Latest / Today’s AI News / DeepMind's CEO on AGI by 2030, Jensen Pushes Back on AI-Proof Careers, AI Hiring Is Biased

Transcript

DeepMind's CEO on AGI by 2030, Jensen Pushes Back on AI-Proof Careers, AI Hiring Is Biased · Today’s AI News

• 0:00
  Building AI right now is like constructing A supersonic jet
• 0:04
  while we're midair. We're still figuring out how
• 0:06
  gravity works, but we're we're already arguing about where
• 0:10
  we're going to land. And according to today's
• 0:12
  sources, well, that landing is coming much sooner than we
• 0:15
  thought. A.
• 0:15
  Lot sooner, yeah. Welcome back to today's AI News,
• 0:18
  brought to you by 9X Productions.
• 0:20
  Today, we're taking you on a deep dive from Google Deepmind's
• 0:23
  new, very concrete timeline for AGI straight into the real,
• 0:28
  honestly pretty messy impact AI is having on the job market
• 0:31
  today. And.
• 0:32
  We're also going to breakdown a practical workflow for
• 0:34
  automating your weekly reporting, plus a bunch of quick
• 0:36
  hits across the industry. Yeah.
• 0:38
  It is a dense stack of sources today.
• 0:40
  So let's just get right into this DeepMind interview.
• 0:42
  Let's do it. So Demis Sasabas, the CEO of
• 0:44
  Google DeepMind just gave this massive exclusive interview and
• 0:48
  he put a real stake in the ground.
• 0:50
  He is confidently predicting that we will achieve AGI by
• 0:54
  20-30. You know, give or take a year.
• 0:56
  Which is wild to hear on record. Right.
• 0:58
  But looking at where our models are today, that feels almost, I
• 1:02
  don't know, premature. What is he actually seeing that
• 1:04
  bridges the gap between the chat bots we use now and actual
• 1:08
  general intelligence? Well, he's looking really
• 1:10
  closely at the specific technical hurdles separating
• 1:13
  current LLMS from that ultimate goal.
• 1:16
  He actually called out for missing pieces explicitly.
• 1:19
  OK, what are they? So it's world physics, memory
• 1:22
  consistency and continual learning, He's saying once those
• 1:26
  four are solved, the entire architecture fundamentally
• 1:29
  changes. OK, let's stick on world physics
• 1:31
  and memory for a second. Because the way I look at
• 1:33
  current models, it's like, it's like we've built a genius with
• 1:37
  amnesia in a sensory deprivation tank.
• 1:39
  Oh, that's a good way to put. It can pass the bar exam right,
• 1:42
  but it doesn't remember what you told it yesterday, and it
• 1:45
  doesn't fundamentally understand how a ball bounces.
• 1:48
  Solving those two things bridges the gap between a really good
• 1:51
  calculator and a digital brain. Yeah, that analogy is close.
• 1:54
  But I mean, the reality is even stranger.
• 1:56
  It's not just that the AI has amnesia, its memory is entirely
• 2:01
  stateless. Stateless meaning what exactly?
• 2:04
  Meaning every single time you open a new prompt window, the
• 2:08
  model is essentially born anew. It has no continuous episodic
• 2:12
  memory of its interactions with you or, you know, the world over
• 2:16
  time. So it's just reacting in a
• 2:17
  vacuum every single time. Exactly.
• 2:19
  And then regarding the world physics part, current models are
• 2:23
  strictly just predicting the next most likely token based on
• 2:27
  massive text data sets. Right, the autocomplete on
• 2:29
  steroids thing. Yeah.
• 2:31
  So if you type I dropped a glass cup on a tile floor, the model
• 2:35
  outputs the word shatters. But it does that because those
• 2:38
  words are statistically correlated in his training data.
• 2:41
  Not because it knows what glasses.
• 2:43
  Right. It does not possess A spatial 3
• 2:44
  dimensional understanding of gravity or fragility or physical
• 2:47
  impact. So solving world physics
• 2:49
  requires moving beyond just text prediction to multimodal systems
• 2:54
  that actually simulate physical environments natively.
