and then
that robotic arm can resort those punch
cards, and your mom would be fine, right?
And that's right.
Yeah, exactly.
We should. That's right.
I need, I need the, the coding
assistant, but for, for punch cards, that
actually would be an awesome project.
I agree with Chris that it's going to be a
hybrid world where we have these people who
have no clue about coding, but they're still
being able to use this vibe coding to create
really nice things for rapid prototyping
or proof of concepts or applications maybe
that have been mature by these, uh, tools.
But then we still need the people who
really understand what's, what's happening
behind the scenes, who can know how
to debug, who can know to optimize.
It's going to be specialization
at these different levels.
For sure.
Yeah. And I, I got to believe, I mean, some of
these debates happened when like object
oriented programming came around, right?
People were like, ah, you don't
understand the DNR workings of the system.
It's like this battle happens almost at
each layer of abstraction, uh, arguably.
I'm going to move us on to our next topic.
Uh, we wanted to do a quick segment to
talk a little bit about scaling laws
and this segment kind of puts together.
I think a couple of things that
we've touched on last few episodes,
particularly when it comes to DeepSeek.
Um, and I think Kaoutar, great to have you on the
show because I think you suggested this topic.
Um, the background here of course is, uh,
scaling laws are really the idea that we
have kind of this interesting relationship
in machine learning where sort of the, the
more compute that you're using in doing
pre-training the better the
capabilities are that kind of come out.
There's kind of this rough relationship
between like how much kind of like
muscle you're putting in and the model that kind
of comes out of it. And you know, this has been
kind of like the thing that has motivated the
entire thesis of these companies, particularly
the kind of frontier model companies raising
enormous amounts of money, which is to say
well if we want really powerful systems, we
really need lots and lots of data lots and
lots of compute and we need to do the biggest
possible sort of pre-training run that we
need to do and I know Kaoutar, you wanted
to bring up this topic because you want to
talk a little bit about how you think that
DeepSeek kind of doesn't necessarily break this
idea, but kind of nuances it a little bit.
Do you want to talk a little bit about that?
Yeah, definitely.
So of course, as you mentioned, in traditional
AI development, there is this general belief
that bigger models and larger data sets lead to
better performance following the scaling laws.
And this often translates into these massive
investments in hardware and infrastructure.
But what this DeepSeek really
demonstrated, uh, they're challenging
this traditional AI scaling laws.
They're demonstrating that smaller,
more cost efficient models, they
can achieve competitive performance.
And they're, you know, even this titan
existing business models, reliance
on these large scale infrastructures.
So they, they used a lot of techniques, uh,
like, uh, uh, quantization and distillation
and, uh, they even, you know, did some
optimizations at the PTX level, given the
limitations that they had with the H100 GPUs.
So a lot of focus on efficiency,
oversize, so, uh, emphasizing efficiency
rather than the sheer model size
and optimizing at different levels in the stack
and also looking at, you know,
enhancing the data quality and
employing better training strategies.
So I mean, the, the, the key idea here is how do
we leverage smarter techniques, uh, or smarter
training techniques, for example, for their, in
their example, that achieves better performance
with fewer parameters and reduce these
computational costs and computational needs.
And I was really fascinated by the wide
range of techniques that they have used.
You know, they had these
data centric approaches.
They also had the hardware wear optimizations.
They were considering also sustainability.
Um, and I think this has, you know,
implications on the AI community.
So where the focus here is shifting from the
scaling by size to scaling by efficiency.
I think this is actually, I mean, like
there's a lot here that we could get into.
Um, I think in some ways for me, the kind of
scaling law question is kind of interesting
because it has kind of turned out that
people have meant a couple of things when
they say scaling laws, um, in, in the
popular, uh, kind of discussion of AI.
And,
you know, one of them is just
like how much compute you need.
Um, and I guess in some ways
DeepSeek doesn't really change that.
It certainly changes the kind of platforms you
need to get high performance out of the models.
Um, but I guess it doesn't really
necessarily kind of eliminate
the idea that like more compute,
like, equals better performance.
Is that, is that right?
Is that the right way of thinking about it?
I don't know, Vyoma, if you want to jump in.
With the scaling laws, with DeepSeq,
with etc. I see a new shift in
people trying to optimize GPUs.
Or, uh, the ways in which they can
revolutionize this entire field.
So I don't know if, uh, people know about
this, but a month ago, Meta and NVIDIA
came up with a paper and with some,
and they said something called as warp
specialization will be a part of PyTorch.
