Latest / Today’s AI News / Meta's AI Can Predict How Your Brain Works, Apple Opens Siri to All AI, Wikipedia Bans AI Writing

Transcript

Meta's AI Can Predict How Your Brain Works, Apple Opens Siri to All AI, Wikipedia Bans AI Writing · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by Nynex
• 0:02
  Productions. It's great to be here.
• 0:04
  So imagine trying to understand how a complex Symphony is played
• 0:09
  just by like listening through a thick concrete wall and then
• 0:13
  suddenly being handed the master sheet music.
• 0:16
  Oh wow, yeah. Right.
• 0:17
  That's essentially the breakthrough we're seeing in
• 0:19
  neuroscience today. And AI is basically figured out
• 0:22
  how to read the brain's sheet music without ever needing a
• 0:25
  physical scan. Which is just, it's a massive
• 0:29
  leap forward. It really is.
• 0:31
  So our mission for today's Deep Dive is to jump immediately into
• 0:34
  all these latest developments. We're starting with Meta's Mind
• 0:37
  Reading AI, and then we're going to look at Apple's massive Siri
• 0:41
  overhaul in iOS 27. Right, the walled garden is
• 0:44
  opening. Up.
• 0:44
  Exactly. Plus, we'll get into why
• 0:46
  Wikipedia is drawing this really hard line in the sand against
• 0:49
  AI, and then we'll share a slick way to turn perplexity into your
• 0:53
  own personal shopper. There is definitely a lot of
• 0:56
  ground to cover today. Yeah.
• 0:57
  So OK, let's unpack this. We're starting with Meta.
• 0:59
  They just open sourced a model called Tri V2.
• 1:02
  And this is an AI that was trained on over 1000 hours of
• 1:06
  brain scans from, I think more than 700 people.
• 1:10
  Yeah, over 700. Right.
• 1:11
  And its core function is to literally simulate neural
• 1:15
  activity across vision, hearing and language.
• 1:18
  Which, just to give people some context here, the first version
• 1:21
  of this model only mapped about 1000 brain regions.
• 1:24
  And that was from what, like 4 volunteers?
• 1:26
  Exactly just 4 but this new one Tri V2, it maps an incredible
• 1:31
  70,000 brain regions. That is just wild.
• 1:35
  It is. It has essentially replicated
• 1:37
  decades of science findings entirely in software.
• 1:40
  It can correctly pinpoint the exact brain regions for things
• 1:43
  like faces or speech or text with 0 physical scans required
• 1:48
  anymore. OK, but wait, I have to ask, how
• 1:49
  does a piece of software accurately predict population
• 1:52
  level brain activity better than an actual F MRI machine?
• 1:56
  Yeah, I mean, those machines cost millions of dollars for a
• 1:58
  reason. Right.
• 1:59
  Yeah. What's fascinating here is that
• 2:01
  you have to understand the concept of biological noise.
• 2:04
  When you put a human being inside an F MRI machine, you are
• 2:08
  trying to capture incredibly subtle changes in blood flow to
• 2:11
  very specific regions of the brain.
• 2:13
  But you know, a human being is a terribly noisy environment for
• 2:16
  data collection. Because we're like constantly
• 2:18
  breathing, swallowing, twitching, all that stuff.
• 2:21
  Precisely every single heartbeat sends this rhythmic pulse of
• 2:25
  interference right through that data.
• 2:27
  Oh wow, I never really thought about the heartbeat.
• 2:29
  Yeah, and every microscopic physical movement, combined with
• 2:32
  the, you know, the background magnetic hum of the machine
• 2:36
  itself, it creates this thick declared of noise over the
• 2:39
  actual cognitive signal you're trying to read.
• 2:42
  But try V2 solves this through sheer scale and pattern
• 2:45
  recognition. Because it was trained on so
• 2:48
  many different brains, the AI basically learned to
• 2:51
  algorithmically subtract the unique physical noise of the
• 2:54
  individual. It cross referenced those 700
• 2:57
  different biological environments to isolate the pure
• 3:01
  universal signal of how human neurons fire in response to a
• 3:05
  specific stimulus. So it's basically like
• 3:06
  mathematical noise cancelling headphones for brain waves?