• 2:56
  And Hassabus mentioned that once we crack the simulation
• 2:59
  environments, timelines for things like drug discovery just
• 3:02
  hardened significantly. Right?
• 3:04
  Moving from theory to concrete reality.
• 3:06
  Yeah. And DeepMind is aggressively
• 3:08
  targeting oncology and immunology first.
• 3:10
  The mechanism here is just fascinating.
• 3:12
  How? Would an AGI actually do that
• 3:14
  though? Well, it wouldn't just, you
• 3:16
  know, read medical journals to guess it cures.
• 3:18
  It would simulate biological processes at a molecular level.
• 3:22
  It could test millions of complex protein folds or
• 3:24
  chemical interactions in a digital sandbox.
• 3:27
  Oh wow, so you take the physical trial and error process of a
• 3:30
  lab, which usually takes what, decades?
• 3:32
  Decades, yeah, yeah, and you can press it into days of
• 3:35
  computation. The end game he described is an
• 3:37
  AI engine capable of curing any disease.
• 3:40
  Opens up a completely wild philosophical question for you
• 3:43
  listening. If AI handles all the heavy
• 3:46
  analytical lifting, Hasabis argues that human traits,
• 3:49
  specifically our taste, our original thinking and emotional
• 3:52
  connections, will become the most valuable assets we have.
• 3:55
  Yeah, it's suggested in a post AGI world.
• 3:57
  Flawless execution is basically commoditized, so our subjective
• 4:01
  human imperfections become our unique currency.
• 4:04
  And that philosophical view from Hasabis is actually colliding
• 4:07
  right now with some advice coming from the hardware side of
• 4:10
  the industry. Jensen Huang, the CEO of NVIDIA,
• 4:13
  just weighed in on what people should actually be focusing on
• 4:17
  as the job market shifts. And this is huge because there
• 4:21
  is massive panic among parents and students right now.
• 4:24
  Everyone is desperately searching for an AI proof major.
• 4:28
  Right, everyone wants to know what's safe.
• 4:29
  Exactly. But Huang is telling people to
• 4:32
  stop trying to dodge the technology.
• 4:34
  He says the focus should be on how AI can elevate your specific
• 4:38
  craft instead. He used journalism as a prime
• 4:41
  example, which is interesting, but I mean, isn't writing
• 4:43
  articles one of the primary things LLMS are currently
• 4:46
  automating away? Routine copy generation, yeah,
• 4:48
  absolutely being automated. But Huang separates it's that
• 4:51
  from the core of journalism, which he says requires dynamic
• 4:54
  listening. Dynamic listening.
• 4:55
  Right when a human journalist conducts a high stakes
• 4:58
  interview, they read the room. They pick up on hesitation,
• 5:01
  underlying anxieties, microexpressions, shifts in
• 5:04
  tone. They pivot dynamically based on
• 5:06
  human empathy. Exactly.
• 5:08
  And AI can instantly generate a list of 20 interview questions
• 5:11
  based on a topic, but it cannot currently read the subtle,
• 5:14
  unspoken context of a human interaction.
• 5:18
  And Huang actually tied this back to a Japanese aesthetic
• 5:21
  concept, didn't he? Wabi sabi.
• 5:22
  Yes, wabi sabi. It's the acceptance of
• 5:24
  transience and the beauty of imperfection, valuing the
• 5:28
  weathered, the asymmetric, the uniquely human.
• 5:31
  So the idea is that as AI makes everything perfectly polished in
• 5:35
  uniform, we'll crave the imperfections.
• 5:37
  Right society will naturally prize wabi sabi across all
• 5:40
  domains. It's a nice thought honestly,
• 5:43
  but he also called the narrative tying AI to job cuts lazy,
• 5:47
  claiming AI has just arrived so it can't be causing job losses
• 5:50
  yet and the data completely contradicts him.
• 5:53
  Oh, 100%. We've seen over 80,000 jobs
• 5:56
  explicitly cut this year alone in favor of AI restructuring.
• 5:59
  It's easy for the guy selling the shovels to say the gold rush
• 6:02
  won't ruin the farming industry, you know?
• 6:04
  Yeah, that's a very convenient position for NVIDIA.
