Latest / Journey to the Fringe / Fringy Mini: Braaaaaaaaaa AI iiiiiiiiiinnnnnnnnnssssssssssss

Transcript

Fringy Mini: Braaaaaaaaaa AI iiiiiiiiiinnnnnnnnnssssssssssss · Journey to the Fringe

• 0:00
  Okay, Chelsea, I think I found some AI -like
• 0:02
  news that is actually very good coming out right
• 0:06
  around Halloween. It's good news? It is not good
• 0:10
  news. It is good for this time of year. Okay.
• 0:13
  As in, like, it makes sense to be a Halloween
• 0:16
  article. Okay. So this article comes from The
• 0:19
  Straight Times, spelt S -T -R -A -I -T -S, or
• 0:23
  The Straight Times, written by Li Liying on...
• 0:28
  August 16th, 2026. Article title. Forget silicon
• 0:32
  chip servers. Singapore's newest data centers
• 0:34
  need to be fed. Singapore. Tucked inside the
• 0:39
  National University of Singapore Center for Life
• 0:41
  Science is a data center with a pulse. Unlike
• 0:44
  traditional server farms running on silicon chips,
• 0:47
  this facility processes data using living human
• 0:51
  brain cells. Every three days, a laboratory technician
• 0:55
  feed the brain cells a cocktail of sugar, micronutrients,
• 0:59
  and pH buffers. A gas mixer pumps in carbon dioxide,
• 1:03
  oxygen, and nitrogen, providing a dedicated life
• 1:05
  support system. The brainchild of Australian
• 1:08
  biotech startup Cortical Labs has joined forces
• 1:10
  with NUS, University of Singapore, and data center
• 1:14
  operator Day One to bring the concept to Singapore
• 1:17
  for joint internal research and development.
• 1:21
  Their tie -up was first announced in March. It
• 1:23
  has commercialized a similar biological data
• 1:25
  center in Melbourne, its first in the world.
• 1:28
  Chonghong Wen, a Cortical Labs founder and chief
• 1:31
  executive, said that the biological data centers
• 1:34
  are better suited for applications where data
• 1:36
  sets may be limited and where conditions are
• 1:38
  highly unpredictable. He cited the use of humanoid
• 1:41
  robots in public spaces, homes, offices, or industries
• 1:44
  where conditions are ever -changing. there is
• 1:47
  not enough data in the world to train humanoid
• 1:50
  robots to do everything that we need the robots
• 1:52
  to do using traditional servers end quote said
• 1:55
  Chong during a facility tour on August 7th for
• 1:59
  instance teaching a robot to navigate from one
• 2:01
  point to another may require vast amounts of
• 2:03
  training data that mirrors the real world environment
• 2:05
  biological data centers are also better suited
• 2:08
  for cyber security applications and spotting
• 2:10
  anomalies without a huge training data set he
• 2:13
  claimed quote so the solution is to ensure limited
• 2:16
  data can be used efficiently end quote he added
• 2:19
  this is similar to how humans can learn from
• 2:22
  just a handful of examples and apply what they
• 2:24
  have learned even when the circumstances change
• 2:26
  singapore's first biological data center comprises
• 2:29
  20 units of biological computers also called
• 2:33
  cl1s each cl1 unit houses at least 200 000 lab
• 2:38
  -grown neurons sitting on an electrode fitted
• 2:41
  silicon chip These living neurons, derived from
• 2:45
  blood cells programmed into stem cells, exchange
• 2:47
  electrical signals with a computer where the
• 2:50
  neural activity is translated into raw computing
• 2:52
  power. There are plans to house up to 1 ,000
• 2:55
  CL1s here, following regulatory approval as well
• 2:58
  as energy efficiency and safety tests. Its Melbourne
• 3:00
  facility currently operates 120 CL1s, with around
• 3:04
  20 paying customers including corporate, research,
• 3:07
  arms, universities, experimenting with robotics,
• 3:09
  and gaming. Customers pay US $2 ,200 a month
• 3:14
  to access the computing resources of a CL1. This
• 3:17
  is about half of the $4 ,300 monthly fee charged
• 3:20
  by major cloud platforms for renting a high -end
• 3:23
  AI chip for training and running AI models. But
• 3:26
  Chong recognizes that traditional silicon chips
• 3:28
