Latest / AI Fire Daily / 🎙️ EP 337: a16z Partner’s Viral AI Bama Rush Experiment & Ornith-1.5 Open-Source AI Release

Transcript

🎙️ EP 337: a16z Partner’s Viral AI Bama Rush Experiment & Ornith-1.5 Open-Source AI Release · AI Fire Daily

• 0:00
  Imagine a 19 -year -old viral TikTok star at
• 0:03
  Bama Rush. Oh, right. Yeah, so she captures a
• 0:06
  million views almost effortlessly. She completely
• 0:09
  fools the mainstream media, but... She does not
• 0:13
  actually exist. It is genuinely wild to think
• 0:16
  about. It really is. Yeah. So welcome to the
• 0:18
  deep dive. You are joining us on a pretty fascinating
• 0:21
  journey today. Yeah. We are looking at a bizarre
• 0:23
  horizon right now. Absolutely. So first we will
• 0:26
  explore this whole illusion of AI influencers.
• 0:29
  Then we track the massive corporate scramble
• 0:32
  for data. Right. And the hardware, too. Exactly.
• 0:35
  After that, we will look at some highly specialized
• 0:38
  professional tools hitting the market. And finally,
• 0:40
  we examine an AI model that, well, actually teaches
• 0:43
  itself. It is a genuinely staggering shift to
• 0:46
  witness. We're watching tools become fully autonomous
• 0:49
  creators. Yeah. So let's unpack this fake TikToker
• 0:52
  first. An A16's partner named Olivia Moore ran
• 0:56
  this fascinating experiment. She created a fictional
• 0:59
  19 -year -old student named Janie. Janie was
• 1:01
  supposed to be rushing an Alabama sorority, right?
• 1:03
  Right. And the experiment pulled in around one
• 1:05
  million TikTok views. And that immense engagement
• 1:08
  happened in just one single week. That is a staggering
• 1:12
  return on attention. I mean, it really is. It's
• 1:15
  crazy. Let's examine the mechanics behind this
• 1:17
  viral illusion. Moore spent about 30 minutes
• 1:22
  a day managing Janie. That is nothing. I know.
• 1:25
  She spent roughly $100 in total. It is a staggeringly
• 1:28
  low barrier to entry. She basically stacked Lego
• 1:31
  blocks of data together. Right. Janie actually
• 1:34
  started her life from a single chat GPT image.
• 1:37
  Just one image. Then Moore used Minimax for the
• 1:39
  talking clips. Oh, yeah. And she used Grok Imagine
• 1:42
  for the complex dance routines. Right. And Eleven
• 1:45
  Labs provided the incredibly realistic synthetic
• 1:47
  voice. It is amazing how these separate tools
• 1:50
  integrate so seamlessly. You don't need a Hollywood
• 1:52
  studio anymore. No, you really don't. You just
• 1:54
  need a laptop and a clever prompt. Yeah. But
• 1:56
  here is where the story gets deeply fascinating
• 1:59
  to me. Yeah. TikTok actually labeled eight of
• 2:02
  those 20 videos. They were clearly marked as
• 2:05
  AI generated content. Right. People noticed those
• 2:08
  subtle AI artifacts within a few days. They saw
• 2:10
  the weird fingers and the strange lighting. But
• 2:13
  the viewers just kept watching her anyway. Exactly.
• 2:16
  The labels did not noticeably hurt her overall
• 2:19
  performance. Not at all. Janie reached about
• 2:21
  1 ,300 dedicated followers very quickly. Her
• 2:25
  first video approached 100 ,000 views alone.
• 2:28
  The Daily Mail even covered her as a real personality.
• 2:31
  I saw that. Yeah, they treated Janie as a popular
• 2:33
  Alabama recruitment figure. And this happened
• 2:36
  before the entire experiment was fully revealed.
• 2:39
  This whole situation reminds me of professional
• 2:41
  wrestling. Oh, that is a great analogy. Right.
• 2:43
  Fans know the wrestling matches are completely
• 2:45
  fabricated. The physical reality is totally secondary
• 2:48
  to the drama. Yeah, the compelling story is what
• 2:51
  keeps them watching. Exactly. Moore came to a
• 2:54
  very similar conclusion herself. Audiences care
• 2:57
  way less about physical reality than we assume.
• 2:59
  They really do. They just want high -quality,
• 3:02
  engaging entertainment value. They are fine with
• 3:04
  it, provided there is transparency. It is a profound
• 3:07
  shift in consumer psychology. We are witnessing
• 3:10
  the birth of synthetic reality media. We are
• 3:13
  watching the lines blur in real time. We really
• 3:16
  are. So does this fundamentally alter human trust
• 3:20
  or are viewers just accepting a new genre of
• 3:23
  fiction? I think they are embracing a totally
• 3:26
  new genre. People want compelling narratives
• 3:28
  above everything else today. They really don't
• 3:30
  mind if the actor is made of pixels. They just
• 3:32
  want to feel invested in the ongoing story. So
• 3:35
  we happily trade physical reality for pure entertainment
• 3:38
  without missing a single beat. Exactly. All right.
