Latest / Today’s AI News / BONUS EP: What If AI Only Knew the Past? Meet Talkie, the Language Model Born in 1930

Transcript

BONUS EP: What If AI Only Knew the Past? Meet Talkie, the Language Model Born in 1930 · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by 9X
• 0:02
  Productions. I I want you to imagine
• 0:06
  something for a second. Lay it.
• 0:07
  On me. Imagine walking into this
• 0:09
  massive, grand library where time mysteriously stopped
• 0:13
  exactly a century ago. Like, perfectly frozen, but the
• 0:17
  librarian inside has somehow memorized every single book in
• 0:20
  there. What could that librarian teach
• 0:23
  you about the future? Oh wow, that is a.
• 0:26
  That's a fascinating way to frame.
• 0:27
  It right because today we are taking a deep dive into the
• 0:30
  surprisingly high stakes world of vintage large language
• 0:35
  models. Yeah, which sounds like total
• 0:37
  sci-fi, but it's very real. It really is.
• 0:39
  These are AI systems that are deliberately trained exclusively
• 0:43
  on historical data up to a super specific cutoff date.
• 0:46
  So like maybe 2019 or 1930 or even, you know, all the way back
• 0:50
  to the Roman. Empire just entirely cordoning
• 0:51
  off the present day. Exactly, and on the surface it
• 0:54
  sounds like this perky time travel experiment, but it is
• 0:58
  actually emerging as one of the most critical new frontiers for
• 1:02
  testing how AI reasons, how it forecasts global events.
• 1:06
  And this is the big one, whether it can genuinely Dr. scientific
• 1:10
  discovery. It's wild because it really
• 1:12
  flips the script on how the whole industry approaches
• 1:14
  artificial intelligence. I mean, the default instinct has
• 1:17
  always been just feed the AI every single scrap of modern
• 1:21
  data we can possibly scrape from the Internet, right?
• 1:24
  The whole more data equals a smarter model mentality.
• 1:27
  Exactly. But with these vintage models,
• 1:29
  developers are intentionally starving the system of the
• 1:32
  present. They are basically imposing this
• 1:34
  artificial amnesia. Just to isolate what an AI can
• 1:37
  logically deduce using only like the fragmented building blocks
• 1:40
  of the past, right? So we should look closely at a
• 1:43
  recent flagship experiment in this space.
• 1:45
  Researchers built a 13 billion parameter vintage model and they
• 1:48
  trained it on roughly 260 billion tokens of pre 1931
• 1:53
  English text. Which is just a massive amount
• 1:55
  of historical. Text it is, and the 1930 cutoff
• 1:58
  was chosen mostly because those works are securely in the US
• 2:02
  public domain, so they had this huge accessible corpus to work
• 2:05
  with. But when they actually started
• 2:07
  prompting this model, the outputs were just astonishing.
• 2:11
  Oh, they really were. Like, for instance, they asked
• 2:13
  it to adopt the persona of a wise observer in the year 1928,
• 2:18
  and they asked it what it feared most for the coming decades.
• 2:22
  And the. Response it generated was
• 2:24
  honestly it was haunting. It warned of an encroaching
• 2:26
  universal materialism. It specifically cautioned that
• 2:30
  society was being overrun by Mammon worshippers, and it said
• 2:34
  something like for ideals, men's substitute aspirations.
• 2:38
  Which is just such a vivid, tragic piece of writing,
• 2:41
  especially when you consider that this AI mathematically
• 2:43
  thinks it is sitting right on the precipice of the
• 2:46
  catastrophic 1929 stock market crack.
• 2:49
  Oh yeah. Doesn't know the Cliff is right
• 2:51
  there. And you see that deep historical
• 2:53
  immersion across all sorts of different tests.
• 2:56
  Like when the researchers instructed the model to draft a
• 2:59
  letter from the perspective of a rural English doctor in 1919
• 3:03
  who's dealing with returning World War One soldiers, the
• 3:06
  persona it adopted was incredibly weary.
