Latest / Connecting the Dots / DeepSeek V4 Unveiled, Million-Token Context, and The AI Race Intensifies

Transcript

DeepSeek V4 Unveiled, Million-Token Context, and The AI Race Intensifies · Connecting the Dots

• 0:00
  Today's tech news moves fast, but understanding
• 0:03
  it shouldn't be hard. That's why Snarfle 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. Imagine waking up to find out that the
• 0:19
  smartest intern in your office... now works for
• 0:22
  like 14 cents a day, never sleeps, and has a
• 0:25
  photographic memory of literally every document
• 0:28
  your company has ever produced. Right, which
• 0:30
  sounds totally crazy. It sounds like science
• 0:32
  fiction, but as of this morning, it's the new
• 0:35
  reality of the global economy. I'm Alex. And
• 0:38
  I'm Morgan. And it is Friday, April 24th, 2026.
• 0:42
  Welcome everyone to today's deep dive right here
• 0:45
  on connecting the dots brought to you by snarfle
• 0:48
  solutions group in Sacramento, California So
• 0:50
  glad you're here with us. We've got a really
• 0:52
  dense highly disruptive Morningstar report for
• 0:55
  you today But first let's get a quick snapshot
• 0:58
  of the world outside your window and the numbers
• 1:00
  on your screens Yeah, so looking out over Sacramento
• 1:02
  overnight. It's mostly cloudy sitting at a very
• 1:05
  calm 47 degrees Fahrenheit with just you know,
• 1:09
  the gentle two mile per hour wind Sounds nice.
• 1:13
  Very peaceful on the local front, yeah. That
• 1:15
  piece is highly localized, though. I mean, if
• 1:17
  you zoom out and look at the national map, there
• 1:19
  are currently 378 active weather alerts across
• 1:22
  the country. Wow. Yeah, with 141 of those over
• 1:25
  land. It's a massive contrast, right? The calm.
• 1:29
  47 degrees here versus the 378 alerts nationwide,
• 1:33
  it kind of feels like sitting in the eye of a
• 1:34
  hurricane. It really does. And honestly, that
• 1:37
  atmospheric contrast is, well, it's the perfect
• 1:40
  metaphor for the tech world today. On the surface,
• 1:42
  if you're just checking your email or your calendar,
• 1:45
  everything feels totally normal. Right. Business
• 1:47
  as usual. Exactly. But underneath, there are
• 1:50
  massive storms brewing that are actively reshaping
• 1:53
  how human beings interact with computers. And
• 1:55
  the markets are definitely feeling some of that.
• 1:59
  Atmospheric pressure and they try to price in
• 2:01
  these technological shifts. Oh, absolutely. Let's
• 2:03
  look at Thursday's close So the Dow Jones ended
• 2:05
  the day at forty nine thousand three hundred
• 2:07
  and ten point three two Which is a just a hair
• 2:09
  zero point zero five percent barely moving and
• 2:12
  the S &P 500 closed down slightly at seven thousand
• 2:14
  one hundred eight point four zero Which is a
• 2:16
  drop of zero point one five percent. Mm -hmm.
• 2:19
  Meanwhile looking at crypto this Friday morning
• 2:21
  Bitcoin is sitting at seventy eight thousand
• 2:23
  one hundred and four point seven two. That's
• 2:25
  down about zero point two two percent for the
• 2:28
  day. But, you know, the broader weekly trend
• 2:30
  shows it up nearly six percent since Monday.
• 2:33
  Investors are clearly just holding their breath
• 2:35
  right now. There is a lot of capital waiting
• 2:37
  on the sidelines to see how how this next generation
• 2:41
  of artificial intelligence is really going to
• 2:43
  filter through the broader economy. Which brings
• 2:46
  us to the core of today's tech report. DeepSeek,
• 2:50
  the AI lab that essentially turned the industry
• 2:52
  upside down a year ago, just released the preview
• 2:54
  versions of their new V4 models. Huge news. It's
• 2:57
  sending absolute shockwaves through Silicon Valley,
• 3:01
  Wall Street, and honestly, global hardware supply
• 3:04
  chains. Yeah, the implications touch literally
• 3:07
  everything from the fundamental cost of software
• 3:09
  development to the geopolitical balance of power
• 3:12
  in microchip manufacturing. It's a lot to take
• 3:15
  in. Okay, let's unpack this, but before we get
• 3:17
  into the massive of DeepSeek V4, we have a quick
• 3:20
  word from our sponsor. Today's episode is brought
• 3:22
  to you by Strong Coffee, the premium instant
• 3:24
  coffee that tastes like a traditional brew but
• 3:27
  is packed with 15 grams of protein, MCTs, and
• 3:30
  adaptogens like ashwagandha. Get clean, sustained
• 3:33
  energy without the jitters. Use promo code SNARFL
• 3:36
  for 20 % off at strongcoffeecompany .com. All
• 3:40
  right, let's look closer at what actually dropped
• 3:41
  today. DeepSeek released two preview models.