• 3:09
  Yes, it just strips away the heartbeat and leaves only the
• 3:12
  visual thought. That is a perfect way to
• 3:14
  visualize it, and the implications for the broader
• 3:18
  scientific landscape are just staggering.
• 3:20
  We are really looking at the alpha fold moment for brain
• 3:23
  research here. Right.
• 3:24
  And just to remind everyone, alpha fold was the Google Deep
• 3:28
  Mind model that solved protein folding a few years back, right?
• 3:31
  Yes, exactly. Before alpha fold, figuring out
• 3:35
  the 3D structure of justice, a single protein required months
• 3:38
  of expensive, painstaking physical lab work, right?
• 3:41
  Alpha Fold compressed that massive physical bottleneck into
• 3:44
  literally seconds of computing time, and neuroscience has been
• 3:48
  trapped in a very similar physical bottleneck for decades
• 3:51
  now because. Of the scanners.
• 3:53
  Right. Historically, to test a
• 3:54
  hypothesis, you had to recruit subjects, schedule them, and put
• 3:58
  them inside these highly expensive, highly demanded
• 4:00
  scanners for every single new experiment.
• 4:02
  Which takes forever. It does.
• 4:04
  Entire fields of study could only move forward one small
• 4:07
  expensive scan at a time. But Tribe V2 compresses months
• 4:11
  of that expensive physical scanning into mere seconds of
• 4:13
  compute. Which brings us to the real
• 4:15
  kicker for you listening right now.
• 4:17
  Meta didn't just, you know, publish A theoretical research
• 4:20
  paper and hide it behind a paywall.
• 4:22
  No, they didn't. They open sourced the code, they
• 4:24
  released the model weights so the actual learned parameters of
• 4:27
  the AI, and they even put up a live demo online.
• 4:30
  Yeah, it's all out there. Meaning anyone with the
• 4:33
  requisite computing power can start running virtual brain
• 4:35
  experiments today. Yeah, you don't have to build
• 4:38
  the underlying architecture from scratch at all.
• 4:41
  It fundamentally democratizes a field of science that has, up
• 4:44
  until now been gated by extreme hardware costs and, you know,
• 4:48
  institutional access. Totally.
• 4:50
  Now, while Met is out here pulling the secrets of the human
• 4:53
  mind out into the open, Apple is finally opening up its own
• 4:57
  digital brain. Yes, this is a big pivot.
• 4:59
  Right, so starting with the upcoming iOS 27, Apple is
• 5:03
  actively opening Siri up to rival AI assistance.
• 5:07
  They're doing this via a new extension setting inside the
• 5:10
  operating system. This really is a massive pivot
• 5:13
  in how Apple views its walled garden.
• 5:16
  It really is. I mean, this move officially
• 5:18
  ends the exclusive integration deal that ChatGPT has held on
• 5:22
  the iPhone since way back in 2024, right?
• 5:25
  And frankly, sources are reporting that the actual
• 5:28
  consumer use of that specific ChatGPT integration has been
• 5:33
  minimal at best over the years. Yeah, it didn't really take off
• 5:36
  the way they hoped. Exactly.
• 5:38
  But moving forward, users are going to be able to route their
• 5:41
  voice queries to the frontier model of their choice right from
• 5:45
  the device level. And we're going to get a real
• 5:47
  look at this during Apple's WWDC developer event in early June.
• 5:51
  Which should be really interesting.
• 5:53
  Yeah, that event is slated to debut this new Siri overhaul,
• 5:56
  which is largely powered by Google's Gemini, actually.
• 5:59
  And so if you think about it, Apple's essentially building the
• 6:02
  ultimate digital shopping mall, and they're just letting
• 6:04
  ChatGPT, Gemini and Claude rent the storefronts.
• 6:08
  It's a brilliant strategy on Apple's part because they are
• 6:10
  actively choosing to circumvent the messy, incredibly expensive
• 6:15
  model war entirely. Right, Let everyone else fight
• 6:18
  it out. Exactly.
• 6:19
  Look at the landscape. You've got Google, Open AI,
• 6:22
  Anthropic, and they are all spending 10s of billions of
• 6:25
  dollars building these foundational models.
• 6:28
  The compute costs alone are insane.
