Latest / AI Fire Daily / 🎙️ EP 362: Claude Discovers Novel Enzyme System & AI Fingerprints in Academic Writing

Transcript

🎙️ EP 362: Claude Discovers Novel Enzyme System & AI Fingerprints in Academic Writing · AI Fire Daily

• 0:00
  Imagine a scientist hunting through ancient DNA,
• 0:02
  right? They're looking for totally new biological
• 0:05
  systems. Now, imagine that scientist isn't actually
• 0:10
  human. Yeah. And at the exact same time, it's
• 0:12
  doing something completely different. Right.
• 0:14
  It is quietly changing the punctuation human
• 0:16
  scientists use. I mean, it's leaving these invisible
• 0:19
  marks on our entire culture. It's just such a
• 0:21
  striking contrast to observe. The very same technology
• 0:24
  hunting for cures is editing our grammar. Welcome
• 0:28
  to the deep dive. Today, we're exploring a massive
• 0:30
  shift in AI. We're looking at sources that reveal
• 0:33
  a huge transition. AI is moving way past being
• 0:37
  a simple chatbot. It's becoming a truly autonomous
• 0:40
  agent. It's discovering new biology entirely
• 0:42
  on its own. Exactly. It's running our daily chores
• 0:45
  and our finances. It is even leaving measurable
• 0:47
  stylistic fingerprints on human culture. We have
• 0:50
  a lot of fascinating ground to cover. The evolution
• 0:52
  from chatbot to autonomous agent is profound.
• 0:55
  We're really talking about a massive shift in
• 0:57
  agency. So let's begin right at the bleeding
• 1:00
  edge of science. We want to see what actual computational
• 1:02
  autonomy looks like. So we're looking closely
• 1:04
  at Antropic today. They recently launched a new
• 1:07
  life sciences research group. They actually built
• 1:09
  a physical wet lab. Yeah, that physical wet lab
• 1:12
  is a crucial detail here. Software companies
• 1:14
  usually stay entirely in the digital realm, but
• 1:17
  here they're bridging a massive gap. They're
• 1:19
  connecting digital intelligence to real biological
• 1:23
  experiments. And one of their first projects.
• 1:26
  produced a wild result. Oh, it really did. Claude
• 1:29
  helped identify a previously uncharacterized
• 1:31
  biological enzyme system. This system has these
• 1:35
  unusual CRISPR -like repeating DNA structures.
• 1:38
  Right, and human researchers had completely overlooked
• 1:41
  this hidden pattern. The mechanism of this discovery
• 1:44
  is absolutely fascinating, you know? Anthropic
• 1:46
  didn't just ask Claude a simple question. They
• 1:49
  deployed a massive swarm of autonomous Claude
• 1:51
  agents. These agents searched a massive biological
• 1:54
  sequence. database. The database contained over
• 1:57
  200 ,000 reverse transcriptases. Which, just
• 1:59
  for context, a reverse transcriptase is an enzyme
• 2:02
  that turns RNA back into DNA. Exactly. An enzyme
• 2:04
  that turns RNA back into DNA. So the GLOD agents
• 2:07
  analyzed this massive ocean of genetic data.
• 2:10
  Right. They were looking for structural anomalies
• 2:12
  in the code. They initially identified roughly
• 2:14
  3 ,500 candidate systems. Then they narrowed
• 2:17
  those down to 20 promising candidates. They investigated
• 2:20
  those final 20 candidates even further. The scale
• 2:23
  of this computation effort is staggering. It
• 2:26
  really is. Whoa. Imagine spinning up 950 AI agents
• 2:31
  to chew through 210 million tokens in a day.
• 2:34
  It really makes you pause and think. They ran
• 2:37
  these agents continuously for 21 hours. Yeah.
• 2:40
  A single human researcher could never process
• 2:42
  that data. The AI just doesn't get tired or lose
• 2:45
  focus. So one specific agent spotted an unusual
• 2:48
  repeating DNA pattern. It was sitting right beside
• 2:51
  a known reverse transcript. Human researchers
• 2:54
  had apparently stared right past it for years.