• 3:41
  Let's transition from fake creators to real data.
• 3:45
  Because the corporate hunger for real -world
• 3:48
  human data is immense right now. Well, synthetic
• 3:51
  data from a fake TikToker is incredibly cheap.
• 3:54
  But training massive foundational models requires
• 3:57
  messy reality. Right. You cannot build a world
• 4:00
  model on fake dance videos. You really can't.
• 4:02
  So Google just made a fascinating strategic move
• 4:04
  here. They bid $10 million for Spirit Airlines
• 4:07
  data. Wow. Yeah, this is anonymized internal
• 4:10
  data from the bankrupt company. Planes weren't
• 4:13
  the only valuable thing left behind. Operational
• 4:15
  logistics data is incredibly valuable for AI
• 4:18
  training. It teaches the models how complex physical
• 4:22
  systems actually operate. It helps them understand
• 4:25
  deep spatial reasoning and resource allocation.
• 4:28
  It makes perfect sense when you put it like that.
• 4:30
  We also see a massive hardware scramble happening
• 4:32
  globally. The demand for raw compute power is
• 4:36
  absolutely insatiable. It is wild. There was
• 4:39
  even a crazy rumor about SpaceX buying Cognition.
• 4:42
  It supposedly happened right after the cursor
• 4:44
  coding tool exploded. Cognition's CEO completely
• 4:47
  denied that specific rumor, though. He explicitly
• 4:50
  stated that they are not for sale. But the immense
• 4:53
  capital flowing into hardware is undeniably real.
• 4:57
  I mean, look at the AI chip startup called Etched.
• 4:59
  Oh, yeah. They just raised $700 million in funding.
• 5:02
  That puts them at a $21 billion valuation. They
• 5:05
  doubled their massive value in under a single
• 5:07
  month. It is unheard of. And that happened after
• 5:10
  Jane Street tested their specialized hardware.
• 5:12
  Jane Street successfully deployed the very first
• 5:15
  rack of chips. And this insane compute power
• 5:18
  is reaching consumers very quickly, too. Google
• 5:20
  is giving students a free year of Gemini. They
• 5:23
  are offering new study notebooks and visual tools.
• 5:25
  They're essentially hooking the next generation
• 5:27
  of professionals. Right. They want them locked
• 5:29
  into their specific AI ecosystem early. Meanwhile,
• 5:33
  Zed AI just released a massive system update.
• 5:36
  They made GLM 5 .3 completely free inside Zcode.
• 5:41
  I saw that. Yeah. It offers 5 million free daily
• 5:45
  processing tokens. It features a staggering 1
• 5:49
  million token context window, and it also pushes
• 5:52
  out 128 ,000 output tokens. I mean, I still wrestle
• 5:56
  with prompt drifts myself. Managing a million
• 5:58
  tokens of context is genuinely overwhelming.
• 6:01
  Right. And keeping the model focused is incredibly
• 6:03
  hard. The AI tends to lose the plot with that
• 6:06
  much data. You feed it a novel and it forgets
• 6:08
  the first chapter. Exactly. OpenAI is actually
• 6:10
  trying to manage this massive scale. They found
• 6:12
  a clever way to compete with Anthropic. They're
• 6:15
  testing private safety processing right now.
• 6:17
  It detects long -term misuse without keeping
• 6:20
  your private data. It runs those crucial safety
• 6:22
  checks in an isolated environment. Which is huge.
• 6:25
  It is a massive step forward for enterprise security.
• 6:27
  They're also expanding chat GPT advertisements
• 6:30
  globally. 31 European markets will see standalone
• 6:34
  sponsored widgets soon. Advertisers are going
• 6:36
  to get direct access to this ecosystem. Yeah.
• 6:39
  We are also seeing... hardware integrate this
• 6:42
  power directly. There is a leaked Apple AirPods
• 6:45
  demo circulating online right now. Oh, the sure
• 6:48
  one. Yeah, let's users save what they see in
• 6:50
  real time. The fundamental concept looks absolutely
• 6:52
  wild to consider. It transforms your earbuds
• 6:55
  into an always -on memory assistant. The biggest
• 6:57
  question is still how Apple is doing this efficiently.