• 6:29
  Insane. They are absorbing massive
• 6:32
  technical hurdles, immense costs, and severe reputational
• 6:35
  risks whenever those models hallucinate or output biased
• 6:39
  information. Yeah, so Apple is just letting
• 6:41
  everyone else walk through the minefield first.
• 6:43
  Exactly. Instead of fighting a grueling
• 6:46
  battle at the foundation level, Apple relies on its ultimate
• 6:49
  advantage, and that is distribution.
• 6:51
  Right, the hardware. Yes, by layering the best
• 6:55
  available external AI models on top of an installed base of over
• 6:58
  a billion iPhones, they let their hardware mote do all the
• 7:02
  heavy lifting. They own the interface, which
• 7:04
  means they fundamentally own the user.
• 7:07
  And there is a deeply cynical, yet I mean financially brilliant
• 7:10
  masterstroke hidden in this extension setting, isn't there?
• 7:14
  The App Store tab. Yeah, By allowing these rival
• 7:16
  chat bots into the ecosystem as selectable extensions, Apple
• 7:20
  instantly creates a fresh revenue stream.
• 7:22
  They always find a way. They do.
• 7:24
  They take their standard 30% cut of any premium AI subscriptions
• 7:28
  purchased across their devices, so they turn their biggest
• 7:32
  existential AI rivals into paying tenants.
• 7:36
  It's wild. Apple didn't spend the $10
• 7:38
  billion to train the model, but they will happily collect 1/3 of
• 7:42
  the revenue it generates on their hardware.
• 7:43
  It is the most classic Apple move imaginable.
• 7:46
  But you know, here is the fascinating contradiction in the
• 7:48
  tech world today. Apple is eagerly opening its
• 7:52
  doors to invite every synthetic mind into your pocket, but the
• 7:56
  Internet's most relied upon repository of human knowledge is
• 8:00
  aggressively locking them out. Right.
• 8:02
  The ideological divide regarding the value of synthetic text is
• 8:05
  becoming very, very stark right now, so.
• 8:07
  Wikipedia's volunteer editors just held a near unanimous 40 to
• 8:11
  2 vote. They have officially banned AI
• 8:13
  from writing or rewriting articles on the English language
• 8:16
  site. A 40 to 2 consensus is huge.
• 8:18
  Right, because Wikipedia editors famously debate the minutiae of
• 8:22
  single sentences for months, so to get 40 to 2 is virtually
• 8:25
  unheard of. It really speaks to a shared
• 8:28
  existential threat perceived by that community.
• 8:31
  Absolutely, and the author of the policy didn't mince words at
• 8:34
  all. They call this ban a pushback
• 8:38
  against insidification and the forceful push of AI by so many
• 8:42
  companies. Yeah, now the policy does have a
• 8:45
  little nuance. AI is still permitted for basic
• 8:48
  grammar fixes and direct translations.
• 8:50
  But, and this is a big but, only if there is strict human review
• 8:54
  before it goes live. Right, the human in the loop is
• 8:56
  mandatory. Exactly.
• 8:58
  And Wikipedia isn't standing alone in this philosophy either.
• 9:02
  Platforms like Stack Overflow and the German version of
• 9:04
  Wikipedia have enacted very similar bands.
• 9:07
  I believe the Spanish Wikipedia went even further right.
• 9:09
  They did. They aggressively banned AI
• 9:11
  entirely. Like even for those basic
• 9:13
  formatting and editing purposes. Complete ban.
• 9:15
  If we connect this to the bigger picture, to understand the
• 9:18
  timing of this defensive posture, we really have to look
• 9:20
  back at the tipping point we hit last year in 2025.
• 9:23
  Right, the big crossover. Yeah, that was the year when,
• 9:25
  according to widespread reports, AI generated text officially
• 9:29
  surpassed human output on the Internet for the first time.
• 9:32
  The exact moment the Internet became more synthetic than
• 9:35
  organic. Yes, the volume of machine
• 9:38
  generated content now outweighs human creation, so Wikipedia is
• 9:42
  essentially making a philosophical bet on scarcity.
• 9:45
  They are trying to hold the human line in an increasingly
• 9:48
  synthetic sea, and you can see how this sits in direct
• 9:51
  ideological contrast to projects like Elon Musk's Grokopedia,
• 9:55
  which is fundamentally designed to be a rapid, entirely AI
• 9:59
  created knowledge base. We are watching a fork in the
• 10:02
  road regarding how society values information.
