Latest / Connecting the Dots / From Possibility to Reality — A Year of Recalibration

Transcript

From Possibility to Reality — A Year of Recalibration · Connecting the Dots

• 0:00
  Today's tech news moves fast, but understanding
• 0:03
  it shouldn't be hard. That's why Snarfl Solutions
• 0:06
  Group brings you Connecting the Dots, where we
• 0:09
  break down what's happening in tech and why it
• 0:11
  matters to you. Now, let's dive in with your
• 0:14
  hosts. Welcome back to Connecting the Dots. I'm
• 0:18
  Alex. And I'm Morgan. And this is our special
• 0:20
  year -end edition where we try to step back from
• 0:23
  the day -to -day noise and really look at the
• 0:27
  big currents that reshape technology and, frankly,
• 0:30
  our daily lives over the last 12 months. It's
• 0:32
  good to be here for this one. I feel like this
• 0:35
  year, 2025, it wasn't just another year of incremental
• 0:40
  updates. No. It felt like a real inflection point,
• 0:43
  a fundamental shift. It absolutely did. And I
• 0:45
  think to understand it, You have to look at the
• 0:48
  two years before, 2023, 2024. They were all about
• 0:53
  what we call the unbridled exuberance of discovery.
• 0:57
  Well, totally. That was the era of just build
• 0:59
  bigger. The birth of generative AI, that frantic,
• 1:02
  almost hysterical gold rush for GPUs. The prevailing
• 1:05
  belief was that scale solved every problem. Just
• 1:08
  throw more compute, more money at it. And that
• 1:10
  assumption is precisely what broke this year.
• 1:13
  2025 is, I think, going to be remembered as the
• 1:15
  year of recalibration. That's a great way to
• 1:17
  put it. It's the year that all those theoretical
• 1:19
  boundaries, all those dreams of just infinite
• 1:22
  AI growth, they collided violently with the physical
• 1:27
  and economic realities of actually deploying
• 1:29
  this stuff. And the security realities, too.
• 1:31
  Absolutely. The security realities. I find that
• 1:33
  framing so important. We didn't just see progress
• 1:35
  this year. We saw a fundamental restructuring
• 1:38
  of everything, of the entire tech stack, of the
• 1:41
  regulatory landscape globally, and maybe most
• 1:44
  importantly, how these digital systems actually
• 1:46
  touch the physical world. Exactly. So our mission
• 1:50
  today is to to synthesize all of that. We're
• 1:52
  not just reading headlines. We're going to connect
• 1:55
  these seemingly separate threads, the breakthroughs,
• 1:58
  the crises, the policy shifts into one story.
• 2:01
  The year of implementation. The year of implementation.
• 2:03
  The goal is that you walk away with a really
• 2:06
  actionable understanding of the new landscape
• 2:08
  we're all operating in now. We're talking about
• 2:10
  everything from a $5 million AI model that just
• 2:13
  humbled trillion dollar companies. To a single
• 2:16
  bug in a configuration file that took 20 % of
• 2:19
  the internet offline. It's a story of efficiency,
• 2:22
  fragility, and the sudden massive split in how
• 2:25
  the world's governments think about technology.
• 2:28
  It's the story of the rubber meeting the road.
• 2:30
  In a big way. It really is. So before we jump
• 2:33
  into that first massive theme, the one that really
• 2:35
  reset the economics of AI, we want to thank our
• 2:38
  sponsor for making this year -end special possible.
• 2:41
  We are supported today by Old Glory, the iconic
• 2:44
  music and sports fan merch store. The classic.
• 2:47
  A true classic. If you're looking for that vintage
• 2:49
  concert t -shirt, a rare throwback jersey, or
• 2:53
  just gear for your favorite team, Old Glory is
• 2:55
  the place. And they have an incredible offer
• 2:57
  for our listeners. They do. If you head over
• 2:59
  to OldGlory .com and use our promo code, you
• 3:02
  get 15 % off your entire purchase. That code
• 3:05
  is SNARFUL. S -N -A -R -F -U -L. SNARFUL. All
• 3:09
  one word. It gets you 15 % off everything at
• 3:12
  OldGlory .com. Go check them out. All right,
• 3:14
  let's do it. OK, let's unpack this. We have to
• 3:16
  start with the biggest shock to the system this
• 3:18
  year, the thing that completely redefined AI.
• 3:21
  Let's talk about the algorithmic paradigm shift.
• 3:23
  The efficiency shock. The efficiency shock. This
• 3:26
  is the story that changed the fundamental assumption
• 3:28
  of how you build and price artificial intelligence.
• 3:32
  For years, the dogma, the so -called scaling
• 3:35
  laws, was just incredibly simple. Bigger is better.
• 3:39
  Bigger is better. You needed more data. more
• 3:41
  parameters, and an exponential amount of coupon
• 3:43
  power. We're talking over a hundred billion dollars
• 3:46
  in CapEx for GPUs and electricity. The brute
• 3:50
  force hypothesis, it was the entire business
• 3:53
  model for the big US hyperscalers. Just keep
• 3:55
  building bigger data centers. And intelligence
• 3:57
  will just sort of emerge from the pile. Right.
• 3:59
  That entire hypothesis just completely collapsed
• 4:02
  in January of 2025. Yeah. And it collapsed because
• 4:05
  of a single model. DeepSeek R1. DeepSeek R1.
• 4:08
  And it didn't come from Silicon Valley. It came
• 4:09
  from a research lab in Hangzhou, China. And this
• 4:12
  was a genuine Sputnik moment for the entire industry.