• 3:44
  Pretty good, yeah. The heavy way is DeepSeek
• 3:47
  V4 Pro, which boasts a staggering 1 .6 trillion
• 3:50
  total parameters. Then there's the smaller, faster
• 3:53
  version V4 Flash with 284 billion parameters.
• 3:58
  Both models feature what they call a 1 million
• 4:00
  token context window. Just pausing here, 1 .6
• 4:04
  trillion parameter sounds less like a software
• 4:06
  program and more like, I don't know, a national
• 4:08
  economy? It's an unfathomable number. How does
• 4:11
  a computer even load something that massive without
• 4:13
  catching fire? Well, the secret is that it doesn't
• 4:15
  load all of them at once. Okay. The architecture
• 4:17
  here relies on a mechanism called mixture of
• 4:19
  experts. Okay. So out of those 1 .6 trillion
• 4:22
  parameters in the pro model, only 49 billion
• 4:25
  are active at any given time. Wait, only 49 billion?
• 4:28
  Yeah, just a fraction. And with the flash model,
• 4:31
  Out of 284 billion, only 13 billion are active.
• 4:35
  Okay, I need a translation for that. What does
• 4:37
  it actually mean for a parameter to be inactive?
• 4:41
  So think of a parameter as a tiny specialized
• 4:45
  piece of knowledge or like a specific connection
• 4:49
  between ideas. If we use a corporate analogy,
• 4:52
  imagine a gigantic sprawling global enterprise
• 4:56
  with 1 .6 trillion employees. That's a big HR
• 5:00
  department. Right. But if you need a new logo
• 5:02
  designed, you don't call an all -hands meeting
• 5:04
  and ask all 1 .6 trillion people for their input.
• 5:07
  No, that would be chaos. Exactly. It would paralyze
• 5:09
  the company and cost an absolute fortune in hourly
• 5:12
  wages. Instead, you just route the request to
• 5:14
  the specific marketing and design department,
• 5:16
  the 49 billion specialists who actually know
• 5:19
  how to do that specific job. And the rest of
• 5:21
  them? The rest of the company stays asleep. Oh,
• 5:23
  wow. So you have the latent brain power of the
• 5:25
  massive company, but the real -time computing
• 5:27
  costs of just the small specialized team you
• 5:29
  actually need for that specific second. Precisely.
• 5:32
  It routes the prompt through the neural network,
• 5:35
  only activating the pathways that are relevant
• 5:38
  to the user's request. OK, that makes the efficiency
• 5:41
  make a lot of sense. Let's talk about the memory
• 5:43
  side of this, though. The 1 million token context
• 5:47
  window. Yeah, that's huge. Because a token is
• 5:49
  basically a chunk of a word. A million tokens
• 5:52
  is a lot of words. Is that like the AI being
• 5:55
  able to hold an entire epic fantasy book series
• 5:58
  in its head at once without forgetting what happened
• 6:00
  to a minor character back in chapter one? Yeah,
• 6:03
  it's equivalent to handing the AI a giant stack
• 6:06
  of novels plus several thick financial reports
• 6:08
  and a massive code base, and it can actively
• 6:11
  reference all of it simultaneously. It doesn't
• 6:13
  lose the plot halfway through a long conversation.
• 6:16
  Exactly. And the benchmark testing shows this
• 6:18
  V4 Pro model beats every single other open source
• 6:21
  model currently available in math, stem, and
• 6:24
  coding tasks. I do want to clarify one thing,
• 6:26
  though, because our sources are pretty specific
• 6:28
  here. It's beating the open source models. Right.
• 6:30
  Open source. But when you stack it up against
• 6:33
  the absolute top tier closed source frontier
• 6:36
  models like, you know, OpenAI's GPT 5 .4 or Google's
• 6:40
  Gemini 3 .1 Pro, it's not taking the crown. No,
• 6:44
  it's not. Deepseek's own documentation openly
• 6:47
  admits they are trailing those absolute bleeding
• 6:50
  edge models by about three to six months in terms
• 6:53
  of pure complex reasoning and broad world knowledge.
