Latest / AI Fire Daily / #666 Neil: Gemini Desktop Takes On Claude Cowork With New AI Skills

Transcript

#666 Neil: Gemini Desktop Takes On Claude Cowork With New AI Skills · AI Fire Daily

• 0:00
  Imagine your computer actually doing your job
• 0:02
  while you sleep. Right. Not just setting alarms
• 0:04
  or backing up files. I mean reading your emails,
• 0:07
  running your spreadsheets, and messaging your
• 0:09
  team. The real work. Exactly. Imagine if your
• 0:12
  screen stopped being just a piece of glass you
• 0:13
  stare at. Imagine it became a co -worker. A colleague
• 0:17
  sitting quietly on your desktop doing the actual
• 0:20
  heavy lifting. Welcome to the Deep Dive. Today
• 0:23
  we are exploring a massive shift in personal
• 0:25
  computing. It really is a massive shift. We are
• 0:28
  unpacking the new Google Gemini desktop update.
• 0:32
  And this isn't just another incremental software
• 0:34
  patch. It is a fundamental change in our relationship
• 0:37
  with machines. Yeah, Google is aggressively pushing
• 0:39
  back here. They are going right after competitors
• 0:42
  like Cloud Cowork and ChatGPT Work. We're going
• 0:45
  to map out this entire journey for you today.
• 0:47
  We are moving away from simply chatting with
• 0:49
  an AI. We are moving toward delegating real complex
• 0:53
  labor. Right. So we'll look at connecting external
• 0:56
  apps, building reusable workflows called skills,
• 0:59
  and finally putting your digital life on autopilot.
• 1:02
  It's a profound transition. I mean, we are moving
• 1:04
  from a passive tool to an active agent. Which
• 1:06
  is a huge leap. It is. For years, we've treated
• 1:09
  AI like a very smart encyclopedia. You ask a
• 1:12
  question, it gives an answer. But now you are
• 1:15
  no longer just asking questions. Exactly. You're
• 1:17
  assigning actual labor. That's a totally different
• 1:20
  paradigm. Before automating anything, we really
• 1:23
  have to understand this fundamental shift. It's
• 1:25
  the difference between asking for advice and
• 1:27
  delegating a task. Yeah, that distinction is
• 1:30
  everything. When you open the new desktop app,
• 1:32
  there is a new interface switch right at the
• 1:34
  top of the prompt box. You literally have to
• 1:36
  choose between chat and task. That visual switch
• 1:40
  is super important. Chat is for brainstorming.
• 1:44
  Right. You have to write ideas. You explore abstract
• 1:46
  concepts. It uses stochastic processing. It's
• 1:49
  basically predicting the next best word to sound
• 1:51
  smart. But task mode is fundamentally different.
• 1:55
  Entirely different. It's built for deterministic
• 1:57
  execution. You're giving the AI a strict mandate
• 2:00
  to go do something specific. Let's break down
• 2:04
  how that actually looks in practice. Say you
• 2:06
  have a massive zip folder full of business files,
• 2:10
  spreadsheets, messy raw data, text documents.
• 2:13
  You drop that entire zipped folder into the task
• 2:15
  box. What actually happens under the hood? This
• 2:19
  is where it gets fascinating. The AI essentially
• 2:21
  mounts that zip file into a secure local sandbox.
• 2:24
  Like a digital clean room. Exactly. The AI is
• 2:28
  locked inside with your files. It can read everything,
• 2:31
  but the doors to the outside internet are bolted
• 2:33
  shut. So data can't leak out. Right. It automatically
• 2:36
  unzips the folder, then it digs into the internal
• 2:39
  sales reports, scans your expense sheets, and
• 2:42
  pulls the marketing data. So it's reading all
• 2:44
  these disparate file types simultaneously. Yeah,
• 2:46
  all at once. And it's not just summarizing them,
• 2:49
  it's actually doing math. It calculates total
• 2:51
  revenue, expenses, and net profit. It's like
• 2:54
  stacking Lego blocks of data. You give it the
• 2:56
  scattered pieces, it builds the complete structure.
