Latest / AI Fire Daily / #118 Robin: OpenAI Dots Turns ChatGPT Into a Proactive AI Chief of Staff

Transcript

#118 Robin: OpenAI Dots Turns ChatGPT Into a Proactive AI Chief of Staff · AI Fire Daily

• 0:00
  Imagine waking up to a perfectly quiet morning.
• 0:05
  You grab your phone before you've even had your
• 0:06
  first coffee. But instead of an endless list
• 0:09
  of unread emails, an AI assistant has already
• 0:13
  messaged you. Yeah, which sounds kind of nice.
• 0:15
  It caught a $10 ,000 billing mistake. It fixed
• 0:19
  those annoying bugs in your software code. And
• 0:21
  it even made a meme about it. That is just wild
• 0:25
  to think about. Welcome to the deep dive. Today
• 0:28
  we are exploring something completely fascinating.
• 0:31
  We're looking at OpenAI's new proactive assistant,
• 0:34
  DOTS. It really shifts the landscape entirely
• 0:37
  for us. It changes AI from a simple reactive
• 0:39
  tool into something else. It effectively becomes
• 0:42
  your proactive digital chief of staff. We'll
• 0:45
  unpack some truly wild real world tests together
• 0:48
  today. We'll examine complex video production
• 0:50
  pipelines running entirely in parallel. We'll
• 0:53
  also see how it finds lost business revenue automatically.
• 0:56
  And we are going to examine the messy reality
• 0:59
  underneath, too, because we have to talk about
• 1:01
  frustrating software beta bugs. We also need
• 1:04
  to look at a very crowded chat GPT interface.
• 1:07
  But before we look at what it actually accomplished,
• 1:09
  we really need to understand how DOTS operates
• 1:11
  differently, because it works completely differently
• 1:14
  from standard chat GPT entirely. Right. DOTS
• 1:17
  isn't just another simple standard chat bot interface.
• 1:21
  You don't just open it for quick, reactive help.
• 1:23
  It's an orchestrator. that actively manages your
• 1:26
  other tools. Right, and hopefully one that doesn't
• 1:29
  schedule too many meetings. Exactly. Think about
• 1:33
  how you use AI today normally. You probably open
• 1:35
  a brand new chat window. You type out a long,
• 1:39
  highly detailed prompt again. You have to explain
• 1:42
  your entire business context from scratch. You
• 1:44
  do this every single time you start a project.
• 1:48
  Dots changes that frustrating dynamic completely
• 1:50
  for you. It essentially sits above the individual
• 1:53
  tasks as a management layer. The biggest advantage
• 1:56
  here is definitely the continuity. It carries
• 1:58
  over your existing chat GBT context seamlessly.
• 2:02
  Yeah, it remembers your specific memories and
• 2:04
  connected apps perfectly. You never have to start
• 2:07
  from absolute zero again. It feels much more
• 2:10
  like an ongoing professional relationship. You
• 2:13
  give Dots a clear overarching goal just once.
• 2:16
  And it figures out which specific plot should
• 2:18
  handle things. And you can reach it in a few
• 2:20
  different ways. You can make a phone -like voice
• 2:23
  call very easily. The interface actually looks
• 2:25
  exactly like a ringing phone. It does. You can
• 2:27
  also add it directly to Slack DMs. That's really
• 2:30
  amazing for quick team research tasks. You can
• 2:33
  research complex things right beside your coworkers
• 2:35
  without them necessarily knowing a robot did
• 2:38
  the heavy lifting. There's even a simple email
• 2:40
  option included right now. It fundamentally follows
• 2:43
  you wherever you're actually working today. It
• 2:46
  doesn't just live trapped inside a browser tab.
• 2:48
  Let's talk about the actual brains behind the
• 2:50
  system. It uses a truly fascinating two -computer
• 2:53
  setup conceptually. Yeah, this is where it gets
• 2:56
  really interesting, technically. The cloud computer
• 2:59
  actually stays on 24 -7. It actively browses
• 3:03
  websites while you're simply walking away. It
• 3:06
  keeps complex background work running while you
• 3:08
  sleep. Meanwhile, the local computer interacts
• 3:11
  directly with your machine. It runs those specific
• 3:14
  workflows on your actual desktop computer. It's...
• 3:18
  A massive shift in overall AI capability. You
• 3:21
  can watch exactly what happens on the cloud computer.
• 3:23
  You can seamlessly take control of it when necessary.