• 6:06
  Exactly how do we square this idea of embracing human
• 6:10
  imperfection when companies are actively deploying automated
• 6:13
  gatekeepers to aggressively trim their workforces?
• 6:17
  You really can't easily square it, especially when those
• 6:20
  automated gatekeepers are fundamentally flawed.
• 6:23
  We just got the results of a massive new Stanford study
• 6:26
  analyzing AI hiring bias, and it shows exactly what happens when
• 6:30
  you replace human judgement with algorithmic sorting at an
• 6:33
  industrial scale. Let's get into those numbers
• 6:36
  because they are incredibly stark.
• 6:38
  They are The Stanford researchers analyzed 4 million
• 6:41
  job applications. 4 million. That is a huge sample size.
• 6:45
  Huge. And it was across 156 employers
• 6:48
  using data from π metrics, they found clear measurable racial
• 6:51
  bias. What were the specific metrics?
• 6:53
  Specifically, they found a 10.62% adverse impact against
• 6:57
  Black applicants and a 5.32% adverse impact against Asian
• 7:01
  applicants. These demographics are being
• 7:03
  disproportionately screened out by these tools.
• 7:05
  What is the actual mechanism causing this?
• 7:08
  Because I think a lot of people assume it's just, you know, a
• 7:10
  single bias programmer writing bad code.
• 7:13
  But how does the AI actually learn to screen out specific
• 7:17
  demographics? It comes down to the training
• 7:19
  data and proxy variables. Proxy variables, yeah.
• 7:22
  These AI models are trained on historical hiring data to
• 7:25
  recognize what a successful candidate looks like for a given
• 7:29
  role. But historical hiring data is
• 7:32
  full of past human biases. So it just replicates what we've
• 7:36
  already done. Exactly.
• 7:37
  So the model looks for patterns and starts latching onto proxy
• 7:40
  variables. What does a proxy variable look
• 7:42
  like on a resume? Like practically speaking it.
• 7:45
  Could be a specific zip code, participation in certain
• 7:47
  extracurricular activities, the name of the university, or even
• 7:51
  subtle linguistic patterns and how a candidate describes their
• 7:53
  experience. And those correlate with race.
• 7:56
  Heavily. They correlate heavily with race
• 7:58
  or socioeconomic background. The the AI doesn't know it's
• 8:01
  discriminating, it is just relentlessly optimizing for the
• 8:04
  historical mold of a good fit based on the flawed data it was
• 8:08
  fed. And the scale of this is
• 8:10
  terrifying because of the shared infrastructure.
• 8:12
  The study pointed out that across those 156 employers,
• 8:16
  there were only 42 AI models being utilized.
• 8:19
  That is the compounding danger right there.
• 8:21
  Yeah. When companies buy off the shelf
• 8:23
  hiring tools from major vendors, they inherit that vendors model
• 8:27
  and all of its unseen blind spots.
• 8:30
  You get rejected by 1 employer because the AI mistakenly flags
• 8:33
  your resumes formatting or heavily weights a proxy variable
• 8:37
  against you. And ten other companies use that
• 8:39
  exact same model. You don't just get rejected once
• 8:42
  you get shadow banned from the industry.
• 8:43
  Wow. And the sources noted that 4% of
• 8:46
  applicants who applied to 10 different positions were
• 8:48
  rejected from all of them. Just completely locked out.
• 8:50
  If you are a job seeker listening to this right now,
• 8:53
  that has to be the most frustrating thing in the world.
• 8:55
  The philosophical debate about wabi sabi doesn't pay the rent
• 8:57
  when an invisible shared algorithm is blocking you.
• 9:00
  No, it doesn't. To add a little context, the
• 9:02
  researchers did know this data covers 2018 to 2022, and today's
• 9:07
  LLM driven hiring tools process information quite differently.
• 9:11
  But the structural danger is still there.
• 9:14
  Exactly when a bias exists in a centralized foundational model,
• 9:18
  it skills instantly across thousands of companies.
• 9:22
  And the reality is, whether we like it or not, these tools are
• 9:25
  integrating deeply into corporate workflows.
