Latest / Today’s AI News / Perplexity Becomes a Finance App, Amazon Reveals Its AI Revenue, Oxford’s Heart Failure Breakthrough

Transcript

Perplexity Becomes a Finance App, Amazon Reveals Its AI Revenue, Oxford’s Heart Failure Breakthrough · Today’s AI News

• 0:00
  Imagine trying to build a faster horse, but then suddenly
• 0:03
  realizing you could just build a teleporter instead.
• 0:05
  Oh, I love that. Right.
• 0:06
  I mean, that is exactly what is happening in the AI space right
• 0:09
  now. Companies are stopping trying to
• 0:12
  just, you know, answer our questions, and they're starting
• 0:14
  to try to actually do our chores, manage our money, and
• 0:17
  even diagnose our hearts. Yeah, it's a completely
• 0:20
  different ball game. It really is.
• 0:21
  Welcome back to today's AI News, brought to you by 9X
• 0:25
  Productions. We are taking a deep dive into a
• 0:28
  stack of fresh updates today, and if you are listening to this
• 0:32
  on your commute or at your desk, what we are about to unpack
• 0:35
  fundamentally changes your relationship with the technology
• 0:39
  sitting in your pocket. Absolutely.
• 0:40
  We are looking at this massive industry wide shift from passive
• 0:45
  chat bots to active autonomous agents.
• 0:48
  Yeah, because you know, we have spent the last couple of years
• 0:51
  just fascinated by digital thought.
• 0:53
  We type a prompt and AI generate some text at some surprises, an
• 0:57
  article or reads a poem, but it is a closed loop of information.
• 1:02
  Exactly, just giving you answers, right?
• 1:05
  But today's deep dive is really about watching AI cross that
• 1:10
  threshold from digital thought into tangible, real world
• 1:14
  action. We are entering the agentic
• 1:16
  economy. Software doesn't just tell you
• 1:19
  how to do something anymore. It reaches out into the digital
• 1:22
  world and actually does it for you.
• 1:24
  Well, let's guide straight into Perplexity because they are kind
• 1:28
  of the poster child for this pivot right now.
• 1:30
  For sure. For a long time, the narrative
• 1:32
  was just that Perplexity was a really smart search engine
• 1:35
  trying to chip away at Google's market share.
• 1:37
  But they are going after a completely different pie now.
• 1:40
  They aren't just targeting the search bar anymore, they are
• 1:43
  going after utility apps like Mint or TurboTax.
• 1:47
  Which is huge. It is.
• 1:48
  They recently rolled out their computer agent and it now
• 1:51
  features a direct integration with Plaid.
• 1:53
  And for those so don't know, Plaid is essentially the
• 1:56
  underlying plumbing for the financial Internet, right?
• 1:58
  Right. Like when you connect a
• 1:59
  budgeting app to your bank account, you are almost always
• 2:03
  routing through Plaid. It connects to over 12,000
• 2:06
  financial institutions. 12,000 so with this new integration
• 2:10
  perplexities agent can just pull your checking your credit, loan
• 2:14
  and brokerage data. Wow.
• 2:16
  Yeah, now it is a read only view, but you can use simple
• 2:20
  natural language text prompts to have the agent build customized
• 2:24
  budgets, net worth trackers, debt payoff plans.
• 2:28
  Even interactive retirement dashboards.
• 2:30
  Exactly. And this comes right on the
• 2:32
  heels of their U.S. tax integration that autonomously
• 2:35
  fills out IRS forms. Which people hate doing
• 2:38
  themselves. Oh, they despise it.
• 2:39
  And the market is clearly responding.
• 2:42
  This pivot just pushed Perplexity's annual recurring
• 2:45
  revenue past $450 million in March.
• 2:48
  That's a massive number. It is a 50% jump in a single
• 2:52
  month. Honestly, to see 50% jump in a
• 2:54
  30 day window at that scale, yeah, it points to a fundamental
• 2:58
  shift in consumer behavior. Because search is inherently
• 3:01
  transient, right? How?
