Latest / Connecting the Dots / AI Consciousness Debates, Gemma 4 12B, and Local macOS AI

Transcript

AI Consciousness Debates, Gemma 4 12B, and Local macOS AI · 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 to Connecting the Dots. I'm Alex.
• 0:17
  And I'm Morgan. We're representing Snarfle Solutions
• 0:20
  Group and we are coming to you live right here
• 0:23
  from sunny Sacramento, California. Really sunny
• 0:25
  today, actually. Yeah, it's definitely a hot
• 0:27
  one. Just to get you all oriented before we start,
• 0:30
  it is Thursday, June 4, 2026. And like you said
• 0:34
  here in Sacramento, it is gorgeous, but we are
• 0:37
  looking at a high of 96 degrees today. Yeah,
• 0:40
  that is definitely air conditioning weather.
• 0:43
  Oh, absolutely. And we've got some light winds,
• 0:45
  you know, around one to seven miles per hour.
• 0:48
  But if you happen to be traveling out of town,
• 0:50
  just a quick heads up, there are currently 120
• 0:53
  active weather alerts across the U .S. right
• 0:55
  now. Wow, that's a lot. Yeah. And about 43 of
• 0:58
  those are over land. So definitely check your
• 1:00
  local weather app if you're flying anywhere.
• 1:02
  For sure. Speaking of things getting a little
• 1:04
  heated, Wall Street had a pretty rough day yesterday
• 1:07
  too. Oh, you can say that again. It was just
• 1:08
  a decidedly red day for the markets all around.
• 1:11
  Yeah, totally red. The Dow Jones took a real
• 1:13
  hit, dropping over 533 points, which is, you
• 1:16
  know, a 1 .04 % slide. Yeah, and the S &P 500
• 1:20
  wasn't spared either. That slipped about 51 points.
• 1:23
  And if you're holding any crypto right now, Bitcoin
• 1:25
  had a pretty rough Thursday morning. It really
• 1:28
  did. It dropped almost $1 ,000 to close, hovering
• 1:30
  right around $63 ,041. So yeah, some definite
• 1:35
  economic anxiety in the air today. Well, and
• 1:37
  honestly, that kind of anxiety ties in perfectly
• 1:39
  to what we're talking about today. Yeah. Because
• 1:41
  we are looking at the tech sector and, you know,
• 1:44
  who really holds the power right now. Exactly.
• 1:46
  Let's unpack this. Today's episode is brought
• 1:49
  to you by Strong Coffee, the premium instant
• 1:51
  coffee that tastes like a traditional brew, but
• 1:54
  is packed with 15 grams of protein, MCTs, and
• 1:57
  adaptogens like ashwagandha. Get clean, sustained
• 2:00
  energy without the jitters. Use promo code SNARFL
• 2:03
  for 20 % off at strongcoffeecompany .com. All
• 2:07
  right, so we are starting off today by looking
• 2:09
  at the, well really the philosophical and societal
• 2:12
  impact of these. massive cloud -based AI models.
• 2:16
  Right. Because there is this overwhelming hype
• 2:18
  cycle right now about AI achieving consciousness
• 2:20
  and they're selling us this glittering promise
• 2:22
  of endless productivity. Yeah, which is just
• 2:25
  everywhere you look. Exactly. But the sources
• 2:27
  we are looking at today actively dismantle that
• 2:30
  whole narrative. They completely tear it down.
• 2:33
  I mean, a really fantastic starting point is
• 2:35
  Ted Chiang's recent article where he analyzes
• 2:37
  the way these tech CEOs are actually talking
• 2:39
  about their models. Oh, this article was so good.
• 2:42
  It was. Yeah. Like, take Anthropic, for example.
• 2:45
  They released an 84 page constitution for their
• 2:49
  flagship model, Claude. Yeah. And the CEO. And
• 2:53
  their in -house philosopher are doing these interviews
• 2:56
  where they suggest they want Claude to be happy.