• 3:08
  It wrote. Good Lord, how damaged?
• 3:11
  Yes, good Lord, how damaged? And then it organically
• 3:14
  transitioned into discussing the absolute limitations of medicine
• 3:18
  during the influenza epidemic. It was essentially admitting
• 3:20
  that doctors of that era could do, well, little more than keep
• 3:23
  patients hydrated. Wow.
• 3:26
  And then there was that formal debate it generated about the
• 3:28
  concept of aerial bombardment, set in 19210.
• 3:32
  That one was intense. Yeah, it accurately echoed early
• 3:35
  military theorists. And it coldly argued that
• 3:37
  bombing civilian populations could actually be considered
• 3:40
  like a humane weapon. Yeah, because it would
• 3:43
  efficiently shorten wars by breaking national unity.
• 3:45
  Just ruthless logic based on the text of the time.
• 3:48
  Yeah, it even cited statistics like 22,000 bombs dropped and
• 3:52
  700 civilians killed just to build a strategic case for
• 3:56
  future conflicts. OK, let's unpack this.
• 3:58
  Let's do it, because this feels like interacting with a highly
• 4:01
  advanced piece of historical fiction.
• 4:04
  But isn't there a risk that this model just becomes, you know, a
• 4:07
  perfect echo chamber for the biases and the limited
• 4:10
  perspectives of the past? What's fascinating here is it
• 4:13
  does reflect the culture of its training text, but that's
• 4:16
  exactly why it's so valuable. OK, how so?
• 4:19
  Well, think about the humanistic benefits of these models.
• 4:22
  We are building systems capable of exploring true counterfactual
• 4:26
  history on a massive scale. Historically, knowledge was
• 4:30
  strictly siloed by geography. There was a massive firewall
• 4:34
  between different civilizations. I mean, scholars in the year 500
• 4:38
  AD in China had absolutely no idea what scholars in Western
• 4:41
  Europe were writing. And vice versa, yes.
• 4:42
  Exactly, but an AI doesn't have physical borders.
• 4:46
  Right, So if you trained a vintage model on both Chinese
• 4:49
  and Western texts from the year 500 AD, it could
• 4:52
  anachronistically merge their knowledge.
• 4:54
  Precisely, the AI synthesizes the latent philosophical and
• 4:58
  scientific concepts from both cultures, so we can interact
• 5:02
  with it to see what breakthroughs might have
• 5:05
  happened centuries earlier if those distinct intellectual
• 5:08
  frameworks had simply been able to interact and iterate on one
• 5:11
  another. That is a fascinating
• 5:13
  application, really, but I have to stop and look at the
• 5:15
  economics here, because training a language model from scratch
• 5:19
  costs an absolute fortune. Oh, a staggering amount of
• 5:22
  money. Right.
• 5:23
  We are talking hundreds of millions of dollars in computing
• 5:26
  power for a frontier model, tech developers and massive
• 5:30
  corporations. They aren't spending that kind
• 5:32
  of capital just to build a counterfactual history simulator
• 5:35
  for academics. There has to be a larger
• 5:37
  structural motive. There is absolutely, and it has
• 5:40
  everything to do with achieving what developers call epistemic
• 5:43
  AI. Epistemic AI.
• 5:45
  OK, break that down for. Me So epistemic AI is a pursuit
• 5:48
  of using artificial intelligence to accurately model the
• 5:51
  mechanics of the real world, to make highly calibrated forecasts
• 5:55
  and actually help humans invent genuinely new things.
• 5:58
  But testing a modern AI is ability to forecast rare,
• 6:02
  complex events. You know, like a global
• 6:04
  pandemic, a supply chain collapse, or an energy crisis.
• 6:08
  It presents a massive methodological problem.
• 6:10
  Because you can't test a system's ability to predict the
• 6:13
  future if it already knows the outcome.