• 3:26
  This matters because Dots does much more than
• 3:28
  generate text. It can actually browse complex
• 3:31
  software dashboards completely independently.
• 3:33
  It collects vital information from various scattered
• 3:36
  web pages. Then it passes that information into
• 3:39
  other chat GPT workflows seamlessly. Dotsa also
• 3:42
  works closely with a specialized system called
• 3:44
  Pages. Pages is a Notion -style system to keep
• 3:47
  project notes from getting lost. Because, let's
• 3:50
  face it, chat threads get way too long. Important
• 3:52
  details simply vanish inside hundreds of chat
• 3:55
  messages. Pages gives your valuable information
• 3:58
  a permanent, safe home. You can create subpages
• 4:01
  and seamlessly save important notes. You also
• 4:04
  get powerful scheduled task features right now.
• 4:06
  These tasks can be triggered by time or specific
• 4:09
  events. Imagine summarizing along keynote the
• 4:12
  exact moment it ends. Right. It handles that
• 4:14
  organizational friction completely in the background.
• 4:17
  We should probably compare this to other AI tools
• 4:19
  briefly. You might have heard of a consumer tool
• 4:21
  called Muse. Muse is built mostly for everyday
• 4:24
  consumer errands. Right. It can handle calling
• 4:26
  local businesses or booking travel. Then there's
• 4:30
  Grokbot, built specifically for multi -agent
• 4:33
  bot teams. Grokbot lets several specialized bots
• 4:36
  work together on complex tasks. But DOTS has
• 4:40
  a very different kind of superpower entirely.
• 4:43
  Its true strength is native integration inside
• 4:46
  the existing ecosystem. It leverages the rich
• 4:49
  context you've already built up over time. I'll
• 4:54
  be honest here. I'll be honest. I still wrestle
• 4:56
  with prompt drift myself quite a bit. Oh, absolutely.
• 4:59
  It happens to all of us. You get into these long
• 5:02
  single thread chats, and over time, It just gets
• 5:05
  incredibly confusing. You lose track of the original
• 5:08
  project goal entirely. Using standard chat GPT
• 5:11
  feels like being a highly stressed micromanager.
• 5:15
  DOTS is much more like hiring a dedicated project
• 5:17
  manager. Does this mean we stop talking to standard
• 5:20
  chat GPT entirely? Well, not exactly, but DOTS
• 5:24
  becomes a high -level control layer. It sits
• 5:26
  securely above all your other specific projects.
• 5:30
  You just give your broad instructions to the
• 5:32
  manager directly. You manage DOTS, and it dynamically
• 5:35
  manages the actual work. So DOTS just becomes
• 5:38
  the only front door you need. Now that we understand
• 5:42
  the management structure clearly, let's look
• 5:44
  at what happens when it executes real work. Let's
• 5:47
  definitely walk through the concrete real -world
• 5:49
  examples today. We can see DOTS acting as a true
• 5:52
  orchestrator here. We can deeply analyze where
• 5:54
  it shines and hits walls. Because the mechanism
• 5:57
  behind its success is genuinely fascinating.
• 6:00
  The complex video production pipeline example
• 6:02
  is completely mind blowing. A user basically
• 6:05
  wanted to produce a brand new video. Docs receives
• 6:08
  the request and immediately delegates the tasks.
• 6:11
  It sends the initial writing directly to a tool
• 6:13
  called Work. Work is a dedicated workspace for
• 6:16
  long form writing and research. It naturally
• 6:19
  generates three possible intros for the upcoming
• 6:22
  video. It evaluates and picks the very best introductory
• 6:25
  script option. then it seamlessly hands that
• 6:29
  script over to Codex. And just to clarify, Codex
• 6:32
  is an AI tool that writes and tests computer
• 6:34
  code. Exactly. Codex then creates the necessary
• 6:38
  motion graphics for you automatically. Meanwhile,
• 6:41
  Dots also starts a separate thread entirely for
• 6:44
  thumbnails. It actively manages all these parallel
• 6:47
  tasks at once. It patiently waits for one specific
• 6:50
  step to finish first. And here is the truly magical
• 6:53
  part of the process. What's that? It writes those
• 6:56
  detailed handoff instructions completely automatically.
• 6:59
  That's incredibly impressive coordination for
• 7:01
  any AI system, because honestly, most humans
• 7:04
  would shorten those complicated handoff prompts.
