Latest / Today’s AI News / AI Voice Agents Can Finally Reason, Fitbit Becomes a Google AI Play, Anthropic’s Intelligence Bomb

Transcript

AI Voice Agents Can Finally Reason, Fitbit Becomes a Google AI Play, Anthropic’s Intelligence Bomb · Today’s AI News

• 0:00
  For years, talking to AI has felt like using a laggy
• 0:05
  walkie-talkie with a highly distracted intern.
• 0:08
  But today, the new developments in voice AI feel like handing an
• 0:11
  earpiece to a fluent multitasking expert who actually
• 0:15
  listens and thinks while they speak.
• 0:17
  Yeah, the difference is just night and day now.
• 0:19
  Welcome back to today's AI News, brought to you by Nynex
• 0:22
  Productions. We are skipping the typing today
• 0:24
  and diving head first into an era where AI, you know, builds
• 0:27
  itself, manages our health through screenless trackers, and
• 0:30
  seamlessly speaks our language. OK, let's unpack this.
• 0:33
  I mean, if you look at the stack of sources we're going through
• 0:35
  for this deep dive, the overarching theme points to one
• 0:39
  really unavoidable conclusion. The interface as we know it is
• 0:42
  kind of justice vanishing. Right, like the keyboard is on
• 0:45
  its way out. Exactly.
• 0:46
  We are moving from this world where we command machines
• 0:49
  through keyboards and, you know, glowing rectangles, to a world
• 0:53
  where they operate right alongside us in real time.
• 0:56
  And Austin? They're completely invisible.
• 0:58
  Which makes sense, right? Because for AI to actually be
• 1:00
  useful in the real world, the interface simply has to escape
• 1:03
  the keyboard. And that actually brings us to
• 1:06
  the latest developments from Open AI.
• 1:08
  Oh yeah, the new models. Yeah, they just dropped a trio
• 1:11
  of new real time voice models, so we're looking at GPT Real
• 1:15
  Time 2, GPT Real Time Translate, and GPT Real Time Whisper.
• 1:19
  And this is huge because up until now, interacting with
• 1:23
  voice AI had this inherent, unavoidable friction.
• 1:26
  It was just how the underlying architecture was built.
• 1:29
  It was so clunky. It really was.
• 1:30
  It was this sequential pipeline. You would speak, and the system
• 1:34
  had to like, process your audio waves into text first.
• 1:37
  Then it fed that text into a language model to generate a
• 1:40
  text response, right? And then finally it sent that
• 1:42
  text to a synthesizer to turn it back into that sort of robotic
• 1:45
  audio. Every single step just added
• 1:48
  latency. Plus, turning your voice into
• 1:50
  text stripped away all the context you know, like your
• 1:52
  tone, your sarcasm, your pauses. Yeah, all the human elements
• 1:56
  were just gone. It's basically the difference
• 1:58
  between a terrible zoom call where everyone accidentally
• 2:00
  talks over each other because of the lag.
• 2:02
  The worst? Right, versus a fluid in person
• 2:06
  conversation where you can seamlessly interrupt without the
• 2:09
  other person completely losing their train of thought.
• 2:11
  If we connect this to the bigger picture, Real Time 2 bypasses
• 2:16
  that old pipeline completely. It brings GPT 5 level reasoning
• 2:21
  directly to the live audio stream.
• 2:23
  So it's not converting it to text at all.
• 2:25
  Nope. The model is processing the raw
• 2:28
  audio waves natively. It literally talks while it
• 2:31
  thinks, and it can use multiple digital tools simultaneously
• 2:34
  while doing so. That is wild.
• 2:36
  It gets better. There's this metric called Big
• 2:37
  Bench Audio. It basically tests an AI's
• 2:40
  ability to handle human interruptions, complex audio
• 2:43
  reasoning, and tonal accuracy when it's under pressure.
• 2:46
  OK, so how did it do? Well, the predecessor model
• 2:48
  scored an 81.4% real time too just hit 96.6%.
• 2:53
  Really a 15 point jump in real time reasoning that is massive.
• 2:57
  It's a game changer. It means the system understands
• 3:00
  nuance and like context shifts. Almost.