• 3:02
  Do you mean? Well, you ask for a recipe, you
• 3:04
  get the recipe and you close the tab.
• 3:05
  Yeah, the platform has to constantly fight to get you
• 3:08
  back, but by locking into your financial data via platform had
• 3:12
  perplexity is moving from being a search tool to a like a
• 3:17
  fiduciary utility. OK, let's unpack this, though,
• 3:19
  because I have to push back a little.
• 3:21
  Sure, giving a budgeting app my credit card data years ago felt
• 3:24
  invasive. Giving an AIA window into my
• 3:27
  entire financial soul feels like hiring a brilliant accountant
• 3:32
  who just lives permanently inside my laptop, watching every
• 3:35
  dime I spend. It does feel a little creepy,
• 3:37
  right? Like I get the the read only API
• 3:39
  means the AI cannot magically wire my money way, but the
• 3:43
  psychological barrier of an AI judging my impulse purchases
• 3:47
  that is steep. Is the public ready for that?
• 3:49
  Well, trust is always the ultimate friction point for
• 3:51
  autonomous agents. But look at those revenue
• 3:53
  numbers you just mentioned. Yeah, 450 million.
• 3:56
  Exactly. The convenience is rapidly
• 3:58
  outweighing the apprehension. Imagine logging in and the agent
• 4:02
  has already analyzed your auto loan rate, your current credit
• 4:04
  card balance, and the APY on your checking account.
• 4:08
  And then it instantly tells you the mathematically optimal way
• 4:10
  to distribute your next paycheck to save on interest.
• 4:13
  Factoring in inflation and everything.
• 4:15
  Exactly. Factoring in inflation your
• 4:17
  historical spending habits. When a system provides that
• 4:20
  level of personalized, actionable intelligence, that
• 4:25
  creepiness factor tends to just evaporate for the average user.
• 4:29
  So it moves AI from a parlor trick to an actual financial
• 4:33
  partner. Precisely.
• 4:34
  And if people are ready to hand over their bank accounts,
• 4:37
  handing over their business workflows is obviously the next
• 4:40
  step. Oh yeah, it's bleeding directly
• 4:41
  into the enterprise sector now with Notion.
• 4:43
  Right, so let's talk about notion.
• 4:45
  So Notion is positioning itself to be your entire back office
• 4:49
  staff. Their new custom agents are
• 4:51
  designed to automate recurring business tasks, inbound leads,
• 4:55
  accounting, marketing campaigns. And without the user needing a
• 4:59
  degree in computer science to set it up for everything.
• 5:01
  Exactly. No coding required.
• 5:02
  Because the sources outline exactly how you set this up, and
• 5:05
  it's surprisingly straightforward.
• 5:06
  You just go into Notion and create 2 simple databases. 1 is
• 5:10
  called Tasks which has columns for names, source, priority and
• 5:14
  status and the second database is called Reports.
• 5:17
  And that database structure is critical.
• 5:19
  Why is that? Because you have to give an
• 5:21
  autonomous agent a clearly defined input space and a
• 5:25
  rigidly defined output space. Otherwise it will hallucinate or
• 5:30
  perform actions totally outside of its intended scope.
• 5:33
  So they act like guardrails. Exactly.
• 5:36
  The databases are the physical guardrails for the AI's
• 5:39
  reasoning. Got it.
• 5:40
  So once those guardrails are up, you open Notion AI, click Create
• 5:45
  Custom Agent, and just prompt it in plain English.
• 5:48
  You type something like Create a weekly planner agent that reads
• 5:52
  my last seven days of emails every Monday morning, adds
• 5:55
  actionable items into the tasks database and writes A1 paragraph
• 5:58
  summary into the reports database.
• 6:00
  And then you just connect it to your Gmail.
• 6:01
  Right, yeah. Connect to Gmail, confirm the
• 6:02
  schedule, and you are done. You can clone that for virtually
• 6:05
  any recurring job. I mean 10 years ago setting up
• 6:08
  an automated pipeline that parsed raw emails and populated
• 6:12
  a dashboard. Oh, we use a nightmare.
