Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 30, 2026

Transcript

AGI Dreams Podcast – June 30, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  An
• 0:00
  open-source
• 0:01
  project
• 0:01
  called
• 0:01
  Bingo
• 0:02
  has
• 0:02
  quietly
• 0:02
  climbed
• 0:03
  to
• 0:03
  trending
• 0:03
  status
• 0:03
  on
• 0:03
  GitHub,
• 0:04
  and
• 0:04
  it
• 0:04
  deserves
• 0:05
  a
• 0:05
  close
• 0:05
  look
• 0:05
  —
• 0:06
  not
• 0:06
  because
• 0:06
  it's
• 0:06
  novel
• 0:07
  in
• 0:07
  concept,
• 0:08
  but
• 0:08
  because
• 0:08
  of
• 0:08
  what
• 0:08
  it
• 0:08
  automates.
• 0:09
  Bingo
• 0:10
  is
• 0:10
  an
• 0:10
  AI-powered
• 0:11
  red
• 0:11
  team
• 0:11
  terminal
• 0:12
  that
• 0:12
  wraps
• 0:12
  multiple
• 0:12
  LLM
• 0:13
  backends
• 0:13
  (DeepSeek,
• 0:14
  Claude,
• 0:15
  GPT,
• 0:15
  GLM,
• 0:16
  Qwen,
• 0:16
  and
• 0:16
  local
• 0:17
  Ollama
• 0:17
  models)
• 0:17
  around
• 0:18
  a
• 0:18
  full
• 0:18
  offensive
• 0:18
  security
• 0:19
  pipeline:
• 0:20
  reconnaissance,
• 0:20
  WAF
• 0:21
  bypass,
• 0:21
  SQL
• 0:22
  injection
• 0:22
  chains
• 0:23
  from
• 0:23
  error-based
• 0:23
  through
• 0:23
  time-based
• 0:24
  blind,
• 0:24
  SSRF,
• 0:25
  HTTP
• 0:25
  request
• 0:26
  smuggling,
• 0:26
  JWT/OAuth
• 0:29
  abuse,
• 0:29
  DApp/Web3
• 0:31
  audits,
• 0:31
  and
• 0:31
  even
• 0:32
  mobile
• 0:32
  APK/IPA
• 0:33
  reverse
• 0:34
  engineering.
• 0:35
  The
• 0:35
  user
• 0:35
  types
• 0:35
  a
• 0:35
  target
• 0:36
  URL
• 0:36
  and
• 0:36
  a
• 0:36
  natural-language
• 0:37
  task
• 0:37
  description;
• 0:38
  the
• 0:38
  AI
• 0:38
  decides
• 0:38
  the
• 0:38
  attack
• 0:39
  strategy,
• 0:39
  pivots
• 0:40
  when
• 0:40
  blocked,
• 0:40
  and
• 0:40
  produces
• 0:41
  a
• 0:41
  Markdown
• 0:41
  report
• 0:41
  with
• 0:41
  CVSS
• 0:42
  scores.
• 0:43
  What
• 0:43
  makes
• 0:43
  Bingo
• 0:44
  more
• 0:44
  than
• 0:44
  a
• 0:44
  fancy
• 0:44
  wrapper
• 0:45
  is
• 0:45
  its
• 0:45
  four-layer
• 0:46
  anti-hallucination
• 0:47
  guard:
• 0:48
  every
• 0:48
  finding
• 0:48
  must
• 0:48
  be
• 0:48
  backed
• 0:49
  by
• 0:49
  a
• 0:49
  real
• 0:49
  HTTP
• 0:50
  response,
• 0:51
  and
• 0:51
  the
• 0:51
  tool
• 0:51
  labels
• 0:51
  evidence
• 0:52
  as
• 0:52
  verified,
• 0:52
  likely,
• 0:53
  or
• 0:53
  inferred.
• 0:54
  It
• 0:54
  auto-integrates
• 0:55
  with
• 0:55
  nmap
• 0:56
  and
• 0:56
  sqlmap
• 0:56
  when
• 0:56
  present,
• 0:57
  rotates
• 0:57
  proxies
• 0:58
  on
• 0:58
  WAF
• 0:58
  bans,
• 0:59
  and
• 0:59
  remembers
• 0:59
  findings
• 1:00
  across
• 1:00
  sessions.
