Latest / AI Fire Daily / 🎙️ EP 338: GEN-1.5 Physical Prompting Robotics & Claude Autonomously Designs Functional Proteins

Transcript

🎙️ EP 338: GEN-1.5 Physical Prompting Robotics & Claude Autonomously Designs Functional Proteins · AI Fire Daily

• 0:00
  Imagine showing a robot what to do for just three
• 0:02
  seconds. Beat. And then, well, he just does it.
• 0:06
  No coding. Just watching. Yeah, it's wild. Let's
• 0:10
  jump right in. Welcome to today's Deep Dive.
• 0:13
  We're taking a stack of new sources and extracting
• 0:16
  the most important signals from the noise, just
• 0:18
  for you. Exactly. Lots to cover today. Today,
• 0:21
  we are exploring how AI is stepping out of the
• 0:23
  chat box and, you know, into the physical world.
• 0:26
  We'll unpack a new robotics model called GEN
• 0:29
  1 .5. Right. And we'll sift through today's rapid
• 0:32
  -fire AI updates, including a bizarre data center
• 0:37
  cooling idea. Yeah, that one is weird. And we'll
• 0:39
  look at how Claude is autonomously running complex
• 0:41
  biological labs, I'll be honest. Beep! I still
• 0:44
  wrestle with getting my smart thermostat to listen
• 0:46
  to me, so handing a robot a tool feels a bit
• 0:48
  surreal. I mean, the pace is definitely disorienting,
• 0:51
  but the mechanics behind this shift are like...
• 0:54
  Absolutely fascinating. Okay, let's unpack this.
• 0:56
  Generalists just introduced GEN 1 .5, a robot
• 0:59
  model that learns a physical task from a single
• 1:01
  3 to 12 second demonstration. Right, just a tiny
• 1:04
  video clip. Yeah. It watches the human perform
• 1:08
  the task. And then it attempts the task almost
• 1:11
  immediately. Over 10 simple tasks, one demo gave
• 1:14
  it a 59 % average success rate. Which is actually
• 1:18
  a huge deal. They call it physical prompting.
• 1:20
  Instead of retraining the robot from scratch,
• 1:22
  you just drop a short video into its 30 -second
• 1:25
  context window. Let me define that real quick.
• 1:27
  A context window is basically the AI's short
• 1:29
  -term memory for a task. Exactly. 12 words. Spot
• 1:32
  on. So instead of, you know, writing complex
• 1:35
  code or mapping strict spatial coordinates for
• 1:37
  every single motor. It just takes forever. Right.
• 1:40
  It just holds that visual in its memory. And
• 1:42
  it infers depth and physics from a flat 2D video.
• 1:45
  But here's where it gets nuts. Okay. With just
• 1:48
  five minutes of extra data and 10 weight updates,
• 1:51
  that success rate jumped to 83%. Wow. Yeah. It
• 1:55
  imitates human hands. It transfers skills from
• 1:58
  simulation to reality seamlessly. And it can
• 2:01
  even combine multiple prompts. Whoa. Imagine
• 2:04
  it just picking up a tool it's never seen during
• 2:06
  training and improvising on the fly. I know.
• 2:08
  It's like bridging the sim to real gap instantly.
• 2:10
  The sim to real gap. That's the difference between
• 2:14
  a perfect digital simulation and the messy real
• 2:17
  world. Precisely. A simulation has perfect physics.
• 2:21
  But the real world has friction and weird lighting.
• 2:24
  GEN 1 .5 handles that transition beautifully.
• 2:27
  That is wild. But wait, what is the actual difference
• 2:31
  between its two learning modes here? The source
• 2:33
  mentions one -shot versus few -shot learning.
• 2:36
  Right, good question. So the one -shot demo is
• 2:39
  in -context learning. It's learning on the fly
• 2:41
  without changing its core programming. Like holding
• 2:44
  a temporary instruction. Yeah, exactly. But the
• 2:47
  few -shot version actually updates the model's
• 2:49
  weights. That makes it closer to continual learning,
• 2:52
  permanently updating its brain with new skills
• 2:54
  like stacking Lego blocks of data. Got it. So
• 2:57
  it either holds the memory temporarily or permanently
• 3:00
  rewires itself. Spot on. And that permanent rewiring
• 3:03
  is, you know, the holy grail for robotics. So
• 3:06
  robots are learning physical tasks, but our everyday
• 3:09
  software agents are also taking more direct action
• 3:12
  on our behalf. Oh, definitely. We're giving them
• 3:13
  broader goals now. Here's where it gets really
• 3:16
  interesting. Chat GPT can now draft replies and
• 3:20
  literally text people for you through an Apple
• 3:23
  Messages plug -in. Right. And Meta AI launched
• 3:26
  a Mac app with system -wide dictation and screen
• 3:29
  awareness. Screen awareness, meaning it, like,
• 3:31
  sees everything on my monitor. Basically, yeah.
• 3:33
  Yeah. Personal custom GPTs are kind of fading
• 3:36
  out while these workspace encoding agents take
• 3:38
  over. They can see your open spreadsheets, your
• 3:40
  emails, everything. That feels incredibly intrusive.
• 3:43
  It is a lot of access. But the market is moving
• 3:45
