Latest / AI Fire Daily / 🎙️ EP 354: Figure Crosses 86K Weekly Active Users & Cognition Releases SWE-2 Coding Model

Transcript

🎙️ EP 354: Figure Crosses 86K Weekly Active Users & Cognition Releases SWE-2 Coding Model · AI Fire Daily

• 0:00
  Whoa. Beat. Imagine ingesting five years of human
• 0:04
  life every single day. Yeah. It is completely
• 0:07
  staggering to think about. Right. But I mean.
• 0:10
  That is exactly what robotics company figure
• 0:12
  is doing today. Welcome to the deep dive. We
• 0:14
  are so glad you are joining us. We have a big
• 0:17
  stack of AI updates to synthesize today. We really
• 0:19
  do. We want to understand the bigger picture
• 0:21
  for you. We are looking at how AI is rapidly
• 0:24
  scaling up everywhere. It is mapping the physical
• 0:28
  world with highly advanced robots. Right. And
• 0:30
  it is crossing major digital thresholds in our
• 0:33
  daily lives. Exactly. And that is triggering
• 0:36
  massive economic shifts across the globe. Finally,
• 0:40
  it is pushing the boundaries of complex coding
• 0:42
  capabilities. We are going to explore all of
• 0:45
  this together. We have some truly fascinating
• 0:47
  updates to unpack. The sheer pace of this technological
• 0:49
  evolution is just incredibly fast. Let us kick
• 0:52
  things off with the physical AI data explosion.
• 0:55
  To teach AI to navigate our physical world, we
• 0:58
  must first translate human life. into data figure
• 1:01
  is tackling this massive challenge head -on right
• 1:03
  now they certainly are they are building an enormous
• 1:06
  network to collect real -world data right figure
• 1:10
  just shared some wild new numbers recently their
• 1:12
  robot data network reached 86 ,000 weekly active
• 1:16
  users hmm let us pause and think about the scale
• 1:20
  of that that is a massive global workforce feeding
• 1:23
  the machine constantly it really is users upload
• 1:27
  robot training data directly through their platform.
• 1:30
  And they nearly doubled from 44 ,000 weekly users
• 1:34
  in no time. Wow. Yeah, that explosive growth
• 1:37
  happened in roughly two and a half weeks. That
• 1:39
  kind of rapid adoption is just hard to fathom.
• 1:41
  We are talking about an unprecedented crowdsourcing
• 1:44
  effort here. Absolutely. At launch, they already
• 1:46
  had 264 ,000 app downloads. Those downloads span
• 1:50
  across 108 different countries around the world.
• 1:52
  And people have uploaded more than 16 million
• 1:55
  videos so far. Let us just visualize that massive
• 1:57
  data pipeline for a second. The system is ingesting
• 2:00
  around 30 minutes of video constantly. It processes
• 2:03
  that much new footage every single second. And
• 2:06
  if you do the math on that ingestion rate, that
• 2:08
  equals roughly 4 .9 years of human activity per
• 2:11
  day. Almost five years of physical life digested
• 2:14
  every 24 hours. Beat. That is how you teach a
• 2:18
  machine to understand our reality. Yeah. And
• 2:21
  Figure is actually paying contributors for this
• 2:24
  data collection. They call these users creators,
• 2:26
  rewarding them for their time. Okay. They have
• 2:29
  already paid them $15 million for this footage.
• 2:31
  That is a highly significant financial investment
• 2:34
  in raw human data. Well, they actually plan to
• 2:37
  spend over $1 billion soon. A billion. Yeah.
• 2:40
  That billion goes toward data and compute over
• 2:43
  the next 12 months. Let us unpack how they actually
• 2:46
  use all this. these everyday videos when figure
• 2:48
  processes this footage it is not just watching
• 2:51
  passively no not at all it analyzes the complex
• 2:53
  kinematics of every single human movement exactly
• 2:56
  they are extracting incredible detail from all
• 2:59
  these mundane chores Per 1 ,000 hours of accepted
• 3:02
  data, they find 373 unique tasks. They also meticulously
• 3:07
  map 11 ,146 manipulated objects from that footage.