• 6:14
  It required a developer complex API keys and those fragile ZAP
• 6:18
  your workflows that broke every single time an app updated.
• 6:21
  Now a solopreneur can spin up a fully functioning digital back
• 6:25
  office in 5 minutes using the exact same language they would
• 6:29
  use to talk to a human assistant.
• 6:30
  It's crazy. It is a massive democratization
• 6:33
  of administrative power. I love a good workflow, but I
• 6:36
  have to ask about the human nuance here.
• 6:39
  I tried setting up an automated e-mail parser last year using an
• 6:43
  older tool, and it completely missed a highly aggressive,
• 6:46
  unhappy tone from a client out there.
• 6:49
  Yeah, it flagged a furious e-mail as a standard check in
• 6:53
  because it only read the literal words, right?
• 6:56
  Right. So if Notions agents are just
• 6:57
  scanning for keywords like urgent or meeting, I am
• 7:00
  terrified to let it manage my actual inbox.
• 7:03
  Well, that is the difference between legacy automation and
• 7:05
  modern agentic AI. OK, older tools relied on
• 7:08
  deterministic keyword logic. If a word equals X, do Y.
• 7:12
  But modern large language models don't read words like that.
• 7:15
  They process semantic vectors. Semantics vectors.
• 7:18
  Yeah, they understand context. Sarcasm, frustration, and
• 7:21
  urgency based on the relationship of words in their
• 7:23
  training data. So it knows if someone is
• 7:25
  actually mad. Exactly.
• 7:26
  The Notion agent doesn't just see the word fine, it
• 7:29
  understands that a client replying fine after three
• 7:32
  delayed deliverables means they are extremely unhappy.
• 7:35
  Wow. It maps that emotional nuance
• 7:38
  and prioritizes the task accordingly.
• 7:41
  The AI is basically acting as a highly intelligent triage nurse.
• 7:46
  Making sure you only deal with the most critical cases.
• 7:48
  Exactly. OK, so we have millions of
• 7:51
  people spinning up personal finance agents on Perplexity and
• 7:54
  millions of small businesses running these custom semantic
• 7:57
  e-mail reading agents on Notion. Yeah.
• 8:00
  The sheer volume of processing power required to run all of
• 8:03
  those continuous complex queries is astronomical.
• 8:06
  Yeah, somebody has to physically build the engine for this
• 8:09
  agentic economy. You.
• 8:10
  Literally cannot have an authentic economy without the
• 8:13
  silicon to run. It right and enter Amazon.
• 8:15
  Yes, because the public narrative for the last two years
• 8:18
  has been entirely focused on software models.
• 8:20
  Like who has the smartest chat bot?
• 8:22
  Exactly, Open AI and Tropic Google and in that model centric
• 8:26
  view Amazon seemed like they were really lagging behind.
• 8:30
  But Amazon CEO Andy Jassy just released his annual shareholder
• 8:34
  letter, and he completely flipped that narrative.
• 8:36
  We really did. Investors have been anxious
• 8:38
  about Amazon's plan $200 billion AI capital expenditure, calling
• 8:44
  it a bubble. Jassy defended that spend by
• 8:46
  revealing that AWS AI ARM has already crossed $15 billion in
• 8:51
  annualized revenue. 15 billion, yeah.
• 8:54
  For context, that is 260 times where AW WS itself was at the
• 8:59
  exact same point in its history. That is just wild.
• 9:02
  But the most revealing metric in that letter was actually about
• 9:04
  their physical infrastructure. The chips.
• 9:06
  Yeah, Amazon's custom chips, Tranium, Graviton and Nitro,
• 9:11
  they just crossed $20 billion in yearly revenue. 20 billion And
• 9:15
  here is where it gets really interesting. 2 unnamed AWS
• 9:19
  customers actually tried to buy Amazon's entire 2026 supply of
• 9:23
  Graviton chips. The entire supply.