• 6:56
  Which, I mean, three to six months in the AI
• 6:58
  world is practically a lifetime. Oh, really is.
• 7:00
  But it's still an incredibly narrow gap historically.
• 7:03
  I'm fascinated by how they manage the reasoning
• 7:05
  process itself, too. The documentation mentions
• 7:07
  three distinct modes. Non -think. Think high
• 7:11
  and think max. Why wouldn't I just want my AI
• 7:14
  to think at the maximum level all the time? Because
• 7:16
  thinking consumes compute. And compute equals
• 7:19
  time and, more importantly, money. Right. Non
• 7:22
  -think is for fast, intuitive responses, like
• 7:24
  asking it to reformat an email or just summarize
• 7:27
  a short article. It just relies on its pre -trained
• 7:29
  reflexes. Makes sense. Think -high introduces
• 7:31
  a conscious logical analysis loop for more complex
• 7:33
  planning, like writing a multi -step project
• 7:36
  proposal. And think -max. Think -max pushes the
• 7:39
  AI's reasoning capabilities to the absolute limit.
• 7:42
  It might spend minutes churning through possibilities,
• 7:45
  verifying its own logic. and exploring dead ends
• 7:48
  before it finally gives you an answer to a highly
• 7:51
  complex math or coding problem. It's a literal
• 7:54
  slider for how hard you want the machine to concentrate.
• 7:58
  Exactly. But here is the part I'm kind of struggling
• 8:00
  with. If I'm listening to this, I'm thinking,
• 8:01
  OK, a Chinese AI company released a very good
• 8:04
  model. It's highly efficient. It has massive
• 8:07
  memory, but it's still technically trailing the
• 8:09
  American tech giants by a few months. Right.
• 8:12
  Why is this dominating the global news cycle
• 8:14
  today? Why do we care so much about the runner
• 8:17
  up? The answer comes down to the price tag. The
• 8:19
  disruption isn't about being the absolute smartest
• 8:22
  model out there. It's about the cost of intelligence
• 8:25
  plummeting to near zero Okay deep seek v4 pro
• 8:28
  costs a dollar and 74 cents per million input
• 8:32
  tokens That is incredibly cheap for a frontier
• 8:35
  class model very cheap but the V4 Flash model
• 8:38
  is a staggeringly low 14 cents per million input
• 8:41
  tokens. 14 cents. That makes it the absolute
• 8:45
  cheapest in its class, severely undercutting
• 8:48
  OpenAI's smallest model, GPT 5 .4 Nano, which
• 8:53
  currently sits at 20 cents. Let's contextualize
• 8:56
  that 14 cents, because for the average professional
• 8:58
  sending, I don't know, a few prompts a day, the
• 9:02
  difference between 20 cents and 14 cents sounds
• 9:04
  like literal pocket change. Like, who cares if
• 9:08
  my prompt costs a fraction of a penny less? A
• 9:11
  human typing prompts into a chat window won't
• 9:13
  notice the difference at all. But we are moving
• 9:15
  away from that chatbot paradigm. The industry
• 9:18
  is shifting entirely toward AI agents. Okay,
• 9:21
  let's unpack this concept of an agent. An agent
• 9:23
  isn't just a search engine that talks back to
• 9:25
  you. An agent is a system that does work for
• 9:27
  you in the background. Exactly. If I ask an AI
• 9:30
  agent to build a basic app, it doesn't just instantly
• 9:32
  spit out a finished product. It enters a loop.
• 9:35
  Right, it iterates. It writes the first draft
• 9:37
  of the code, then it reads its own code. catches
• 9:40
  a syntax error, rewrites it, tests it again,
• 9:42
  realizes the logic is broken, searches for a
• 9:44
  solution, and fixes it. And every single time
• 9:48
  it talks to itself in that loop, it is generating
• 9:51
  and consuming tokens. Right. It might iterate
• 9:54
  50 or 100 times before it ever shows you the
• 9:57
  final result. That is an agentic workflow. And
• 10:01
  it burns through massive amounts of computing
• 10:03
  power. Which perfectly explains the recent news
• 10:07
  about GitHub Copilot. Our sources note they actually
• 10:10
  had to restrict their individual plans recently
• 10:12
  because users were deploying these agents to
• 10:15
  do complex parallel tasks and the compute costs
• 10:18
  were just bankrupting the service. Yeah, the
• 10:21
  AI was thinking too much in the background. Literally
• 10:23
  thinking them out of business. Exactly. When
• 10:25
  an agent loops a thousand times to solve a single
• 10:28
  problem, a difference of a few cents per million
• 10:30
  tokens suddenly scales into millions of dollars
• 10:33
  in overhead for a tech company. It's like the
• 10:35
  cost of electricity suddenly dropping from standard
• 10:37
  rates down to fractions of a penny. Oh, that's
• 10:39
  a great way to put it. You just stop worrying
• 10:41
  about leaving the lights on. In this case, you
• 10:44
  stop worrying about leaving your AI agents running
• 10:46
  24 -7. You can just let them grind away at iterating
• 10:49
  code, organizing schedules, and analyzing data
• 10:52
  all day and night because the underlying electricity
• 10:54
  costs almost nothing. And the mechanism making
• 10:57
  this cheap electricity possible goes back to
• 10:59
  a severe reduction in FLOPs. FLOPs? Yeah, FLOP
• 11:03
  stands for Floating Point Operations Per Second.