  are far superior for fast, precise, and repeatable
• 3:31
  calculations that underpin large language models
• 3:33
  such as ChatGPT. One huge plus for biological
• 3:37
  data centers is their low energy footprint, as
• 3:40
  they are not heat generating servers that require
• 3:42
  cooling, said Chong. In Singapore, the energy
• 3:46
  and water demands for traditional data centres
• 3:47
  were so intense that the country imposed a temporary
• 3:50
  pause on their construction of new facilities
• 3:52
  in 2019 to plan for greener digital growth. In
• 3:55
  2019, data centres accounted for 5 .3 % of the
• 3:58
  country's electricity consumption, a figure that
• 4:01
  rose 7 % in 2020 as the COVID -19 pandemic accelerated
• 4:04
  digitalization. Brain cells, by contrast, use
• 4:07
  very little energy and do not need cooling. At
• 4:10
  30 watts per minute, each CL1 uses less power
• 4:13
  than a handheld calculator. This is even with
• 4:16
  its life support systems factored in, said Chong.
• 4:19
  Comparatively, an AI silicon chip such as an
• 4:22
  NVIDIA's H100SXM can consume up to 700 watts
• 4:27
  when running demanding computing tasks. A typical
• 4:30
  data server containing eight H100 chips can consume
• 4:33
  up to 10 ,200 watts, including power consumed
• 4:37
  by supporting hardware. Quote, Singapore is a
• 4:40
  major fiber connectivity hub in Asia Pacific
• 4:43
  region. So it is naturally a very important place
• 4:46
  for data centers because they need fast connections
• 4:48
  to move vast amounts of data. But Singapore has
• 4:51
  constraints in terms of electricity and water,
• 4:54
  and our technology can help to overcome that,
• 4:56
  end quote, said Chong. The facility at NUS will
• 4:59
  help cortical labs work out the manpower requirements
• 5:02
  and skill sets needed for commercial uses here.
• 5:05
  Quote, by identifying precisely which neurons
• 5:08
  and support cells work best together, we can
• 5:11
  build a scientific case for manufacturing these
• 5:14
  cell types at a scale. Purpose built for data
• 5:17
  center operations. End quote. Said NUS Yonglu
• 5:20
  Lin School of Medicine Neuroscience Professor
• 5:23
  Ricky Patani, who oversees the Biological Data
• 5:25
  Center. That's the end of the article. Well,
• 5:28
  I guess that is kind of good news. I mean, yes,
• 5:33
  there's less electricity being used, but there's
• 5:36
  zombie data centers. Like, that is disturbing.
• 5:39
  I mean... And I just don't know. I don't know.
• 5:43
  It's like... I don't know. It is a Halloween
• 5:45
  article right there. I don't know that there's
• 5:48
  any upsides. really to it. Yeah, I mean, this
• 5:52
  is like Soylent Green, where they're like, oh
• 5:54
  no, they're turning people into food for us to
• 5:56
  eat. Which, sorry, spoiler alert for a movie
• 5:59
  that came out in the 70s, I believe. I can't
• 6:00
  believe you drew in that for me. And a book that
• 6:02
  came out earlier than that. But yeah, it's not
• 6:05
  just, well, we don't need the food yet, but we're
• 6:07
  gonna maybe use your body parts for brain power
• 6:10
  at a certain point. The implications for what
• 6:13
  kind of ghost that would create? Yeah. That is
• 6:16
  a real question, is are these data centers haunted?
• 6:19
  yet not yet probably they might be someone might
• 6:24
  be pissed off enough about a data center that
• 6:27
  they want to haunt it yeah you just wait till
• 6:30
  we figure out that like all the viruses on the
• 6:32
  internet are just these things as ghosts from
• 6:37
  the future holy shit yeah yeah figured it out
• 6:41
  i guess we gotta ask people who know nothing
• 6:43
  about how computers work I'm pretty sure I do
• 6:48
  know. Did you just hear how we put that together?
• 6:50
  Yeah. I mean, it's pretty straightforward compared
• 6:52
  to how computers are supposed to work. Yeah,
• 6:55
  that's incredible. And you guys got 48 hours.
• 6:58
  Prove us wrong on that. I don't know. Don't.
• 7:00
  Yeah, I don't think you can. I don't think you
• 7:02
  can either. Bye. Bye.