• 2:57
  Anthropic calls this newly identified biological
• 2:59
  system ART. That stands for Array Associated
• 3:02
  Reverse Transcriptase. And the structure of this
• 3:04
  ART system is highly specific. It contains a
• 3:07
  reverse transcriptase and an accessory protein.
• 3:09
  Right, plus a long array of space DNA repeats.
• 3:12
  This specific pattern strongly resembles a traditional
• 3:15
  CRISPR array. We definitely must be careful about
• 3:18
  drawing early conclusions here. Anthropic hasn't
• 3:21
  shown that... Art is a CRISPR replacement. Exactly.
• 3:24
  They haven't confirmed that it can actually edit
• 3:26
  DNA. Its biological function remains entirely
• 3:29
  unknown right now. It's like dumping a million
• 3:32
  random Lego blocks. You drop them all on the
• 3:34
  floor at once. A human scientist gets exhausted
• 3:37
  looking for the hidden pattern. Yeah, they spend
• 3:39
  a lifetime looking for pieces that match. But
• 3:42
  the AI just tirelessly clicks pieces together
• 3:45
  at lightning speed. It finds the hidden pattern
• 3:47
  in a massive tapestry. That's a perfect way to
• 3:50
  visualize the search. Claude handled the computational
• 3:53
  discovery process with high -level direction.
• 3:55
  Human scientists then performed all the physical
• 3:58
  wet lab experiments. What does this human -AI
• 4:00
  division of labor mean for the future of scientific
• 4:03
  labs? It fundamentally changes the bottleneck
• 4:06
  of scientific discovery. AI handles the brute
• 4:09
  force pattern recognition across massive data
• 4:11
  sets. Humans then shift their focus entirely
• 4:14
  to physical verification. So AI does the heavy
• 4:17
  lifting. Humans verify the physical reality.
• 4:20
  Precisely. It accelerates the scientific method
• 4:23
  dramatically. Well, if Claude can autonomously
• 4:25
  hunt through complex DNA. It makes sense that
• 4:29
  AI agents are expanding their reach. Oh, absolutely.
• 4:31
  We're starting to trust them with autonomous
• 4:33
  navigation. They're managing our daily lives
• 4:36
  and our finances. Anthropic actually made Claude
• 4:38
  three times faster in two weeks. They achieved
• 4:41
  this by using Claude to improve Claude. Yeah,
• 4:44
  the AI is literally optimizing its own underlying
• 4:47
  architecture. It identifies the bottlenecks in
• 4:50
  its own processing pipeline. Then it writes the
• 4:53
  code to fix those very bottlenecks. Meta is also
• 4:56
  moving aggressively in this direction. Mark Zuckerberg
• 4:58
  announced they're going all in on Muse. The major
• 5:01
  tech companies recognize that agents are the
• 5:04
  future. They want persistent digital entities
• 5:06
  working on your behalf. This agentic shift is
• 5:09
  already transforming the financial sector. There's
• 5:11
  a hedge fund called Bracket 22. It is run almost
• 5:14
  entirely by autonomous AI agents. And the manager,
• 5:17
  Brian Kelly. mostly just presses a button. The
• 5:20
  agents allocate capital and execute the trades
• 5:22
  autonomously. I still wrestle with prompt drift
• 5:25
  myself, you know, and people are letting agents
• 5:27
  run hedge funds. Right. And it's only getting
• 5:29
  faster. They are trusting algorithms with complex,
• 5:33
  high stakes financial decisions. These agents
• 5:36
  are also handling our everyday mundane tasks.
• 5:40
  ChatGPT Voice can now draft documents and summarize
• 5:43
  emails. Yeah, you can run larger work tasks hands
• 5:46
  -free from your phone. You just speak your intent,
• 5:49
  and the agent executes it. We have a perfect
• 5:51
  example from a Delta Airlines passenger. They
• 5:54
  were facing a frustrating seven -hour flight
• 5:56
  delay. A human would normally wait on hold for
• 5:59
  hours. Exactly. You'd listen to terrible hold
• 6:01
  music and get stressed. But this passenger used
• 6:03
  a Muse AI agent instead. The agent navigated
• 6:07
  the customer service bureaucracy autonomously.