  fast. Replit launched a free mode using GPT 5
• 3:50
  .6 Luna. Which is huge for developers. Totally.
• 3:53
  But it's not all flawless. Grok Lite had a rare
• 3:57
  glitch recently where it was sending absolute
• 3:59
  gibberish to people. See, that is exactly my
• 4:01
  fear. Yeah, the company called it a temporary
• 4:03
  glitch, but it highlights the tension right now.
• 4:05
  To stop those hallucinations, you need astronomically
• 4:09
  better training data. Which explains the massive
• 4:11
  data boom. Exactly. To power all this, MicroOne's
• 4:15
  AI training data business jumped from a 100 million
• 4:17
  to 500 million gross annual run rate in just
• 4:21
  eight months. A five -fold increase in eight
• 4:24
  months. The demand is massive. And all that processing
• 4:28
  runs incredibly hot. Which leads to... Liquid
• 4:33
  death joking about cooling data centers with
• 4:35
  pee. Which experts say isn't fully ridiculous.
• 4:38
  Right. I mean, scientifically speaking, liquid
• 4:41
  cooling is vastly superior to air. Water absorbs
• 4:44
  heat exponentially faster. So as the chips get
• 4:48
  denser, traditional fans just can't keep up.
• 4:50
  The servers would literally melt. Exactly. It's
• 4:53
  absurd marketing, but the underlying physics
• 4:56
  are real. But let me circle back to the texting
• 4:58
  thing. With ChatGPT directly texting people for
• 5:01
  you, aren't we handing over a bit too much autonomy?
• 5:03
  Yes, which is exactly why the source heavily
• 5:06
  warned you to keep human approval turned on before
• 5:08
  anything actually gets sent as you. Right. Keep
• 5:11
  human approval on so it doesn't text your boss
• 5:12
  absolute gibberish. The human in the loop is
• 5:14
  still your best defense, sponsor. We just talked
• 5:17
  about handing over autonomy for sending texts,
• 5:19
  but Anthropic is handing over autonomy for scientific
• 5:22
  research. It's a massive escalation. It really
• 5:25
  is. Anthropic tested Mythos Preview and Clawed
• 5:29
  Opus 4 .8 on autonomous protein design campaigns.
• 5:32
  It produced working molecules for 14 out of 15
• 5:35
  targets. The numbers are staggering. The binding
• 5:39
  success rates hit 22 to 35 percent. Anthropic
• 5:43
  notes the industry baseline is usually just 10
• 5:45
  to 15 percent. So it's literally doubling the
• 5:48
  industry standard. Yeah. And protein design is
• 5:51
  basically solving 3D biological puzzles. It's
• 5:53
  the foundation of drug discovery. But Claude
• 5:55
  didn't actually mix the chemicals in a physical
• 5:57
  lab, did it? No, no. Twist Bioscience and Adaptive
• 6:01
  Bio handled the actual real -world lab synthesis.
• 6:04
  But Claude did the heavy lifting. Generating
• 6:07
  the chemical candidates. Right. Generating candidates,
• 6:09
  evaluating results, and revising ideas with minimal
• 6:12
  human direction. He was acting like the lead
• 6:14
  architect. Exactly. And the speed breakthrough
• 6:16
  is incredible. In another experiment, Claude
• 6:19
  Opus 5 analyzed raw instrument files and estimated
• 6:22
  a sample's purity at 96 .4%. And parsing those
• 6:25
  complex data outputs usually takes forever. It
• 6:29
  did this in 19 minutes. The lab's normal report
• 6:31
  takes four days. 19 minutes. Beat. Is Claude
• 6:35
  just acting as a really fast calculator here,
• 6:38
  or is it actually leading the scientific process?
• 6:41
  What's fascinating here is it's moving beyond
• 6:43
  just answering biology questions. It's autonomously
• 6:47
  running multi -step scientific workflows and
• 6:50
  leading the research loop. So it is practically
• 6:52
  running the entire lab loop on its own now. Yes,
• 6:55
  effectively steering the ship. Faster iteration
• 6:58
  means faster breakthroughs. Let's step back.
• 7:01
  We covered a lot today. We've seen AI learn physical
• 7:04
  tasks in a three -second window, embed itself
• 7:07
  into our daily text messages, and cut four -day
• 7:10
  lab analyses down to 19 minutes. If we connect
• 7:13
  this to the bigger picture, AI is shifting. It's
• 7:15
  no longer just a passive chat box you query.
• 7:17
  It is an active agent observing, learning, and
• 7:20
  executing in the real world. So what does this
• 7:22
  all mean? Think about your own day. If AI can
• 7:25
  learn a task from a single three -second video,
• 7:27
  what is the very first task you would demonstrate
• 7:29
  for it? And take it a step further. If Claude
• 7:32
  can design proteins twice as well as industry
• 7:34
  baselines, what happens when Gene 1 .5's improvising
• 7:38
  robotic hands are paired with Claude's scientific
• 7:40
  brain in a fully automated lab? Two sec silence.
• 7:44
  Something to think about. Thanks for joining
• 7:45
  us on this deep dive. Stay curious. Out to your
• 7:47
  own music.