• 9:28
  We are seeing individuals use them to automate massive chunks
• 9:32
  of their own daily tasks just to stay competitive.
• 9:36
  Which is becoming a required survival skill at this point,
• 9:38
  yeah. There is a breakdown in today's
• 9:40
  sources on setting up a repeatable Claude cowork process
• 9:44
  to handle a weekly marketing report.
• 9:47
  I really want to understand how this actually operates under the
• 9:49
  hood, because it sounds far more advanced than just, you know,
• 9:52
  typing a prompt into a chat window.
• 9:54
  Oh, it is entirely different from a standard chat interface.
• 9:57
  It requires setting up dedicated digital infrastructure on your
• 10:00
  local machine or a cloud drive. OK, so how do I actually build
• 10:04
  this? First, you create a main
• 10:05
  repository, let's call it Marketing Metrics Review.
• 10:08
  OK, simple enough folder. Right inside that main folder
• 10:11
  you build 4 distinct subfolders. You name them inputs, working
• 10:15
  skills and outputs. Wait, what exactly goes into the
• 10:18
  skills folder? That sounds like the brain of
• 10:20
  the operation. It is the skills folder houses
• 10:23
  your system prompts. It's the core set of
• 10:25
  instructions guiding the AI agent.
• 10:27
  Text files with instructions. Exactly.
• 10:30
  You write a text file telling Claude to act as a data
• 10:32
  aggregator. You instructed to systematically
• 10:35
  pull summaries of last week's Slack messages, check your Gmail
• 10:38
  for sales updates, review your Google Analytics for revenue
• 10:42
  numbers, and extract meeting notes from Google Drive.
• 10:45
  So it's actively pulling from multiple APIs and software
• 10:48
  integrations all at once, Yes. The agent executes those tasks
• 10:52
  and dumps all that raw data into the inputs folder.
• 10:56
  And then what? Once those inputs are gathered,
• 10:58
  a second skillful another text file in that folder tells Claw
• 11:01
  to analyze that raw data. It drafts the marketing report,
• 11:05
  a leadership brief, and a Slack update, and places those drafts
• 11:08
  in the working folder for human review.
• 11:10
  Oh I see, so it waits for me and working.
• 11:12
  Right, once you verify the numbers, the polished versions
• 11:14
  move to outputs. That is incredibly slick, but
• 11:18
  the guide mentioned a specific pro tip about a feedback loop
• 11:20
  that makes this smarter over time.
• 11:23
  How does that work technically? This is the most important part.
• 11:26
  After you review the drafts in the working folder, you don't
• 11:30
  just fix the errors yourself. You ask the AI to do it.
• 11:33
  You actively ask Claude what did you get wrong and what changed
• 11:38
  about our formatting this week. You take the AI's feedback and
• 11:42
  manually update the text files in the skills folder.
• 11:45
  So you are literally rewriting the agent score instructions
• 11:48
  every single week so it starts making the same mistakes.
• 11:51
  Exactly. It gets better every single
• 11:53
  iteration. It is exactly like hiring an
• 11:55
  eager intern who has the keys to your entire digital life.
• 11:59
  Pretty much, yeah. Is the real skill of the future
• 12:01
  not writing the reports at all, but rather becoming an expert
• 12:04
  manager of these AI agents? The industry is absolutely
• 12:08
  betting on that shift. We are seeing a massive pivot
• 12:11
  toward agentic infrastructure right now from all the major
• 12:14
  players. Well, let's run through the
• 12:16
  quick hits then, because the Asian ecosystem is exploding as
• 12:19
  we speak. Perplexity just rolled out their
• 12:22
  Computer Cloud agent and it is now actively managing entire
• 12:26
  Shopify stores. And it is not just writing
• 12:28
  product descriptions either. It is handling the operational
• 12:31
  back end, inventory checks, customer routing, the whole
• 12:35
  thing. That's wild.
• 12:36
  Anthropic also just pushed an update to their coding agent
• 12:39
  Claude Code. They added a security guidance
• 12:41
  plug in to ensure these autonomous tools don't, you
• 12:44
  know, accidentally write vulnerable software.