• 2:59
  Well, it's doing something even more nuanced
• 3:01
  than just stacking. It's connecting the conceptual
• 3:04
  dots across multiple documents. How so? So in
• 3:07
  one of the main stress tests of this system,
• 3:09
  the AI spotted a really subtle contradiction.
• 3:12
  Oh, interesting. Yeah, it noticed that September's
• 3:14
  raw sales actually increased dramatically. But
• 3:17
  then it cross -referenced that with the marketing
• 3:19
  spend and the customer service logs. And what
• 3:21
  did it find? It realized that higher refund rates
• 3:24
  and massive marketing costs completely wiped
• 3:26
  out the profitability of those new sales. It
• 3:28
  caught that automatically? Completely autonomously,
• 3:31
  without a human pointing it out. Wow. Then it
• 3:33
  outputs a neat little file called monthly business
• 3:36
  report and saves it right there on your machine.
• 3:38
  Yep, directly to your desktop. I have to admit
• 3:40
  something here. I still wrestle with prompt drift
• 3:43
  myself. But we all do. Sometimes I ask an AI
• 3:46
  to do math and halfway through it just starts
• 3:48
  hallucinating numbers to make the formatting
• 3:50
  look pretty. Getting it to follow strict rules
• 3:53
  requires absolute precision. That is the most
• 3:57
  crucial part of operating in task mode. You must
• 4:00
  define strict boundaries. Right. You have to
• 4:03
  explicitly tell it, never invent missing numbers.
• 4:06
  If a spreadsheet has a blank cell, the AI needs
• 4:09
  to know it should label that data as missing.
• 4:12
  It shouldn't try to guess. Exactly. It shouldn't
• 4:14
  extrapolate just to please you. In task mode,
• 4:17
  it is programmed to reference the exact source
• 4:20
  files behind every important finding. Does it
• 4:22
  actually edit or delete my original files when
• 4:25
  it does this? No, it reads them and creates a
• 4:28
  brand new document. Got it. Your original data
• 4:30
  stays perfectly safe. Completely untouched. It's
• 4:33
  a strict read -only process for the source files.
• 4:36
  The AI analyzes the inputs, generates the insights,
• 4:40
  and writes the output into a totally separate
• 4:41
  file. So you never risk corrupting your original
• 4:44
  data. Right. Zero risk there. OK. So my local
• 4:47
  files are safe in this little clean room sandbox.
• 4:50
  But honestly, who keeps files on their hard drive
• 4:53
  anymore? Basically nobody. My entire life is
• 4:56
  in the cloud. I use Google Drive, Dropbox, web
• 4:58
  apps. Does this AI just hit a brick wall if my
• 5:01
  data isn't sitting in a local folder? Not anymore.
• 5:04
  This is where we bring in Gemini Connectors.
• 5:06
  Connectors? Yeah. Connectors break the AI out
• 5:09
  of its isolated desktop sandbox safely. It links
• 5:12
  Gemini directly to the web apps you already use
• 5:14
  every day. So it integrates with things like
• 5:17
  Canva for graphic design, Dropbox for cloud storage,
• 5:20
  Wix for website management? Exactly. My favorite
• 5:23
  integration so far is with Gemini Notebook. Oh,
• 5:26
  that was formerly known as Notebook LM, right?
• 5:28
  Yeah, they rebranded it. Imagine you store all
• 5:31
  your research materials, articles, and passing
• 5:33
  thoughts in a digital notebook. With the connector,
• 5:37
  you can ask the desktop agent to pull that saved
• 5:39
  research. It can synthesize a comprehensive marketing
• 5:42
  strategy directly from those old notes. Oh, wow.
• 5:45
  And you never have to copy paste a single line
• 5:47
  of text into the chat box. It just queries the
• 5:50
  notebook directly. That's incredibly powerful,
• 5:52
  but this raises a massive security concern for
• 5:55
  me. I know exactly where you're going with this.