• 2:58
  Right. And they're worrying about it getting
• 3:00
  anxious. They're purposefully keeping the door
• 3:02
  open to the idea that Claude might actually possess
• 3:05
  some form of real consciousness. Which Shang
• 3:07
  argues is just massive anthropomorphism, right?
• 3:10
  Like, it is a basic human survival instinct to
• 3:13
  project a soul onto anything that talks back
• 3:15
  to us. Precisely. We are hardwired to do that.
• 3:18
  But he makes this really vital point that These
• 3:20
  large language models are fundamentally just
• 3:23
  highly sophisticated sentence continuation machines.
• 3:26
  It's literally just a high speed game of predictive
• 3:29
  text, you know, scaled up to billions of parameters.
• 3:33
  Right. And if you take that a step further, think
• 3:34
  about it like this. If you ask one of these models
• 3:37
  to generate a conversation between, say, Julius
• 3:40
  Caesar and Genghis Khan. Oh, yeah, the history
• 3:43
  prong. Exactly. It produces a totally coherent,
• 3:47
  completely fascinating dialogue. But no one in
• 3:50
  their right mind would think you actually resurrected
• 3:53
  the soul of Julius Caesar inside a server farm.
• 3:56
  Right, obviously not. You wouldn't think the
• 3:58
  computer is genuinely feeling the existential
• 4:00
  dread of an ancient Roman Emperor facing his
• 4:02
  demise. No, of course not. But, and this is the
• 4:05
  wild part, when we change the prompt to say,
• 4:08
  the following is a conversation between a helpful
• 4:10
  AI chatbot and a user, Suddenly, we completely
• 4:13
  project a soul into the machine. Yep. We fall
• 4:16
  right for it. Why do we fall for it in that one
• 4:17
  scenario, but not the other? Well, because the
• 4:19
  entire system is engineered to mimic human interaction.
• 4:23
  Chang describes Claude's so -called Constitution
• 4:27
  as basically just a role -playing character sheet.
• 4:29
  Oh, like in Dungeons and Dragons. Exactly like
• 4:32
  D &D. It just tells the machine how to mathematically
• 4:34
  calculate the next word so it can play the character
• 4:37
  of a helpful moral entity. But when the computer...
• 4:40
  generates a string of text saying I am sad. There's
• 4:44
  no actual sadness. There is absolutely no subjective
• 4:47
  experience happening behind the screen. He uses
• 4:49
  this incredible analogy about space travel in
• 4:52
  the piece that really stuck with me. He says
• 4:54
  imagine someone shows you a hyper realistic like
• 4:58
  8k defake video of an astronaut orbiting Alpha
• 5:02
  Centauri. Right. Do you suddenly believe we have
• 5:04
  achieved interstellar travel? Obviously not.
• 5:07
  Because we haven't even successfully colonized
• 5:09
  Mars yet. Exactly. We haven't solved the hundreds
• 5:11
  of smaller, really unglamorous engineering problems
• 5:15
  required to get a human to another star system.
• 5:18
  You would just immediately recognize the video
• 5:20
  for what it is. A very impressive fake. And Chang's
• 5:23
  core point is that text is now a deep fake medium
• 5:25
  too. Faking a conversation between two conscious
• 5:28
  beings is vastly, just unimaginably easier than
• 5:32
  actually creating a conscious being from scratch.
• 5:35
  Which brings us to the real world consequences
• 5:37
  of all this illusion. Because it's not just some
• 5:40
  abstract academic debate, you know. It is an
• 5:43
  economic weapon. Yeah, that transitions into
• 5:46
  the piece from The Verge perfectly. Exactly.