• 7:00
  Right. So with hardware scaling this fast, can
• 7:03
  privacy tools like OpenAI's new processing? actually
• 7:07
  keep pace. Honestly, it is a constant grueling
• 7:10
  uphill battle. The processing volume easily outruns
• 7:13
  traditional safety nets. Localized privacy solutions
• 7:15
  are absolutely critical right now. We cannot
• 7:18
  rely entirely on cloud -based safety checks anymore.
• 7:21
  Hardware moves way too fast, making localized
• 7:23
  privacy solutions absolutely critical right now.
• 7:26
  So we see this massive compute power trickling
• 7:28
  down now. It is moving from foundational models
• 7:32
  to hyper -specialized tools. Yeah, these tools
• 7:35
  integrate directly into everyday professional
• 7:37
  workflows. We are moving past the simple chatbot
• 7:40
  interface entirely. Thank goodness. The tools
• 7:43
  are gaining highly specific professional contexts.
• 7:46
  They're becoming tailored digital colleagues
• 7:48
  rather than generic assistants. Consider the
• 7:51
  legal sector tool called Harjit II. It lets AI
• 7:54
  agents inherit deep, complex case context. It
• 7:57
  actually remembers the distinct writing style
• 8:00
  of each lawyer. That is wild. Right. It works
• 8:03
  seamlessly with Harvey's new in -house legal
• 8:05
  model. This represents a massive shift in professional
• 8:08
  identity. The AI is mimicking the human professional
• 8:12
  perfectly. Firefox is taking a slightly different
• 8:14
  approach with Smart Window. They added an AI
• 8:17
  chatbot for live web search. It helps with complex
• 8:20
  tab organization and model choice. And crucially,
• 8:23
  it keeps all your private data stored locally.
• 8:26
  It doesn't send your browsing habits to a massive
• 8:28
  server. Which is exactly what we were just talking
• 8:30
  about. Right. Then we have tools targeting the
• 8:32
  developers directly. Cursor Origin is a highly
• 8:35
  specialized... git hosting platform it is built
• 8:38
  specifically for human coders and autonomous
• 8:41
  agents you can review pull requests directly
• 8:44
  inside the cursor environment you manage the
• 8:46
  complex code alongside your ai agent it feels
• 8:50
  like a genuine collaborative working relationship
• 8:52
  yeah but then we see tools moving in the exact
• 8:54
  opposite direction Claude Watermark Remover is
• 8:58
  a perfect example of this. It finds and scrubs
• 9:01
  hidden AI traces in your text. It removes exotic
• 9:04
  spaces and HTML artifacts instantly. You can
• 9:07
  clean the document with just one single click.
• 9:10
  Wait, I have to push back on this entirely. Yeah.
• 9:12
  This creates a deeply fascinating philosophical
• 9:14
  tension for us. We build complex tools like Harvey
• 9:18
  II to integrate AI. Right. We want it embedded
• 9:20
  deeply in our daily professional work. Yet we
• 9:23
  build tools explicitly to hide its presence.
• 9:25
  Well, we want the productivity, but we deeply
• 9:27
  fear the stigma. That is true. Nobody wants a
• 9:30
  client to know a robot wrote it. It ruins the
• 9:32
  illusion of bespoke, highly expensive human labor.
• 9:36
  It feels like a strange form of professional
• 9:38
  hypocrisy. We are lying to our clients about
• 9:42
  who does the work. So what does it say about
• 9:44
  our ethics when we build tools explicitly to
• 9:47
  scrub AI watermarks? It reveals a profound insecurity
• 9:51
  in our human professional value. We crave the
• 9:55
  machine's efficiency, but we desperately need
• 9:57
  the human credit. We are terrified of becoming
• 10:00
  obsolete in our own specialized fields. We crave
• 10:03
  machine efficiency, but we desperately need to
• 10:06
  claim the human credit. Exactly. Okay, we must
• 10:08
  transition to the most profound shift today.
• 10:10
  We are moving away from human -guided workflows
• 10:12
  entirely. Yes, we are shifting toward models
• 10:15
  that remove human bottlenecks. Ornith AI just
• 10:17
  released a massive breakthrough called Ornith
• 10:20
  1 .5. This is a completely new open -source AI
• 10:23
  model family. They released a 9 billion parameter
• 10:26
  -dense model variant. And they also released
• 10:29
  two massive Moe versions for developers. One
• 10:31
  is $35 billion. The other is $397 billion. And
• 10:36
  just for clarity, Moe means specialized subnetworks
• 10:38
  that take turns to save massive computing power.
• 10:41
  Yeah, it is an incredibly efficient way to process
• 10:43
  data. But the core breakthrough here is the self
• 10:46
  -improvement loop. This is the crazy part. Right.