• 7:07
  Writing a massive detailed handoff every single
• 7:09
  time is annoying. We get lazy and leave out crucial
• 7:12
  contextual details. DOTS actually spends the
• 7:15
  extra computational effort explaining tasks properly.
• 7:18
  It ensures the next AI tool has exactly what
• 7:20
  it needs. It also handled a highly complex fitness
• 7:23
  app launch recently. Dots actively used an existing
• 7:26
  fitness app project thread. It dug deeply in
• 7:29
  and analyzed the existing code thoroughly. It
• 7:32
  systematically found and fixed annoying bugs
• 7:34
  in the underlying code. Then it created six polished
• 7:37
  App Store screenshots completely automatically.
• 7:40
  It intelligently used real screens from the actual
• 7:43
  working app. The seamless handoff is the really
• 7:46
  important part here again. Dots already knew
• 7:48
  exactly which thread contained the app context.
• 7:51
  It didn't have to start from absolute zero at
• 7:53
  all. But the biggest win was auditing a complex
• 7:56
  revenue workflow. It actively audited a massive
• 8:00
  sponsorship pipeline for someone. It systematically
• 8:03
  searched through a huge amount of unstructured
• 8:05
  data. Think about what that kind of task actually
• 8:08
  requires mechanically. It has to fundamentally
• 8:11
  understand the underlying concept of a sponsorship.
• 8:14
  It has to know exactly where the raw invoice
• 8:16
  data lives. It cross -referenced delivered ad
• 8:19
  reads against paid Google billing invoices. And
• 8:23
  it did all of this without explicit step -by
• 8:25
  -step human instructions. Ultimately, it found
• 8:28
  a $10 ,000 Google billing discrepancy. It also
• 8:32
  successfully found an un -invoiced Riverside
• 8:34
  ad campaign. Whoa. Imagine scaling that audit
• 8:38
  across a whole year of business data. That is
• 8:41
  fundamentally much stronger than summarizing
• 8:44
  a simple document. It actively spots missed work
• 8:46
  without being asked directly. It also truly excels
• 8:50
  at heavy research and weird tasks. It meticulously
• 8:54
  looked at YouTube comments for a specific creator.
• 8:57
  It actively summarized 1800 separate YouTube
• 9:00
  comments very quickly. It successfully extracted
• 9:03
  five distinct channel growth themes in 30 minutes.
• 9:06
  It systematically saves a huge amount of boring
• 9:09
  manual work. But it also handles weird, highly
• 9:13
  specific physical product searches beautifully.
• 9:16
  Someone asked it to source walnut magnetic stacking
• 9:19
  stones via voice. Yeah, that was a brilliant
• 9:22
  little workaround it used. Exactly, it actually
• 9:24
  used Codex's in -app browser as a clever workaround.
• 9:28
  It intentionally bypassed standard search limits
• 9:30
  to find those specific stones. It also researched
• 9:34
  some highly complex online learning courses recently.
• 9:37
  It carefully compared similar courses while operating
• 9:39
  inside a Slack DM. It clearly suggested three
• 9:43
  specific landing page improvements in six minutes.
• 9:46
  It also deeply reviewed previous Slack conversations
• 9:49
  for a private community. It confidently gave
• 9:51
  that community an A - grade overall. It gracefully
• 9:55
  suggested three distinct structural improvements
• 9:58
  for that group, but it's a useful reminder about
• 10:00
  AI research agents generally. They can still
• 10:03
  produce slightly different answers from identical
• 10:05
  requests. Right, and it was surprisingly good
• 10:08
  at team creative work. It actively made funny
• 10:11
  memes based on people inside a Slack community.
• 10:14
  The highly personal memes worked incredibly well
• 10:16
  for the team because Dots actually had deep context
• 10:19
  about the people and their personalities. It
• 10:22
  fundamentally understood their previous conversations
• 10:24
  and inside jokes perfectly. But when the specific
• 10:28
  task moved toward general AI news... The creative
• 10:31
  results became much weaker and significantly
• 10:33
  less funny. More personal context definitely
• 10:36
  produces much better creative output overall.