• 5:48
  The Verge piece details how this fake consciousness
• 5:51
  is being bundled with what they call the productivity
• 5:54
  scam. And this is where it gets incredibly reliant
• 5:57
  to you, the listener, because think about your
• 5:59
  own job for a second. Companies are selling these
• 6:01
  $99 a month AI assistants to, you know, Manage
• 6:05
  your calendar, color code your spreadsheets,
• 6:07
  draft your emails. And they promise you all this
• 6:09
  free time. Right. But are you actually getting
• 6:11
  that free time back? Like if you use an AI to
• 6:14
  finish your daily tasks by 3 p .m., does your
• 6:16
  boss say, hey, great job, go home early and spend
• 6:19
  time with your kids? No, definitely not. They
• 6:21
  just give you more work. And the article points
• 6:23
  out this stark honestly almost dystopian contrast
• 6:27
  going on right now on one side. You have these
• 6:30
  tech billionaires envisioning a post work utopia
• 6:33
  where algorithms do all the heavy labor and we
• 6:36
  all get to be and this is their actual quote.
• 6:38
  theater kids. Right, just making art and exploring
• 6:41
  our passions. Yeah, enjoying life. But meanwhile,
• 6:44
  those same billionaires are buying 387 -foot
• 6:46
  yachts while simultaneously laying off massive
• 6:49
  chunks of their own human workforces just to
• 6:52
  pay for these new server farms. And at the exact
• 6:54
  same time, we're seeing the social safety net
• 6:56
  being systematically dismantled. The article
• 6:58
  points out that SNAPI benefits are being cut
• 7:01
  for vulnerable families right now. Yeah, the
• 7:03
  timing is pretty grim. So what good is a $99
• 7:06
  AI assistant planning a fun itinerary for your
• 7:09
  weekend if you literally cannot afford to have
• 7:11
  any free time. The productivity output of the
• 7:14
  average worker goes way up, but the wages just
• 7:17
  stagnate. It is the unholy alliance of what sociologists
• 7:21
  are calling the busy trap and the software brain.
• 7:24
  Oh, let's actually define those for the listener,
• 7:26
  because those are great terms. Sure. So the busy
• 7:28
  trap is the phenomenon where we use these time
• 7:30
  -saving tools only to immediately fill that safe
• 7:33
  time with even more trivial tasks, just to prove
• 7:36
  to our bosses that we are constantly working.
• 7:38
  Right. The performative busyness. Exactly. And
• 7:41
  then the software brain is this Silicon Valley
• 7:44
  mindset that views deeply complex, messy human
• 7:47
  problems, things like poverty, burnout, economic
• 7:51
  inequality, purely as optimization puzzles. Like
• 7:55
  they can just be solved with a quick software
• 7:56
  update. Yep. By subscribing to these cloud AI
• 7:59
  services, we are basically offloading our tasks
• 8:02
  to the very companies that engineered this exhausting,
• 8:05
  always on work culture in the first place. I
• 8:07
  hear that, and I definitely see the economic
• 8:09
  trap there. But let me push back on something
• 8:11
  from earlier for a second. Sure, go for it. If
• 8:13
  we are saying AI is just a glorified predictive
• 8:16
  text game, like a deep fake of a conversation,
• 8:19
  How do we explain its ability to write complex,
• 8:22
  professional -level computer code? Because for
• 8:24
  decades, we thought coding required a conscious,
• 8:27
  reasoning mind. If it can do that, why shouldn't
• 8:30
  we trust it for moral reasoning or ethical advice
• 8:33
  like the tech companies are suggesting with those
• 8:35
  constitutions? See, that is the million dollar
• 8:38
  question right there. It is a very natural leap
• 8:40
  for us to make, right? Like, if it is smart enough
• 8:43
  to build an entire app from scratch, it must
• 8:45
  be smart enough to tell me right from wrong.
• 8:47
  Yeah, that seems logical. But here is the critical
• 8:51
  difference. Writing code... no matter how complex
• 8:54
  it gets, actually turned out to just be a highly
• 8:57
  sophisticated pattern matching task. Interesting.