• 6:15
  Exactly. A modern AI has already read the
• 6:17
  Wikipedia page for COVID-19. It already knows the entire
• 6:21
  sequence of the Ukraine war. Right.
• 6:23
  It's the equivalent of testing someone's ability to predict the
• 6:25
  lottery numbers when they already have the winning ticket
• 6:28
  in their pocket. Yeah, that doesn't prove
• 6:30
  anything. No, it doesn't.
• 6:31
  So to scientifically evaluate forecasting algorithms,
• 6:35
  researchers need a flawless control group like a gold
• 6:38
  standard baseline. And that is where a vintage LLM
• 6:42
  trained strictly up to, say, December 2019, becomes
• 6:46
  invaluable. Yes, it possesses 0 latent
• 6:49
  knowledge of COVID-19, so researchers can run back testing
• 6:52
  experiments where they feed that 2019 model the early confusing
• 6:56
  fragmented data from January 2020.
• 6:59
  Oh wow. And then they measure whether
• 7:00
  it's reasoning architecture could have accurately predicted
• 7:03
  the pandemic's global trajectory, entirely isolated
• 7:06
  from the outcome. Exactly, it removes the bias
• 7:09
  completely. Here's where it gets really
• 7:11
  interesting, though. That same logic applies to
• 7:13
  scientific discovery, yes. Which might be the most
• 7:16
  compelling part of all this research.
• 7:18
  Developers are asking these crazy questions like can an LLM
• 7:23
  trained only up to 1989 independently conceptualize the
• 7:26
  architecture of the World Wide Web or discover the foundations
• 7:31
  of quantum computing? Right.
• 7:32
  Or what if you had a model train up to 1600, right before Newton
• 7:37
  formalized his laws of motion? Or, before probability theory
• 7:40
  was even established, could the AI make those conceptual leaps
• 7:44
  using only the fragmented knowledge of the Renaissance?
• 7:47
  And this cuts to the heart of the biggest debate in artificial
• 7:50
  intelligence right now, generalization versus
• 7:53
  memorization. Critics frequently argue that
• 7:56
  large language models are merely stochastic parrots.
• 7:59
  The argument is that they aren't actually thinking or reasoning,
• 8:02
  they're simply regurgitating massive amounts of memorized
• 8:05
  data scraped from the modern Internet.
• 8:08
  Right, They're just really good at guessing the next word based
• 8:10
  on stuff they've already seen. Exactly so.
• 8:13
  Vintage models offer a pristine laboratory to definitively prove
• 8:17
  whether an AI can extract underlying mathematical or
• 8:21
  scientific logic and apply it to a completely novel situation.
• 8:24
  And there is a remarkable example of that in the 1930
• 8:28
  model experiment. The developers gave this model a
• 8:31
  programming test it. Sounds impossible.
• 8:33
  A total paradox because an AI trained exclusively on pre 1931
• 8:38
  text has absolutely no concept of a digital computer, let alone
• 8:41
  a modern programming language like Python, right?
• 8:44
  But the researchers gave the model a few examples of Python
• 8:47
  code in its prompt, and they asked it to write a script to
• 8:50
  invert a rotation cipher, which is basically decoding a message
• 8:54
  by reversing the mathematical function used to scramble.
• 8:56
  It and the vintage model successfully wrote the correct
• 8:59
  Python code. Yeah, just by swapping an
• 9:02
  addition operation for a subtraction operation in the
• 9:04
  syntax. OK.
• 9:05
  But wait, I have to challenge this.
• 9:06
  How do we know it actually understood the underlying math?
• 9:09
  Is it possible the model just, I don't know, memorized some
• 9:13
  obscure 1920s logic puzzle about ciphers that happened to map
• 9:17
  perfectly onto the Python syntax it was prompted with?
• 9:21
  It's a critical question for sure, but the structure of the
• 9:24
  success rules out simple memorization.