• 4:15
  It really was. And you have to remember the context
• 4:17
  here. The US companies were, and still are, operating
• 4:20
  under strict export controls on the most advanced
• 4:23
  hardware. Specifically, the NVIDIA H100s, the
• 4:26
  top of the line ships. Exactly. And DeepSeq proved
• 4:29
  to the world that you don't actually need unrestrained
• 4:32
  access to the best hardware if you just have
• 4:34
  smarter, more efficient algorithms. The numbers
• 4:36
  are just. They're staggering when you compare
• 4:39
  them. You really are. DeepSeek R1 was trained
• 4:41
  for a reported cost of just $5 .6 million. $5
• 4:45
  .6 million. And yet, it achieved reasoning parity
• 4:50
  with models that cost, I mean, orders of magnitude
• 4:53
  more to build. Sometimes it even surpassed them
• 4:55
  on certain benchmarks. It did. And that massive
• 4:58
  gap between cost and performance, that's the
• 5:01
  tension that defined the whole rest of the year
• 5:03
  in tech strategy. So how did they do it? If they
• 5:06
  couldn't just throw bigger hardware at it, what
• 5:08
  was the magic trick? It really comes down to
• 5:10
  two very specific, very powerful technical breakthroughs.
• 5:15
  The first one was the architecture of the model
• 5:16
  itself. The mixture of experts, MoE. Exactly.
• 5:20
  For years, models were dense. That means every
• 5:23
  single parameter, every neuron in the network
• 5:24
  was active every time you generated a single
• 5:27
  word. It's incredibly inefficient. DeepSeq R1
• 5:30
  was a MoE design. It had this enormous total
• 5:33
  parameter count, 671 billion. Which sounds huge,
• 5:37
  but that's not the important number. Not at all.
• 5:40
  Because during any single calculation, only a
• 5:43
  tiny fraction of those parameters were actually
• 5:45
  active. About 37 billion, right? That's right,
• 5:48
  37 billion. So think of it like a giant library
• 5:50
  with 671 billion books. Okay. In a dense model,
• 5:55
  to read one book, you have to turn on every single
• 5:58
  light in the entire library. Which is incredibly
• 6:00
  wasteful. Right. But in a Moe model, only the
• 6:04
  specific shelf, the expert, that holds the book
• 6:06
  you need, lights up. So you get all the knowledge
• 6:08
  of that massive library, but with the energy
• 6:11
  cost of just lighting up one small section. Precisely.
• 6:14
  That sparsity is the key. It let them have this
• 6:16
  vast encyclopedic knowledge base without the
• 6:19
  insane latency and cost you'd normally get from
• 6:21
  a model that size. OK, so that's the architecture.
• 6:23
  What was the second breakthrough? The second
• 6:25
  one was even more fundamental. It was in how
• 6:27
  they trained it. They used a new technique called
• 6:30
  Group Relative Policy Optimization, GRPO. Right.
• 6:34
  And traditional reinforcement learning, the PPO
• 6:36
  technique everyone was using, it requires this
• 6:40
  second critic model. It does. You have an actor
• 6:43
  model that does the work and a critic model that
• 6:45
  judges the actor's work. And that critic model
• 6:47
  essentially doubles the amount of GPU memory
• 6:49
  you need during training, doesn't it? It absolutely
• 6:51
  does. It's a huge memory hog. And that's exactly
• 6:54
  the resource that was being constrained by the
• 6:56
  export controls. So GRPO gets around that. It
• 6:59
  gets around it entirely. Instead of needing a
• 7:01
  separate critic to tell the actor if it did a
• 7:03
  good job, GRPO just has the actor generate a
• 7:06
  group of possible outputs. And then it compares
• 7:09
  them against each other? It compares them against
• 7:11
  each other and uses their relative performance
• 7:14
  to figure out the best way forward. It's reinforcement
• 7:16
  learning without the memory doubling critic.
• 7:19
  And that innovation let them train a world -class
• 7:22
  model on constrained hardware. It did. And the
• 7:25
  market fallout from this was immediate and frankly.
• 7:29
  brutal. The pricing. The pricing. DeepSeek's
• 7:31
  API didn't just compete. It forced a complete
• 7:34
  repricing of what intelligence costs across the
• 7:37
  entire industry. They priced it at 55 cents per
• 7:41
  million input tokens. Right. For context, that
• 7:44
  undercut the market leaders, like Google's Gemini
• 7:46
  3 Pro, by a factor of almost four. That's not
• 7:49
  just competition. That's an existential threat
• 7:52
  to the high margin business models of the big
• 7:54
  U .S. hyper stalers. It completely changed the
• 7:57
  game. So the Western response couldn't just be
• 8:00
  to match on price. They couldn't. It would have
• 8:02
  destroyed their margins. So they had to differentiate.
• 8:04
  They had to pivot hard to where sheer costs couldn't
• 8:08
  compete. And that was the frontier of what we
• 8:10
  call aogenic power and massive context. This
• 8:13
  is the strategic pivot of 2025. It is. Google
• 8:16
  responded with its Gemini 3 family of models.
• 8:20
  And they focused on two things, context and multimodality.
• 8:23
  They expanded that context window to over a million
• 8:27
  tokens. A million tokens. Which is, I mean, it's
• 8:30
  hard to even comprehend. It's like feeding it
• 8:31
  hundreds of thousands of pages of text in a single
• 8:33
  go. So you can reason over an entire company's
• 8:36
  internal wiki or years of legal documents or
• 8:39
  a massive software code base all at once. It
• 8:42
  shifts the value from just answering a question
• 8:44
  to synthesizing an enormous amount of data. And
• 8:47
  they also introduced Google anti -gravity. Right.