• 6:09
  It secured a $250 credit. And it also rebooked
• 6:12
  the entire flight in just five minutes. The AI
• 6:15
  agent doesn't get tired or frustrated. It levels
• 6:17
  the playing field against a massive corporate
• 6:19
  bureaucracy. The healthcare sector is seeing
• 6:21
  similar massive adoption. Anthropic teamed up
• 6:24
  with a platform called Open Evidence. They provide
• 6:26
  free AI clinical support to doctors. Yeah, U
• 6:29
  .S. clinicians made 42 million queries in August
• 6:32
  alone. Doctors are relying on these agents for
• 6:35
  critical medical information. We also see huge
• 6:38
  investments in AI drug discovery. Like a company
• 6:41
  called Invada just raised $311 million. They're
• 6:45
  pushing AI -discovered drugs from plants into
• 6:47
  human trials. This includes treatments for severe
• 6:50
  skin conditions and weight management. New agentic
• 6:53
  tools are launching constantly to empower users.
• 6:56
  Tesseract lets AI agents edit videos with compositions
• 6:59
  and keyframes. It turns a highly technical creative
• 7:02
  process into a guided conversation. OpenMUSE
• 7:05
  lets you self -host a personal, persistent AI
• 7:07
  agent. It has a browser and access to local files.
• 7:11
  Sherpa crafts complex story arcs for full -length
• 7:13
  fiction. We're seeing specialized agents for
• 7:15
  every domain of knowledge. OpenRadar is even
• 7:18
  cutting AI token costs roughly in half. They
• 7:21
  use a batch API across 70 different models. With
• 7:23
  tools like OpenMUSE and autonomous hedge funds,
• 7:25
  are we losing control or just delegating the
• 7:28
  mundane? We're not necessarily losing control
• 7:30
  over our lives. We're elevating our role from
• 7:33
  task doers to strategic managers. We set the
• 7:36
  high -level goals and the agents figure out the
• 7:38
  steps. We're moving from typing manual commands
• 7:41
  to managing digital employees. That is exactly
• 7:44
  the paradigm shift we're experiencing right now.
• 7:47
  Welcome back to the Deep Dive. So we've seen
• 7:50
  AI act as an autonomous scientist. We've seen
• 7:53
  it work as a daily digital assistant. Now let's
• 7:56
  look at something much more subtle. Right. The
• 7:59
  invisible ways AI interacts with us. It's actually
• 8:01
  rewiring human behavior and human culture. A
• 8:04
  fascinating new study analyzed a massive trove
• 8:07
  of scientific writing. They looked at 69 ,632
• 8:11
  Medrekfer preprints. They wanted to find AI's
• 8:15
  stylistic fingerprints in human text. That is
• 8:18
  a highly robust sample size for linguistic analysis.
• 8:21
  They have enough data to track very subtle behavioral
• 8:23
  shifts. They discovered a booming rise in the
• 8:25
  use of the MDash. This sharp increase started
• 8:28
  right after ChatGPT became widely available.
• 8:30
  The numbers in this study are quite revealing
• 8:33
  about human habits. Before ChatGPT, the MDash
• 8:36
  was relatively rare in these papers. Only 4 .23
• 8:40
  % of discussion sections used one. After ChatGPT
• 8:44
  launched, that number rose to 11 .58 % overall.
• 8:49
  The usage stayed near 4 % through 2023. It reached
• 8:53
  8 % in 2024. Then, you know, it skyrocketed to
• 8:57
  20 .3 % in 2025. The growth curve perfectly tracks
• 9:01
  the adoption of large language models. But we
• 9:04
  have to understand why AI uses the MDash so much.