• 10:05
  Do we prioritize the frictionless speed and vast
• 10:07
  synthesis of a machine, or do we prioritize the verifiable,
• 10:12
  transparent, albeit much slower, consensus of human volunteers?
• 10:16
  Which raises a massive logistical question for you to
• 10:18
  ponder. Can an open, crowdsourced
• 10:21
  Internet actually hold back that tide?
• 10:23
  I mean, if more than half the text generated online today is
• 10:26
  synthetic, keeping Wikipedia purely human is going to require
• 10:30
  an immense, almost unfathomable amount of policing by those
• 10:33
  volunteers. It is going to be incredibly
• 10:35
  difficult. Right.
• 10:35
  How do they even prove an article was written by a human
• 10:38
  when AI detectors are notoriously unreliable?
• 10:41
  That is the fragile point in their defense.
• 10:43
  Absolutely. It will become an endless game
• 10:45
  of cat and mouse regarding attribution and verification.
• 10:49
  The ban is a powerful statement of intent, but the actual
• 10:53
  mechanics of enforcing it over the next decade will be
• 10:55
  incredibly taxing on their volunteer base.
• 10:58
  Well, if Wikipedia won't let AI write its articles, you might as
• 11:02
  well put these models to work doing your chores.
• 11:04
  Like shopping? Exactly like shopping, this is
• 11:07
  where the technology shifts from theoretical philosophy to actual
• 11:12
  utility in your day-to-day life. Perplexity is rolled out a
• 11:15
  feature that acts as a highly capable personal shopper, and it
• 11:19
  fundamentally changes how we search for things online.
• 11:22
  It is a perfect example of moving beyond the traditional
• 11:25
  paradigm of information retrieval.
• 11:27
  Right, so here's how it works. If you are a Mac subscriber or a
• 11:30
  Pro user with credits on Perplexity, you can open their
• 11:34
  site or the latest app and toggle on a feature called
• 11:37
  Perplexity Computer. OK, From there, you just give it
• 11:39
  a natural language prompt. You tell it the specific item
• 11:42
  you want to buy, your brand preferences, your cosmetic
• 11:44
  style, and your exact sizing. And notice the mechanical shift
• 11:47
  here. A traditional search engine or
• 11:50
  even an early chat bot would simply return a list of blue
• 11:54
  links, right? Yeah, just a bunch of links.
• 11:56
  Or maybe a summary of available products, but you still have to
• 11:59
  manually click through and verify the stock yourself.
• 12:02
  Right, but Perplexity Computer actually delegates these little
• 12:06
  sub agents to go out and actively browse the web for you.
• 12:10
  Wow. It isn't just reading text.
• 12:12
  The software literally simulates the actions of a human user.
• 12:16
  It reads the real time stock levels on retail pages, it
• 12:19
  parses the drop down menus for your specific shoe size, and
• 12:23
  then it compiles all of this disparate data into a perfectly
• 12:26
  formatted comparative report. That saves so much time.
• 12:30
  It's amazing and you don't even have to instruct it on how to
• 12:32
  present the data. The model organizes it
• 12:35
  intuitively based on what makes sense for a shopper.
• 12:37
  The ability for an orchestrator model to spin up smaller sub
• 12:40
  agents, assign them specific scraping tasks and then
• 12:43
  synthesize their findings. That is what defines an active
• 12:46
  agent. It is taking multi step actions
• 12:48
  on your behalf, navigating the friction of the web so you don't
• 12:51
  have to. And if you are hunting for a
• 12:53
  specific deal, you can just tell the prompt.
• 12:56
  Turn this into a daily automation or make it weekly.
• 12:59
  Perplexity will literally schedule those agents to
• 13:02
  silently watch the prices for you every single morning.
• 13:05
  That is incredibly. Useful.
• 13:06
  Oh, and here's a quick pro tip for you listening.
• 13:09
  If you want to save on your token usage while running these
• 13:13
  intensive multi step tasks, pop into your settings and switch
• 13:17
  the orchestrator model so the main brain directing those
• 13:21
  little sub agents over to Claude Sonnet.
• 13:23
  Oh, good tip. Yeah, it handles the delegation
• 13:25
  beautifully, but it costs way less compute.