• 3:12
  And they carefully capture 116 different environments
• 3:15
  to teach their system. They are using this specific
• 3:18
  data to teach their helix system. It learns how
• 3:21
  everyday tasks are performed across different
• 3:23
  homes. It studies workplaces, varied objects,
• 3:26
  and complex spatial environments. They are essentially
• 3:29
  building a massive physics engine from scratch.
• 3:31
  Think of it exactly like stacking Lego blocks
• 3:34
  of data. One video of a person picking up a mug
• 3:36
  is a single block. Right. But you stack 16 million
• 3:40
  of those tiny blocks together. Suddenly, the
• 3:43
  AI builds a three -dimensional intuition of our
• 3:45
  physical world. That is a perfect analogy. Figure
• 3:48
  is effectively building an ImageNet scale data
• 3:51
  engine here. Yeah. They're using tens of thousands
• 3:54
  of humans for physical AI. They desperately want
• 3:57
  to solve robotics. Biggest data problem with
• 3:59
  consumers. Is this consumer scale data enough
• 4:02
  for reliable physical AI? Gives them a massive
• 4:05
  foundational understanding of human movement.
• 4:08
  But real world physics are incredibly unpredictable
• 4:10
  and often very messy. They will still need specific
• 4:14
  training for highly unusual complex edge cases.
• 4:17
  Quantity helps, but real world chaos still requires
• 4:19
  massive edge case training. Precisely. The foundational
• 4:23
  data is just the starting line today. So as robots
• 4:27
  learn to mimic our physical movements so closely,
• 4:30
  software AI is simultaneously crossing major
• 4:32
  thresholds in our digital lives. It is rapidly
• 4:35
  entering our cognitive workspaces in very profound
• 4:38
  ways. We are definitely seeing automation hit
• 4:41
  complex technical roles lately. An electronics
• 4:44
  engineer just reported a strange milestone this
• 4:46
  week. Oh! Yeah, they explicitly stated AI can
• 4:50
  now do 100 % of their job. Hmm. That has to be
• 4:55
  a very surreal milestone to experience. I still
• 4:57
  wrestle with bump drift myself on a daily basis.
• 5:00
  Beat. The technology can still be quite finicky
• 5:02
  to control sometimes. It definitely can be tricky
• 5:05
  to manage for most people. But for this specific
• 5:07
  engineer, the automation is completely total
• 5:09
  now. We might see many more roles reach this
• 5:12
  exact point soon. It makes you wonder about the
• 5:14
  shifting nature of technical work. We're also
• 5:16
  seeing AI mimic human creativity almost perfectly
• 5:19
  now. Oh, definitely. A Redditor recently asked
• 5:22
  ChatGPT to create a highly realistic image. It
• 5:25
  nearly pulled off the absolute perfect photographic
• 5:27
  illusion for them. It was incredibly close to
• 5:30
  a genuine real -world photograph. But one tiny
• 5:34
  major detail gave the entire image away completely.
• 5:38
  It shows how AI still struggles with certain
• 5:40
  physical logic internally. It does not truly
• 5:43
  understand how light and space interact naturally.
• 5:46
  Exactly. It is just predicting pixels without
• 5:48
  fully grasping the underlying physics. Because
• 5:51
  AI is crossing these human thresholds so rapidly
• 5:54
  lately, these models are also facing strict new
• 5:57
  societal guardrails right now. Claude just got
• 6:00
  much stricter about user age requirements recently.
• 6:02
  They are rolling out strict new age gating for
• 6:05
  all users. Claude can now actively flag suspected
• 6:08
  minors using the platform. It will firmly pause
• 6:11
  their access until they verify they are 18. If
• 6:13
  you are setting up Cloud, expect a strict age
• 6:15
  check. It is interesting to see them enforcing
• 6:18
  these societal boundaries so rigidly. It shows
• 6:21
  they take the immense power of these tools seriously.