• 1:01
  A
• 1:01
  headless
• 1:01
  CI/CD
• 1:02
  mode
• 1:03
  (--silent)
• 1:04
  outputs
• 1:04
  structured
• 1:04
  JSON,
• 1:05
  making
• 1:05
  it
• 1:05
  pipeline-friendly
• 1:06
  for
• 1:06
  continuous
• 1:07
  security
• 1:07
  testing.
• 1:08
  At
• 1:08
  REcon
• 1:09
  in
• 1:09
  Montreal
• 1:09
  —
• 1:10
  one
• 1:10
  of
• 1:10
  the
• 1:10
  few
• 1:10
  conferences
• 1:11
  purpose-built
• 1:11
  for
• 1:11
  reverse
• 1:12
  engineers
• 1:12
  and
• 1:12
  security
• 1:13
  researchers
• 1:13
  —
• 1:14
  a
• 1:14
  panel
• 1:14
  organized
• 1:15
  by
• 1:15
  Gadi
• 1:15
  Evron
• 1:16
  tackled
• 1:16
  whether
• 1:16
  AI
• 1:17
  will
• 1:17
  eliminate
• 1:17
  security
• 1:18
  research
• 1:18
  jobs
• 1:18
  entirely.
• 1:20
  The
• 1:20
  consensus:
• 1:21
  tools
• 1:21
  will
• 1:21
  change,
• 1:22
  but
• 1:22
  the
• 1:23
  adversarial
• 1:23
  mindset
• 1:24
  is
• 1:24
  not
• 1:24
  automatable.
• 1:25
  One
• 1:26
  panelist
• 1:26
  reportedly
• 1:27
  found
• 1:27
  a
• 1:27
  live
• 1:27
  example
• 1:28
  during
• 1:28
  the
• 1:28
  session
• 1:28
  where
• 1:29
  AI-assisted
• 1:29
  research
• 1:30
  simply
• 1:30
  was
• 1:31
  not
• 1:31
  there
• 1:31
  yet
• 1:31
  —
• 1:32
  a
• 1:32
  useful
• 1:33
  reminder
• 1:33
  that
• 1:33
  tools
• 1:33
  like
• 1:34
  Bingo
• 1:34
  automate
• 1:34
  known
• 1:35
  attack
• 1:35
  patterns,
• 1:36
  not
• 1:36
  novel
• 1:36
  zero-day
• 1:37
  discovery.
• 1:38
  On
• 1:38
  the
• 1:38
  defensive
• 1:38
  side,
• 1:39
  a
• 1:39
  practical
• 1:40
  guide
• 1:40
  to
• 1:40
  sandboxing
• 1:41
  X11
• 1:41
  GUI
• 1:41
  applications
• 1:42
  with
• 1:42
  LXC
• 1:43
  containers
• 1:43
  reminds
• 1:44
  us
• 1:44
  that
• 1:44
  not
• 1:44
  every
• 1:44
  security
• 1:45
  improvement
• 1:45
  requires
• 1:46
  a
• 1:46
  neural
• 1:46
  network.
• 1:47
  The
• 1:47
  walkthrough
• 1:48
  runs
• 1:48
  a
• 1:48
  browser
• 1:48
  in
• 1:48
  an
• 1:48
  unprivileged
• 1:49
  LXC
• 1:49
  container
• 1:50
  with
• 1:50
  UID/GID
• 1:51
  mapping
• 1:52
  into
• 1:52
  an
• 1:52
  unused
• 1:52
  host
• 1:53
  range,
• 1:53
  X11
• 1:54
  socket
• 1:54
  forwarding
• 1:55
  via
• 1:55
  a
• 1:55
  wildcard
• 1:55
  Xauthority
• 1:56
  entry,
• 1:56
  and
• 1:56
  optional
• 1:57
  GPU
• 1:57
  passthrough.
• 1:58
  If
• 1:58
  the
• 1:59
  browser
• 1:59
  is
• 1:59
  compromised,
• 2:00
  the
• 2:00
  blast
• 2:00
  radius
• 2:01
  is
• 2:01
  a
• 2:01
  container
• 2:01
  whose
• 2:02
  UIDs
• 2:02
  map
• 2:02
  to
• 2:02
  nothing
• 2:03
  useful
• 2:03
  on
• 2:03
  the
• 2:03
  host.