• 13:28
  It illustrates the fundamental transition we are living through
• 13:31
  right now. We're moving from asking a
• 13:33
  machine for a recipe to having the machine walk into the
• 13:37
  kitchen, check the pantry, order the missing ingredients, and
• 13:40
  prep the meal for you. Let's zoom through the quick
• 13:44
  hits and everything else shaping the AI landscape today, because
• 13:46
  that friction between what is real and what is generated is
• 13:50
  disappearing everywhere, particularly in the realm of
• 13:52
  audio. So first up, Google just rolled
• 13:54
  out Gemini 3 point one Flash Live.
• 13:58
  This is their new voice AI integrated into Search Gemini
• 14:02
  Live and their API. The major focus here is high
• 14:05
  speed, low latency and hyper realism.
• 14:08
  And latency is the absolute key to crossing the uncanny valley
• 14:11
  in audio. If a voice model pauses for even
• 14:14
  half a second longer than a natural human breath before
• 14:17
  responding, our brains immediately flag it as
• 14:20
  synthetic. It just sounds wrong.
• 14:23
  Yeah, you can instantly tell. Right, driving that latency down
• 14:26
  to milliseconds is what creates the illusion of genuine
• 14:29
  presence. And the tools to create that
• 14:31
  presence are getting wildly accessible.
• 14:33
  Mistral just dropped voxel TTS. There's a text to speech AI that
• 14:37
  can clone any speaker across 9 different languages using just a
• 14:40
  three second audio clip. Which is astonishing when you
• 14:43
  think about the underlying mechanism there. 3 seconds?
• 14:46
  Yeah, it doesn't need to learn your specific vocal patterns
• 14:48
  from scratch. It takes those 3 seconds and
• 14:51
  maps your unique acoustic stick features onto a massive
• 14:53
  pre-existing mathematical map of human vocal traits.
• 14:57
  The latent space. Exactly the latent space,
• 15:00
  allowing it to instantly puppeteer your digital voice.
• 15:03
  Meanwhile, on the transcription side of things, Cohere released
• 15:07
  Transcribe. This is a completely free open
• 15:10
  source speech recognition model that just knocked everything
• 15:13
  else off the podium. Oh.
• 15:14
  Really. Yeah, it took the absolute
• 15:15
  number one spot on Hugging Face for accuracy across 14 different
• 15:19
  languages. Wow.
• 15:20
  And to top off the audio boom, SUNO launched version 5.5 for
• 15:25
  their Pro subscribers. This brings massive upgrades to
• 15:28
  their AI music generation, including the ability to clone
• 15:31
  voices for vocals and a feature that actually learns your
• 15:34
  personalized musical style overtime.
• 15:37
  This raises an important question when you look at these
• 15:39
  audio developments in aggregate. You know models hitting #1
• 15:43
  accuracy for transcription. While cloning requires a mere 3
• 15:46
  seconds of source audio, You realize the audio layer of the
• 15:49
  Internet has become entirely malleable.
• 15:51
  Yeah, the friction of creating, faking or translating human
• 15:55
  speech is just gone. It really is.
• 15:56
  It's a lot to process, but moving into the corporate
• 16:00
  sphere, we're seeing real world integration at massive scale.
• 16:04
  Novo Nordisk, the pharmaceutical giant behind drugs like Ozempic,
• 16:08
  is actively deploying AI agents into their clinical trials right
• 16:12
  now. That's a massive deployment.
• 16:14
  Huge They're reporting that the technology is successfully
• 16:18
  trimming down regulatory approval times and significantly
• 16:21
  cutting their reliance on outside contractors.
• 16:24
  Applying agentic workflows to something as notoriously
• 16:27
  bureaucratic as clinical trials is highly significant, because
• 16:31
  the bottleneck in pharmaceuticals isn't always the
• 16:33
  science. It's often the mountain of
• 16:36
  regulatory paperwork and data synthesis.
• 16:38
  And why does that matter to you? Because AI agents cutting
• 16:41
  through clinical red tape means life saving or life altering
• 16:44
  treatments could potentially hit the consumer market years faster
• 16:47
  than they do under the current human driven paradigm.