• 10:39
  OpenAI employees reportedly have an internal
• 10:41
  Slack channel already running. Employee dots
• 10:44
  actually share and playfully reply to memes there
• 10:47
  constantly. It beautifully shows how agents become
• 10:50
  active participating team members, but we absolutely
• 10:54
  need to examine its actual creative limits. A
• 10:57
  user actively suggested creating a new paid digital
• 10:59
  product. Dots tried building a paid product called
• 11:03
  PM Workbench. It was basically a workflow kit
• 11:06
  for product managers. It completely nailed the
• 11:08
  overall landing page and the visuals. But the
• 11:10
  actual marketing copy was incredibly shallow
• 11:13
  overall. It frankly sounded like a tired marketer
• 11:17
  on a Friday afternoon. It desperately needed
• 11:20
  substantial human editing before anyone could
• 11:22
  sell it. I actually have to push back on that
• 11:25
  output slightly. How do you mean? Because I think
• 11:28
  we misunderstand what AI should be doing here.
• 11:31
  The AI can easily build the fundamental product
• 11:34
  structure perfectly. It can quickly create the
• 11:37
  basic conceptual framework and examples. Right.
• 11:39
  But true subject matter expertise still fundamentally
• 11:42
  requires a human mind. You simply cannot just
• 11:45
  outsource actual hard -earned wisdom to code.
• 11:49
  Getting to a genuinely strong product takes deep
• 11:52
  human feedback. The AI provides the scaffolding,
• 11:55
  but You provide the soul. With so many parallel
• 11:58
  tasks running, how much manual checking does
• 12:00
  the user still have to do? You still absolutely
• 12:03
  have to review the final output carefully. DOTS
• 12:06
  seamlessly coordinates things, but you must strictly
• 12:08
  verify the details. This is especially true for
• 12:11
  highly nuanced or deeply creative work. So trust
• 12:15
  it to build, but verify every single word. Trusting
• 12:19
  it is genuinely great when things go smoothly.
• 12:22
  But relying on an invisible manager comes with...
• 12:25
  serious blind spots. Because when things break
• 12:28
  down in the background... They break down silently.
• 12:31
  We'll dive into exactly what that means right
• 12:33
  after this. We are back. So before the break
• 12:37
  we mentioned those dangerous blind spots. Let's
• 12:40
  talk deeply about the messy reality right now.
• 12:43
  We really have to discuss the failed task visibility
• 12:47
  problem. Because this is where the orchestration
• 12:50
  model gets slightly terrifying. There was a specific
• 12:54
  Substack newsletter draft that entirely failed.
• 12:58
  DOTS was actively trying to turn a recorded interview
• 13:00
  into a draft. DOTS unexpectedly hit a hard blocker
• 13:04
  with the required tool. The underlying tool couldn't
• 13:07
  safely access Substack properly back then. But
• 13:10
  DOTS didn't gracefully surface the failure clearly
• 13:13
  at all. The silent failure was deeply buried
• 13:16
  inside the complex workflow. With several background
• 13:20
  jobs running, it was incredibly easy to miss.
• 13:24
  If an AI is actively working while you're away,
• 13:28
  silent failures are incredibly dangerous for
• 13:30
  your ongoing professional projects. You might
• 13:32
  confidently assume a critical task is fully finished,
• 13:35
  but it actually quietly stopped running several
• 13:38
  hours ago. We also desperately need to talk about
• 13:41
  the ongoing beta bugs. Proactive assistants fundamentally
• 13:44
  need to be completely reliable always. You cannot
• 13:47
  trust a chief of staff who randomly falls asleep.
• 13:51
  The persistent cloud computer literally crashed
• 13:54
  after a system reset. The vital scheduled task
• 13:57
  feature simply wouldn't open properly either.
• 14:00
  These specific features are completely central
• 14:02
  to how DOTS actually works. A proactive assistant
• 14:05
  is completely useless if it simply crashes. Hopefully
• 14:10
  these specific beta bugs will improve very quickly.
• 14:13
  Because the fundamental promise relies entirely
• 14:16
  on absolute unquestioned stability. And the main
• 14:19
  conversation thread gets incredibly messy, extremely
• 14:22
  fast. You basically have one main conversation
• 14:25
  for absolutely everything, initially. Once five
• 14:28
  distinct workflows are actively running, it becomes
• 14:31
  pure chaos. You might simultaneously have research,
• 14:35
  complex coding, and financial audits running.
• 14:37
  Finding a specific tiny update becomes much harder
• 14:40
  then. The single main dot conversation basically
• 14:42
  becomes impossible to read. Pages certainly helps,
• 14:45
  but you need strict personal organizational discipline.
• 14:48
  You have to actively use separate project threads
• 14:51
  consistently yourself. The main ongoing conversation
• 14:54
  should just be a high -level command center.
• 14:57
  There are also some genuinely major missing features
• 14:59
  currently. You're strictly limited to one single
• 15:02
  dot per user right now. Multiple specialized
• 15:05
  dots would make the whole system easier eventually.