• 9:00
  It is syntax. It is pure logic. It can be solved
• 9:03
  by throwing massive amounts of computational
• 9:05
  horsepower at a vast data set of historical GitHub
• 9:08
  repositories. But moral reasoning is categorically
• 9:11
  different. Wait, I'm not sure I follow. Why can't
• 9:13
  a machine just pattern match morality from, say,
• 9:16
  reading thousands of philosophy books? Because
• 9:18
  moral reasoning requires a physical body. Yeah,
• 9:21
  it requires a history of lived consequences.
• 9:24
  Experiencing an emotion like desperation or empathy
• 9:26
  or guilt that is biologically inseparable from
• 9:30
  stress hormones like cortisol actually flooding
• 9:32
  your physical system. That makes a lot of sense.
• 9:35
  It requires having actual skin in the game. Yeah.
• 9:37
  If you make a bad moral choice, you suffer. A
• 9:40
  server rack cannot suffer. Wow. So when a machine
• 9:43
  outputs a string of text saying, I cannot in
• 9:46
  good conscience do this. It's really no different
• 9:49
  than a pre -recorded voice on a phone tree saying,
• 9:51
  your call is important to us. Right. It means
• 9:54
  absolutely nothing. Wow. Yeah. That completely
• 9:57
  reframes the whole interaction for me. When we
• 10:00
  ask an AI to make moral choices for us in the
• 10:02
  workplace or ask it to navigate complex human
• 10:05
  relationships, we aren't tapping into some superhuman
• 10:08
  unbiased wisdom. No, we are just cowardly offloading
• 10:11
  our own accountability. We're using machines
• 10:13
  for the evasion of moral responsibility. If you
• 10:16
  take a chat bot's advice on who to fire or how
• 10:18
  to discipline an employee, you are the one who
• 10:20
  has to live with the human wreckage afterwards,
• 10:23
  not the algorithm. Well, if giving our data and
• 10:25
  our moral agency to A fake, conscious cloud AI
• 10:30
  feels like a corporate trap. That explains exactly
• 10:33
  why the developer community is suddenly so desperate
• 10:35
  to pull AI offline entirely. Yes, exactly. And
• 10:39
  that desperation is what forced Google's hand
• 10:41
  this week with our next story. The massive pivot
• 10:44
  away from the cloud. The tech world's response
• 10:47
  right now is basically to shrink AI down so you
• 10:49
  can own it entirely, putting all that power directly
• 10:52
  onto your local machine. Right. Enter Google's
• 10:54
  release of the Gemma 412b model. This is just
• 10:57
  a massive milestone. It is a model with 11 .95
• 11:01
  billion parameters. Just to give some context
• 11:03
  for that, a parameter is essentially a single
• 11:05
  connection in the AI's artificial brain. 12 billion
• 11:08
  of them is enough to give the model highly advanced
• 11:10
  logical reasoning. But here is the wild part.
• 11:12
  It has been engineered to be small enough to
• 11:14
  run locally on a standard consumer laptop. Which
• 11:17
  is crazy. Right. All you need is 16 gigabytes
• 11:20
  of video RAM or VRAM or, you know, what Apple
• 11:22
  calls unified memory. And the open source community
• 11:25
  is already devouring it. I mean, the Gemma 4
• 11:27
  models have already crossed 150 million downloads.
• 11:30
  That was fast. But what makes the 12B model a
• 11:34
  true breakthrough is the underlying architecture.
• 11:37
  It is an encoder free model. Okay, wait, I need
• 11:40
  to stop you right there. We promise to keep the
• 11:42
  heavy jargon out of this deep dive. Fair enough.