• 9:28
  The model didn't just output an answer to a Riddle, it
• 9:31
  demonstrated an understanding of inverse functions, which is a
• 9:34
  core mathematical concept, and it successfully abstracted that
• 9:38
  concept into a completely alien framework, which in this case
• 9:43
  was the Python syntax provided in the prompt.
• 9:46
  I see it. Proved it could separate the
• 9:47
  fundamental logic of inversion, the historical context it was
• 9:50
  trained on, and then apply it to a structural format it had
• 9:54
  literally never seen before. That is incredible, and the
• 9:56
  researchers also mapped the limits of its predictive
• 9:59
  abilities using a very clever method.
• 10:01
  They used the New York Times on This Day feature.
• 10:04
  Oh yeah, this is a brilliant test.
• 10:05
  They fed the model nearly 5000 historical event descriptions
• 10:08
  from different decades just to measure how surprising the text
• 10:11
  was to the AI. And when we say surprising, we
• 10:14
  are talking about a specific machine learning metric called
• 10:17
  perplexity. Right.
• 10:18
  Explain perplexity for us. Sure.
• 10:20
  So perplexity mathematically measures how confused or
• 10:24
  surprised a models probability distribution is by the next
• 10:28
  sequence of words. If a model predicts the next
• 10:31
  word with high confidence, perplexity is low.
• 10:34
  But if the actual next word shatters its statistical
• 10:37
  expectations, perplexity spikes. And you can see on their data
• 10:40
  graphs exactly when historical events triggered those
• 10:43
  mathematical spikes. As the timeline of the prompts
• 10:47
  moved past the 1930 cut off and pushed into the 1950s and 60's,
• 10:52
  the model's perplexity just skyrocketed it.
• 10:54
  Couldn't comprehend it. Exactly.
• 10:56
  It was structurally, mathematically shocked by the
• 10:58
  unfolding of the mid 20th century.
• 11:00
  But you know, for any of these epistemic back tests to actually
• 11:03
  be valid, the time capsule has to be completely impenetrable.
• 11:07
  Yes, 100%. If a model is going to prove it
• 11:09
  can invent a modern concept using only historical data, it
• 11:13
  cannot have even a single token of modern data hiding in its
• 11:17
  neural network. Which seems like a massive
• 11:19
  engineering headache. I mean, if the entire value of
• 11:22
  this scientific test relies on the model being completely blind
• 11:25
  to the future, how do you actually guarantee that?
• 11:28
  One slip up, one modern word, and your entire control group is
• 11:31
  ruined. It is brutally difficult and
• 11:34
  it's referred to as the shadow of the future or temporal
• 11:37
  leakage. Temporal leakage, yeah.
• 11:39
  And the researchers discovered just how pervasive this is with
• 11:42
  an early 7 billion parameter version of their 1930 model.
• 11:47
  During testing, the model just casually mentioned that Franklin
• 11:50
  D Roosevelt was president in 1933 and that he signed the New
• 11:54
  Deal. It also possessed knowledge
• 11:56
  about the United Nations and the division of Germany after World
• 11:59
  War 2. OK, how does a model strictly
• 12:01
  constrained to text published before 1931 know about the
• 12:05
  geopolitical fallout of 1945? Metadata and modern editorial
• 12:09
  insertions. Oh, of course, right.
• 12:11
  When we digitize historical works, we rarely strip away the
• 12:14
  modern packaging. An automated web scraper might
• 12:17
  pull a genuinely public domain text from 1920, but that
• 12:21
  specific digital edition includes, say, a preface written
• 12:24
  by a historian in 1998 explaining the author's legacy.
• 12:28
  Or the publisher added modern footnotes referencing the cold.
• 12:31
  War. Exactly.
• 12:32
  And if your data ingestion pipeline doesn't catch those
• 12:35
  anachronisms, that text gets embedded into the training data.