• 8:50
  A suite of tools. pushing toward these fully
• 8:53
  agentic systems that could collaborate with developers,
• 8:55
  not just a system. And Gemini 3 really cemented
• 8:59
  its lead in superior multimodal handling. Meaning
• 9:02
  it could process video, audio, and text all at
• 9:05
  the same time, natively. Exactly. You could literally
• 9:08
  watch a video, listen to the audio track, and
• 9:11
  read the transcript simultaneously to understand
• 9:13
  the content. Meanwhile, Anthropic was carving
• 9:15
  out its own niche. They doubled down on reliability
• 9:18
  and control. Their Claude Opus 4 .5 series had
• 9:21
  a major breakthrough in what they called computer
• 9:23
  use. Because it was huge. It was. The AI could
• 9:26
  literally look at a screenshot of your computer
• 9:28
  screen. understand the software you're using,
• 9:31
  and then manipulate a virtual cursor to click
• 9:33
  buttons and type in fields. It moved the AI out
• 9:35
  of the chat window and onto your actual desktop.
• 9:37
  It did. It could now update a spreadsheet for
• 9:40
  you or navigate some ancient internal ERP system
• 9:43
  or manage your cloud console. That was the missing
• 9:46
  link for the agentic era. And all of this fed
• 9:49
  into the big cognitive trend of the year. the
• 9:52
  move from predictive models to true reasoning
• 9:55
  models. We finally moved past what they call
• 9:57
  system one thinking. That fast, intuitive, but
• 10:01
  often hallucination -prone style of the early
• 10:03
  GPTs. Right. And we moved into what really felt
• 10:05
  like system two thinking, deliberative logical
• 10:09
  reasoning. And the key technology there was the
• 10:11
  native use of chain of thought processing, Coup.
• 10:14
  That's right. These new models started generating
• 10:16
  a hidden internal monologue, a sort of thought
• 10:19
  trace. So it would Plan its approach, critique
• 10:22
  its own logic, and self -correct its mistakes
• 10:24
  before it ever showed you the final answer. And
• 10:27
  that self -verifying, self -correcting nature
• 10:29
  was the absolute missing link for enterprise
• 10:31
  adoption. All those companies that got burned
• 10:33
  by confident -sounding hallucinations in 2024
• 10:36
  finally had a reason to trust the output for
• 10:39
  high -stakes work. Legal analysis, financial
• 10:41
  modeling, things where a mistake is not an option.
• 10:44
  And that leap in quality, kicked off by DeepSeek's
• 10:46
  efficiency shock, really sets the stage for our
• 10:49
  next major theme. the birth of the agentic economy.
• 10:52
  That's the perfect transition. That move to system
• 10:55
  two reasoning is exactly what unlocked this second
• 10:59
  massive theme of 2025. The agentic economy. This
• 11:02
  is where the industry moved decisively past the
• 11:05
  chat bot. As we like to say, a chat bot answers
• 11:08
  a question. An agent pursues a goal. It's about
• 11:11
  closing that loop between perception, cognition,
• 11:15
  and action. Right. Giving the AI hands and feet.
• 11:17
  Right. An agent can perceive new data coming
• 11:19
  in, think about it using that chain of thought
• 11:22
  logic, and then take action. And not just one
• 11:24
  action. It can plan a whole multi -step workflow.
• 11:27
  Exactly. It can research a market. then draft
• 11:30
  a proposal based on that research, access an
• 11:32
  internal database for sales figures, write a
• 11:35
  little code patch, and then execute a final transaction.
• 11:38
  All autonomously. Or semi -autonomously, that's
• 11:40
  the key. And this wasn't just research anymore.
• 11:42
  The infrastructure for this became standardized
• 11:44
  in 2025. We saw orchestration frameworks like
• 11:47
  Microsoft's AutoGen, LandGraph, and CrewAI become
• 11:50
  the go -to tools for developers. These frameworks
• 11:52
  let you build essentially Digital org charts.
• 11:55
  You could create a team of agents. So you'd have
• 11:57
  a researcher agent that finds the data, a writer
• 12:00
  agent that drafts the summary, and a manager
• 12:03
  agent to coordinate them and report the final
• 12:06
  result to a human. The promise was just immense
• 12:09
  productivity gains, right out of the box. In
• 12:11
  theory, yes. The anticipation was huge. But 2025
• 12:15
  is where we ran headfirst into the corporate
• 12:17
  adoption gap. The productivity versus transformation
• 12:20
  gap. Exactly. A big PWC survey found that 79
• 12:24
  % of companies claimed they were adopting AI
• 12:26
  agents. But when you dug into it, most were stuck
• 12:29
  in the pilot phase. Or they were just using them
• 12:31
  for really low stakes internal tasks, like summarizing
• 12:35
  meetings. So what was the friction? If the tech
• 12:38
  was so capable, Why was the rollout of true transformations
• 12:42
  so slow? In a word, trust. The ghost of hallucinations
• 12:46
  passed. Precisely. The memory of those confident,
• 12:49
  completely wrong answers from 2024 was still
• 12:51
  very fresh. So while having an agent draft an
• 12:54
  internal email is pretty low risk. Giving that
• 12:57
  same agent right access to your production customer
• 12:59
  database. No way. The IT and compliance departments
• 13:02
  would absolutely balk at that or letting it deploy
• 13:05
  new code to customers or execute a million dollar
• 13:09
  financial transaction. The risk of that point
• 13:11
  one percent failure rate was just too high. It
• 13:14
  was operationally and legally unacceptable. So
• 13:17
  the real world outcome was that agents were widely
• 13:20
  deployed as co -pilots. The human had to stay
• 13:22
  in the loop for that final approval. Which prevented
• 13:25
  them from being truly autonomous digital workers,
• 13:27
  at least for anything mission critical. And this
• 13:30
  need for human oversight created a brand new
• 13:32
  job title, didn't it? It did. We saw the rapid
• 13:35
  emergence of the agent orchestrator. A human
• 13:37
  worker whose whole job is to manage, debug, and
• 13:40
  audit a fleet of these digital employees. Right.