• 6:23
  It really does. Right. Cloud also added a new
• 6:25
  technical capability for developers recently.
• 6:27
  Yeah. Cloud can now evaluate plugins natively
• 6:30
  within the system itself. Okay, wow. Yeah. You
• 6:33
  can use a specific command cloud plugin, Evil,
• 6:35
  to test them. That sounds like a highly useful
• 6:37
  tool for AI developers. It helps them debug the
• 6:41
  often mysterious black box of AI. It lets you
• 6:44
  directly compare results with and without a plugin.
• 6:47
  You can carefully follow their guide to test
• 6:49
  each step properly. Right. It brings a lot more
• 6:51
  transparency to AI tool development overall.
• 6:54
  Will our jobs shift entirely to just managing
• 6:57
  these AI systems? It looks like that is the clear,
• 7:00
  inevitable trajectory right now. We will spend
• 7:03
  less time doing the raw creation ourselves moving
• 7:05
  forward. We will spend way more time actively
• 7:08
  guiding and correcting the AI. Yes. We become
• 7:11
  editors and supervisors rather than the primary
• 7:13
  creator. Exactly. We direct the intelligence
• 7:15
  instead of generating the initial output. Because
• 7:18
  AI is crossing these major human thresholds so
• 7:21
  incredibly quickly, massive capital and institutional
• 7:25
  structures are shifting to fully support it.
• 7:27
  The financial landscape of AI is changing incredibly
• 7:29
  fast right now. Government contracts are a perfect
• 7:32
  example of this sudden institutional shift. U
• 7:35
  .S. agencies are officially losing their special
• 7:37
  $1 OpenAI deal. That was a very unique arrangement
• 7:41
  for those government agencies. It practically
• 7:44
  gave them unlimited access for almost nothing
• 7:47
  at all. It really was. But now OpenAI has definitively
• 7:51
  ended that annual deal. Agencies will now pay
• 7:54
  based on their actual daily chat GPT usage. Right.
• 7:58
  They will still get a pretty generous 50 % discount
• 8:01
  overall. But the absolute free ride is officially
• 8:04
  done for them today. This shows how astronomical
• 8:07
  the underlying compute costs have become recently.
• 8:10
  Massive funding rounds are also happening globally
• 8:13
  at this very moment. China Air's ZAI just raised
• 8:16
  $5 billion over in Hong Kong. That is a truly
• 8:19
  staggering amount of capital to secure so quickly.
• 8:22
  $2 billion came from new shares sold at a 10
• 8:24
  % discount. And $3 billion came from convertible
• 8:27
  bonds to round out the massive funding. Issuing
• 8:30
  convertible bonds is a very interesting financial
• 8:32
  strategy here. It helps investors hedge their
• 8:35
  massive risk in this volatile market. Exactly.