• 12:46
  Which is very necessary. Plus we are seeing back end
• 12:49
  tools like Parse 2 Point O launching which extends document
• 12:53
  parsing API specifically built so these AI agents can read and
• 12:57
  comprehend incredibly dense complex files.
• 13:00
  That kind of fundamental reasoning capability is exactly
• 13:03
  what developers are chasing, which actually ties directly
• 13:06
  into the math breakthrough from Anthropics technical team.
• 13:08
  Yes, the Erdi's problem. Right, Sholto Douglas revealed
• 13:11
  that their Mythos model solved the Erdi's problem #90 now Open
• 13:15
  AI cracked this recently, but Anthropics model matched it with
• 13:18
  a simpler proof. For context on why the industry
• 13:21
  is paying so much attention to this, the Erda's problem #90 is
• 13:25
  a notoriously difficult puzzle in number theory.
• 13:28
  It involves unit fractions and subsets.
• 13:30
  Right. Yeah, calculating sums of unit
• 13:32
  fractions. It has stumped human
• 13:34
  mathematicians for decades. But why is a simpler proof the
• 13:37
  gold standard here? Why does that matter more than
• 13:39
  just solving it? Because it demonstrates genuine
• 13:42
  deductive reasoning rather than just brute force computation.
• 13:46
  A complex proof might just exhaust every possible
• 13:49
  mathematical combination until it stumbles onto the answer.
• 13:52
  Like a machine just trying every key on the key ring.
• 13:55
  Exactly. But a simple, elegant proof
• 13:57
  means the model discovered a direct, fundamentally sound
• 14:01
  logical pathway. It shows a real leap in abstract
• 14:04
  thinking. And that level of abstract
• 14:05
  reasoning is being treated like a national security asset at
• 14:08
  this point. Bloomberg is reporting that
• 14:10
  China is now imposing overseas travel restrictions on top AI
• 14:14
  researchers from major labs like Alibaba and DeepSeek.
• 14:17
  It is a very aggressive talent retention strategy.
• 14:21
  China is trying to close the capability gap with the US tech
• 14:24
  sector, and restricting travel prevents brain drain.
• 14:28
  They just lock them in. Pretty much they do not want
• 14:31
  their top minds being recruited by Western labs or attending
• 14:35
  international conferences where they might share proprietary
• 14:38
  techniques. It's like a Cold War space race,
• 14:40
  but instead of building expensive rockets, the weapons
• 14:43
  are getting exponentially cheaper every single day.
• 14:46
  Unbelievably cheap. Look at Xiaomi in China.
• 14:49
  They just permanently slashed the pricing for their MIMO V 2.5
• 14:54
  API by up to 99% while simultaneously increasing token
• 14:58
  allowances by 5 to 8 times. They're just flooding the market
• 15:02
  with cheap compute. If you are a developer listening
• 15:04
  to this right now, a 99% place drop fundamentally changes your
• 15:07
  math. You don't even need a local
• 15:09
  model anymore. Exactly.
• 15:10
  You don't have to decide whether to build a localized smaller
• 15:12
  model. You just rely on Xiaomi's
• 15:14
  massive cloud infrastructure for pennies.
• 15:17
  They are giving away access just to capture developer market
• 15:19
  share. And developers are consuming it
• 15:22
  at an unbelievable rate. Open Router, which is the
• 15:25
  platform that routes traffic between all these different AI
• 15:28
  models so developers can easily switch between them.
• 15:30
  They just raised $113 million. And their usage stats are
• 15:35
  insane. Yeah, they hit a 1.5 quadrillion
• 15:39
  token annual run rate with 8 million developers on the
• 15:42
  platform. A quadrillion.
• 15:43
  A quadrillion tokens. It is hard to even conceptualize
• 15:47
  that that number. It represents the sheer
• 15:49
  overwhelming volume of API calls happening globally every single
• 15:54
  second and. Those tokens aren't just
• 15:55
  generating text anymore either. 11 Labs just released Music V2.
• 15:59
  It features better vocals, improved instrumentation,
• 16:02
  multilingual support, and they added track level in painting.
• 16:06
  Track level in painting is a massive workflow upgrade for
• 16:09
  audio engineers. How did they?