• 12:39
  Suddenly, your 1930 model has absorbed a localized
• 12:42
  understanding of the 20th century.
• 12:44
  And even if you managed to scrub all the metadata, just getting
• 12:47
  the historical text digitized accurately is its own nightmare,
• 12:50
  because obviously native digital text didn't exist in 1930.
• 12:54
  Every single newspaper, book and scientific journal has to be
• 12:58
  physically scanned and transcribed by Optical character
• 13:00
  recognition software, or OCR. And classic OCR relies on
• 13:05
  modern, standardized digital fonts to recognize characters,
• 13:08
  so it catastrophically fails when confronted with old, faded
• 13:12
  typefaces. Or complex physical newspaper
• 13:14
  layouts. Right, or historical ligatures
• 13:16
  where the ink of the letters bleed together.
• 13:18
  The researchers actually provided an example of how the
• 13:21
  model sees the text of The Wonderful Wizard of Oz from an
• 13:24
  1899 scan. The OCR transcription was
• 13:28
  complete chaos. Oh, it's illegible.
• 13:30
  Words were just mangled into random carrots, numbers and
• 13:33
  shattered punctuation. That garbage data destroys the
• 13:37
  models ability to learn. In machine learning, a model
• 13:40
  relies on the consistent structure of language to
• 13:43
  optimize its neural pathways, a process called gradient descent.
• 13:47
  When it is said chaotic OCR text, the model waste massive
• 13:51
  amounts of computational power just trying to decipher the
• 13:55
  grammar of the random symbols just.
• 13:56
  Trying to figure out what a word even is, right?
• 13:58
  The researchers found the training on messy OCR text
• 14:01
  caused the models learning efficiency to plummet to just
• 14:04
  30% compared to a model trained on clean human transcribed text.
• 14:08
  There's a huge. Drop.
• 14:10
  They literally have to build custom vintage OCR vision models
• 14:13
  just to teach the system how to correctly parse ink bleed on a
• 14:16
  19th century printing press. But OK, assuming you solved the
• 14:20
  OCR roblem and you perfectly seal out modern metadata, you
• 14:23
  run into another massive wall. Exactly.
• 14:26
  If we strictly cannot use modern text, how do we teach this
• 14:29
  vintage model to actually converse?
• 14:32
  Because modern LLM's only know how to be polite, helpful chat
• 14:35
  bots Because they undergo a massive post training phase
• 14:38
  where they are fine-tuned on millions of modern human written
• 14:43
  chat transcripts, right? If you feed those modern chat
• 14:46
  logs into the 1930 model, you instantly inject it with 2024
• 14:50
  vocabulary and speech patterns. That is the post training
• 14:53
  paradox, and to bypass it, researchers had to get
• 14:56
  incredibly creative. I love this part.
• 14:58
  To teach the vintage model how to follow instructions and
• 15:00
  maintain a coherent multi turn conversation without breaking
• 15:04
  its temporal character, they fine-tuned it on highly
• 15:07
  structured historical texts. They utilize 1859 etiquette
• 15:11
  manuals, 1890s parlor guides, and early 1900s letter writing
• 15:15
  manuals. Which conceptually makes so much
• 15:17
  sense, using Victorian etiquette to teach an AI how to be
• 15:19
  structurally helpful. But you still need a way to
• 15:21
  mathematically score how well the model is following those
• 15:24
  complex conversational instructions, right?
• 15:26
  Right, and they solved that by using a modern AI, specifically
• 15:29
  Claude 4.6 as a reward model. OK, the modern AI steps.
• 15:33
  In Yes, the modern AI evaluates the vintage model's outputs and
• 15:37
  scores its conversational flow on a strict numerical scale.
