Latest / Chief in Tech Deep Dive Shortcast / Chief in Tech S1E65 - Varsha Agarwal : Universal AI – Building by Everyone for Everyone

Transcript

Chief in Tech S1E65 - Varsha Agarwal : Universal AI – Building by Everyone for Everyone · Chief in Tech Deep Dive Shortcast

• 0:00
  Welcome to the deep dive. You know, if you look
• 0:02
  at the tech headlines today, there's this very
• 0:04
  specific, almost universal consensus, really,
• 0:07
  about the ultimate goal of artificial intelligence.
• 0:09
  Oh. Absolutely. Right. It's everywhere. Right.
• 0:12
  It's framed as this high -speed race by just
• 0:14
  a handful of tech giants to build quote -unquote
• 0:17
  AI for everyone. We're constantly sold this vision
• 0:20
  of a perfectly finished omnipotent product that'll
• 0:22
  just be, you know, handed down from the cloud.
• 0:24
  And the rest of us are just supposed to passively
• 0:26
  use it to solve all our problems. Exactly. We
• 0:28
  just sit back and consume it. It is the ultimate
• 0:30
  top -down model. I mean, the core assumption
• 0:32
  there is that if you just throw enough computing
• 0:34
  power and, like, mountains of data at a centralized
• 0:37
  system, you can somehow create a single tool
• 0:40
  that navigates every single nuance of human work
• 0:43
  and life. Okay. Let's unpack this. Because our
• 0:45
  mission today is to explore a presentation titled
• 0:48
  Universal AI Building by Everyone. for everyone.
• 0:52
  And the core thesis here suggests that this entire
• 0:55
  top -down premise, the one everybody is chasing,
• 0:59
  might actually be fundamentally flawed. Yeah,
• 1:01
  it completely flips the script. It really does.
• 1:04
  We're going to look at the critical difference
• 1:05
  between creating these one -size -fits -all AI
• 1:07
  solutions versus actually enabling you, the listener,
• 1:10
  to build your own AI tools. We are talking about
• 1:13
  a total shift from being passive users to active
• 1:15
  builders. And if we connect this to the bigger
• 1:17
  picture, you know, we aren't just talking about
• 1:19
  a nice democratic ideal here. This isn't just
• 1:22
  philosophy. No, not at all. This is a hard -nosed
• 1:24
  business and innovation case. The material we're
• 1:28
  analyzing argues that when you shift from AI
• 1:31
  for everyone to AI by everyone, you fundamentally
• 1:34
  accelerate how technology evolves. Because the
• 1:37
  people actually dealing with the issues are the
• 1:39
  ones fixing them. Exactly. When the people closest
• 1:42
  to a problem are given the agency to build the
• 1:44
  solution themselves, you unlock this incredible
• 1:47
  resilience that centralized systems simply cannot
• 1:50
  manufacture. But I want to challenge the premise
• 1:53
  right out of the gate here, because I mean, the
• 1:55
  entire selling point of modern AI is its vast
• 1:58
  universal scale, right? Right, the massive language
• 2:00
  models. Yeah, we hear about trillions of parameters,
• 2:03
  unfathomable processing power. Are we really
• 2:06
  saying that scale isn't enough to solve everyday
• 2:08
  problems? Well, that is the crux of the argument.
• 2:10
  Scale is impressive, obviously. But the presentation
• 2:13
  introduces a principle that, quite frankly, context
• 2:16
  matters more than scale. Context over scale.
• 2:20
  Yes. And to explain why, the speaker introduces
• 2:24
  something called the 360 -degree context framework.
• 2:27
  OK, 360 -degree context. Right. It posits that
• 2:30
  there are five distinct dimensions of context
• 2:33
  that dictate whether an intelligent system actually
• 2:36
  survives contact with the real world. I want
• 2:39
  to slow down here because we hear the word context
• 2:42
  thrown around a lot in tech and it gets a bit
• 2:44
  buzzwordy. What are these five specific dimensions?