• 3:04
  Florida State. Honestly, the AI industry's
• 3:07
  current fixation on text based agents is actually a bit short
• 3:11
  sighted when you look at these numbers.
• 3:12
  Because text is just too slow. Yeah, text is a low bandwidth
• 3:15
  bottleneck for human expression. The next wave of AI will be
• 3:19
  spoken to, not typed at voice. When it actually works with this
• 3:23
  level of reasoning and tonal realism just removes the barrier
• 3:27
  between human intent and digital execution.
• 3:30
  And we are already seeing the real world applications rolling
• 3:33
  out today like Zillow is using this for real estate AI agents,
• 3:37
  Priceline is using it for voice managed travel.
• 3:39
  And Deutsche Telekom has it for customer support?
• 3:42
  Right. And add to that the live
• 3:43
  translator they shipped alongside it that covers over 70
• 3:45
  languages in real time. So you have a full case agent
• 3:48
  toolkit now? Just think about what that
• 3:50
  actually looks like in practice. I mean, you could theoretically
• 3:52
  be on a phone call with a real estate agent in Berlin speaking
• 3:55
  in English while the AI translates back and forth
• 3:58
  seamlessly. And it would pick up on the
• 4:00
  hesitation in your voice exactly while pulling up property
• 4:03
  records in the background. It's crazy.
• 4:04
  But, you know, while Open AI is removing the keyboard to reduce
• 4:08
  friction, Google is sort of taking that philosophy a step
• 4:11
  further. Yeah, they seem to believe the
• 4:13
  ultimate low friction interface is no interface at all.
• 4:17
  Right, which is why they are actively killing the screen.
• 4:19
  Google has officially opened its Gemini powered AI health coach
• 4:23
  to the public after all those months in beta.
• 4:26
  They are integrating AI directly into our physical health.
• 4:30
  They're. Actively consolidating the
• 4:32
  Fitbit app Health Connect. Apple Health, various wearable
• 4:36
  data streams, and even US medical records into a single
• 4:39
  Google Health hub. But the wild part of this
• 4:42
  rollout is the actual hardware. Have you seen?
• 4:44
  This the new Fitbit. A $99 Fitbit Air.
• 4:47
  It weighs just 12 grams, carries heart rate, oxygen and
• 4:50
  temperature sensors and it has 0 screen.
• 4:53
  None at all. None.
• 4:54
  No clock, no step counter to look at.
• 4:56
  You just wear it and it transmits body data to the AI
• 4:59
  coach, and then that coach tailors weekly workouts,
• 5:02
  interprets your medical records, and it can even identify what
• 5:04
  you ate just from a photo on your phone.
• 5:07
  And the system is designed to eventually open up to non Google
• 5:10
  hardware too, like the Apple Watch, Garmin and Aura ranks.
• 5:14
  OK, wait, let me push back on this design choice for a second.
• 5:17
  Sure. Are consumers actually ready for
• 5:20
  a completely screenless tracker? I mean, think about how you
• 5:23
  operate right now. We are so heavily conditioned to
• 5:26
  the immediate visual feedback of our devices.
• 5:29
  Like close using your rings on an Apple Watch.
• 5:31
  Exactly. Or glancing down to see your
• 5:34
  heart rate during a run. Won't the lack of a screen
• 5:37
  alienate users who want that immediate hit of dopamine?
• 5:41
  Like Is invisible AI guidance alone enough of a draw?
• 5:45
  Well, by intentionally removing the screen, Google makes the AI
• 5:50
  layer the core product. They aren't selling you a
• 5:52
  digital watch anymore, they're selling you a centralized brain
• 5:55
  for your health. OK, I see that.
• 5:57
  Because traditional trackers give you raw passive data,
• 6:00
  right? They tell you that you slept
• 6:01
  poorly, but they lack the context of why you slept.
• 6:04
  Poorly, right? They just give you a bad sleep
• 6:05
  score and leave you to figure it out.
• 6:07
  Exactly. But because the Fitbit Air is
• 6:11
  feeding continuous biometric data into the same Gemini AI
• 6:14
  that can read your medical records and see the photos of
• 6:17
  your meals, it can synthesize all those disparate data points.