• 13:43
  Their job isn't just to use the agent. It's to
• 13:45
  make sure its actions are safe, traceable, and
• 13:48
  compliant with company policy. It's a brand new
• 13:50
  role born directly from that trust gap. So while
• 13:53
  companies were hesitant to let agents run wild
• 13:55
  internally, the race to build the best agent
• 13:58
  platform was just frantic. Oh, the external race
• 14:01
  was an all -out war. This led to the explosion
• 14:03
  of the acquirer war where you're not buying a
• 14:06
  company for its product. You're buying it for
• 14:07
  its people exactly Talent that understood how
• 14:11
  to build complex reliable agents became the new
• 14:14
  gold The tech giant saw this as the next great
• 14:18
  competitive moat met his acquisition of menace
• 14:21
  that Singapore -based startup, for $2 billion.
• 14:25
  That was a huge signal. It was. It showed that
• 14:28
  geography was irrelevant. They were buying the
• 14:30
  talent. Menace was focused on building these
• 14:32
  general -purpose agents that could work across
• 14:34
  different software environments. Meta basically
• 14:36
  bought their entire research team. They did.
• 14:39
  And it wasn't just Meta. OpenAI acquired StatSig
• 14:41
  for over a billion dollars. And that wasn't for
• 14:43
  a consumer product. Not at all. It was to get
• 14:46
  their engineering experimentation platform. It
• 14:49
  was about securing the ability to test and iterate
• 14:52
  on agent behaviors at scale faster than anyone
• 14:55
  else. So the strategy was clear. Even if we can't
• 14:58
  fully deploy them yet, we need to own the talent
• 15:00
  that can build them for the future. Absolutely.
• 15:03
  2025 made it clear that while DeepSeq might have
• 15:06
  commoditized basic reasoning, the ability to
• 15:09
  turn that reasoning into reliable action, that
• 15:11
  agent layer, was the new premium commodity. And
• 15:15
  that reliance on reliable action on a system
• 15:17
  that's always on and always correct, just threw
• 15:20
  a harsh spotlight on how fragile the foundations
• 15:22
  of the internet really are. That's the perfect
• 15:25
  segue. The eugenic economy demands 100 % uptime.
• 15:28
  and in 2025 the underlying infrastructure violently
• 15:31
  disagreed with that demand. That's putting it
• 15:34
  mildly. This whole agentic future, totally dependent
• 15:36
  on real -time data and connectivity, was just
• 15:40
  rocked by a series of failures that exposed the
• 15:42
  alarming fragility of the centralized web. I
• 15:45
  remember the feeling during those outages. It
• 15:46
  wasn't just an inconvenience for a lot of businesses.
• 15:49
  It was a complete shutdown. And what was so striking
• 15:51
  about the catastrophic cloud outages of 2025
• 15:54
  is that they weren't caused by some sophisticated
• 15:57
  nation state cyber attack. No, they were internal
• 16:00
  configuration errors, human errors, systemic
• 16:02
  errors, the kind of thing that just shouldn't
• 16:04
  happen at that scale. Let's start with the biggest
• 16:06
  one, the AWS DynamoDB outage. on October 20th.
• 16:11
  This wasn't a blip. This was over 15 hours of
• 16:13
  downtime. In the us -east -one region, which
• 16:15
  is basically the heart of the internet for a
• 16:17
  huge chunk of the world. The number of affected
• 16:19
  users was historic. 17 million people. Think
• 16:23
  about the services that run there. Netflix, Snapchat,
• 16:26
  major banks. Everything just stopped. It was
• 16:29
  a global pause button. And the mechanism behind
• 16:31
  it was just this perfect lesson in technical
• 16:34
  humility. It was caused by a non -malicious internal
• 16:37
  race condition. We should probably explain what
• 16:40
  a race condition is. It sounds technical, but
• 16:41
  the concept is actually pretty simple. It's just
• 16:43
  a timing problem. Imagine you have two automated
• 16:46
  workers, a DNS planner and a DNS enactor, and
• 16:49
  they're both trying to update the same critical
• 16:51
  set of instructions. The planner figures out
• 16:53
  the new configuration, the enactor applies it.
• 16:55
  Right. But in this case, the enactor was too
• 16:58
  fast. It executed an old instruction to remove
• 17:01
  a bunch of critical infrastructure before the
• 17:04
  planner could finish giving it the new instructions.
• 17:06
  It's like deleting the old version of a file
• 17:08
  a split second before the new version is finished
• 17:11
  saving. The file just disappears. That's exactly
• 17:14
  what happened. The system effectively deleted
• 17:16
  its own address book, making the entire DynamoDB
• 17:19
  service invisible to the rest of the internet.
• 17:22
  And here's where it gets really scary. The problem
• 17:25
  of the circular dependency. This is the real
• 17:28
  lesson. DynamoDB is a foundational service. A
• 17:32
  lot of other AWS services use it to store their
• 17:34
  own operational data. So when DynamoDB went down...
• 17:37
  A ton of other things went down with it. But
• 17:39
  here's the kicker. The AWS Status Dashboard.
• 17:43
  The very tool they use to tell customers what's
• 17:46
  wrong. It relies on DynamoDB. It does. and so
• 17:49
  did many of their internal tools needed to fix
• 17:51
  the problem. So the engineers couldn't even use
• 17:53
  their own tools to fix the broken service because
• 17:55
  the tools themselves were broken by the outage.