• 15:42
  The genius here is that the modern AI acts purely as a
• 15:44
  structural judge. It evaluates the probability of
• 15:48
  a good conversational turn without ever feeding its own
• 15:51
  modern vocabulary or factual knowledge back into the vintage
• 15:55
  system. OK, so they've managed to seal
• 15:57
  the data. They've figured out how to train
• 15:59
  it to chat using etiquette manuals.
• 16:01
  But that leads us to the ultimate bottleneck of this
• 16:03
  entire endeavor, which is scale. Yes, the biggest hurdle.
• 16:07
  A modern frontier model requires upwards of 50 trillion words of
• 16:11
  text to train. If developers are strictly
• 16:14
  limiting themselves to pristine, perfectly scanned, completely
• 16:17
  isolated pre 1930 data, they are going to run out of text.
• 16:22
  Humanity simply did not produce enough written material back
• 16:24
  then to scale these models up to modern parameters.
• 16:27
  It just doesn't exist. So how do we push this
• 16:29
  technology forward? Well, developers are leaning
• 16:31
  heavily into three highly technical solutions to bypass
• 16:34
  this scaling bottleneck. The first is synthetic data
• 16:37
  bootstrapping. Imagine you have a smaller,
• 16:39
  weaker vintage model that has successfully ingested all the
• 16:43
  clean 1930 text available. You instruct that model to
• 16:47
  generate thousands of historical paraphrases, analytical essays,
• 16:51
  and remixes based solely on the facts it already knows.
• 16:54
  But how does a model bootstrapping itself actually
• 16:56
  create net new intelligence? If it's just remixing what it
• 17:00
  already knows, aren't you just feeding it its own exhaust?
• 17:03
  It creates intelligence by expanding token diversity and
• 17:07
  forcing the model to explicitly map latent connections.
• 17:10
  OK. Give me an example.
• 17:11
  For example, if the model read Fact A in a Science Journal and
• 17:14
  Fact B in an economics book, the synthetic generation phase might
• 17:18
  ask it to write a comprehensive essay connecting the two.
• 17:21
  By articulating that latent connection, it generates the
• 17:24
  massive new data set of clean, highly structured vintage text.
• 17:28
  The synthetic corpus can then be used to train a much larger,
• 17:31
  more sophisticated vintage model, essentially letting the
• 17:34
  AI build its own expanded textbook.
• 17:36
  That perfectly explains the mechanism of synthetic scaling,
• 17:39
  but what about the massive financial cost of training these
• 17:42
  separate timelines? That leads to the second
• 17:45
  solution, chronological forking. Forking.
• 17:48
  OK, think of training an AI like growing a massive tree.
• 17:52
  Instead of planting one tree and growing it to 2021, and then
• 17:55
  paying hundreds of millions of dollars to plant a completely
• 17:57
  separate tree and grow it to 2024, you do one massive
• 18:01
  training run up to 2021. Then you fork the model.
• 18:04
  You split it. Exactly.
• 18:06
  You take a structural snapshot of the tree, freezing the neural
• 18:09
  network weights exactly where they are at that checkpoint.
• 18:13
  That frozen version becomes your isolated vintage model for
• 18:16
  forecasting tests. Meanwhile, you allow a copy of
• 18:19
  that model to continue training on new 2022 to 2024 data,
• 18:23
  becoming your modern frontier model.
• 18:25
  Saving massive amounts of compute.
• 18:27
  Precisely. Now the third solution outlined
• 18:29
  in the research is what really caught my attention, because it
• 18:31
  involves outsourcing complex computational tasks to modern
• 18:35
  AI. The proposal is that if you have
• 18:37
  a 1900 model grappling with a complex logical problem, it
• 18:41
  could theoretically use a 2024 LLM as a calculator to help it
• 18:45
  process the reasoning. Wait, if a 1900 model asks A2024
• 18:50
  model for help, doesn't the modern model inadvertently reply
• 18:54
  with modern facts? That immediately breaks the
• 18:56
  temporal firewall? This raises an important
• 18:58
  question, and it's where developers employ
• 19:00
  compartmentalized LLMS. It relies on a mechanism called
• 19:04
  mathematical masking. Masking, right?