• 6:21
  Yeah. So it can say you slept poorly
• 6:23
  because of that late high sodium meal you photographed combined
• 6:26
  with your historically low oxygen levels when pollen counts
• 6:30
  are this high. Wow, so it's shifting from a
• 6:33
  passive data display on your wrist to an active personalized
• 6:38
  health guide just whispering in your ear Or I guess sending a
• 6:41
  brief morning brief to your phone?
• 6:43
  Yeah, it positions Google as the overarching intelligence,
• 6:46
  regardless of what piece of plastic or metal you happen to
• 6:48
  be wearing. The hardware becomes a dumb
• 6:50
  sensor. The AI becomes the actual
• 6:52
  product. OK, so if Google is trying to be
• 6:54
  the overarching intelligence for our physical health, how do we
• 6:57
  know which intelligence to use for actual work?
• 7:00
  That is the billion dollar question right now.
• 7:03
  Right, because with AI suddenly becoming our personal trainers,
• 7:06
  our travel agents, and our customer service reps, it is
• 7:10
  getting increasingly critical to know which model is actually the
• 7:14
  smartest for our specific day-to-day needs.
• 7:17
  It is the ultimate modern dilemma for knowledge workers.
• 7:19
  I mean, capabilities are shifting daily.
• 7:22
  Relying on marketing buzz from these massive tech companies is
• 7:25
  just a losing strategy at this point.
• 7:27
  Well, the answer according to our next source is the Ultimate
• 7:31
  AI by taste test. Have you seen this guide on
• 7:34
  using a platform called Open Router Fusion?
• 7:36
  Yes, it's brilliant. It lets you test the exact same
• 7:39
  prompt across multiple AI models simultaneously.
• 7:43
  Think of it like a blind wine tasting for tech workers.
• 7:46
  That's a perfect way to describe it.
• 7:47
  Instead of guessing based on brand loyalty to like Open AI or
• 7:51
  Anthropic, you just line up the outputs and see which vintage
• 7:54
  actually solves your problem best.
• 7:56
  Demystifying the process is so crucial right now.
• 7:59
  You are testing empirical utility instead of just trusting
• 8:02
  A benchmark score published by the company that's trying to
• 8:05
  sell you the product. Absolutely.
• 8:06
  So here's how it actually works under the hood.
• 8:09
  You create an account, choose your payment method, either open
• 8:12
  a router, credits or your existing API keys and then you
• 8:14
  pick the specific models you want to compare side by side.
• 8:17
  In the source demo, I think they tested Opus 4.7 versus GPT 5.4
• 8:22
  versus Grok. Yeah, that's the one.
• 8:25
  And the specific parameters of the prompt are what make this
• 8:28
  testing method so effective, right?
• 8:29
  Yeah, the prompt they used was highly structured.
• 8:31
  It was. You are advising a twenty person
• 8:34
  sauce company deciding whether to replace its weekly status
• 8:37
  meeting with an async written update.
• 8:40
  Write a recommendation memo with three benefits, 3 risks and a
• 8:43
  two week implementation plan. Keep it concise and practical.
• 8:46
  Right, and by giving the models very strict constraints like 3
• 8:50
  benefits, 3 risks a specific timeline, you are testing their
• 8:54
  reasoning and formatting capabilities, not just their
• 8:57
  ability to generate creative text.
• 8:59
  And the cost efficiency of doing this is really what makes it
• 9:01
  accessible for everyone. Oh the cost is crazy.
• 9:04
  Running about 10 side by side comparisons using Open Router
• 9:07
  Fusion cost the tester around $0.40.
• 9:09
  Just $0.40. $0.40 to basically pit the world's smartest
• 9:14
  supercomputers against each other to solve your middle
• 9:17
  management meeting problem. I mean, think about how you work
• 9:20
  right now. You probably default to whatever
• 9:23
  AI tab you happen to have open in your browser.
• 9:25
  This flips that entirely. It highlights the absolute
• 9:28
  necessity of building a CHEAT SHEET for your daily workflows,
• 9:32
  because true critical thinking in this space means evaluating
• 9:35
  price and speed right alongside the output quality.