• 17:58
  It's the ultimate circular dependency risk and
• 18:01
  it dramatically extended the outage. It became
• 18:03
  the poster child for the danger of single points
• 18:06
  of failure in the cloud. And then, less than
• 18:08
  a month later, it happened again in a different
• 18:10
  way. The Cloudflare bot management blackout on
• 18:13
  November 18th. Cloudflare is the front door for
• 18:16
  something like 20 % of the entire web. This one
• 18:19
  hit the AI services hard. OpenAI, Anthropic,
• 18:24
  X. They all went dark because traffic couldn't
• 18:26
  even reach them. And again, it wasn't a hack.
• 18:28
  It was a software deployment failure. A bug in
• 18:31
  a new configuration file caused their edge software,
• 18:34
  the actual computers running in data centers
• 18:36
  all over the world, to just crash. So when the
• 18:39
  software crashed, it just returned that generic
• 18:42
  fatal HTTP 500 error to millions of people, a
• 18:47
  global blockade. And the irony was just profound.
• 18:50
  The very system they designed to prevent denial
• 18:53
  of service attacks became the source of a global
• 18:55
  denial of service, all because of one bad configuration
• 18:58
  file. These two events back to back were a brutal
• 19:01
  wake up call for the industry. It showed that
• 19:03
  the internet's plumbing is alarmingly fragile,
• 19:06
  especially when you're trying to build a future
• 19:07
  of autonomous agents on top of it. And while
• 19:10
  the infrastructure was failing from internal
• 19:11
  errors, the security situation was failing from
• 19:14
  just basic negligence. The data breach epidemic.
• 19:16
  This is where we saw trust just completely erode.
• 19:19
  The Ticketmaster and Snowflake breach was the
• 19:22
  big one. 560 million customer records stolen
• 19:25
  by a group called Shiny Hunters. But again, this
• 19:27
  wasn't some sophisticated zero -day hack. It
• 19:30
  was a failure of the most basic access control.
• 19:33
  It was a credential stuffing attack, right? It
• 19:35
  was. The attackers used a username and password
• 19:38
  they harvested from a third -party contractor's
• 19:40
  machine. And the detail that just shocked everyone.
• 19:44
  The contractor's account did not have multi -factor
• 19:46
  authentication enabled. MFA. The front door was
• 19:49
  just wide open. All you needed was the username
• 19:51
  and password. That one detail, no MFA on an account
• 19:55
  with access to 560 million records. It triggered
• 19:59
  an immediate industry -wide mandate. Every CISO
• 20:02
  in the world started enforcing MFA on every single
• 20:05
  account, internal or external, the next day.
• 20:07
  It really highlighted that the most devastating
• 20:10
  attacks can be stopped by the simplest security
• 20:12
  measures that we often ignore. And it pointed
• 20:14
  to the bigger credential crisis. In June, researchers
• 20:18
  found a single consolidated database with 16
• 20:20
  billion stolen credentials in it. 16 billion.
• 20:23
  Which just proves that the username and password
• 20:26
  model is fundamentally broken. It's not a secure
• 20:28
  system anymore. Which is why we saw such a massive
• 20:31
  accelerated pivot to pass keys and biometrics
• 20:34
  this year. It became an existential necessity.
• 20:37
  We also saw ransomware get much more targeted
• 20:40
  and frankly more cruel. The attack on PowerSchool,
• 20:44
  the education software provider. They used a
• 20:46
  terrifying form of double extortion. They didn't
• 20:48
  just threaten to leak. sensitive student data
• 20:51
  to the public internet. They threatened to send
• 20:53
  it directly to individual school districts to
• 20:56
  maximize the pressure and the pain. It's a highly
• 20:58
  targeted, highly psychological attack. And then
• 21:00
  the most alarming development was ransomware
• 21:03
  targeting critical physical infrastructure. The
• 21:06
  attack on the Romania energy complex. It forced
• 21:08
  them to shut down all the IT systems at the country's
• 21:11
  largest coal power producer. Which highlights
• 21:13
  the threat to operational technology, the OT
• 21:16
  systems that actually run the power grid. When
• 21:18
  you shut down the IT, You compromise the ability
• 21:21
  to safely operate the physical plant. The stakes
• 21:24
  become life and death. So if you put it all together,
• 21:27
  2025 was the year the world realized that this
• 21:30
  amazing, agentic era is only as strong as its
• 21:33
  weakest, most mundane link. Whether that's a
• 21:36
  timing bug in a DNS server, a missing MFA token,
• 21:40
  or a single circular dependency. And as all these
• 21:43
  failures mounted, it forced governments around
• 21:46
  the world to react. Which brings us to our next
• 21:48
  feature. The governance fracture. That infrastructure
• 21:52
  instability really set the stage for the geopolitical
• 21:54
  story of the year. The governance fracture. This
• 21:57
  is what created these long -term strategic headaches
• 21:59
  for every single multinational tech company.
• 22:02
  Absolutely. Because in 2025, the governance of
• 22:06
  technology just split down the middle. We saw
• 22:08
  a sharp ideological divergence between the US
• 22:10
  and the European Union. They took fundamentally
• 22:12
  opposite paths completely opposite Let's start
• 22:14
  with the US the new administration came in in
• 22:17
  January and immediately reversed the federal
• 22:19
  AI policy that had been building up They issued
• 22:21
  executive order 14179 and the title says it all
• 22:25
  Removing barriers to American leadership in artificial
• 22:28
  intelligence. It was a very clear very rapid
• 22:31
  pivot to deregulation. It explicitly revoked
• 22:35
  the previous administration's AI safety mandates.
• 22:38
  So things like mandatory safety testing, pre
• 22:41
  -deployment audits for these big frontier models.
• 22:44
  All suspended at the federal level. And the stated
• 22:46
  rationale was all about national security and
• 22:49
  economic competition. Right. The argument was
• 22:51
  that the old rules were creating too much friction.