• 8:37
  It gives them a safety net while still chasing
• 8:39
  huge upsides. Enterprise deployment is also expanding
• 8:42
  rapidly across the board lately. GPT -Roslyn
• 8:44
  is now officially out of its initial research
• 8:47
  preview. It is available for eligible organizations
• 8:50
  to start using right away. Teams can start testing
• 8:53
  it through ChatGPT Enterprise right now. It is
• 8:56
  also fully available through Codex and the official
• 8:58
  API endpoints. We're also seeing highly empowered
• 9:02
  local AI tools popping up everywhere. Resurf
• 9:05
  is a fascinating new tool that helps organize
• 9:07
  your context. It stores your links, images, and
• 9:10
  PDFs in one private library. Right. And it gives
• 9:13
  your AI secure access through MCP and CLI. For
• 9:17
  those wondering, MCP is a standard way for AI
• 9:19
  to securely access your private files. Be it,
• 9:23
  it keeps your local data highly organized and
• 9:25
  ready for AI. Another highly useful new tool
• 9:28
  is called Perplexity Hybrid Compute. This specific
• 9:31
  tool splits AI tasks between the cloud and your
• 9:34
  Mac. It keeps your sensitive private files completely
• 9:37
  secure on your local machine. That is huge. It
• 9:40
  is, but it still leverages powerful cloud AI
• 9:42
  for the heavy processing tasks. ScreenCursor
• 9:45
  is another fascinating new tool built for digital
• 9:47
  creators. It turns your raw screen recordings
• 9:49
  into highly polished finished videos. It automatically
• 9:52
  adds smooth zooms for every single click and
• 9:55
  drag. It even zooms in perfectly for your specific
• 9:59
  keystrokes automatically. Wait, if we have tools
• 10:01
  like Perplexity running hybrid compute locally,
• 10:04
  it saves a massive amount of server costs for
• 10:07
  everyone involved. Sure. Are these massive funding
• 10:09
  rounds actually sustainable in the long term?
• 10:12
  Well... The raw compute costs for these frontier
• 10:15
  models are absolutely astronomical today. They
• 10:18
  desperately need these billions just to keep
• 10:21
  the servers running constantly. The enterprise
• 10:23
  tools have to generate massive ongoing efficiency
• 10:26
  to justify it. Only if the efficiency gains outweigh
• 10:29
  the staggering compute costs over time. Exactly.
• 10:32
  The underlying math absolutely has to make sense
• 10:34
  in the long run. We are going to take a quick
• 10:36
  break to hear from our sponsors. We will be right
• 10:38
  back. Welcome back to the Deep Dive. We are exploring
• 10:43
  how massive investments are driving aggressive
• 10:45
  technical breakthroughs today. To justify those
• 10:48
  billions in massive funding and new government
• 10:50
  contracts, AI developers are desperately chasing
• 10:52
  much cheaper and more efficient compute. They
• 10:55
  want flawless agentic coding and ultimate self
• 10:58
  -improving digital models. Wild rumors are spreading
• 11:00
  like absolute wildfire about Google DeepMind
• 11:03
  recently. People say they have finally reached
• 11:05
  a massive breakthrough with RSI internally. RSI
• 11:09
  is AI that rapidly improves its own intelligence
• 11:12
  without human help. Beat. That is essentially
• 11:15
  the holy grail of artificial intelligence research
• 11:17
  right there. DeepMind, Demis Hassabis, and Sergey
• 11:20
  Brin are suddenly circling around RSI developments.
• 11:23
  Right. The entire tech industry is watching these
• 11:26
  subtle movements very closely right now. The
• 11:28
  implications of recursive self -improvement are
• 11:30
  just incredibly profound to consider. Absolutely.
• 11:33
  Meanwhile, Cognition released their SW2 model
• 11:36
  back on September 10th. It is their most highly
• 11:39
  capable coding model released yet to date. They
• 11:41
  rolled it into Devon Desktop and CLI right out
• 11:44
  of the gate. It is also coming to Devon Web and
• 11:46
  Fusion very soon. SW2 is post -trained directly
• 11:49
  from Moonshot AIR's Kimi K3 model. That is a
• 11:52
  massive 2 .8 trillion parameter model they actively
• 11:55
  use as a foundation. Parameters are the billions
• 11:58
  of digital brain connections shaping how AI thinks.
• 12:02
  So let us carefully break down the actual benchmark
• 12:04
  performance for SW2. Okay. On the Frontier Code
• 12:07
  1 .1 main benchmark, SW2 scored a solid 50 .0%.