So it kind of optimizes the GPU performance on
the, any of the hopper architectures, like the H100s that they do by assigning like some
sort of distinct role to each one of these warps.
And what is one warp? Like a group
of 32 threads that are running.
So it kind of pivots to this entire point that
we are looking into how do we optimize
all of these hardware specs, which are
also available based on the
previous information that we've got.
And I think that had it also came
into picture because of some of the
scaling loss that we've been seeing.
So I don't see that as a way in
which it would limit us, etc. I
feel we've come up with better ways.
Right.
It's not necessarily about magnitude.
It's more like how we're
treating the GPUs basically.
Exactly.
Exactly. I think the hardware where optimizations
are becoming increasingly important.
If you also see the work that's
happening around these state space
models and, uh, the flex attentions.
Every now and then we hear about, you
know, different algorithms around flash.
You know, how do you do these transformer
attention computations more efficiently?
Like there is a flash attention.
There are various versions
of these flash attentions.
There is a flex attention.
Now the Mamba and the Bamba models.
They're also doing a lot of optimizations by
understanding the underlying architecture,
especially the GPUs right now, and then figuring
out how to restructure the computations,
and especially the data movement so you can
drive more efficiency from the hardware.
Um, so also other things
like the test time computes.
which is something that is
also becoming very important.
So instead of focusing solo
solely on pre-training computes.
So, and this is an example also that DeepSeek,
uh, uh, pointed out, which can we focus on
inference time computes, which is really
more critical, meaning smaller models can
compute more at test time, longer reasoning.
Tree search, Monte Carlo
inference, and things like that.
And this also reduces the need
for enormous parameter count.
There was a very interesting paper about
test time computes, which showed all of
these techniques that really focuses on how
do we bring more from the model during test
time, not during the pre the training time.
And also the distillation, this creating these
compact models with large model capabilities.
So, of course, you still need to have the
large model, but we can create a variety
of distilled versions that do really
better and can also inherit knowledge
and reasoning from much larger models.
And of course, you know, I think the high
quality training is also something that
is outperforming this raw scaling, smart
data selections, better fine tuning,
reinforcement learning with self improvement.
So well trained models can also outperform
these poorly trained massive models,
especially if you focus on the data quality.
Yeah, I was just going to say I think actually
we've came from a world of vibe training,
which is really, if you think about what
was going on in the beginning, which is just
like, we're going to take some transformers,
and we're just going to throw a bunch of
data and get it to NextSoak and predict.
And actually we're in this stage now
where it's really about honing the
algorithms, honing the chain of thoughts,
as you say, starting to engineer things.
I mean, Kaoutar, you made some really
good points on DeepSeek, right?
So actually one of the interesting things
they did, I think it was last week, is
they open sourced a whole bunch of their,
uh, code bases that you use to train.
So, and that's everything from data frameworks,
they, they even engineered themselves a new
file system, a distributed file system, etc. So,
actually, all of these engineering techniques,
anything you can get more efficiency, anything
on a better training, I loved your point about
the high quality chain of thoughts and inference
time compute, that makes a huge difference.
That allows you to, to then.
start getting smaller models, higher
quality models, and I really think we've
moved into this kind of engineering phase.
So, but I'm going to, I'm going to,
like Karpathy, I'm going to, I'm going
to call vibe training and see if, see
if I can, see if I can get myself a
Wikipedia entry off the back of that.
Yeah, you heard it here first.
Um, I guess Kaoutar, maybe I'll throw it
to you for the last kind of question here.
Do we think that scaling laws no longer matter?
Like, do we care about scaling laws anymore?
Given this kind of new era of
optimization that we're now in?
Yeah, I think we should really
shift from just bigger to more
smarter and more efficient models.
So I, we should really redefine the
traditional scaling laws as they were
defined by just bigger, better, but
I think it should be about
smarter and more efficient.
Well, I'm gonna move us on to our final topic.
I want to end on sort of a, a kind of
fun, sort of odd story that kind of
came across our um, uh, sort of cues.
Um.
There was an announcement recently that,
uh, Deutsche Telekom and Perplexity, um,
were going to work together to announce
and launch, uh, what they call an AI phone,
which would be a phone that integrates
a bunch of, I guess, AI features for
less than 1,000 and coming out in 2026.
Um, and this news kind of struck everybody as
like a little bit sort of surprising in some
ways, because if you know Perplexity, the
company, um, they largely have been in the world
of, uh, AI search, AI powered search, right?