• 22:54
  that they were stifling the innovation the U
• 22:56
  .S. needed to stay ahead of China. And this federal
• 22:59
  deregulation created a massive policy vacuum.
• 23:02
  Which the state legislatures rushed to fill.
• 23:04
  And that's how we ended up with this regulatory
• 23:06
  patchwork, a total mess for compliance. New York
• 23:08
  passed its Raise Act. Right. The RAISE Act mandated
• 23:12
  that developers of these big frontier models
• 23:14
  had to register their safety protocols with the
• 23:16
  state and report any critical incidents within
• 23:18
  72 hours. A safety first approach. And then California
• 23:22
  followed almost immediately with SB 53, which
• 23:24
  did something very similar. Which put developers
• 23:27
  in this impossible position. Do you follow the
• 23:30
  state rules and slow down, or do you follow the
• 23:33
  federal mandate to move fast and risk getting
• 23:35
  sued by the states? And this set up a massive
• 23:38
  constitutional legal battle. A huge one. The
• 23:41
  federal government tried to preempt the state
• 23:43
  laws, arguing that they had sole authority, and
• 23:47
  the states argued it was their right to protect
• 23:49
  their citizens. And that legal fight over who
• 23:51
  has the right to regulate technology is going
• 23:54
  to be the dominant story for the next year at
• 23:56
  least. No question. While the US story was deregulation
• 24:01
  and fragmentation, the EU story was the complete
• 24:04
  opposite. It was all about enforcement and solidification.
• 24:07
  They moved into the active enforcement phase
• 24:09
  of their landmark AI Act. This was years in the
• 24:12
  making, and now it had teeth. Real deadlines,
• 24:15
  real prohibitions. In February, the bans on unacceptable
• 24:18
  risk AI came into force. Things like social scoring
• 24:21
  systems, predictive policing tools outright banned
• 24:24
  in the EU. And then in August, the big one. The
• 24:26
  governance rules for general purpose AI models
• 24:29
  took effect. This meant the providers of those
• 24:32
  big systemic models, your GPT -5s, your Claude
• 24:34
  Opuses, they now had to provide incredibly detailed
• 24:38
  technical documentation, submit results from
• 24:41
  rigorous adversarial testing, and prove, absolutely,
• 24:45
  that their training data complied with EU copyright
• 24:47
  law. So the result of this massive divergence
• 24:50
  was that Every big tech company had to adopt
• 24:53
  what they called a bifurcated strategy. They
• 24:55
  had to. They were basically forced to build two
• 24:57
  versions of their products, an unrestricted move
• 25:00
  fast version for the U .S. And a separate, highly
• 25:03
  compliant and often slightly less capable version
• 25:05
  for the European market to avoid these massive
• 25:08
  fines. And the cost of that compliance was so
• 25:11
  high that some companies like Apple and Metta
• 25:14
  just delayed releasing their most advanced AI
• 25:16
  features in the EU altogether. They decided it
• 25:18
  just wasn't worth the risk. created this noticeable
• 25:21
  gap in functionality between what a user in the
• 25:24
  US could do versus a user in Europe. And speaking
• 25:27
  of governance, the US antitrust philosophy also
• 25:29
  saw a really interesting shift. Right. The Department
• 25:32
  of Justice cleared Google's $32 billion acquisition
• 25:35
  of WiZ, the cloud security company. Under any
• 25:39
  prior administration, a deal that big a tech
• 25:42
  giant buying a leader in a critical adjacent
• 25:44
  market, it would have faced an enormous, probably
• 25:47
  fatal antitrust challenge. Absolutely. The approval
• 25:50
  signaled a clear philosophical shift. It showed
• 25:53
  a prioritization of national competitiveness
• 25:56
  over traditional market concentration concerns.
• 25:59
  The argument was that to compete globally and
• 26:01
  defend against foreign cyber threats, U .S. champions
• 26:04
  like Google needed to be able to acquire best
• 26:07
  -in -class security. And that national security
• 26:09
  argument won out over the antitrust concerns.
• 26:12
  So this whole segment just shows that technology
• 26:14
  is no longer just a business story. It's a story
• 26:16
  of diverging national priorities. And companies
• 26:18
  are now caught in the middle of this geopolitical
• 26:20
  chessboard trying to decide whether to optimize
• 26:23
  for speed like the US or for safety like the
• 26:26
  EU. And that geopolitical pressure, that need
• 26:29
  to spend billions on AI infrastructure to stay
• 26:31
  competitive, is exactly what drove the massive
• 26:34
  corporate restructuring we saw in 2025. It was
• 26:38
  a bizarre year, wasn't it? Record -breaking acquisitions
• 26:40
  happening at the exact same time as huge targeted
• 26:44
  layoffs. The M &A Renaissance was all about one
• 26:47
  thing, building the complete end -to -end AI
• 26:50
  stack. The Google and Wiz deal we just talked
• 26:53
  about. $32 billion. That was about integrating
• 26:56
  security so deeply into the cloud that it wasn't
• 26:59
  an add -on anymore. It was part of the foundation.
• 27:01
  And then we saw Salesforce buy Informatica for
• 27:04
  $8 billion. This was a direct play for the agentic
• 27:07
  era. Absolutely. It was all about data hygiene.
• 27:09
  Your autonomous agents are useless if they're
• 27:11
  running on garbage data. Right. By buying Informatica,
• 27:14
  Salesforce bought the piping. They secured the
• 27:17
  tools needed to make sure clean, reliable, compliant
• 27:20
  data was flowing into their agent force platform.
• 27:23
  They bought the fuel for their future agents.