• 9:25
  Yeah, Amazon had to decline the offer just to protect their
• 9:28
  other clients access to the hardware.
• 9:30
  They are even floating the idea of selling racks of these chips
• 9:33
  directly to third parties. Everyone treats NVIDIA like they
• 9:37
  own the only power grid in town, but Amazon on quietly built a
• 9:41
  massive network of localized nuclear reactors in their own
• 9:44
  backyard. That is the perfect way to look
• 9:46
  at it because Nvidia's had this virtual monopoly on the compute
• 9:50
  market, right? But their GPU's are unparalleled
• 9:53
  for training massive models. Training requires parallel
• 9:57
  processing, like doing billions of math problems all at once.
• 10:00
  But running an agent that is managing your notion tasks or
• 10:04
  balancing your budget, that is completely different.
• 10:07
  How so? That relies on inference, The AI
• 10:10
  applying its training to a specific sequential user prompt,
• 10:14
  and Amazon's Graviton and Tranium chips are specifically
• 10:17
  architected for inference. I see.
• 10:19
  They run cooler, they use way less power, and they are
• 10:21
  significantly cheaper for sequential tasks.
• 10:24
  So Amazon isn't even trying to beat NVIDIA training, they're
• 10:27
  just securing the foundational hardware layer for the entire
• 10:30
  agentic economy. So Amazon is building the
• 10:32
  physical engine, this massive web of compute.
• 10:35
  But when you apply that level of processing power to pixel by
• 10:38
  pixel data rather than just text or spreadsheets, you move from
• 10:42
  organizing emails to literally predicting the future of the
• 10:45
  human body. That is what Oxford University
• 10:48
  researchers just achieved. And this is where the tech stops
• 10:52
  being an administrative convenience and fundamentally
• 10:55
  alters human mortality. The researchers at Oxford have
• 10:59
  built an AI system that catches heart failure up to five years
• 11:02
  early using nothing more than routine chest CT stance.
• 11:06
  Five years early, that's incredible.
• 11:08
  The mechanism here is what got me.
• 11:10
  The AI detects microscopic changes in the fat surrounding
• 11:14
  the heart, specifically texture shifts caused by muscle
• 11:18
  inflammation, and these are textural patterns that are
• 11:21
  physically invisible to the human eye on a standard scan.
• 11:25
  Right, because the human eye processes visual information in
• 11:28
  continuous gradients, we. Look for obvious things.
• 11:31
  Yeah, we look for anomalies like a tumor or a blocked artery.
• 11:34
  But the AI does not see the scan as a picture.
• 11:37
  It processes it as a dense matrix of numbers.
• 11:39
  Just raw data. Exactly.
• 11:41
  It's analyzing the exact density value pixel by pixel.
• 11:44
  So when heart muscle becomes inflamed, it subtly alters the
• 11:47
  density of the surrounding fat. And the AI sees that math
• 11:51
  change. Yes, it detects those
• 11:53
  mathematical texture years before the inflammation causes
• 11:57
  actual structural damage to the heart itself.
• 12:00
  They tested this on 72,000 patients and it achieved an 86%
• 12:05
  accuracy rate. Wow, and here is the statistic
• 12:08
  that literally stopped me in my tracks.
• 12:10
  In the highest risk bucket identified by this AI, one in
• 12:14
  four patients ended up developing heart failure within
• 12:16
  five years. One in four.
• 12:18
  That is a 20 times gap compared to the patients the AI flagged
• 12:22
  as safe. Oxford is already working with
• 12:24
  regulators to bring this to the National Health Service and
• 12:27
  plans to extend it to all chest CT scans within months.
• 12:31
  Because heart failure's biggest enemy isn't the lack of
• 12:33
  treatment options, you know? What is it?
• 12:36
  It is timing. By the time a human cardiologist
• 12:39
  can see structural damage on an echocardiogram, the heart muscle
• 12:42
  is already irreversibly compromised.
• 12:44
  The patient is already symptomatic.