• 2:47
  Let's break them down mechanically so we understand
• 2:49
  why a multi -billion dollar AI might fail where
• 2:52
  a tiny local solution succeeds. Sure, let's walk
• 2:54
  through them. So the first is spatial context,
• 2:58
  meaning literally where is the user physically
• 3:00
  located? Okay, like their actual environment.
• 3:03
  Exactly. An AI designed for a quiet, climate
• 3:06
  -controlled office in Silicon Valley fails completely
• 3:09
  if it's deployed on a noisy, poorly -lit factory
• 3:11
  floor where a worker is wearing heavy protective
• 3:14
  gloves. The physical environment dictates the
• 3:17
  interface. Oh, that makes total sense. The centralized
• 3:19
  developers aren't thinking about the thickness
• 3:21
  of a welder's glove when they design a touchscreen.
• 3:24
  Not at all. Which brings us to the second dimension,
• 3:27
  temporal context. When is this interaction happening?
• 3:31
  Like, time of day? More like the urgency of the
• 3:33
  moment. Is the user casually browsing data at
• 3:36
  the end of the month? Or are they on an assembly
• 3:39
  line where a decision has to be made in like
• 3:41
  three seconds to prevent a machine failure? Wow,
• 3:44
  yeah. Three seconds changes everything. It completely
• 3:47
  changes how the AI must process and deliver that
• 3:49
  information. You can't give them a paragraph
• 3:51
  to read. Right. So space and time. What's number
• 3:54
  three? Operational context. And this is my favorite
• 3:57
  because it's the messy reality of how work actually
• 3:59
  gets done. As opposed to how the manual says
• 4:02
  it gets done? Exactly. Every single team has
• 4:05
  undocumented workflows, workarounds, specific
• 4:08
  local rules. If an AI doesn't understand that
• 4:11
  unofficial operational reality, it just becomes
• 4:13
  an obstacle. Oh, I've seen that firsthand. Corporate
• 4:16
  rolls out some shiny new system, but the actual
• 4:18
  team just keeps a secret spreadsheet on the side
• 4:20
  because the new system doesn't account for, say,
• 4:23
  their specific vendor's weird billing cycle.
• 4:26
  Precisely. The secret spreadsheet is the ultimate
• 4:28
  proof of operational context failing, which leads
• 4:32
  right into the fourth dimension, cognitive context.
• 4:35
  Cognitive context. So what's going on in the
• 4:37
  user's head? Basically, yeah. What is the technical
• 4:40
  baseline of the end user? And more importantly,
• 4:43
  what is their mental load in that specific moment?
• 4:46
  Right, if they're stressed or distracted. Yeah,
• 4:48
  you cannot give a highly complex, text -heavy
• 4:51
  dashboard to someone who is simultaneously managing
• 4:54
  heavy machinery and dealing with customer complaints.
• 4:57
  Their cognitive load is maxed out. That makes
• 4:59
  total sense. And the final dimension? Resource
• 5:02
  context. What are the actual constraints on the
• 5:04
  ground? Does the user have high -speed, uninterrupted
• 5:07
  internet? Or battery life, right? Exactly. Are
• 5:11
  they working on a 10 -year -old tablet in a rural
• 5:13
  area? A huge cloud AI is useless if you don't
• 5:16
  have the bandwidth to reach it. Let me try to
• 5:18
  visualize how all five of these dimensions interact,
• 5:21
  because it really feels like current top -down
• 5:23
  AI development completely ignores them. It's
• 5:26
  like, OK, it's like a massive centralized global
• 5:28
  kitchen trying to cook one single meal for the
• 5:30
  entire world. I like that analogy. Right. Like,
• 5:33
  sure, they have the biggest ovens and the absolute
• 5:35
  most expensive ingredients, but they are trying
• 5:37
  to feed a standardized dish to billions of people,
• 5:41
  completely ignoring local dietary needs, local
• 5:43
  weather, and what ingredients are actually in
• 5:45
  season in that specific town. Yes. The central
• 5:48
  length kitchen can produce at an incredible scale,
• 5:52
  but it lacks all five dimensions of local context.