• 12:13
  That is directly compared with 50 .9 % for the
• 12:15
  Fable 5 .1 model. And GPT -6 Astra scored a slightly
• 12:19
  higher 53 .3 % on that same test. So it is basically
• 12:23
  running neck and neck with the absolute top Frontier
• 12:25
  models. But Cognition says SW2 reaches that performance
• 12:29
  at a 64 % lower cost. That massive price drop
• 12:32
  is the real main story for SWE2 today. Making
• 12:36
  powerful AI coding dramatically cheaper is a
• 12:38
  huge deal for the industry. Driving down the
• 12:40
  cost of reasoning changes the entire economic
• 12:42
  equation entirely. On the DeepSW 1 .1 test, SW2
• 12:46
  scored an impressive 73 .0%. That is actually
• 12:50
  slightly above Fable 5 .1 Tritus' score of 67
• 12:53
  .4%. It also scored an amazing 92 .6 Asterisk's
• 12:56
  high score of 74 .1 % there. It also scored an
• 12:59
  amazing 92 .8 % on Terminal Bench 2 .1 recently.
• 13:02
  That puts it significantly ahead of both Fable
• 13:04
  5 .1 and GPT -6 Astra. But there's one very important
• 13:08
  massive caveat we have to mention here. Right.
• 13:11
  SWE2 does not blindly match Frontier models everywhere
• 13:15
  they are rigorously tested. This is where the
• 13:18
  technical nuance becomes incredibly important
• 13:19
  to understand. On the much newer Terminal Bench
• 13:22
  4, it scored only 27 .3%. Yeah. That falls incredibly
• 13:27
  far behind Fable 5 .1, sitting at 55 .8%. And
• 13:32
  GPT -6 Astra scored 57 .9 % on that demanding
• 13:36
  newer test. It still falls remarkably far behind
• 13:39
  on some harder agentic coding tests. If these
• 13:42
  models are getting cheaper but failing hard on
• 13:44
  newer tests, are we hitting a temporary ceiling
• 13:47
  in complex coding logic? Why does SW2 fail so
• 13:50
  hard on Terminal Bench 4? The newer tests require
• 13:53
  much more complex multi -step logical reasoning
• 13:55
  capabilities. Agentic workflows force the AI
• 13:58
  to plan, execute, fail, and adapt autonomously.
• 14:00
  The model was likely optimized heavily just for
• 14:03
  the older standard benchmarks. It severely struggles
• 14:06
  when faced with entirely new, highly complex
• 14:08
  autonomous paradigms. It optimizes for older
• 14:10
  tests, struggling with newer, highly complex
• 14:12
  workflows. Exactly. It shows we have not quite
• 14:14
  hit a permanent hard ceiling yet. Right. But
• 14:17
  we are clearly seeing the limits of this specific
• 14:19
  optimization architecture. Two sec silence. We
• 14:23
  have covered a tremendous amount of complex ground
• 14:25
  in this deep dive today. We really have. If we
• 14:28
  carefully synthesize this overarching... theme
• 14:30
  today. We are watching a simultaneous aggressive
• 14:32
  push on two massive automated fronts. On one
• 14:37
  side, we are brute forcing data collection in
• 14:39
  the real physical world. Figures capturing literal
• 14:42
  years of human life to train physical robots.
• 14:44
  On the other side, we are relentlessly driving
• 14:47
  down digital computing costs. Models like SWE2
• 14:50
  are making highly capable digital reasoning.
• 14:52
  Much cheaper. Yeah. It is a powerful dual assault
• 14:55
  on the current boundaries of automation. Thank
• 14:57
  you so much for taking this fascinating deep
• 14:59
  dive with us. We deeply appreciate you exploring
• 15:02
  these complex topics alongside us today. We want
• 15:05
  to leave you with one final, highly provocative
• 15:07
  thought. Beat. If an AI can perfectly mimic human
• 15:11
  motion through figure as massive data. beat and
• 15:14
  if it perfects its own digital code entirely
• 15:16
  via rsi beat where does that leave the human
• 15:19
  in the loop 10 years from now beat are we simply
• 15:21
  building new tools for our own everyday use beat
• 15:24
  or are we just the biological bootloader for
• 15:26
  a new digital ecosystem two sec silence