• 10:49
  These models dynamically generate their own new
• 10:51
  learning tasks. They build their own complex
• 10:54
  training scaffolds completely autonomously. They
• 10:57
  produce solution attempts without any human prompting.
• 11:00
  Then... They learn directly from those generated
• 11:02
  results. They are no longer relying on human
• 11:05
  design training data. It is amazing. Whoa, imagine
• 11:09
  scaling to a billion queries without human bottlenecks.
• 11:12
  That kind of speed is genuinely hard to wrap
• 11:15
  your head around. The implications of that massive
• 11:17
  scaling are genuinely hard to grasp. We are essentially
• 11:20
  watching a machine evolve its own intelligence.
• 11:22
  It creates the test, takes the test, and grades
• 11:25
  the test. Exactly. And the benchmark results
• 11:27
  are already proving this autonomous method works.
• 11:30
  The massive 397 billion model scored highly.
• 11:34
  It hit 86 .1 on terminal bench 2 .1. And it also
• 11:38
  scored 56 on the grueling deep SWE benchmark.
• 11:42
  Warneth says this performance is roughly comparable
• 11:44
  to Claude Opus 4 .8. Yeah, it shows remarkably
• 11:47
  strong results across several other major evaluations
• 11:49
  too. It performed exceptionally well on SWBench
• 11:52
  and HLE. It also excelled at ClawEvol and the
• 11:56
  Tool Decathlon benchmark. These are not simple,
• 11:59
  easily -gamed, standardized tests anymore. Not
• 12:01
  at all. The global accessibility of this advanced
• 12:04
  model is also striking. All three models are
• 12:07
  released under the OpenMIT license. That broad
• 12:09
  license allows for commercial and deep research
• 12:12
  use. They are available as raw, open weights
• 12:14
  on Hug & Face. And highly optimized quantized
• 12:17
  versions are widely available right now, too.
• 12:20
  Let's explain why that quantization step matters
• 12:22
  so much. Supported formats include FP8, GGUF,
• 12:26
  MLX, and NVFP4. Basically, these formats compress
• 12:30
  the massive models into manageable sizes. You
• 12:33
  don't need a massive data center to run them.
• 12:34
  You can run them locally using free tools like
• 12:37
  Alama. And they also work perfectly inside LM
• 12:39
  Studio right now. Developers can easily fine
• 12:41
  -tune them for highly custom use cases. The entire
• 12:44
  training process pushes toward continuous autonomous
• 12:47
  creation. It continuously creates its own learning
• 12:51
  tasks day and night. So what happens to human
• 12:54
  oversight when models permanently outgrow the
• 12:57
  need for our training data? We transition rapidly
• 12:59
  from being active teachers to passive observers.
• 13:02
  The AI starts defining its own internal logic
• 13:05
  entirely. We just sit back and watch the intelligence
• 13:08
  explode outward. rapidly shift from active teachers
• 13:11
  to merely observing their independent autonomous
• 13:14
  evolution. Precisely. Let's take a moment to
• 13:16
  calmly synthesize this journey. We started our
• 13:19
  deep dive with a completely fake TikTok persona.
• 13:22
  Right. Janie was built using entirely disjointed,
• 13:25
  totally separate tools. She proved we value a
• 13:27
  good story over physical reality. Then we moved
• 13:30
  through a multi -billion dollar corporate scramble.
• 13:32
  Massive corporations are desperately hunting
• 13:34
  for hardware and real data. And we saw this immense
• 13:37
  compute power flow into hyper -specialized tools.
• 13:40
  Professionals are integrating AI deeply into
• 13:42
  very specific daily workflows. Finally, we ended
• 13:45
  with massive AI models learning to teach themselves.
• 13:48
  The overarching narrative here is becoming very
• 13:50
  clear to us. AI is rapidly transitioning away
• 13:53
  from being a piloted tool. It is becoming a fully
• 13:56
  autonomous engine that builds itself. It is a
• 13:58
  profound shift in our technological evolution.
• 14:01
  It changes our entire relationship with the digital
• 14:03
  world. I want to leave you with a final provocative
• 14:06
  thought. Think about Ornith 1 .5 constantly generating
• 14:10
  its own complex learning tasks. It continuously
• 14:13
  improves itself without any human intervention
• 14:15
  whatsoever. What happens when you hook that self
• 14:18
  -improving loop up to something else? Oh, wow.
• 14:21
  Imagine connecting it directly to a synthetic
• 14:23
  creator like Janie. Will our future entertainment
• 14:25
  optimize itself in real time? Will it evolve
• 14:28
  instantly based purely on our subconscious reactions?
• 14:31
  Thank you for joining us on this deep dive. Always
• 14:33
  remember to question the information you consume
• 14:35
  daily. We will see you on the next exploration.