• 5:58
  Does connecting Dropbox mean Gemini is just indiscriminately
• 6:01
  scraping everything I've ever saved? I really
• 6:04
  don't want an AI constantly reading my personal
• 6:06
  tax returns just because they sit in the same
• 6:08
  cloud ecosystem as my work files. Right. And
• 6:10
  that's a very common fear. But the architecture
• 6:13
  actively prevents that. How does it stop it?
• 6:16
  The connector doesn't give the AI blanket access
• 6:18
  to your cloud. It uses a permission -based retrieval
• 6:22
  system. OK. Think of it like giving a hotel concierge
• 6:25
  a temporary key card. That card only opens one
• 6:28
  specific drawer in your room for exactly five
• 6:31
  minutes. Ah, I see. It only accesses specific
• 6:34
  files when your prompt explicitly requires them.
• 6:37
  It never ingests your entire cloud drive into
• 6:40
  its training memory. Do I have to migrate my
• 6:43
  Dropbox files into Gemini's ecosystem? Not at
• 6:45
  all. The connector just peeks into Dropbox when
• 6:48
  you ask it to. So it visits your apps instead
• 6:51
  of moving your data? Precisely. It acts as a
• 6:54
  temporary bridge. It reads what it needs, completes
• 6:57
  the specific task you assigned, and then burns
• 7:01
  the bridge. That is deeply reassuring. We'll
• 7:04
  be right back after this quick break. See you
• 7:06
  in a minute. Mid -roll sponsor read. And we are
• 7:10
  back. Let's jump right back in. OK, so we've
• 7:13
  established that Gemini can read local files
• 7:15
  and pull specific data from your cloud apps.
• 7:17
  Right. But typing out a massive 50 -word prompt
• 7:20
  with perfectly strict rules every single morning
• 7:23
  sounds exhausting. It is exhausting. If I have
• 7:27
  to tell it exactly how to read my Dropbox every
• 7:29
  time, that defeats the whole purpose of efficiency.
• 7:31
  We need a way to save the instructions permanently.
• 7:34
  Yeah, nobody wants to write a novel just to get
• 7:36
  their daily tasks done. Exactly. That's why Google
• 7:39
  introduced Gemini skills. Think of skills as
• 7:42
  saved reusable workflows. OK. You define the
• 7:46
  complex logic just once. then you trigger that
• 7:49
  entire sequence with a simple keyboard shortcut.
• 7:51
  Let's walk through a practical scenario, setting
• 7:53
  up an email inbox manager skill. Oh, this is
• 7:56
  a brilliant use case. You essentially write a
• 7:59
  rule book for how the AI should treat your messy
• 8:01
  inbox. Yeah. You instruct it to classify every
• 8:04
  incoming message. You define the categories yourself.
• 8:07
  Like what? Well, urgent means it requires your
• 8:10
  immediate attention today. Important means it
• 8:13
  needs a response eventually. Right. Newsletter
• 8:16
  is just informational reading, and low priority
• 8:19
  requires no action at all. And crucially, it
• 8:21
  doesn't just sort them into folders silently.
• 8:23
  No, it tells you why an email needs a reply.
• 8:26
  That's the part that blows my mind. It analyzes
• 8:29
  the context of the conversation. It explains
• 8:32
  the required action. But we really must highlight
• 8:35
  the strict boundaries here. Yes, the boundaries
• 8:38
  are vital. When setting up this skill, you have
• 8:40
  to explicitly tell it to never delete emails.
• 8:43
  Never. And you tell it to never send a reply
• 8:45
  without your final approval. So you're building
• 8:47
  a digital safety net. Exactly. Mechanically,
• 8:51
  you are granting the AI read access to your inbox
• 8:54
  API, but you are restricting its send access.
• 8:57
  That makes perfect sense. You define the workflow,
• 9:00
  but you also define the hard constraints. To
• 9:03
  trigger all of this, you just type the at symbol,
• 9:06
  then email inbox manager. And it instantly runs
• 9:09
  the rules and generates your daily summary. It's
• 9:12
  exactly like training a new human assistant.
• 9:15
  That's a perfect analogy. You spend hours writing
• 9:17
  the operational manual once. From then on, you
• 9:20
  just look at them and say, do the inbox routine,
• 9:22
  and they know exactly what that entails. Right.