• 11:06
  It's essentially the raw number of mathematical
• 11:08
  calculations the computer has to perform. Okay.
• 11:11
  DeepSeq managed to engineer V4 Pro. so that requires
• 11:15
  only 27 % of the FLOPs of their previous V3 .2
• 11:19
  model when handling those massive million token
• 11:23
  context tasks. Wait, they cut the required math
• 11:25
  by over 70 %? They figured out how to get the
• 11:28
  same exact output with vastly fewer calculations.
• 11:31
  That's why. And fewer calculations mean the physical
• 11:33
  servers in the data center draw less electricity,
• 11:35
  generate less heat, and require less cooling.
• 11:38
  The entire physical infrastructure required to
• 11:40
  run the model shrinks. Which brings us nicely
• 11:42
  to the physical infrastructure, the third major
• 11:45
  thread in today's report. The geopolitics of
• 11:47
  hardware. Alright, this is where it gets complicated.
• 11:50
  Because there is a massive contradiction here
• 11:52
  that is currently making Wall Street lose its
• 11:54
  mind. To train a 1 .6 trillion parameter model,
• 11:59
  even a highly efficient one, requires an unbelievable
• 12:02
  amount of specialized hardware. We're talking
• 12:05
  about massive clusters of top -tier AI chips.
• 12:09
  Yeah, and the United States has strict export
• 12:10
  controls in place specifically designed to prevent
• 12:13
  the most advanced AI chips, like the high -end
• 12:17
  hardware, from companies like Nvidia from reaching
• 12:20
  China. Right. The whole goal of those sanctions
• 12:22
  is to physically starve the computing clusters
• 12:25
  needed to train models exactly like this one.
• 12:27
  So I'm looking at these numbers and I'm genuinely
• 12:29
  confused. If you need a warehouse of specialized
• 12:32
  restricted chips to train 1 .6 trillion parameters
• 12:35
  and the U .S. blocked the good stuff from going
• 12:38
  to China. How does DeepSeek even exist? It's
• 12:42
  a million dollar question. There were rumors
• 12:44
  they somehow smuggled in restricted Nvidia Blackwell
• 12:47
  chips, but the latest reporting completely pivots
• 12:49
  away from that. Right. Well, Huawei officially
• 12:51
  stepped into the spotlight this morning. Oh,
• 12:53
  wow. They announced that their domestic AI computing
• 12:55
  cluster, which is powered entirely by their own
• 12:58
  proprietary Ascend 950 processors, provides full
• 13:03
  native support for running these new DeepSeek
• 13:06
  V4 models. OK, we need to look at this impartially
• 13:08
  because the geopolitical fallout here is incredibly
• 13:11
  complex. From the Western perspective, there
• 13:14
  are immediate national security concerns. Absolutely.
• 13:17
  DeepSeek was already banned from being used on
• 13:19
  US federal devices and government networks over
• 13:22
  data privacy and security fears. And South Korea
• 13:25
  recently paused downloads of the DeepSeek app
• 13:27
  for similar reasons. Those are the stated security
• 13:30
  policies from allied governments. But the financial
• 13:33
  markets are telling a very specific, very different
• 13:36
  story about how investors interpret the Huawei
• 13:39
  announcement. How so? When the news broke that
• 13:41
  Huawei chips were natively running V4, Chinese
• 13:44
  domestic chipmakers saw a massive surge. SMIC,
• 13:48
  which is a major semiconductor manufacturer in
• 13:50
  China, spiked by 10 % almost immediately. I'm
• 13:53
  looking at that S -Mix bike, but at the exact
• 13:55
  same time, Chinese AI application companies like
• 13:59
  the companies building software on top of AI
• 14:01
  actually saw their stock drop. Yeah, they did.