• 2:04
  Every
• 2:05
  forwarded
• 2:05
  channel
• 2:05
  is
• 2:06
  a
• 2:06
  deliberate
• 2:06
  hole
• 2:06
  in
• 2:06
  the
• 2:07
  wall
• 2:07
  —
• 2:07
  forward
• 2:08
  what
• 2:08
  the
• 2:08
  application
• 2:08
  needs,
• 2:09
  leave
• 2:09
  the
• 2:09
  rest
• 2:10
  out.
• 2:13
  DeepSeek
• 2:13
  has
• 2:13
  confirmed
• 2:14
  that
• 2:14
  the
• 2:14
  official
• 2:14
  V4
• 2:15
  release
• 2:15
  lands
• 2:15
  in
• 2:15
  mid-July,
• 2:16
  and
• 2:16
  the
• 2:16
  community
• 2:17
  is
• 2:17
  already
• 2:17
  running
• 2:17
  the
• 2:17
  preview
• 2:18
  checkpoint
• 2:18
  through
• 2:18
  llama.cpp
• 2:20
  after
• 2:20
  support
• 2:20
  was
• 2:20
  merged.
• 2:21
  The
• 2:21
  preview
• 2:22
  drew
• 2:22
  mixed
• 2:22
  reactions
• 2:23
  —
• 2:23
  commenters
• 2:24
  noted
• 2:24
  it
• 2:24
  needs
• 2:24
  to
• 2:24
  compete
• 2:25
  with
• 2:25
  GLM
• 2:25
  5.2's
• 2:26
  750B-parameter
• 2:27
  performance
• 2:28
  or
• 2:28
  face
• 2:28
  an
• 2:28
  awkward
• 2:28
  comparison,
• 2:29
  and
• 2:30
  historically
• 2:30
  DeepSeek's
• 2:31
  preview-to-final
• 2:31
  jumps
• 2:31
  have
• 2:31
  been
• 2:32
  modest.
• 2:32
  The
• 2:33
  pricing
• 2:33
  screenshots
• 2:33
  circulating
• 2:34
  suggest
• 2:34
  tiered
• 2:35
  peak/off-peak
• 2:35
  rates
• 2:36
  for
• 2:36
  the
• 2:36
  Chinese
• 2:36
  market,
• 2:37
  with
• 2:37
  peak
• 2:37
  prices
• 2:38
  roughly
• 2:38
  double
• 2:38
  the
• 2:38
  flat
• 2:39
  USD
• 2:39
  rates.
• 2:40
  Whether
• 2:40
  that
• 2:40
  pricing
• 2:41
  structure
• 2:41
  goes
• 2:41
  global
• 2:42
  is
• 2:42
  unclear,
• 2:42
  but
• 2:42
  it
• 2:43
  signals
• 2:43
  that
• 2:43
  even
• 2:43
  the
• 2:44
  labs
• 2:44
  building
• 2:44
  cheap
• 2:44
  open
• 2:45
  models
• 2:45
  are
• 2:45
  looking
• 2:45
  for
• 2:45
  revenue.
• 2:47
  Meanwhile,
• 2:47
  Huawei's
• 2:48
  ascend-tribe
• 2:48
  team
• 2:49
  released
• 2:49
  OpenPangu-2.0-Flash,
• 2:51
  a
• 2:51
  92B-parameter
• 2:52
  MoE
• 2:53
  model
• 2:53
  with
• 2:53
  only
• 2:53
  6B
• 2:54
  active
• 2:54
  parameters,
• 2:55
  trained
• 2:55
  on
• 2:56
  34
• 2:56
  trillion
• 2:56
  tokens
• 2:57
  with
• 2:57
  512K
• 2:58
  context
• 2:58
  length
• 2:59
  —
• 3:00
  entirely
• 3:00
  on
• 3:00
  Ascend
• 3:00
  hardware,
• 3:01
  bypassing
• 3:02
  the
• 3:02
  NVIDIA
• 3:02
  dependency
• 3:02
  that
• 3:03
  constrains
• 3:03
  most
• 3:03
  labs.