I think they're one of the first to market
in terms of, you know, you ask a human
language query and it gives you sort of
results, um, that, you know, attempted
to kind of be better than what you'd get
from the sort of quote unquote sort of 10
blue links or 10 links from, from Google.
Um, and so I think the first question,
which is that it's a very kind of
perplexing announcement for Perplexity,
uh, to be getting into the phone space.
Um, Chris, you're already smiling,
so maybe I'll throw it to you first.
Why is Perplexity doing this at all?
Tim, when you're on your mobile
phone and you're doing your Google
searching, do you ever go to
www.google.com
and then type in your query there?
Is that your action?
I do never,
I never do that.
What, what is your action, Tim?
How do you
search on a regular mobile phone?
Uh, I would say open the browser and then
I type my search term into the browser bar.
Exactly, so this is what this is really
about is controlling the browser bar, right?
So, at the end of the day whoever
controls the browser bar means
that they can direct those queries
to their search engine.
So, if I was Perplexity, I would
absolutely launch a mobile phone where
I'm in control of the browser bar.
That's, that's my opinion of what
they're doing, and that is a smart
move.
Uh, do you all agree?
Vyoma, Kaoutar?
Uh, curious if you are like,
brilliant move by Perplexity.
This is exactly what should happen.
And I think this is also marking a shift
in the user interface, you know, how these,
how we're interacting with our phones to
shift to a more voice centric AI driven
user experience, potentially reducing also
the reliance on the app based interactions,
like, you know, going to apps and so on.
I think maybe this is going to shift
and change with the, these AI phones.
It's going to be maybe a completely different
experience that is mostly voice centric.
And probably we'll see the
disappearance of the apps.
And more of these agents in the background
working together to satisfy whatever we need.
Yeah, the interface part of this I
think is really super, super interesting
and it's something I actually want
to dig into a little bit more.
You know, one thing that has been said about a
lot of these AI search features, right, whether
it is perplexity or what Google is doing.
is that increasingly they're kind of
moving to a world where you don't have
to go to the underlying website, right?
It kind of curates the result for you.
And so I guess Chris, to your original question,
it's like, it's, it's kind of very weird to see.
It's almost like the whole feature
has been turned inside out, which
is you are going to a browser bar.
To not browse the web, but instead
to get the results of a chatbot.
That's, that's really
strange, from my perspective.
And that chatbot is gonna launch its own
browser instance, and then Google somewhere,
go to somewhere else, look up that website,
and then come back with the answer.
It, it's gonna be weird.
Definitely.
And I think there's almost kind of like, uh,
the, the interesting debate here is I think the
business rationale for Perplexity makes sense.
It's almost a question though of
like, how valuable is that browser
bar going to be in the future?
It of course has been like, I mean, it's
been the source of a lot of litigation
on like, say, Apple working with Google
to have it as the default search engine.
But you can almost imagine a future
where, you know, maybe the app is
actually the more powerful thing.
Like when I go to want to know something, I
actually don't go to the browser bar anymore.
I just go to perplexity.
Or in Kaoutar's world, it's like,
I just speak into my phone, and the
phone just does what I want it to do.
Um, you know, I guess, uh, Vyoma, maybe I'll
ask you, is like, is there a world
where almost like, uh, Perplexity is trying
to seize this real estate on the phone, which
actually might not be so valuable in the future?
Like, maybe browser bars are just
like, going to be a thing of the past?
Just FYI, uh, it, it has taken
over my browser bar for sure.
I've been using Perplexity for months now.
And so I have many of my friends, you
want to research anything, like in my past
days, I would go back and like be like,
Hey, I want to order these new headphones.
space, Reddit.
Now I no longer have to do that.
Like, I, my, it's on my home screen.
It's right there on my mostly used
apps, because that's all I use now.
I don't research anything no more
about, so it has totally taken all that.
And AI layered integration at the OS
level is much better than anything,
any standalone app that exists.
So I think Perplexity has hit it outside
the park there, and the only one thing that
I sometimes struggle with in Perplexity
is when I actually want to buy something.
So let's say I want to buy a mattress or
like a particular lampshade, then I will
go in and then I'm researching about it.
It's not getting that kind of consumer
knowledge about me, like Google has,
because of course Google's integrated
at OS levels for that context for years.
So this is going to be a great pivot for
them to make their product or their large
language model or their app more context
aware, which is the need of the hour now.
So it's going to feed all
these usage patterns back.
And I think I'm never going to
lose perplexity from my phone.
I love it.
Genuinely.
I no longer have Google search
bar on my phone anymore.
That's a, that's a big deal.