• 5:54
  It doesn't know if the customer has allergies
• 5:56
  or if they even have a fork to eat the meal with.
• 6:00
  Exactly. No fork. Whereas this new approach,
• 6:03
  the buy everyone approach, it feels more like
• 6:05
  giving everyone their own community garden. A
• 6:08
  community garden, yeah. You're giving people
• 6:10
  the soil and the seeds so they can grow exactly
• 6:13
  what their local neighborhood needs, adapting
• 6:15
  to their specific climate and their own operational
• 6:18
  reality. That's spot on. Because a system isn't
• 6:21
  truly intelligent if it cannot adapt to its environment.
• 6:25
  If an AI doesn't have that 360 -degree context,
• 6:28
  it's really just a powerful engine running in
• 6:30
  a vacuum. And the only people who truly understand
• 6:33
  those five dimensions are the people standing
• 6:35
  on the ground. Right. Therefore, they need to
• 6:37
  be the ones building. the solutions. Which naturally
• 6:41
  makes me wonder, who actually figured this out?
• 6:44
  Because usually, you know, the loudest voices
• 6:47
  arguing against top -down tech deployment are
• 6:49
  industry outsiders or activists. Sure, usually.
• 6:53
  But the architect behind this entire paradigm
• 6:55
  is the furthest thing from an outsider. We are
• 6:59
  looking at the insights of Varsha Agarwal, and
• 7:01
  her background is just a master class in contrast.
• 7:04
  But what's fascinating here is that contrast.
• 7:06
  It's what makes her perspective so incredibly
• 7:09
  valuable. Yeah, you really don't see this kind
• 7:11
  of resume often. No, you rarely see someone who
• 7:14
  has operated at the absolute highest ivory towers
• 7:17
  of corporate enterprise, but who also spends
• 7:19
  her time working in the grassroots mud, so to
• 7:21
  speak. Let's look at that contrast because it
• 7:24
  proves why this methodology isn't just theory.
• 7:27
  So Varsha Agarwal is the co -founder and CEO
• 7:29
  of Bodi IQ, which focuses heavily on AI transformation
• 7:33
  and leadership. But for over 25 years, she was
• 7:36
  in the enterprise trenches, driving tech transformation
• 7:39
  at corporate titans like VMware and Sapient.
• 7:43
  Huge, massive -scale companies. Massive. And
• 7:46
  she advises startups through the Atverrin Accelerator,
• 7:49
  which is this collaboration between UCL, IIT
• 7:52
  Delhi, and IIT Madras. Plus, she holds these
• 7:55
  heavy -hitting certifications from the Indian
• 7:57
  School of Business, the Kellogg School of Management,
• 8:00
  IIM Bangalore, and Stanford. Stanford, Kellogg,
• 8:03
  yes. She has seen the very top. She knows exactly
• 8:06
  how the biggest companies in the world attempt
• 8:07
  to deploy technology. She has seen the centralized
• 8:10
  kitchen at its absolute peak. She understands
• 8:13
  the enterprise model completely, inside and out.
• 8:16
  But she also saw its limitations. Right. She
• 8:19
  realized that all the enterprise scale in the
• 8:21
  world simply doesn't matter if the tool shatters
• 8:23
  the second it hits those five dimensions of context
• 8:26
  we just discussed. And that realization is so
• 8:28
  clearly reflected in the other half of her resume,
• 8:30
  which is entirely on the ground. Yeah, the grassroots
• 8:33
  work. She is the Karnataka State Vice President
• 8:36
  of the WikiAIML Council, and through the Linostom
• 8:39
  Foundation, she has personally delivered over
• 8:41
  10 ,000 hours of hands -on AI training to underrepresented
• 8:45
  groups. 10 ,000 hours is a massive commitment.