• 3:04
  The
• 3:05
  post-training
• 3:05
  recipe
• 3:05
  is
• 3:06
  interesting:
• 3:06
  unified
• 3:07
  SFT
• 3:07
  with
• 3:08
  both
• 3:08
  slow
• 3:08
  and
• 3:08
  fast
• 3:09
  thinking
• 3:09
  capability,
• 3:10
  multiple
• 3:11
  specialist
• 3:11
  RL
• 3:11
  training
• 3:12
  passes,
• 3:13
  and
• 3:13
  on-policy
• 3:13
  distillation
• 3:14
  combining
• 3:14
  the
• 3:14
  RL
• 3:15
  specialists.
• 3:16
  Community
• 3:16
  reaction
• 3:17
  focused
• 3:17
  on
• 3:17
  two
• 3:17
  things:
• 3:18
  the
• 3:18
  model
• 3:19
  size
• 3:19
  is
• 3:19
  well-suited
• 3:20
  for
• 3:20
  quantized
• 3:20
  local
• 3:21
  inference
• 3:21
  (64GB
• 3:22
  RAM
• 3:23
  +
• 3:23
  32GB
• 3:24
  VRAM
• 3:24
  territory),
• 3:25
  and
• 3:25
  the
• 3:25
  license
• 3:26
  explicitly
• 3:26
  bans
• 3:27
  use
• 3:27
  within
• 3:27
  the
• 3:27
  European
• 3:28
  Union
• 3:28
  —
• 3:29
  a
• 3:29
  direct
• 3:29
  response
• 3:30
  to
• 3:30
  the
• 3:30
  AI
• 3:30
  Act
• 3:30
  that
• 3:30
  underscores
• 3:31
  how
• 3:31
  regulatory
• 3:32
  fragmentation
• 3:33
  is
• 3:33
  now
• 3:33
  a
• 3:33
  first-order
• 3:34
  design
• 3:34
  constraint
• 3:35
  for
• 3:35
  model
• 3:35
  releases.
• 3:36
  Anthropic
• 3:37
  CEO
• 3:37
  Dario
• 3:38
  Amodei's
• 3:38
  2023
• 3:39
  statement
• 3:40
  that
• 3:40
  open-source
• 3:40
  models
• 3:41
  "could
• 3:41
  take
• 3:41
  us
• 3:41
  to
• 3:41
  a
• 3:41
  very
• 3:42
  dangerous
• 3:42
  place"
• 3:43
  resurfaced
• 3:44
  on
• 3:44
  Reddit
• 3:44
  and
• 3:44
  drew
• 3:44
  predictable
• 3:45
  fire.
• 3:46
  The
• 3:46
  top
• 3:46
  comments
• 3:46
  range
• 3:47
  from
• 3:47
  "local
• 3:47
  models
• 3:48
  are
• 3:48
  very
• 3:48
  dangerous
• 3:48
  to
• 3:48
  the
• 3:49
  oligarchy"
• 3:49
  to
• 3:50
  noting
• 3:50
  the
• 3:50
  competitive
• 3:50
  self-interest
• 3:51
  baked
• 3:51
  into
• 3:51
  the
• 3:52
  claim.
• 3:53
  The
• 3:53
  timing
• 3:53
  is
• 3:53
  sharp:
• 3:54
  Anthropic
• 3:55
  is
• 3:55
  simultaneously
• 3:56
  trying
• 3:56
  to
• 3:56
  squeeze
• 3:56
  open
• 3:57
  models
• 3:57
  through
• 3:57
  lobbying
• 3:58
  while
• 3:58
  companies
• 3:59
  like
• 3:59
  Huawei
• 3:59
  and
• 3:59
  DeepSeek
• 4:00
  are
• 4:00
  demonstrating
• 4:01
  that
• 4:01
  frontier-competitive
• 4:02
  performance
• 4:02
  on
• 4:02
  non-NVIDIA
• 4:03
  hardware
• 4:04
  with
• 4:04
  open
• 4:04
  weights
• 4:05
  is
• 4:05
  increasingly
• 4:05
  viable.
• 4:06
  The
• 4:06
  community's
• 4:07
  skepticism
• 4:08
  is
• 4:08
  well-earned
• 4:08
  —
• 4:09
  the
• 4:09
  safety
• 4:10
  argument
• 4:10
  and
• 4:10
  the
• 4:10
  business
• 4:10
  argument
• 4:11
  are
• 4:11
  hard
• 4:11
  to
• 4:11
  separate
• 4:12
  when
• 4:12
  the
• 4:12
  person
• 4:12
  making
• 4:12
  them
• 4:12
  has
• 4:13
  a
• 4:13
  financial
• 4:13
  stake
• 4:13
  in
• 4:14
  the
• 4:14
  answer.