• 9:38
  Right, because one model might give you a brilliantly nuanced
• 9:41
  memo, but it cost 10 times as much and takes 30 seconds longer
• 9:45
  to generate. Exactly, and another might be a
• 9:47
  lightning fast and dirt cheap, which is perfectly suitable for
• 9:49
  just summarizing a long e-mail chain.
• 9:52
  Finding the right tool for the right job is basically the new
• 9:55
  required skill for modern work. But you know, we are sitting
• 9:58
  here testing these models against each other, otimizing
• 10:01
  our 40 cent prompts. But according to our next source
• 10:04
  from Anthropic, the models might very soon be testing and
• 10:07
  improving themselves. This is where things get really
• 10:09
  intense. We are moving from a paradigm
• 10:12
  where models are our tools to a paradigm where the models are
• 10:16
  essentially their own developers.
• 10:18
  Yeah, Anthropic's newly formed research arm, the Anthropic
• 10:22
  Institute, or Tai, just published a formal research
• 10:25
  agenda explicitly preparing for self improving AI systems.
• 10:29
  And Tai sits inside Anthropic so they can study internal usage
• 10:34
  workflows and security signals before any of this actually hits
• 10:37
  the public. And their agenda is kind of
• 10:40
  wild. It includes planning for massive
• 10:42
  economic disruption, governance issues and what they are calling
• 10:45
  a sudden intelligence explosion. They're even proposing internal
• 10:48
  fire drills for sudden capability surges.
• 10:51
  Yeah, and. Cold War style hotlines between
• 10:54
  AI labs and world governments. The language they are using is
• 10:56
  very deliberate. It draws directly from Anthropic
• 10:59
  Co founder Jack Clark's previous warnings about the sheer speed
• 11:02
  of AI advancement. OK, wait.
• 11:04
  Let me pause you there. Hotlines and fire drills?
• 11:06
  That sounds like intense geopolitical framing.
• 11:08
  It does sound extreme. Are they just normalizing the
• 11:11
  idea of AI as a global weapon to build hype for their new
• 11:14
  research arm? Honestly it feels a little like
• 11:17
  sci-fi fear mongering to me. This raises an important
• 11:20
  question, and it is a totally fair pushback, especially given
• 11:24
  the history of tech companies using dramatic language for
• 11:27
  marketing. Right?
• 11:28
  But to understand why Anthropic is using this specific
• 11:31
  terminology, we really have to look at the mechanics of self
• 11:35
  improving AI and intelligence. Explosion isn't being treated as
• 11:39
  some distant sci-fi scenario, it's a near term logistical
• 11:43
  reality. OK, how so?
• 11:45
  Well. Let's say an AI model becomes
• 11:47
  capable enough to write better code than its human creators.
• 11:50
  It then uses that ability to optimize its own underlying
• 11:54
  architecture, like tweaking its own algorithms to be more
• 11:57
  efficient. Right, OK.
• 11:58
  The next iteration of that model is inherently smarter, which
• 12:02
  means can optimize itself even faster.
• 12:04
  Oh, I see. It creates a massive
• 12:06
  exponential. Exactly.
• 12:08
  The capability surges they are worried about wouldn't happen
• 12:10
  over years. They could happen over a
• 12:12
  weekend. That's why they want fire
• 12:13
  drills. Wow, over a weekend.
• 12:15
  If you internal testing system suddenly jumps from basic
• 12:19
  reasoning to something entirely unmapped overnight, you need a
• 12:23
  pre established protocol to lock it down or communicate the
• 12:26
  breakthrough. That's what the hotlines are
• 12:28
  for. So it's not necessarily about
• 12:29
  weapons. No, it's about managing the
• 12:31
  sheer speed of disruption. And, you know, to their credit,
• 12:35
  Tai is committing to publishing economic index data, monthly
• 12:39
  worker surveys and threat research just to keep this
• 12:42
  process transparent. This idea of an AI autonomously
• 12:46
  navigating a code base and fixing things isn't just a
• 12:49
  theoretical concern for Anthropic either.
• 12:52
  Which actually brings us to the rapid fire practical tools
• 12:55
  hitting the market literally today.
• 12:57
  We have a lot of quick hits today.