• 12:46
  But if a machine is detecting texture shifts that I literally
• 12:49
  cannot see, as a doctor, what do I do with that information?
• 12:53
  A really tough question it. Creates a strange dynamic.
• 12:56
  A patient comes in feeling fine. I look at their scan and see a
• 12:59
  healthy heart, but the machine flags them as high risk.
• 13:03
  How does a doctor verify a black box?
• 13:06
  Well, it forces a complete paradigm shift from observable
• 13:09
  diagnostics to predictive statistical diagnostics.
• 13:14
  The doctor cannot visually verify the microscopic shift,
• 13:18
  but they verify the statistical validity of the AIS training
• 13:21
  data. It transforms the medical
• 13:23
  approach from reacting to a crisis in the ER to implementing
• 13:27
  preventive lifestyle interventions half a decade
• 13:30
  before the patient even feels a symptom.
• 13:32
  The doctor's clinical intuition now just has to incorporate the
• 13:35
  machine's predictive foresight. From predicting heart failure
• 13:39
  half a decade early to managing our checking accounts and
• 13:42
  reading our client emails. It's moving so fast.
• 13:44
  The scope of what agents are doing right now is dizzying.
• 13:47
  We need to keep the energy high and run through the remaining
• 13:50
  news because the periphery of the AI space is moving just as
• 13:53
  fast. Oh, the ecosystem is expanding
• 13:55
  rapidly. Let's hit the day's major
• 13:57
  updates. OK, first up in the tool space,
• 14:00
  Anthropotics, Claude cowork desktop agent is now generally
• 14:04
  available. Pushing that, enterprise
• 14:06
  automation we talked about even further into the operating
• 14:08
  system level. Exactly.
• 14:10
  Meta also launched News Spark, which is a multimodal reasoning
• 14:14
  AI that features a dedicated multi agent mode.
• 14:18
  And on the fintech side, a platform called Meow launched
• 14:22
  infrastructure that allows AI agents to open bank accounts and
• 14:26
  issue corporate cards. Unbelievable.
• 14:28
  So perplexity. Reading your bank account is
• 14:30
  just step one. Autonomous agents owning and
• 14:34
  operating their own checking accounts is Step 2.
• 14:36
  And we are also seeing some really significant moves from
• 14:39
  Open AI. They built a new advanced
• 14:41
  cybersecurity model similar to Anthropics Mythos, which they
• 14:45
  are quietly releasing to a small set of enterprise partners right
• 14:48
  now. Interesting.
• 14:49
  Additionally, they launched a $100 per month Pro tier,
• 14:53
  offering five times more Codex usage than their standard Plus
• 14:56
  tier. Wow.
• 14:57
  So that is clearly a play for developers who are building
• 15:00
  those heavy agentic coding frameworks.
• 15:02
  In corporate news, XAI is undergoing a major
• 15:05
  reorganization of its engineering division, right?
• 15:08
  Their CFO, Anthony Armstrong, is departing and they are bringing
• 15:12
  in executives directly from SpaceX ahead of a planned IPO.
• 15:16
  Which really just demonstrates that deep cross pollination of
• 15:20
  engineering talent within the Elon Musk ecosystem.
• 15:22
  Yeah, they are sharing the talent.
• 15:24
  Pool They are solidifying Xai's infrastructure and executive
• 15:27
  team as they prepare to face the public markets.
• 15:30
  And finally, in the legal sector, Flores attorney general
• 15:33
  has subpoenaed Open AI to investigate whether ChatGPT was
• 15:37
  used to plan a campus shooting at Florida State University.
• 15:40
  Yeah, and that is a stark example of the incredibly
• 15:43
  complex legal questions emerging right now around liability and
• 15:47
  user intent. Right.
• 15:48
  And to be clear, we're just reporting the facts of the
• 15:50
  subpoena here, not taking any stance on the probe itself.
• 15:53
  Exactly. It's just a crucial developing
• 15:55
  story about the legal frameworks.