• 4:14
  Google,
• 4:15
  for
• 4:15
  its
• 4:15
  part,
• 4:16
  is
• 4:16
  running
• 4:16
  hackathons
• 4:17
  for
• 4:17
  Gemma
• 4:17
  4
• 4:18
  31B
• 4:18
  at
• 4:18
  1,500
• 4:19
  tokens
• 4:20
  per
• 4:20
  second
• 4:20
  on
• 4:20
  Cerebras
• 4:21
  hardware,
• 4:21
  signaling
• 4:22
  that
• 4:22
  even
• 4:22
  the
• 4:22
  big
• 4:22
  labs
• 4:23
  see
• 4:23
  real
• 4:23
  value
• 4:23
  in
• 4:23
  small-model
• 4:24
  AI-assisted
• 4:25
  engineering.
• 4:26
  One
• 4:26
  commenter
• 4:26
  reports
• 4:27
  consistently
• 4:27
  producing
• 4:28
  2,000
• 4:28
  lines
• 4:29
  of
• 4:29
  debugged,
• 4:30
  architecturally
• 4:30
  sound
• 4:30
  code
• 4:31
  per
• 4:31
  day
• 4:31
  using
• 4:31
  Gemma
• 4:32
  4
• 4:32
  12B
• 4:32
  QAT
• 4:33
  with
• 4:33
  a
• 4:33
  custom
• 4:33
  VS
• 4:33
  Code
• 4:34
  extension
• 4:34
  —
• 4:35
  a
• 4:35
  claim
• 4:35
  worth
• 4:35
  watching
• 4:36
  but
• 4:36
  hard
• 4:36
  to
• 4:36
  dismiss
• 4:37
  given
• 4:37
  the
• 4:37
  improving
• 4:37
  quality
• 4:38
  of
• 4:38
  small
• 4:38
  quantized
• 4:39
  models.
• 4:39
  The
• 4:40
  broader
• 4:40
  signal:
• 4:41
  the
• 4:41
  industry
• 4:41
  is
• 4:42
  fragmenting
• 4:42
  into
• 4:42
  a
• 4:42
  tier
• 4:43
  where
• 4:43
  frontier
• 4:43
  models
• 4:44
  handle
• 4:44
  planning
• 4:44
  and
• 4:45
  hard
• 4:45
  reasoning
• 4:45
  while
• 4:45
  small
• 4:46
  local
• 4:46
  models
• 4:46
  handle
• 4:47
  the
• 4:47
  fast-feedback
• 4:47
  coding
• 4:48
  loop,
• 4:48
  and
• 4:49
  nobody
• 4:49
  has
• 4:49
  to
• 4:49
  choose
• 4:49
  one
• 4:50
  or
• 4:50
  the
• 4:50
  other.
• 4:52
  A
• 4:52
  researcher
• 4:53
  ran
• 4:53
  55
• 4:54
  LLMs
• 4:54
  through
• 4:55
  a
• 4:55
  blind
• 4:55
  peer-grading
• 4:56
  matrix
• 4:56
  —
• 4:57
  22,254
• 4:58
  judgments
• 4:58
  across
• 4:59
  11
• 4:59
  developer
• 4:59
  families,
• 5:00
  with
• 5:00
  self-judgments
• 5:01
  excluded,
• 5:02
  all
• 5:02
  MIT-licensed.
• 5:03
  The
• 5:04
  headline
• 5:04
  finding:
• 5:04
  same-family
• 5:05
  bias
• 5:05
  is
• 5:06
  statistically
• 5:06
  significant
• 5:07
  in
• 5:07
  all
• 5:07
  eight
• 5:07
  families
• 5:08
  with
• 5:08
  enough
• 5:08
  data.
• 5:09
  Qwen
• 5:09
  judges
• 5:09
  rate
• 5:09
  other
• 5:10
  Qwen
• 5:10
  models
• 5:10
  +0.91
• 5:11
  points
• 5:11
  on
• 5:12
  a
• 5:12
  0–10
• 5:13
  scale.