• 27:25
  And the spending spree continued, meta -buying
• 27:27
  Manishu for $2 billion. The message was clear.
• 27:31
  Secure the talent and the foundational tech to
• 27:33
  automate labor, no matter the cost. And yet at
• 27:35
  the same time, the tech sector shed over 126
• 27:39
  ,000 jobs in 2025. It's an unsettling contradiction.
• 27:43
  It is. And these weren't your typical recession
• 27:45
  -driven cuts. These were strategic reallocations.
• 27:48
  Exactly. Profitable giants like Microsoft and
• 27:51
  Amazon cut thousands of jobs. 15 ,000 and 14
• 27:54
  ,000, respectively. But they cut them in their
• 27:56
  legacy divisions. Hardware, physical retail,
• 27:59
  older IT functions. And the reason they made
• 28:01
  those cuts was explicitly to free up cash to
• 28:05
  fund the massive capital expenditure for AI.
• 28:08
  For new data centers, for buying more GPUs, for
• 28:11
  hiring expensive AI engineers. The message to
• 28:13
  the workforce was stark. If your job doesn't
• 28:16
  directly contribute to the AI pipeline, it's
• 28:18
  at risk. For a company like Intel, their 22 ,000
• 28:22
  job cuts were just existential. They were trying
• 28:25
  to preserve capital for their manufacturing turnaround.
• 28:28
  On a more positive note, the IPO market finally
• 28:30
  thawed out. Thank goodness. After a long freeze,
• 28:33
  we saw big, mature companies like Stripe and
• 28:35
  Databricks finally go public. Stripe at a $65
• 28:38
  billion valuation, Databricks at $43 billion.
• 28:41
  And that provided this crucial injection of liquidity
• 28:43
  back into the venture capital ecosystem, which
• 28:46
  was starving for returns. That cash gets recycled
• 28:49
  into the next generation of startups. So while
• 28:52
  the corporate world was shuffling its balance
• 28:54
  sheets, consumers were out there running their
• 28:56
  own real -world tests on all this new technology.
• 28:59
  It's the consumer verdict. This is where the
• 29:01
  rubber really meets the road. And nothing illustrates
• 29:04
  the limits of fragile automation better than
• 29:06
  the fast food drive crew. Oh, the Taco Bell saga.
• 29:09
  Taco Bell had to scale back their ambitious AI
• 29:12
  voice ordering systems after all those viral
• 29:14
  videos of spectacular failures. We all saw the
• 29:17
  one where the customer was just... desperately
• 29:20
  trying to correct an order that the AI had misinterpreted
• 29:23
  as a request for 18 ,000 cups of water. It's
• 29:26
  hilarious, but the business lesson is dead serious.
• 29:28
  It is. The last mile of automation, that messy,
• 29:32
  chaotic interaction with a real human, is the
• 29:34
  hardest part. In a high -variance environment
• 29:36
  with accents, background noise, people changing
• 29:39
  their minds, a flexible human is often just cheaper
• 29:42
  and more reliable than a brittle AI. We also
• 29:45
  saw this really stark immediate A -B test in
• 29:48
  the wearables market. It was a perfect case study.
• 29:50
  On one side, you had Meta's Ray -Ban Smart Glasses,
• 29:53
  which were a clear success. And the key was subtlety.
• 29:57
  They looked like stylish glasses that just happened
• 29:59
  to have a camera and an AI you could talk to
• 30:01
  if you wanted. They didn't demand your attention.
• 30:03
  And on the other side, you had the Friend AI
• 30:05
  Necklace, a total failure. It was marketed as
• 30:09
  a cure for loneliness and always on companion,
• 30:12
  but everyone just saw it for what it was. A surveillance
• 30:15
  device. The reviews were brutal. People said
• 30:17
  it was socially intrusive, constantly recording
• 30:20
  conversations, and it didn't even work well in
• 30:22
  noisy places. And that experience established
• 30:24
  a very clear boundary for consumers. AI in social
• 30:28
  settings needs to be passive. It needs to wait
• 30:31
  to be called upon. It can't just be listening
• 30:32
  all the time. Finally, we saw a huge hardware
• 30:35
  pivot from Apple, driven entirely by their AI
• 30:38
  ambitions. The iPhone 17. For the first time,
• 30:41
  the base model included the 120Hz promotion display.
• 30:45
  A feature that was always reserved for the expensive
• 30:47
  Pro models before. And that wasn't just them
• 30:49
  being generous, it was a technical necessity.
• 30:52
  To run the powerful on -device AI features in
• 30:55
  Apple Intelligence smoothly, without lag, you
• 30:59
  need that high refresh rate display. And it worked!
• 31:01
  It drove a massive consumer upgrade super cycle
• 31:05
  because people wanted access to those new on
• 31:08
  -device AI features. It proved that software
• 31:11
  is now dictating hardware requirements. So on
• 31:13
  the corporate side, it was all about reallocating
• 31:15
  capital to the AI stack. And on the consumer
• 31:18
  side, it was about rejecting fragility and rejecting
• 31:21
  intrusion. When you step back and look at all
• 31:23
  these themes together, the efficiency shock,
• 31:25
  the agentic trust gap, the infrastructure failures,
• 31:28
  the governance fracture, it really paints a clear
• 31:30
  picture. Year of recalibration. It really was.
• 31:33
  If we try to synthesize everything we've talked
• 31:35
  about, it feels like the industry learned three
• 31:37
  core, inescapable lessons this year. I agree.