• 8:48
  It's unbelievable. She partners with organizations
• 8:51
  like Enable India and Youth for Jobs, and she
• 8:54
  co -founded Together We Can, which is a consortium
• 8:57
  driving diversity and inclusion. So she isn't
• 8:59
  just theorizing about local context from a boardroom.
• 9:02
  Exactly. She's out there working directly with
• 9:05
  groups that traditional centralized tech development
• 9:07
  completely ignores. I mean, just think about
• 9:09
  the mechanics of delivering 10 ,000 hours of
• 9:12
  hands -on tech training to underrepresented groups.
• 9:15
  You are constantly running into friction points.
• 9:18
  Oh, the resource context alone. Right. You are
• 9:20
  dealing with severe constraints regarding internet
• 9:22
  access, hardware limitations, diverse educational
• 9:25
  backgrounds. A developer sitting in a pristine
• 9:28
  climate -controlled office in San Francisco never
• 9:31
  experiences that kind of friction. Which leads
• 9:33
  us directly into her first specific framework
• 9:35
  for how to actually build universal AI. It's
• 9:39
  this concept of building from the edge. Building
• 9:41
  from the edge, yes. The idea is that instead
• 9:43
  of designing a tool for the so -called average
• 9:46
  user, you start by designing for the extreme
• 9:49
  edge cases. But, and here's where it gets really
• 9:52
  interesting. I have to push back on behalf of
• 9:54
  the listener for a second. Go for it. If you
• 9:56
  design something for an extreme highly constrained
• 9:59
  edge case, how does that possibly end up working
• 10:01
  for the average user. It just seems like you'd
• 10:04
  be building a hyper niche product that only works
• 10:06
  for one tiny group. It is a totally logical assumption
• 10:10
  to think that an edge case tool is useless for
• 10:12
  the mainstream. We're sort of trained to think
• 10:14
  that way. But the reality is the exact opposite.
• 10:18
  Really? Yeah. How so? The key takeaway from Agarwal's
• 10:21
  presentation is a phrase that completely reorients
• 10:24
  how we should think about engineering. And the
• 10:26
  phrase is constraints reveal universal truths.
• 10:31
  Constraints reveal universal truths. Okay, break
• 10:34
  down how that actually works mechanically, because
• 10:37
  it sounds a bit philosophical. Let's use an analogy
• 10:41
  outside of software first. Think about Formula
• 10:43
  One racing. Okay, fast cars. Right. Engineers
• 10:46
  build those cars for the absolute extreme edge
• 10:49
  of automotive performance. They face insane constraints
• 10:53
  regarding weight, aerodynamics, heat dissipation,
• 10:56
  things a normal driver never thinks about. Because
• 10:58
  they're pushing the absolute limits of physics.
• 11:00
  Exactly. Now, the average person is never going
• 11:02
  to drive a Formula One car to the grocery store.
• 11:04
  Definitely not. But the technologies forged under
• 11:07
  those extreme constraints, things like anti -lock
• 11:10
  brakes, advanced aerodynamics, energy recovery
• 11:13
  systems, those eventually make their way into
• 11:15
  the standard commuter cars that we drive every
• 11:17
  single day. The constraint actually forced a
• 11:20
  superior innovation. Oh, I see the connection.
• 11:23
  So how does that apply to building artificial
• 11:25
  intelligence? Well, When people who are operating
• 11:28
  under severe constraints build a solution themselves,
• 11:31
  they cannot afford bloated features. Right. They
• 11:34
  can't have a bunch of unnecessary bells and whistles.
• 11:37
  Exactly. Let's say you are empowering a community
• 11:40
  in a rural area to build an AI tool for local
• 11:44
  agriculture. They have low bandwidth and varying
• 11:46
  literacy levels. So the cognitive and resource
• 11:49
  contexts are highly constrained? Precisely. They
• 11:52
  aren't going to build a complex, text -heavy,
• 11:55
  high latency dashboard. It wouldn't work. They're
• 11:57
  going to build something that relies on lightning
• 11:59
  fast, highly accurate voice recognition, and
• 12:02
  simple visual outputs. They strip away all the
• 12:05
  assumptions. And because they had to make the
• 12:07
  voice recognition flawless just to make the tool
• 12:10
  function at all in that environment, that core
• 12:12
  technology ends up being vastly superior to the
• 12:15
  clunky text -based interfaces the centralized
• 12:17
  companies are building for everyone else. Exactly.