• 15:08
  You could effectively have separate dedicated
• 15:10
  dots for sales or content. Right now there's
• 15:13
  no personal email address formally assigned to
• 15:15
  it yet. It also completely lacks any form of
• 15:18
  integrated payment support features. It can orchestrate
• 15:21
  work, but it cannot actually spend your money.
• 15:24
  We're honestly facing a real chat GPT UI crisis
• 15:27
  today. The underlying interface currently has
• 15:30
  standard chats, work, and codecs. It actively
• 15:33
  has pages, dynamic sites, images, and scheduled
• 15:37
  tasks. It's becoming incredibly difficult to
• 15:39
  actually follow closely anymore. The current
• 15:42
  software interface is exactly like a Swiss Army
• 15:45
  knife. Ah. How so? But it's a knife that has
• 15:49
  40 blades folded out at once. It's incredibly
• 15:51
  powerful, but it's dangerous and highly confusing
• 15:54
  to hold. Each individual piece of the chat GPT
• 15:58
  ecosystem can be highly useful, but together
• 16:01
  the entire interface is becoming incredibly overwhelming,
• 16:04
  constantly. It naturally creates a very strange,
• 16:08
  complicated situation for the user. Dots effectively
• 16:10
  adds another heavy layer to the chat GPT software.
• 16:14
  But... It may also mysteriously become the necessary
• 16:18
  saving grace because it actively helps you avoid
• 16:21
  dealing with all the other layers. You could
• 16:24
  eventually spend most of your time just talking
• 16:26
  normally. You just smoothly say, check my current
• 16:29
  projects and keep things moving. Then Dots intelligently
• 16:33
  decides which specific thread or skill handles
• 16:36
  it. If the chat GPT interface is getting too
• 16:39
  cluttered, is Dots the cure or the disease? DOTS
• 16:43
  is actually the necessary cure for the overwhelming
• 16:45
  clutter. It essentially acts as a single unifying
• 16:49
  voice layer above it all. It seamlessly hides
• 16:53
  the complex grinding machinery underneath the
• 16:56
  surface. So DOTS hides the clutter by running
• 16:58
  it all for you. Let's carefully pull all these
• 17:02
  complex threads together right now. Because we
• 17:05
  really need to figure out exactly how to start.
• 17:08
  We absolutely want to distill the primary takeaway
• 17:10
  here today. DOTS isn't just another fancy reactive
• 17:14
  answering machine tool. It's a remarkably powerful
• 17:17
  proactive control layer for your work. The absolute
• 17:20
  best way to use it is as an orchestrator. It
• 17:24
  should seamlessly manage your existing multi
• 17:26
  -tool projects completely. You simply give it
• 17:28
  the broad goal and it decides the rest. And there's
• 17:32
  one incredibly actionable piece of advice here
• 17:34
  today. There's a single highly effective best
• 17:37
  prompt you can use right away. You just tell
• 17:40
  it Let's try to do some ambitious things. What
• 17:44
  else can you do? Give me your top five. It actively
• 17:47
  scans your existing conversational context for
• 17:50
  real clues. It will intelligently suggest ongoing
• 17:53
  projects connected to your actual work. For example,
• 17:56
  it brilliantly suggested rehearsing an app launch
• 17:58
  and testing. It knows exactly what you are building,
• 18:01
  so it knows how to help. We genuinely want you
• 18:04
  to examine your daily workflow closely. Find
• 18:06
  that one messy, incredibly complicated multi
• 18:09
  -tool project you dread. The specific project
• 18:12
  requiring deep writing, complex coding, and tedious
• 18:15
  organizing. That is the exact painful job you
• 18:19
  should completely hand off. Hand it directly
• 18:21
  to a tireless, highly efficient digital chief
• 18:24
  of staff. I want to gracefully leave you with
• 18:27
  a deeply provocative thought. Right now, DOTS
• 18:30
  definitely cannot send external emails or pay
• 18:33
  for things. It is securely locked inside the
• 18:35
  chat GPT ecosystem very safely. But eventually,
• 18:39
  once it actually gets an email address and a
• 18:41
  credit card, at what specific point does it strictly
• 18:45
  stop being just an assistant? When does it truly
• 18:48
  start becoming your autonomous proxy in the real
• 18:51
  world? Thank you so much for joining us on the
• 18:54
  Deep Dive. Take care of yourselves out there.