• 9:25
  You invest time upfront to build the operational
• 9:27
  manual. The AI memorizes that manual as a skill.
• 9:31
  It applies your specific logic flawlessly without
• 9:34
  you ever having to repeat the underlying rules.
• 9:36
  Could a glitch cause the skill to accidentally
• 9:39
  email my boss? No. The strict rules you set completely
• 9:43
  block unapproved outgoing replies. Excellent.
• 9:46
  It organizes the drafts, but you pull the trigger.
• 9:50
  Exactly. You remain the final authority. The
• 9:53
  AI acts as a highly sophisticated filter. And
• 9:56
  prepares the work. Right. It removes the daily
• 9:58
  friction. But the final execution, the actual
• 10:01
  sending of the message, remains safely in your
• 10:04
  human hands. Typing at symbol email inbox manager
• 10:07
  is definitely fast. Very fast. But what if you
• 10:10
  don't want to type anything at all? What if I
• 10:13
  want to wake up? and just have the work already
• 10:15
  done. Let's talk about taking the human completely
• 10:18
  out of the trigger loop. This is the natural
• 10:20
  evolution of the technology. First, we figured
• 10:22
  out how to save the instructions. Now we automate
• 10:25
  the execution. Gemini Automation allows you to
• 10:28
  schedule these recurring tasks and skills. The
• 10:31
  AI acts without you ever prompting it. A great
• 10:33
  example of this is setting up a daily AI news
• 10:36
  brief. You schedule it to run entirely in the
• 10:39
  background every weekday. It triggers exactly
• 10:42
  at 8 a .m. your local time. Yeah, you set the
• 10:44
  parameters once. You tell it to research AI industry
• 10:47
  news from the last 24 hours. Okay. You instruct
• 10:49
  it to pick the top five most impactful stories.
• 10:52
  It writes a catchy headline for each. It writes
• 10:55
  a simple jargon -free summary. And it explains
• 10:58
  why the news actually matters to your specific
• 11:00
  industry. Exactly. And it provides a verifiable
• 11:03
  source link for everything. And you can even
• 11:05
  tell it not to repeat old stories. It actively
• 11:08
  checks its previous outputs from yesterday before
• 11:10
  drafting today's brief. It does all of this autonomously.
• 11:13
  While you are sleeping, it is working. Well,
• 11:17
  imagine scaling to a billion queries. Just picture
• 11:21
  the sheer computing power required. It's massive.
• 11:24
  Millions of users having distinct personalized
• 11:27
  agents running heavy background tasks simultaneously
• 11:31
  across the globe. The computational logistics
• 11:33
  of that are absolutely staggering. It requires
• 11:36
  massive data center coordination. You're managing
• 11:39
  queues and server loads in real time. Which brings
• 11:42
  up a good point. Yeah, that's actually why there's
• 11:44
  a specific caveat to this feature. These scheduled
• 11:47
  tasks run at approximate times. Right. You can't
• 11:52
  expect to the millisecond precision. Will this
• 11:54
  report definitely hit my screen exactly at 8
• 11:57
  a .m.? It runs approximately then, depending
• 12:00
  on server load and your usage limits. Right,
• 12:03
  so think of 8 a .m. as a rough target time. Yeah,
• 12:06
  consider it a window. It might arrive at 8 .02
• 12:08
  or 8 .05. The global system queues millions of
• 12:11
  these automated requests and it processes them
• 12:14
  as compute power becomes available in the data
• 12:16
  centers. To power these scheduled tasks and complex
• 12:19
  workflows smoothly, you actually have to dictate
• 12:22
  how much brain power the AI should use. Right.
• 12:25
  Because not all tasks are created equal and computing
• 12:28
  power isn't infinite. This is a critical concept
• 12:31
  in modern AI usage. Compute is a finite, expensive
• 12:35
  resource. It costs real electricity. Yeah, it
• 12:38
  really does. You don't need massive, deep -thinking
• 12:42
  reasoning power for every tiny administrative
• 12:45
  task. So the desktop app lets you toggle between
• 12:48
  two distinct models, Argon and Flash. Exactly.