Yeah.
And, and, and I live in the Bay Area and
there are many, many people who use that.
So believe me, like you'd see them pop
it out on their phones all the time.
I have friends who use that.
So I feel that is, um, one of the things
that I'll see, but the one thing that they
didn't speak about in that entire blog
post was the hardware specifications of it.
So what is that inference, um,
layer that they're going to use?
Is it going to be sound?
Robust local interference with like the
typical cloud or are they going to be
using on prem such as the NPUs or TPUs,
etc. So that is going to be the deciding
factor, whether it is here to stay or not.
That's a very good point, Vyoma, because this
definitely depends on how mature or how powerful
the edge AI models, especially the on device AI.
So, as we're getting, uh, basically these
LLMs to become smaller and more efficient,
more AI processing can be done on the device.
And this provides, of course, faster responses
and better privacy.
And also what you did is context
that allows customization.
So I, I see this evolution.
I go in hand in hand as the edge AI becomes
really more mature and more powerful.
We can do more with these AI phones.
Exactly.
Yeah, the price angle I think is,
I hadn't really thought about that.
That's, that's very interesting is, you
know, almost how, how cheaply can you pack
these features into the phone is going to
be this really, really interesting question.
Cause right now it's almost
like a luxury feature.
If you want to run kind of a more sophisticated
model, then we got to have all the power and all
the build out and all the hardware at the edge.
It just makes for a much more expensive phone.
And they're promising for like
less than a thousand dollars.
And again, it's already very funny
to be like, it's a phone, but.
It's going to be cheap.
It's going to be less than 1, 000.
But even still, like, I think to kind of pull
that off is like, pretty interesting in terms
of how far you can democratize this tech.
Yeah, it is going to start this whole
wave of having these specialized devices.
I hope we are not going back to the era of
the Amazon, Amazon Firephones, how it kind
of detached itself and wasn't that great.
But I hope that this kind of breaks
that curse and we are able to see
something greater and better on this.
I agree Vyoma, I think I just, I want to see
that new experience as you say there, right?
And you know, yeah, a native integrated
experience and, you know, have all your
contacts but it be private and, you know, and
I loved your point, Kaoutar, about voice, etc.
I think there's so many different modalities
that can kind of come into this and, you
know, and again, back to the camera as well.
So I just, I just hope that we get
a different and new experience.
But I, as I said earlier, right, is if
you want to, if you want to control that
search experience, you need a device there.
So I, I do think it's a brilliant move.
Yeah, this is all happening while
Apple is delaying its AI features.
I don't know your take on this.
Uh, is that because they want to make sure
that they have a very well curated, secure,
because Apple in general has been conservative
about, you know, the security, Or, you know,
this is opening the space for more, for
example, for Perplexity and so on to take over
some of the market that, um, Apple phones have.
Yeah, I think that the Apple part, when they
put out Apple intelligence and they hear, like,
they got a lot of backlash about that as well.
So maybe they are field testing it way,
way more before coming into production.
I don't know if you know about this, but that
Apple, uh, came up with the Apple Kids Watch.
Because again, they're as it is like the
point that you made, Kaoutar, they're known as
the company which respects privacy and has
it integrated in all aspects and they've come
them coming up with the kids watch kind of
showcases their, um, commitment towards it.
So I feel they are looking into several
avenues before coming out with something.
Public
yeah, and I still think I think weirdly the
kind of hardware mentality actually might
be working against them a little bit in
implementing some of these features because
sort of unlike you know building a phone which
you can really kind of like I feel like part
of the problem with these models is they still
You know, unreliable and kind of probabilistic.
And, you know, I think like in some
ways the discipline of like launching
features is a little bit more risk
loving than that Apple might be used to.
And I think it's actually
holding them back in the market.
Yeah, but I ain't giving up
my iPhone for anything, Tim.
So I'm okay with that.
Yeah, that's right.
I mean, I think the counter argument as well.
Uh, they can just keep trying because
everybody's on their phone and
they're not going to throw it away.
So they can just keep going until they get it.
Um, but something to keep an eye on.
Um, we'll definitely be keeping
an eye on this Perplexity project
and, um, a lot more to come there.
Um, so, uh, that's all the
time we have for today.
Uh, Vyoma, Kaoutar, Chris, uh,
thanks for joining us as always.
Um, and, uh, thanks for
joining us, all you listeners.
If you enjoyed what you heard, you
can get us on Apple Podcasts, Spotify,
and podcast platforms everywhere.
And we will see you next
week on Mixture of Experts.