• 15:57
  As these models become deeply integrated into daily planning,
• 16:01
  figuring out where the tools responsibility ends and the
• 16:03
  user's intent begins, he's going to be heavily tested.
• 16:06
  Absolutely. So to bring all of these massive
• 16:09
  industry shifting stories back down to earth, I want to look at
• 16:13
  how everyday people are utilizing this fast moving tech
• 16:16
  right now. Yes, let's do it.
• 16:18
  We have a workflow shared by a listener named AM from New
• 16:21
  Zealand that perfectly illustrates this new agentic
• 16:24
  era. The most grounded way to
• 16:26
  understand a technology is always to look at the community
• 16:28
  use cases. Right, so AM was prepping for a
• 16:32
  job interview for a role that was a massive leap up in their
• 16:35
  career. Instead of just asking ChatGPT
• 16:38
  for a list of common interview questions, they loaded the
• 16:41
  specific job description into the AI and instructed it to
• 16:45
  adopt the persona of the hiring manager.
• 16:47
  Oh, that's smart. Yeah, AM used the mobile voice
• 16:50
  mode to conduct a live spoken mock interview.
• 16:53
  They would answer out loud, the AI would process the verbal
• 16:56
  response, provide immediate structural feedback, and they
• 16:59
  repeated this loop until AM was scoring highly.
• 17:02
  Using the AI as a localized real time verbal tutor is brilliant.
• 17:06
  It really is. But the workflow didn't stop at
• 17:08
  the initial screening stage, did it?
• 17:10
  No, it gets better. AM made it to the second round
• 17:13
  of interviews, which required presenting to the company's five
• 17:16
  person board of directors. Oh wow, high stakes.
• 17:19
  So AM researched the professional backgrounds of all
• 17:22
  5 board members on LinkedIn, fed that specific biographical data
• 17:26
  into chat, and had the AI simulate the boardroom.
• 17:30
  That is next level. The AI actively helped AM tailor
• 17:34
  their answers to appeal directly to the specific roles,
• 17:37
  historical backgrounds, and likely interests of those
• 17:40
  individual board members. That workflow is the perfect
• 17:43
  encapsulation of everything we've discussed today.
• 17:45
  How so? The AI is not functioning as a
• 17:48
  static encyclopedia anymore. It is functioning as an
• 17:50
  interactive proxy. Yeah, AM engaged in a dynamic,
• 17:54
  dynamic agentic process to refine a physical world skill.
• 17:58
  It is like having a world class debate coach and a high end PR
• 18:01
  strategist right in your pocket. Exactly.
• 18:04
  The AI mapped the semantic data of those board members and
• 18:07
  instantly adapted AMS messaging to ensure maximum impact.
• 18:10
  And you know, as we process all of the developments today from
• 18:14
  perplexity, balancing our budgets to Amazon's custom
• 18:17
  silicon to Oxford's medical foresight, yeah, I'm going to
• 18:21
  leave you with a final thought to ponder.
• 18:24
  As AI moves from simply answering our questions to
• 18:28
  managing our bank accounts, running our business tasks and
• 18:31
  predicting our health, we really have to ask ourselves a
• 18:34
  fundamental question. What's that?
• 18:36
  At what point does AI stop being our tool and start becoming our
• 18:40
  proxy in the physical world? When an agent is acting on your
• 18:43
  behalf, speaking for you in an e-mail, and making financial or
• 18:46
  medical decisions before you even realize there is a problem,
• 18:50
  the line between human agency and artificial autonomy becomes
• 18:53
  incredibly thin. Are we building faster horses or
• 18:56
  have we already stepped into the teleporter without realizing it?
• 18:59
  That is the question. Thank you for exploring all of
• 19:02
  this with us today. If you enjoyed this deep dive,
• 19:05
  please hit that subscribe button to stay updated.
• 19:07
  And we would love it if you could take a second to rate the
• 19:09
  show 5 stars on whatever app you are using.
• 19:12
  It makes a huge difference for us.
• 19:15
  We'll see you right back here for tomorrow's Deep Dive.