• 9:07
  It's about how the models actually generate text,
• 9:10
  right? Exactly. Large language models generate
• 9:12
  text token by token. They simulate conversational
• 9:15
  flow by constantly appending new thoughts. And
• 9:18
  the em dash is the perfect punctuation mark for
• 9:21
  that appended structure. It allows the thought
• 9:23
  to expand naturally without starting a new sentence.
• 9:26
  The researchers ran multiple sensitivity and
• 9:28
  falsification tests. The pattern survived all
• 9:31
  of their rigorous statistical checks. They also
• 9:33
  looked at boilerplate sections of these scientific
• 9:35
  papers. These are sections like acknowledgements
• 9:37
  and data availability statements. Those boilerplate
• 9:40
  sections serve as a perfect statistical control
• 9:42
  group. The punctuation in those boilerplate sections
• 9:45
  barely changed at all. This proves the shift
• 9:49
  came from changes in actual writing style. It
• 9:51
  isolates the variable directly to the creative
• 9:54
  composition process. Scientists are actively
• 9:56
  changing how they formulate their complex arguments.
• 9:59
  We must be careful about what this data actually
• 10:02
  means, though. The researchers note this doesn't
• 10:05
  prove AI directly caused the increase. Right,
• 10:08
  an MDash cannot tell you if a specific paper
• 10:10
  used AI. You can't accuse a specific researcher
• 10:13
  of cheating just for using a dash. It's a population
• 10:16
  -level shift happening alongside LLM adoption.
• 10:19
  These are broad behavioral trends, not tools
• 10:22
  for individual detection. It's a statistical
• 10:23
  pattern across the entire scientific community.
• 10:26
  It's like tracking a new slang word spreading
• 10:28
  across a city. You can't predict exactly who
• 10:31
  will use the new dialect today, but you can clearly
• 10:34
  see the broader culture changing in real time.
• 10:36
  That linguistic analogy captures the phenomenon
• 10:38
  perfectly, actually. We're developing a hybrid
• 10:41
  human -AI dialect without even realizing it.
• 10:44
  AI models inherently favor a conversational appended
• 10:48
  sentence structure. And humans read that algorithmic
• 10:51
  output, absorb the rhythm, and subconsciously
• 10:54
  adopt it. So if we can't use this to catch individual
• 10:58
  researchers using AI, why does tracking this
• 11:01
  punctuation matter? It matters because it provides
• 11:03
  hard empirical evidence of cultural integration.
• 11:06
  It proves that AI is deeply weaving into our
• 11:09
  collective intellectual output. It makes the
• 11:12
  invisible cultural shift of AI collaboration
• 11:14
  totally visible. It takes a deeply philosophical
• 11:17
  concept and turns it into measurable data. Let's
• 11:20
  step back and look at the big picture here. AI
• 11:22
  is no longer just a tool we query for facts.
• 11:25
  No, it's an active, autonomous participant in
• 11:28
  our physical world. It has crossed the boundary
• 11:30
  from passive database to active agent. We see
• 11:32
  it discovering the art DNA system alongside human
• 11:35
  wet labs. We see it confidently running financial
• 11:37
  hedge funds and executing trades. We see it aggressively
• 11:40
  negotiating with airlines to book Delta flights.
• 11:43
  And it's subtly changing how... human scientists
• 11:46
  punctuate their very sentences. The spectrum
• 11:49
  of its impact is truly unprecedented in human
• 11:52
  history. We're watching a fundamental reorganization
• 11:55
  of human effort. We're redefining what it means
• 11:57
  to conduct complex scientific research. We're
• 12:00
  changing what it means to manage our mundane
• 12:02
  daily lives. Thank you for joining us on this
• 12:05
  deep dive. I want to leave you with one final
• 12:07
  thought to consider. If AI is already leaving
• 12:10
  measurable stylistic fingerprints on how we write
• 12:12
  today, what invisible fingerprints will it leave
• 12:15
  on how we actually think and solve problems 10
• 12:17
  years from now?