• 16:10
  Do it before. Historically, if an AI generated
• 16:12
  A2 minute song but messed up a single guitar riff or hit a
• 16:15
  weird high note, you had to regenerate the entire track and
• 16:18
  just hope for the best. Oh.
• 16:19
  That sounds misery. It was, but with in painting you
• 16:23
  can highlight that specific 2 second segment and tell the AI
• 16:26
  to regenerate just that part, blending it seamlessly into the
• 16:30
  existing audio. So with quadrillions of tokens,
• 16:33
  geopolitical maneuvering, and AGI timelines flying around, it
• 16:37
  is incredibly easy to get lost in the sheer scale of this
• 16:41
  technology. Very easy, yeah.
• 16:42
  I want to bring it all back down to earth with today's community
• 16:45
  AI workflow. This is a great practical use
• 16:48
  case from a listener named Todd K out of Chicago.
• 16:50
  Right. Todd and his best friend, who
• 16:52
  lives in Minnesota, use a shared ChatGPT conversation, literally
• 16:56
  just titled Travel, to coordinate all their meetups.
• 17:00
  It's so simple but so effective. They used it to plan a Cubs
• 17:03
  weekend in Chicago. They are in the shared chat,
• 17:06
  collaboratively growing prompts at the AI from different states.
• 17:10
  They're asking it to cross reference BBQ spots near Wrigley
• 17:13
  Field. Right, yeah, checking The
• 17:14
  Walking distance to post game bars, hunting down the best
• 17:17
  obscure donut shops on the north side.
• 17:19
  The AI basically acts as a shared hyper intelligent scratch
• 17:22
  pad. It remembers the context of
• 17:25
  everything they ask over weeks of planning.
• 17:27
  And the brilliant part is the execution at the end.
• 17:30
  The day before the trip they just asked the AI to look at all
• 17:34
  their chaotic brainstorming and organize it into a perfectly
• 17:37
  optimized hour by hour itinerary.
• 17:39
  There is a deep, profound irony here.
• 17:42
  What's that? We are talking about the most
• 17:44
  advanced computational technology in human history?
• 17:47
  The same architecture that is solving Airdo's math problems
• 17:51
  and simulating molecular protein folds for drug discovery is
• 17:54
  being used by two buddies to coordinate a hunt for the
• 17:57
  perfect glazed doughnut. It's amazing.
• 18:00
  It grounds the technology in everyday utility.
• 18:02
  But if we pull this all the way back to Jensen Wong's concept of
• 18:06
  wabi sabi, it introduces a really fascinating friction
• 18:10
  point. Oh, where do you see the
• 18:12
  friction? Well, as these AI agents get
• 18:14
  closer to AGI, they're going to slawlessly handle all our
• 18:17
  planning. They will optimize the itinerary
• 18:19
  perfectly. Right.
• 18:20
  The dinner reservations will align Florida State with The
• 18:22
  Walking distance to the bar. Exactly.
• 18:25
  But will that complete removal of friction also remove the
• 18:28
  human imperfection, the wabi sabi?
• 18:30
  I see what you're saying. If the AI perfectly optimizes
• 18:34
  your weekend, you might lose the magic of taking a wrong turn,
• 18:38
  stumbling into the wrong dive bar, and having the happy
• 18:41
  accidents that make travel memorable in the 1st place.
• 18:43
  Wow. Are we optimizing away the happy
• 18:46
  accidents? That is a phenomenal question.
• 18:48
  It's something to think about. Definitely something for you to
• 18:51
  chew on while you eat your perfectly AI curated doughnut.
• 18:55
  That wraps up our deep dive for today.
• 18:57
  If you enjoyed the conversation, make sure you hit subscribe so
• 18:59
  you stay updated on all things AI.
• 19:02
  And hey, if you really liked it, please drop us a five star
• 19:04
  rating. It helps us keep doing this for
• 19:06
  you. We appreciate you spending your
• 19:08
  time with us. We are building the supersonic
• 19:11
  jet midair, but we are figuring it out together.
• 19:14
  We will see you back here for tomorrow's deep dive.