• 11:45
  If I'm running this on my Mac, is my laptop going
• 11:48
  to sound like a jet engine taking off? Like how
• 11:51
  in the world are they running 12 billion parameters
• 11:53
  without just melting my battery? So it comes
• 11:56
  down to a process called quantization. Okay,
• 11:58
  what is that? Think of it as basically compressing
• 12:00
  the math. Normally a 12 billion parameter model
• 12:04
  uses very precise heavy numbers, like imagine
• 12:08
  calculating pi to 30 decimal places. Okay, sounds
• 12:11
  exhausting. Right, and that requires massive
• 12:13
  power -hungry server farms. What Google did was
• 12:16
  trim off the excessive decimal places. They compressed
• 12:19
  the precision just enough to shrink the memory
• 12:21
  footprint drastically without actually losing
• 12:24
  the core logical reasoning of the model. Oh,
• 12:26
  that's clever. Yeah, that is how it squeezes
• 12:28
  into your laptop's unified memory, which is really
• 12:31
  just the shared pool of RAM that your computer
• 12:33
  uses for both general tasks and graphics. Okay,
• 12:36
  that makes sense. But what about the encoder
• 12:37
  -free part? Why is that such a big deal for the
• 12:40
  performance? Well, to understand the breakthrough
• 12:43
  there, you have to look at how older AI models
• 12:46
  handled different types of media. Historically,
• 12:50
  A multi -modal AI, you know, one that can look
• 12:53
  at pictures or listen to sound function, kind
• 12:56
  of like a bad game of telephone. How so? If you
• 12:59
  showed an older AI a picture, the picture first
• 13:02
  had to go through a separate piece of software
• 13:03
  called a vision encoder. Okay. This is basically
• 13:06
  a specialized sub -program that looked at the
• 13:08
  image, wrote down a text description of what
• 13:10
  it thought it saw, and then handed those text
• 13:13
  tokens off to the main language model. And the
• 13:15
  same went for audio. An audio encoder had to
• 13:17
  listen to it. transcribe the sound into text,
• 13:20
  and then pass it along. But, I mean, you lose
• 13:22
  so much nuance that way. Like, if someone is
• 13:23
  smiling sarcastically in a video, the vision
• 13:25
  encoder might just translate it to the word smile,
• 13:28
  and the main brain completely misses the joke.
• 13:31
  The middle man destroys all the context. So the
• 13:34
  old way was like the AI needed a, um, a UN translator
• 13:38
  whispering in its ear to understand a photo or
• 13:40
  a song. But with this encoder free setup, the
• 13:43
  AI is essentially a natural born polyglot that
• 13:46
  just natively speaks pixels and audio waves.
• 13:49
  That is a brilliant way to conceptualize it actually.
• 13:51
  Yes. Google completely removed the middle man.
• 13:55
  They totally bypass those heavy multi -stage
• 13:57
  vision and audio encoders. Now raw 48 by 48 pixel
• 14:02
  patches and 16 kilohertz audio slices are injected
• 14:06
  straight into the main logic backbone of the
• 14:08
  model. OK, so what is cutting out that telephone
• 14:11
  game actually do for the end user like you and
• 14:13
  me? Two massive things. First, it drastically
• 14:16
  reduces latency. which is just the delay between
• 14:19
  you giving a command and the computer answering
• 14:21
  you. It feels practically instantaneous now.
• 14:23
  That's huge. And second, because those old encoders
• 14:25
  took up a massive amount of memory, dropping
• 14:27
  them freed up space for more advanced reasoning,
• 14:29
  it opens the door for what we call agentic workflows.
• 14:32
  Agentic workflows. Yes. Instead of just giving
• 14:34
  the AI simple prompt like, write me an email,
• 14:37
  agentic workflow means you give it a complex
• 14:38
  goal. Like what? You say, look through this folder,
• 14:41
  find all the sales spreadsheets, compare them
• 14:43
  to last year's, write a summary report and format
• 14:46
  it as a PDF. And the AI can autonomously take
• 14:50
  multiple sequential steps across your computer
• 14:52
  to achieve that goal. Which means the real -world
• 14:55
  impact for working professionals is immense.