• 14:04
  That seems completely backward to me. If domestic
• 14:07
  AI just became incredibly cheap and powerful,
• 14:11
  shouldn't the app companies be thrilled? Why
• 14:13
  is Wall Street punishing the software guys while
• 14:16
  rewarding the chipmakers? It really comes down
• 14:18
  to commoditization and the threat of a brutal
• 14:20
  price war. The surge in domestic chip stocks
• 14:23
  like SMIC suggests investors believe the hardware
• 14:26
  bottleneck has been broken. They are pricing
• 14:29
  in the reality that China has successfully built
• 14:31
  its own self -sufficient AI hardware ecosystem,
• 14:36
  essentially bypassing the US sanctions through
• 14:39
  domestic innovation. So the chip makers now have
• 14:41
  a captive booming market. And the software companies.
• 14:44
  Well, the drop in AI application companies reflects
• 14:46
  the reality that if intelligence is now virtually
• 14:49
  free, the barrier to entry for building an AI
• 14:51
  app just fell to zero. Ah, I see. If anyone can
• 14:55
  spin up a hyper intelligent agent for pennies,
• 14:57
  the existing software companies no longer have
• 14:59
  a protective mode. Wall Street is anticipating
• 15:02
  a race to the bottom for software pricing. A
• 15:04
  race to the bottom for software, but a gold rush
• 15:06
  for the foundational players. Our report mentions
• 15:10
  major tech conglomerates like Tencent and Alibaba
• 15:13
  are now looking to invest heavily in DeepSeek,
• 15:15
  pushing its valuation to over $20 billion. $20
• 15:19
  billion for a lab that just proved you can build
• 15:21
  and run frontier -level artificial intelligence
• 15:23
  without relying on the traditional Western -dominated
• 15:26
  hardware pipeline. It's wild. It is a fundamental
• 15:28
  restructuring of the global tech stack, from
• 15:31
  the silicon forge in the foundries all the way
• 15:33
  up to the conversational agent running on your
• 15:35
  phone. Before we wrap up this look at the new
• 15:38
  landscape, a quick thanks to Old Glory, an iconic
• 15:40
  music and sports fan merch store, perfect for
• 15:43
  the techie in your life. Use PROMSE code SNARFL
• 15:45
  for 15 % off at oldglory .com. So to bring all
• 15:50
  these threads together, what we are witnessing
• 15:52
  today is a paradigm shift in both capability
• 15:55
  and access. We have open source models that can
• 15:57
  hold a million tokens of context, reason through
• 16:00
  complex math and coding, and operate at near
• 16:02
  state -of -the -art levels. And crucially, they're
• 16:05
  doing it at a price point 14 cents per million
• 16:08
  tokens. That makes widespread, always on AI agents
• 16:12
  viable for everyday people. Exactly. You no longer
• 16:15
  need a massive corporate budget to have a digital
• 16:17
  assistant iterating and working for you around
• 16:20
  the clock. And finally, the software revolution
• 16:23
  is actively accelerating a geopolitical hardware
• 16:26
  shift. proving that an independent non -western
• 16:30
  silicon ecosystem is currently capable of sustaining
• 16:33
  frontier AI development. It is just a massive
• 16:36
  amount of change to absorb in a single morning.
• 16:39
  And I want to leave you with a final thought
• 16:40
  to mull over as you go about your day. Okay.
• 16:42
  We spent a lot of time talking about how cheap
• 16:44
  this intelligence has become and how agents loop
• 16:47
  in the background to do work. Right. If ultra
• 16:49
  -smart long -memory AI models are now practically
• 16:51
  free to run, what happens when personal AI agents
• 16:55
  become the default way we interact with the internet?
• 16:57
  Oh wow. Like what happens when you don't even
• 16:59
  download standard apps or scroll through social
• 17:01
  media feeds anymore because your personal AI
• 17:03
  just gathers, reads, and custom summarizes the
• 17:07
  entire web for you directly. It dismantles the
• 17:09
  entire advertising and engagement business model
• 17:13
  the current internet is Beth on. It really does.
• 17:15
  It's a whole new world out there. We're here
• 17:17
  to help at Snarfful, so reach out on our website
• 17:19
  if you have any more questions. That's a wrap
• 17:25
  on today's top stories. Remember, tech doesn't
• 17:28
  have to be complicated, and we're here to help
• 17:30
  connect the dots. If you've got questions or
• 17:33
  want to learn more, visit us at snarfle .com.
• 17:36
  Thanks for listening, and we'll catch you next
• 17:38
  time.