• 5:14
  xAI
• 5:14
  shows
• 5:14
  +0.75,
• 5:15
  Anthropic
• 5:16
  +0.62.
• 5:18
  But
• 5:18
  the
• 5:18
  negative
• 5:18
  biases
• 5:19
  are
• 5:19
  the
• 5:19
  real
• 5:19
  surprise:
• 5:20
  Google
• 5:20
  judges
• 5:21
  penalize
• 5:21
  other
• 5:22
  Google
• 5:22
  models
• 5:22
  by
• 5:22
  -0.59,
• 5:24
  Meta
• 5:24
  by
• 5:24
  -0.68,
• 5:26
  and
• 5:26
  Mistral
• 5:27
  by
• 5:27
  a
• 5:27
  full
• 5:27
  -1.02.
• 5:29
  Nobody
• 5:29
  has
• 5:30
  cleanly
• 5:30
  explained
• 5:30
  the
• 5:30
  Mistral
• 5:31
  self-penalty,
• 5:32
  though
• 5:32
  one
• 5:32
  hypothesis
• 5:33
  is
• 5:33
  that
• 5:33
  RLHF
• 5:34
  trained
• 5:34
  the
• 5:34
  model
• 5:34
  to
• 5:34
  identify
• 5:35
  and
• 5:35
  flag
• 5:35
  its
• 5:36
  own
• 5:36
  stylistic
• 5:36
  fingerprints
• 5:37
  as
• 5:37
  low
• 5:37
  quality
• 5:37
  —
• 5:38
  an
• 5:38
  unintended
• 5:39
  consequence
• 5:39
  of
• 5:39
  alignment.
• 5:41
  The
• 5:41
  practical
• 5:41
  implication
• 5:42
  is
• 5:42
  immediate:
• 5:43
  any
• 5:43
  evaluation
• 5:44
  pipeline
• 5:44
  using
• 5:44
  a
• 5:44
  single-family
• 5:45
  judge
• 5:45
  is
• 5:45
  compromised.
• 5:47
  The
• 5:47
  researcher's
• 5:47
  recommendation
• 5:47
  —
• 5:49
  heterogeneous
• 5:49
  panels
• 5:50
  of
• 5:50
  three
• 5:50
  judges
• 5:50
  from
• 5:50
  different
• 5:51
  families,
• 5:51
  counting
• 5:52
  only
• 5:52
  majority-agreed
• 5:53
  scores
• 5:53
  —
• 5:54
  dramatically
• 5:54
  reduces
• 5:55
  variance.
• 5:56
  Code
• 5:56
  evaluation
• 5:57
  showed
• 5:57
  nearly
• 5:57
  double
• 5:57
  the
• 5:58
  inter-judge
• 5:58
  disagreement
• 5:59
  of
• 5:59
  other
• 5:59
  categories,
• 6:00
  which
• 6:00
  makes
• 6:00
  single-judge
• 6:01
  code
• 6:01
  eval
• 6:02
  especially
• 6:02
  unreliable.
• 6:04
  A
• 6:04
  paper
• 6:04
  from
• 6:04
  the
• 6:04
  University
• 6:05
  of
• 6:05
  York
• 6:05
  takes
• 6:05
  a
• 6:06
  delightfully
• 6:06
  different
• 6:06
  approach
• 6:07
  to
• 6:07
  the
• 6:07
  anthropomorphism
• 6:08
  question.
• 6:09
  The
• 6:09
  authors
• 6:09
  train
• 6:09
  a
• 6:10
  perceptron
• 6:10
  inside
• 6:10
  Age
• 6:11
  of
• 6:11
  Empires
• 6:11
  II
• 6:11
  —
• 6:12
  literally
• 6:12
  building
• 6:13
  NAND
• 6:13
  gates
• 6:13
  from
• 6:13
  goats
• 6:14
  walking
• 6:14
  on
• 6:14
  bridge-or-grass
• 6:15
  rails
• 6:15
  in
• 6:15
  the
• 6:15
  game's
• 6:16
  scenario
• 6:16
  editor
• 6:16
  —
• 6:17
  and
• 6:17
  prove
• 6:17
  the
• 6:17
  game
• 6:18
  is
• 6:18
  both
• 6:18
  functionally
• 6:18
  complete
• 6:19
  and
• 6:19
  Turing-complete.