• 12:52
  We should clarify what that means for anyone
• 12:54
  unfamiliar. Let's define model. Good idea. The
• 12:58
  underlying brain engine that powers how the AI
• 13:00
  thinks. It's the best way to frame it. So Argon
• 13:03
  is the heavyweight engine. Okay. It's built for
• 13:06
  highly complex tasks. It handles deep document
• 13:09
  analysis. It handles complex coding projects.
• 13:13
  It possesses advanced reasoning capabilities.
• 13:15
  Right. But because it is so thorough, it is noticeably
• 13:19
  slower. It also consumes your monthly usage limits
• 13:22
  much faster. Then you have the Flash model. Yes.
• 13:25
  Flash is for everyday mundane tasks. It handles
• 13:28
  simple code formatting. It handles quick summaries
• 13:30
  of plain text. And it is significantly faster
• 13:33
  to respond. Almost instantaneous, really. Yeah,
• 13:36
  it's highly optimized for speed and low compute
• 13:38
  cost. If you're just brainstorming blog post
• 13:41
  titles or drafting a quick calendar invite, Flash
• 13:43
  is more than capable. You don't need the genius
• 13:46
  engine for that. Exactly. I think of it this
• 13:48
  way. Argon is a heavy duty freight train. I like
• 13:52
  that. Flash is a zippy electric scooter. Both
• 13:55
  are vehicles, but they serve different purposes.
• 13:57
  A freight train holds massive capacity. It can
• 14:00
  carry 10 ,000 pages of text at once. But it takes
• 14:03
  a lot of energy and time to get moving. A scooter
• 14:06
  has a tiny basket. It can only carry a single
• 14:09
  email. But it gets to the destination in one
• 14:11
  second. You don't use a freight train to run
• 14:14
  to the grocery store. Right. And you don't use
• 14:16
  a scooter to move a mountain of financial data.
• 14:18
  Exactly. Choosing the right tool for the job
• 14:20
  prevents you from burning through your premium
• 14:22
  limits unnecessarily. If I just need a quick
• 14:24
  summary of a newsletter, which one? Flash is
• 14:27
  your best bet. It's fast and saves your usage
• 14:29
  limits. Makes sense. Save the heavy lifting for
• 14:32
  the argon model. Exactly. You manage your compute
• 14:35
  budget wisely. You reserve argon's advanced reasoning
• 14:38
  for tasks that genuinely require deep methodical
• 14:41
  analysis. Let's briefly synthesize this entire
• 14:44
  journey. Google is transforming how we interact
• 14:47
  with our computers. They really are. Gemini is
• 14:49
  no longer just a chatbot oracle that answers
• 14:52
  trivia questions. It is becoming an active embedded
• 14:55
  co -worker. Right. It handles local files securely
• 14:59
  in a sandbox. It connects directly to your cloud
• 15:02
  apps using temporary permissions. It learns your
• 15:05
  specific nuanced routines via skills. And it
• 15:09
  runs those routines autonomously on schedule
• 15:11
  via automation. It's a genuine leap toward the
• 15:14
  era of the true desktop assistant. It changes
• 15:17
  our relationship with software entirely. We are
• 15:20
  moving from being machine operators to being
• 15:22
  machine managers. That's a great way to put it.
• 15:25
  The computer does the executing. We do the directing.
• 15:28
  I highly urge you to try handing off just one
• 15:31
  simple task this week. Just one. Pick something
• 15:34
  tedious you normally do manually every morning.
• 15:37
  Set up a basic skill for it. See how it feels
• 15:39
  to manage an agent rather than doing the granular
• 15:42
  work yourself. But all of this leaves me wondering
• 15:46
  about the long -term reality here. Oh, I'm curious.
• 15:49
  If everyone starts automating their emails, their
• 15:51
  scheduling, their negotiations with these desktop
• 15:54
  agents, How long until your automated assistant
• 15:57
  is just negotiating with someone else's automated
• 16:00
  assistant? And what happens to us in the middle?