• 12:58
  Yeah, so Mozilla just published A blog post about using Claude
• 13:01
  Mythos Preview for their own software security.
• 13:04
  And the results they publish perfectly illustrate that
• 13:07
  concept of an intelligence exlosion in miniature.
• 13:10
  The AI patched more bugs in April than Mozilla patched in
• 13:13
  the past 15 months combined. 15 months.
• 13:16
  That's insane. And we have to look at how it
• 13:18
  actually does that. It's not just generating
• 13:20
  snippets of code in a chat window for a human to copy and
• 13:23
  paste, right? Right.
• 13:24
  The AI is actually navigating the complex, fragmented code
• 13:28
  base of a major browser, identifying the obscure
• 13:31
  vulnerabilities and writing the patches autonomously.
• 13:34
  And that's the connection. Right.
• 13:36
  AI fixing 15 months of human code bugs in a single month is
• 13:39
  exactly the the kind of exponential self improving
• 13:42
  capability that Tai is setting up those hotlines for.
• 13:45
  Exactly, because when an AI can do that safely, the speed of
• 13:50
  software development and software security fundamentally
• 13:52
  changes. The bottleneck is no longer
• 13:55
  human typing speed or human QA testing.
• 13:58
  And Speaking of removing bottlenecks, Perplexity is
• 14:00
  rolling out its personal computer feature to all Mac
• 14:03
  users. The key phrase in the source
• 14:05
  material is that it can take agentic action across your local
• 14:08
  files and the comment browser. The term agentic is so critical
• 14:12
  here. And Agentic AI doesn't just
• 14:14
  answer your questions, it has the agency to execute multi step
• 14:17
  tasks on your behalf without waiting for your next prompt.
• 14:20
  OK, So what does that look like on a Mac?
• 14:22
  In this case, it means the AI can dive into your messy desktop
• 14:26
  folders, locate specific documents, extract the data you
• 14:29
  need, and organize it in your browser entirely on its own.
• 14:33
  And the scale of data being organized by AI right now is
• 14:36
  getting massive. For example, Scale AI just
• 14:39
  landed a $500 million Pentagon contract for military data
• 14:44
  analysis. Just for context, that is a five
• 14:46
  times jump from their $100 million deal just last
• 14:50
  September. The sheer scale of investment
• 14:52
  the defense sector is pouring into data structuring shows you
• 14:55
  where the real value of AI lies. The military has endless streams
• 14:59
  of raw data, right? Satellite imagery, Logistics
• 15:01
  Reports, Communications. But raw data without
• 15:03
  intelligence structuring is useless.
• 15:05
  Scale AI provides the framework to make that data readable for
• 15:09
  agentic AI models to actually act upon.
• 15:11
  Totally. Now, on a slightly less
• 15:13
  militaristic note, over in the audio world, Spotify launched a
• 15:16
  personal podcasts feature. Oh, this is a brilliant
• 15:19
  application of voice synthesis for personalized learning.
• 15:21
  It really is. You can feed it your messy class
• 15:24
  notes, a PDF of a daily briefing, or a stack of research
• 15:28
  papers, and Spotify's agents will digest that text, summarize
• 15:32
  the key points, and use voice synthesis to turn it into a
• 15:35
  custom audio show right inside your Spotify library.
• 15:38
  You're basically generating your own educational curriculum on
• 15:41
  the fly. Which is so cool.
• 15:43
  We are also seeing major workflow shortcuts for creators.
• 15:47
  11 labs just dropped studio agents.
• 15:49
  This one is wild for video editors.
• 15:50
  It's an AI editor that actually drafts videos and places sound
• 15:54
  effects for you it. Really.
• 15:55
  Yeah, it reads the visual context of your video file,
• 15:58
  identifies what is happening on screen, and automatically maps
• 16:01
  the correct sound effects to the precise millisecond on your
• 16:04
  editing timeline. Taking that tedious frame by
• 16:07
  frame timeline workout of video editing frees up the human
• 16:10
  creator to focus entirely on the narrative structure of the
• 16:13
  content. Exactly.
• 16:14
  And then Grok also released an Imagine Quality Mode mode for
• 16:18
  much higher realism in image generation, making those
• 16:21
  synthetic visuals harder to distinguish from photography.