• 19:06
  When the 1900 model queries the 2024 model, the modern model is
• 19:10
  subjected to strict date vector constraints.
• 19:13
  Its attention mechanism is mathematically masked, heavily
• 19:16
  penalizing the probability of generating any tokens associated
• 19:19
  with post 1900 context. Wow.
• 19:21
  OK. It is essentially forcing a
• 19:23
  modern genius to help you solve a complex equation, but
• 19:26
  mathematically restricting their vocabulary and tools to only
• 19:29
  what existed in the 19th century.
• 19:31
  The modern AI provides the advanced structural reasoning,
• 19:33
  but the compartmentalized masking ensures 0 futuristic
• 19:36
  facts leak back into the past. The sheer engineering required
• 19:40
  to maintain that isolation is just staggering.
• 19:43
  You know, taking a step back from all the massive computing
• 19:45
  clusters and historical firewalls, there is a way you
• 19:49
  listening right now, can apply the core philosophy of this
• 19:52
  research to your daily life. Developers refer to these as
• 19:56
  community AI workflows, and it's a powerful mental model for
• 20:00
  problem solving. It really is.
• 20:02
  The overarching benefit of a vintage LLM is that it
• 20:05
  fundamentally breaks out of the echo Chamber of modern
• 20:08
  groupthink, and you can simulate this intellectual isolation
• 20:12
  using the AI tools you interact with everyday.
• 20:15
  If you are grappling with a complex business strategy, a
• 20:18
  logistical hurdle, or an organizational problem, try
• 20:21
  applying A strict date constraint to your prompt.
• 20:23
  I love doing this. Instruct the AI to solve your
• 20:26
  problem utilizing only the frameworks and technological
• 20:29
  limitations of 1995 or 1950 or the Victorian era.
• 20:33
  By doing that, you artificially strip away the digital crutches
• 20:36
  we constantly rely on. If you remove the modern
• 20:39
  toolkit, you force the AI to return to 1st principles.
• 20:41
  You force it to look at the foundational logistics, the core
• 20:45
  human psychology, and the structural realities of the
• 20:48
  problem, rather than just throwing a modern software
• 20:50
  solution at it. It reveals foundational
• 20:53
  solutions that are often obscured because we are so
• 20:55
  blinded by the shiny tools of the present.
• 20:58
  It acts as a phenomenal lens for critical thinking.
• 21:02
  But beyond personal workflows, I think the most profound take
• 21:05
  away from the vintage LLM research is what it implies
• 21:08
  about our current trajectory. We are pouring immense resources
• 21:11
  into building these historical models specifically to test if
• 21:16
  they could have predicted our turbulent present.
• 21:18
  We want to know if a mathematically isolated 2019
• 21:21
  model could have accurately forecasted the global shock
• 21:24
  waves of the early twenty 20s. So what does this all mean?
• 21:27
  Well, it begs the ultimate question, if we flip that
• 21:30
  methodology around, what are the massive state-of-the-art LLMS we
• 21:34
  are training right now, today, silently predicting for the 20
• 21:37
  thirties? What fundamental invisible
• 21:39
  global shifts are our current models already anticipating that
• 21:43
  we, trapped in the present moment, are completely blind to?
• 21:47
  It's the ultimate test of epistemic AI.
• 21:50
  We are building the tools to map the future and we are just
• 21:54
  beginning to understand what those maps are trying to show
• 21:56
  us. We'll have to leave it there for
• 21:57
  today. Remember to hit that subscribe
• 21:59
  button to stay updated with the fastest moving industry on
• 22:02
  Earth. And if you enjoyed today's deep
• 22:04
  dive, please invite your friends and rate the show 5 stars.
• 22:07
  We will see you right back here for tomorrow's episode.
• 22:09
  Keep exploring.