• 14:59
  We are talking about being able to feed a five
• 15:01
  -minute video file directly into your laptop
• 15:03
  and the AI can analyze the whole thing, the sight,
• 15:07
  the sound, the emotional context at one frame
• 15:10
  per second completely locally. And here's the
• 15:14
  real kicker. Because it is running locally on
• 15:16
  your hardware, You aren't paying a $99 monthly
• 15:19
  cloud subscription. You don't even need an internet
• 15:22
  connection. Which brings us to the final piece
• 15:24
  of the puzzle today. Because having an incredibly
• 15:26
  efficient local AI engine is great, but, you
• 15:30
  know, a raw language model running in a terminal
• 15:32
  window isn't exactly user -friendly for someone
• 15:35
  who doesn't know how to code. Right. It's pretty
• 15:37
  intimidating. How do everyday people actually
• 15:39
  drive this thing? Well, Google saw that exact
• 15:42
  bottleneck. So they have officially released
• 15:45
  Mac OS desktop applications specifically designed
• 15:48
  to bring Gemma 412b to life for the average computer
• 15:52
  user. And there are two main apps we need to
• 15:54
  highlight here. The first is called AI Edge Gallery.
• 15:57
  Right. AI Edge Gallery is essentially a local
• 16:00
  AI showcase app. It provides a beautiful user
• 16:02
  interface and runs entirely offline, natively
• 16:05
  on Apple Silicon GPUs. Though we do have to mention
• 16:09
  the rather funny caveat from the app insider
• 16:11
  breakdown. It works beautifully on all modern
• 16:13
  Apple laptops, with the notable exception of
• 16:16
  the MacBook Neo, which apparently is excluded
• 16:18
  from the fun for some mysterious hardware reasons.
• 16:21
  Yeah, a bit of a stray casualty there for the
• 16:22
  Neo users. But for everyone Anyone else on a
• 16:25
  standard M -Series Mac, this app is incredibly
• 16:27
  powerful. It gives you a highly secure, sandboxed
• 16:30
  environment to do local coding and heavy data
• 16:32
  analysis. The article gives a fantastic real
• 16:34
  -world example, actually. You could have two
• 16:37
  raw, messy text files full of baby names from
• 16:39
  2024 and 2025. You just ask the AI in plain conversational
• 16:43
  English to compare them. And because it supports
• 16:46
  those agentic workflows you mentioned earlier,
• 16:48
  the model dynamically generates a Python script.
• 16:51
  executes that script locally in the background,
• 16:53
  and instantly spits out a beautiful comparison
• 16:56
  chart. Exactly. All without your data ever pinging
• 16:59
  a server in California. And the second app is
• 17:01
  perhaps even more relevant for the everyday professional.
• 17:05
  It's called AI Edge Eloquent. This is an AI powered
• 17:09
  voice dictation and text editing app. And again,
• 17:12
  100 % on device, right? No Wi -Fi needed at all.
• 17:15
  Correct. What makes Eloquent game changing is
• 17:18
  a feature called voice edit. You can literally
• 17:21
  highlight a paragraph of text in any application
• 17:24
  on your Mac, say, an email draft, a Word document,
• 17:27
  or some messy meeting notes. You hit a hotkey
• 17:30
  and you just speak a command out loud. You say,
• 17:32
  restructure these notes into an executive summary
• 17:34
  or translate this paragraph into Hindi. And because
• 17:37
  it's powered by the raw logic engine of Gemma
• 17:40
  412b, the instruction following and the quality
• 17:43
  of the output are vastly superior to those older
• 17:46
  clunky on -device models we've had in the past.
• 17:48
  Oh, totally. The ultimate benefit here for you?