• 12:20
  When you strip away the fluff and build a solution
• 12:22
  robust enough to survive at the extreme edge,
• 12:26
  the resulting innovation is incredibly adaptable.
• 12:28
  Wow. It turns out that a stripped -down, hyper
• 12:31
  -efficient AI built to overcome severe constraints
• 12:34
  is usually much more intuitive for the corporate
• 12:36
  workers sitting at a desk, too. Because it hasn't
• 12:39
  been coddled. It's battle -tested. So building
• 12:42
  from the edge creates universal AI. You weren't
• 12:45
  building a niche product at all. You were building
• 12:47
  a deeply grounded fundamentally sound core. Yes.
• 12:51
  But that covers the external societal framework.
• 12:54
  The material also details an internal organizational
• 12:57
  framework. How do companies actually implement
• 12:59
  this mindset inside their own walls? This brings
• 13:02
  us to Agarwal's second design principle, building
• 13:05
  with people. Building with people. OK. The objective
• 13:09
  here is to transform an organization from a group
• 13:11
  of passive AI consumers into a network of active
• 13:14
  AI co -creators. You have to build what she calls
• 13:17
  a contribution culture. Okay, let's pause right
• 13:20
  here because I want to cut through the corporate
• 13:22
  jargon for a second. Contribution culture, and
• 13:25
  another phrase the presentation uses, learning
• 13:27
  loops and the flow of work. They sound great
• 13:30
  in theory, like on a PowerPoint slide. Sure,
• 13:32
  they sound very consultant -heavy. Right. But
• 13:35
  if I am a listener sitting at my desk on Monday
• 13:38
  morning, what does building AI in the flow of
• 13:40
  work actually look like for me? Let's use a very
• 13:43
  concrete example to ground this. Imagine a logistics
• 13:46
  manager at a regional warehouse. OK, got it.
• 13:49
  Every Tuesday, a specific vendor delivers a shipment
• 13:51
  late, which causes a huge bottleneck on the warehouse
• 13:54
  floor. Now, the central supply chain software
• 13:57
  doesn't flag this because it's a hyper local
• 13:59
  issue that only affects this one facility due
• 14:02
  to, say, a weird local traffic pattern. Right.
• 14:05
  The operational and the spatial context is invisible
• 14:08
  to the cloud. And under the old model, that logistics
• 14:11
  manager would just have to suffer or submit an
• 14:14
  IT ticket. Oh, the dreaded IT ticket. They'd
• 14:17
  have to explain the local traffic pattern to
• 14:19
  a developer sitting in another state, wait three
• 14:21
  months for a software patch, and by the time
• 14:23
  the patch actually arrives, the vendor has changed
• 14:26
  their delivery day anyway. The context is totally
• 14:28
  lost. Exactly. That is the passive consumer model.
• 14:32
  It's incredibly slow. Right. But in a contribution
• 14:35
  culture, that logistics manager has access to
• 14:38
  decentralized, low -code or no -code AI tools.
• 14:41
  So they can build it themselves? Yes. On a Monday
• 14:43
  morning, in the normal flow of their work, they
• 14:45
  open their system and use natural language to
• 14:47
  tell the AI, look at the delivery logs for Vendor
• 14:50
  X over the last six months. Notice the Tuesday
• 14:53
  delays. Create a rule that automatically reschedules
• 14:56
  floor staff to Wednesday mornings when Vendor
• 14:58
  X is scheduled. And send an automated alert to
• 15:01
  the floor supervisor. Wait, they just built a
• 15:03
  highly contextual bespoke micro application in
• 15:06
  like 10 minutes. And that is the learning loop.