• 31:39
  The first, driven by DeepSeek, is simply efficiency
• 31:42
  over access. The brute force era is over. It's
• 31:45
  over. DeepSeek proved that smart algorithms can
• 31:48
  trump sheer scale. You can now get state -of
• 31:51
  -the -art reasoning for a tiny fraction of the
• 31:53
  old cost, which opens the door for startups and
• 31:56
  global competition in a way that wasn't possible
• 31:59
  before. Okay, lesson number two. Infrastructure
• 32:01
  matters. The AWS and Cloudflare outages were
• 32:05
  a brutal lesson that your amazing autonomous
• 32:08
  agents are completely useless if the underlying
• 32:11
  plumbing is broken. Reliability and redundancy
• 32:13
  are no longer optional. They're the price of
• 32:16
  admission. They're table stakes for any serious
• 32:18
  enterprise automation. And the third lesson,
• 32:20
  which feels the most human, is that trust is
• 32:22
  the currency. Absolutely. Whether it's an IT
• 32:25
  department refusing to give an agent right access
• 32:27
  to a database or a consumer rejecting an intrusive
• 32:30
  necklace, adoption hinges entirely on reliability.
• 32:33
  And on enforcing basic security hygiene like
• 32:36
  MFA. And on fundamentally respecting user privacy
• 32:38
  and control. Without that trust, none of the
• 32:41
  rest of it matters. Looking back, what surprised
• 32:44
  you the most this year? You know, it was the
• 32:46
  sheer simplicity of why the big infrastructure
• 32:48
  failed. We're all braced for these complex nation
• 32:52
  -state hacks. Right. But AWS and Cloudflare were
• 32:55
  taken down by basic configuration bugs and timing
• 32:59
  issues. It's a sobering reminder that massive
• 33:02
  systems fail in the most mundane human ways.
• 33:06
  For me, it was the speed of the federal deregulation
• 33:11
  in the US. That immediate pivot to a move fast
• 33:15
  mentality really caught a lot of people off guard
• 33:17
  and created that huge state versus federal conflict.
• 33:20
  And what normalized really quickly? The concept
• 33:22
  of the AI agent. Even as a co -pilot, the idea
• 33:25
  of a digital employee pursuing a goal across
• 33:27
  different apps, that just became standard. I
• 33:29
  agree. And chain of thought reasoning went from
• 33:31
  a novel research paper idea to just an expected
• 33:34
  feature of any competitive model. It all just
• 33:37
  shows how deeply technology is now shaping everything,
• 33:40
  from the literal price of intelligence to the
• 33:42
  security steps you have to take to buy a concert
• 33:44
  ticket. It's woven into the fabric of everything.
• 33:47
  Before we look ahead, I think we should thank
• 33:48
  our final sponsor. We should. This special edition
• 33:51
  is also brought to you by Skony, the premium
• 33:54
  golf shoe store. Skony really believes that performance
• 33:57
  and style should go hand in hand. On and off
• 34:00
  the course, their shoes are built for maximum
• 34:02
  comfort and stability, designed for peak performance
• 34:05
  through every swing. So if you're looking to
• 34:07
  elevate your game with some truly high performance
• 34:09
  footwear, you should visit scony .com. And Scony
• 34:12
  is offering our listeners 15 % off their entire
• 34:15
  purchase when they use the promo code SNARFL
• 34:17
  at checkout. That's the code SNARFL, S -N -A
• 34:21
  -R -F -U -L, for 15 % off your entire purchase
• 34:25
  at scony .com. So carrying all these lessons
• 34:27
  from 2025 forward, what grounded trends do you
• 34:31
  see defining the next 12 months? Where does this
• 34:33
  recalibration lead us? I think there are three
• 34:35
  big ones that flow directly from this year. First,
• 34:38
  in quantum computing, I see the quibbit quality
• 34:41
  race. Mirroring the efficiency shock in AI. Exactly.
• 34:45
  After all the skepticism around some of the big
• 34:47
  announcements this year, the focus is shifting.
• 34:49
  It's no longer about who has the most quibbits,
• 34:51
  it's about who has the best quibbits, prioritizing
• 34:54
  fidelity and quality over sheer quantity. It's
• 34:57
  another example of efficiency over excess. What's
• 35:00
  the second one? The second is a direct consequence
• 35:02
  of those outages. The mandate for multi -cloud
• 35:05
  resilience. It's not a preference anymore. It's
• 35:07
  a requirement. It's true. And it taught us that
• 35:10
  technology often moves forward on these two separate
• 35:14
  parallel tracks. What do you mean by that? Well
• 35:17
  on one track you have the high -risk physical
• 35:19
  frontier You have SpaceX finally succeeding with
• 35:23
  the chopstick catch of their super heavy booster
• 35:25
  an incredible engineering achievement It is and
• 35:28
  it's all about prioritizing speed and rapid reusability
• 35:31
  to bring down costs It's a move fast and break
• 35:34
  things. Yep Philosophy applied to giant rockets.
• 35:38
  Okay, so that's one track and on the other track
• 35:40
  you have NASA delaying the Artemis second mission
• 35:43
  again pushing it to 2026 because of Concerns
• 35:46
  about the heat shield safety prioritizing astronaut
• 35:49
  safety above all else exactly technology moves
• 35:52
  forward But it often does so by making a choice
• 35:55
  between those two tracks prioritizing speed or
• 35:58
  prioritizing safety And I think that divergence
• 36:00
  and the necessity of making that choice is what
• 36:03
  really gonna find this year of recalibration
• 36:04
  It was a fascinating year of hard -won lessons.
• 36:07
  It certainly was we're here to help it snarfle
• 36:09
  So reach out on our website if you have any more
• 36:11
  questions That's a wrap on today's top stories.
• 36:18
  Remember, tech doesn't have to be complicated,
• 36:20
  and we're here to help connect the dots. If you've
• 36:23
  got questions or want to learn more, visit us
• 36:25
  at snarful .com. Thanks for listening, and we'll
• 36:28
  catch you next time.