• 16:23
  And finally, for our quick hits, Open AI introduced Trusted
• 16:27
  Contact. It's an opt in ChatGPT feature
• 16:30
  that alerts a designated family member or friend if the AI
• 16:33
  detects a risk of self harm in the user's prompts.
• 16:36
  Which is huge. As these systems have become
• 16:39
  increasingly conversational and empathetic in their tone,
• 16:42
  they're naturally becoming integrated into people's
• 16:44
  private, emotional lives. Yeah, users treat them as
• 16:47
  sounding boards. Right.
• 16:48
  So having a safeguard that can break the digital wall and
• 16:51
  contact a human in a crisis is a necessary step for public
• 16:54
  deployment. It really is.
• 16:55
  You know, it's funny. The Pentagon is spending half a
• 16:57
  billion dollars on AI data, and tech giants are bracing for
• 17:00
  intelligence explosions, but everyday people are using these
• 17:04
  exact same tools to solve much more relatable, deeply human
• 17:08
  problems. Which is honestly often where
• 17:10
  the most profound immediate impacts of tech technology are
• 17:13
  actually felt through society. Very true.
• 17:15
  Our final source today highlights a community AI
• 17:18
  workflow from a reader named Tatiana B.
• 17:20
  She lives in San Francisco. She's the CEO of a startup and a
• 17:24
  mom to a 2 1/2 year old. Sounds like a busy life.
• 17:26
  Extremely. And she uses AI to handle the
• 17:29
  mental load of home management. Because the invisible labor of
• 17:33
  running a household is essentially a massive logistical
• 17:36
  puzzle. She uses AI to compile a
• 17:38
  comprehensive home management document.
• 17:41
  It covers everything. Her daughter's exact meal
• 17:44
  preferences, the step by step daily routine, the specific
• 17:48
  house chores that need to happen on Tuesdays versus Thursdays.
• 17:52
  She treats it like a complex work project.
• 17:54
  So an LLM acts as a structured brain dump in this scenario.
• 17:57
  Exactly. She goes back and forth with the
• 17:59
  AI, using it as a sounding board to think through what actually
• 18:03
  needs to get done. The AI asks her questions to
• 18:06
  fill in her blind spots. That's smart.
• 18:07
  Yeah, it helps her get all that draining, swirling mental load
• 18:10
  out of her own head and onto paper.
• 18:12
  The result is that when new household help comes in, she has
• 18:15
  a seamless handoff ready to go. No starting from scratch.
• 18:19
  No explaining the morning routine for the 5th time.
• 18:21
  This is a brilliant example of applying agentic thinking to
• 18:24
  daily life. We talk so much about AI writing
• 18:27
  emails or analyzing military data, but Tatiana is using it to
• 18:31
  structure chaos. Because she offloaded that
• 18:33
  logistical burden to the AI, she says she is freed up to be fully
• 18:37
  present with her daughter and fully focused at work.
• 18:40
  She is using the machine to reclaim actual human presence.
• 18:45
  The AI just becomes an operational partner, allowing
• 18:47
  her to step out of the logistics and back into her own life.
• 18:50
  It is such a great perspective, and it brings us to a
• 18:52
  fascinating place to wrap up today's deep dive.
• 18:55
  We started by talking about AI graduating from a laggy
• 18:58
  walkie-talkie to a fluent, invisible expert in our ear.
• 19:02
  Yeah, which leaves us with a rather profound question to
• 19:05
  consider. If AI is rapidly becoming our
• 19:08
  screenless health coach, our autonomous code bug fixer, our
• 19:12
  real time language translator, and even our home logistics
• 19:15
  manager, what does the work of being a human actually look like
• 19:19
  when all the operational logistics of life are fully
• 19:22
  outsourced? Wow, that is definitely
• 19:25
  something to Mull over today as you go about your work.
• 19:27
  Thank you so much for joining us on this deep dive.
• 19:29
  Don't forget to subscribe to stay updated on all these rapid
• 19:32
  changes. And if you enjoyed our
• 19:33
  conversation today, please read the show 5 stars.
• 19:36
  We will see you back here for tomorrow's deep dive.