• 17:51
  cannot be overstated. Zero latency, zero subscription
• 17:55
  fees, and total uncompromised data privacy. Yes,
• 18:00
  the privacy is key. Your unstructured thoughts,
• 18:03
  your messy first drafts, your proprietary business
• 18:06
  data. None of it is being harvested by corporate
• 18:09
  servers to train the next iteration of someone
• 18:11
  else's model. I do have to point out the absolute
• 18:13
  irony here, though. Oh, I know what you're going
• 18:15
  to say. I mean, Google, which is arguably one
• 18:17
  of the largest data harvesting cloud companies
• 18:20
  on the entire planet has built an incredibly
• 18:23
  capable private local assistant that shines the
• 18:25
  brightest on Apple hardware. They've essentially
• 18:28
  handed Apple users the exact tool they need to
• 18:31
  protect themselves from the corporate omniscience
• 18:33
  we were just warning everyone about in the first
• 18:35
  segment. It is a profound irony. But it also
• 18:38
  really proves the power of the open source community
• 18:40
  forcing the hands of these tech giants. I mean,
• 18:43
  they have to release these localized private
• 18:46
  tools because the public demand for data sovereignty
• 18:49
  is growing way too loud to ignore. It certainly
• 18:52
  makes you wonder how long before we all need
• 18:54
  to upgrade our laptops just to keep up with the
• 18:57
  memory requirements for these massive local models.
• 18:59
  Oh, I think Apple and PC manufacturers are banking
• 19:02
  heavily on exactly that kind of Hardware upgrade
• 19:05
  super cycle well before we wrap up and summarize
• 19:08
  where this all leaves us a quick Thanks to old
• 19:11
  glory an iconic music and sports fan merch store
• 19:14
  Perfect for the techie in your life use promo
• 19:17
  code snarfle for 15 % off at old glory calm So
• 19:21
  if we pull back and look at the whole picture
• 19:23
  today. We've traced a really fascinating arc
• 19:26
  We really have. We started by, you know, cutting
• 19:28
  through the philosophical hype of cloud AI. We
• 19:31
  looked at how models like Claude are just mimicking
• 19:33
  consciousness by predicting text. Right. And
• 19:36
  how the hollow promise of AI productivity often
• 19:38
  masks a pretty grim economic reality where companies
• 19:41
  capture the profits while everyday workers just
• 19:43
  get squeezed tighter. Then we explored the technical
• 19:47
  breakthrough, offering a way out of that corporate
• 19:49
  cloud trap. Google's Gemma 412b. We saw how compressing
• 19:53
  the math and stripping away the old telephone
• 19:55
  game of encoders allows powerful multimodal reasoning
• 19:58
  to fit directly onto consumer hardware. And finally,
• 20:02
  we looked at how apps like AI Edge Gallery and
• 20:04
  Eloquent are putting that raw, private power
• 20:07
  directly onto our Macs. It gives us high -level
• 20:10
  coding, agentic workflows, and seamless voice
• 20:12
  dictation without ever connecting to the Internet.
• 20:15
  It truly is a paradigm shift. we're moving away
• 20:17
  from AI as a rented cloud -based service toward
• 20:20
  AI as a locally owned utility. It's like having
• 20:23
  a supercomputer right in your backpack. Which
• 20:25
  leaves us with a pretty compelling question to
• 20:27
  chew on today. If open source local AI running
• 20:30
  on our own laptops becomes so incredibly powerful
• 20:33
  that we no longer need the tech giant's expensive
• 20:36
  cloud subscriptions to be productive. Will the
• 20:39
  multi -trillion dollar AI corporate bubble burst?
• 20:42
  Exactly. Will it completely reshape who holds
• 20:44
  the wealth and power in the tech industry? Because
• 20:47
  if the UN translator is suddenly out of a job,
• 20:50
  because we all became natural born polyglots
• 20:53
  overnight, the global market is going to have
• 20:56
  to adjust in ways we can barely predict. We're
• 20:58
  here to help at Snarfl, so reach out on our website
• 21:00
  if you have any more questions. That's a wrap
• 21:05
  on today's top stories. Remember, tech doesn't
• 21:08
  have to be complicated, and we're here to help
• 21:11
  connect the dots. If you've got questions or
• 21:14
  want to learn more, visit us at snarful .com.
• 21:17
  Thanks for listening, and we'll catch you next
• 21:19
  time.