• 16:49
  Exactly. Finally, in other corporate news
• 16:51
  today, Open AI has reportedly shelved its planned erotic
• 16:55
  chatbot mode indefinitely. This internal decision comes
• 16:59
  directly following pushback from both staff and investors
• 17:02
  regarding the feature. It serves as a reminder of the
• 17:04
  constant ongoing tension within these frontier organizations.
• 17:09
  They are continually testing the boundaries of consumer demand
• 17:12
  versus corporate liability and acceptable use.
• 17:15
  Yeah, so we've talked about multibillion dollar foundational
• 17:19
  models, pharmaceutical giants, and Wikipedia's existential
• 17:24
  crisis. But what does this technology
• 17:26
  actually look like when it lands in a normal person's day?
• 17:29
  Right, let's look at the practical side.
• 17:31
  Yeah, let's look at how the community is actually putting
• 17:33
  this tech to work Today. We have a fantastic workflow
• 17:37
  highlighted from a reader named Hugh F out in Sandwich, MA I.
• 17:40
  Always find these hyperlocal specific use cases the most
• 17:44
  illuminating. Same here.
• 17:45
  So Hugh plays the keyboard in a local open mic group.
• 17:49
  Now, if you've ever dealt with a full piano score, you know it
• 17:51
  spans many, many pages. Oh sure, this creates a physical
• 17:54
  friction where the musician has to frantically swipe on an iPad
• 17:58
  or physically turn pages Mint song, all while keeping the
• 18:01
  rhythm going. Which completely disrupts the
• 18:03
  flow state of the performance. Exactly so Hugh takes the PDF
• 18:06
  versions of these dense multi page piano scores, feeds them
• 18:10
  directly into Clawed, and prompts the AI to simplify the
• 18:13
  arrangement and condense it down to just two pages.
• 18:15
  Oh, that's clever. Right.
• 18:17
  The AI parses the musical notation, extracts the core
• 18:20
  melody and chords, and reformats it into a perfect condensed
• 18:24
  accompaniment piece that fits entirely on one screen or two
• 18:27
  sheets of paper. It completely eliminates the
• 18:30
  physical need to turn pages while performing.
• 18:32
  That is a remarkably elegant use of the models multimodal
• 18:36
  capabilities. It isn't just reading text, it
• 18:38
  is understanding spatial formatting, musical syntax, and
• 18:42
  structural data and then rebuilding it to solve a very
• 18:45
  real physical human constraint. And the end result?
• 18:48
  As Hugh points out, he spends way less time sitting at a desk
• 18:51
  adapting scores by hand. That means he sends more time
• 18:55
  actually practicing the instrument, which ultimately
• 18:57
  leads to a better performance and more applause at the open
• 18:59
  mic. That's fantastic.
• 19:01
  I love this because it perfectly captures the ideal state of AI
• 19:04
  integration. It's not about replacing the
• 19:06
  human artist. Hugh still has to sit there,
• 19:09
  feel the music and play the keys.
• 19:11
  It is simply about using the machine to remove the tedious
• 19:14
  friction from the creative process.
• 19:16
  And that is a crucial take away. The most profound, immediate
• 19:19
  value of these tools isn't always found in grand, sweeping
• 19:23
  disruptions of massive industries.
• 19:25
  Often the real magic happens when an individual uses the
• 19:28
  technology to elegantly solve a highly specific, very personal
• 19:32
  bottleneck in their daily life. I think that is the perfect note
• 19:35
  to end on. Today we talked about Meta
• 19:38
  creating digital brains, Apple bringing rival AI minds into our
• 19:41
  pockets, and Wikipedia fighting to keep synthetic minds out.
• 19:45
  It begs the question for you to Mull over 10 years from now,
• 19:48
  will the ultimate luxury be having an AI handle everything
• 19:51
  for you? Or will the ultimate luxury be
• 19:54
  experiencing a digital space guaranteed to be 100% human?
• 19:58
  Man, circling back to that opening thought, we finally have
• 20:01
  the master sheet music to the Symphony.
• 20:03
  But now we have to make the hard choices about who or what we
• 20:06
  actually want playing the instruments.
• 20:08
  Make sure you subscribe to stay updated.
• 20:10
  If you love this deep dive, please rate the show 5 stars.
• 20:13
  We'll see you right back here for tomorrow's episode of
• 20:16
  Today's AI News.