• 15:09
  The manager faced a problem, built a solution
• 15:11
  on the spot, the AI learned a new rule based
• 15:14
  on local context, and the problem is solved immediately.
• 15:18
  That's incredible. The innovation didn't happen
• 15:20
  in an R &D lab. It happened in the flow of work.
• 15:23
  By lowering the barriers to entry, you completely
• 15:25
  remove the friction between the person who sees
• 15:27
  the daily problem and the person who builds the
• 15:30
  solution. because they are now the exact same
• 15:32
  person. Exactly. What really strikes me about
• 15:35
  that warehouse example is that making AI accessible
• 15:38
  isn't just some HR initiative. It's not just
• 15:41
  a box to check for corporate social responsibility.
• 15:44
  Not at all. It is a fundamental, hard -nosed
• 15:47
  strategy to accelerate productivity. I mean,
• 15:50
  if you have thousands of employees constantly
• 15:52
  building these tiny hyper contextual micro solutions
• 15:55
  every single day, the iteration speed of your
• 15:58
  company just goes through the roof. The business
• 16:00
  case is undeniable. You are upgrading the entire
• 16:03
  organization's capability to solve problems in
• 16:05
  real time. You are moving from a fragile static
• 16:08
  system to a dynamic living ecosystem. So what
• 16:12
  does this all mean when we pull these pieces
• 16:13
  together? Well, it's a massive paradigm shift.
• 16:16
  It really is. The ultimate takeaway from Varsha
• 16:19
  Agarwal's insights is that universal AI will
• 16:22
  not arrive in a shiny shrink -wrapped box from
• 16:24
  Silicon Valley. No, it won't. It won't be achieved
• 16:27
  by handing down a finished, one -size -fits -all
• 16:29
  product. True universal AI will only be achieved
• 16:33
  when the tools of creation are fully democratized.
• 16:36
  When the people actively navigating the messy
• 16:38
  reality of the real world are the ones actually
• 16:41
  authoring the solutions. Whether that's an underrepresented
• 16:44
  group at the extreme edge of society. or a logistics
• 16:47
  manager right in the middle of a corporate warehouse.
• 16:50
  Exactly. But this raises an important question.
• 16:53
  It requires a deliberate, intentional change
• 16:56
  in mindset. We have to actively recognize that
• 16:58
  passive consumption of technology is just a dead
• 17:01
  end. It completely reframes how we should interact
• 17:03
  with the digital tools sitting right in front
• 17:05
  of us, which leaves me with the final kind of
• 17:08
  lingering thought for you to chew on as we wrap
• 17:10
  up today's Deep Dive. Ooh, let's hear it. If
• 17:13
  we actually follow this trajectory, I mean, if
• 17:15
  the barriers to building AI truly vanish and
• 17:18
  these continuous learning loops allow absolutely
• 17:20
  anyone to build a bespoke AI solution right in
• 17:23
  the flow of their daily work, what happens to
• 17:26
  the concept of a software company? Wow. That
• 17:30
  is a fascinating thread to pull. Right? Does
• 17:34
  the traditional software developer, the person
• 17:36
  sitting in that centralized kitchen, writing
• 17:38
  code for the masses, do they become obsolete?
• 17:42
  Or does every single employee, regardless of
• 17:44
  whether they are in human resources, marketing,
• 17:47
  or logistics, effectively become their own microdeveloper
• 17:50
  ecosystem? If we truly move from AI for everyone
• 17:54
  to AI by everyone, the entire landscape of who
• 17:59
  builds our reality changes overnight. It forces
• 18:02
  us to entirely redefine what it means to be a
• 18:05
  creator in the modern economy. It really does.
• 18:08
  It's a completely new world. It is. Well, we
• 18:11
  want to warmly thank you for joining us on this
• 18:13
  deep dive today. As you head back to your day,
• 18:16
  take a hard look at the software and the AI tools
• 18:19
  you're using right now and ask yourself, am I
• 18:22
  just passively consuming this or could I be co
• 18:25
  -creating with it? We'll see you next time.