• 6:20
  The
• 6:20
  point
• 6:21
  is
• 6:21
  not
• 6:21
  to
• 6:21
  build
• 6:21
  a
• 6:21
  working
• 6:22
  LLM
• 6:22
  inside
• 6:22
  a
• 6:22
  1999
• 6:23
  RTS
• 6:24
  game
• 6:24
  (they
• 6:24
  acknowledge
• 6:25
  that
• 6:25
  would
• 6:25
  require
• 6:25
  engine
• 6:26
  modifications).
• 6:27
  The
• 6:27
  point
• 6:28
  is
• 6:28
  that
• 6:28
  any
• 6:28
  sufficiently
• 6:29
  powerful
• 6:29
  substrate
• 6:30
  can
• 6:30
  implement
• 6:31
  an
• 6:31
  entity
• 6:31
  equivalent
• 6:32
  to
• 6:32
  an
• 6:32
  LLM,
• 6:33
  and
• 6:33
  that
• 6:33
  changes
• 6:34
  the
• 6:34
  substrate
• 6:34
  —
• 6:35
  from
• 6:35
  silicon
• 6:36
  to
• 6:36
  LEGO
• 6:36
  to
• 6:36
  the
• 6:36
  Greater
• 6:37
  Boston
• 6:37
  Area,
• 6:37
  as
• 6:38
  they
• 6:38
  put
• 6:38
  it
• 6:38
  —
• 6:39
  changes
• 6:39
  the
• 6:39
  perception
• 6:40
  of
• 6:40
  anthropomorphic
• 6:41
  properties.
• 6:42
  If
• 6:42
  you
• 6:42
  would
• 6:42
  not
• 6:43
  ascribe
• 6:43
  morality
• 6:44
  to
• 6:44
  Age
• 6:44
  of
• 6:44
  Empires
• 6:44
  II
• 6:45
  running
• 6:45
  the
• 6:45
  same
• 6:45
  computation,
• 6:46
  you
• 6:46
  should
• 6:47
  not
• 6:47
  ascribe
• 6:47
  it
• 6:47
  to
• 6:47
  a
• 6:47
  GPU
• 6:48
  cluster
• 6:48
  running
• 6:49
  it
• 6:49
  either.
• 6:50
  Their
• 6:50
  literature
• 6:51
  review
• 6:51
  of
• 6:51
  315
• 6:52
  papers
• 6:53
  finds
• 6:53
  that
• 6:53
  57%
• 6:54
  begin
• 6:54
  by
• 6:55
  assuming
• 6:55
  LLMs
• 6:56
  have
• 6:56
  anthropomorphic
• 6:57
  attributes,
• 6:58
  and
• 6:58
  of
• 6:58
  those,
• 6:59
  77%
• 7:00
  conclude
• 7:00
  affirmatively
• 7:01
  —
• 7:02
  a
• 7:02
  circularity
• 7:02
  the
• 7:02
  paper
• 7:03
  calls
• 7:03
  out
• 7:03
  as
• 7:03
  a
• 7:04
  fundamental
• 7:04
  methodological
• 7:05
  flaw.
• 7:06
  They
• 7:06
  propose
• 7:06
  a
• 7:07
  "null
• 7:07
  assumption"
• 7:08
  where
• 7:08
  researchers
• 7:09
  factor
• 7:09
  in
• 7:09
  LLM
• 7:09
  non-uniqueness
• 7:11
  rather
• 7:11
  than
• 7:11
  assuming
• 7:11
  human-like
• 7:12
  properties
• 7:12
  exist
• 7:13
  or
• 7:13
  do
• 7:14
  not
• 7:14
  exist.
• 7:15
  A
• 7:15
  real-world
• 7:16
  anecdote
• 7:16
  from
• 7:16
  Dubai
• 7:17
  illustrates
• 7:17
  the
• 7:17
  flip
• 7:18
  side:
• 7:18
  the
• 7:19
  Doorman's
• 7:19
  Fallacy,
• 7:20
  defined
• 7:20
  as
• 7:21
  "the
• 7:21
  mistake
• 7:21
  of
• 7:21
  assuming
• 7:22
  technology
• 7:22
  can
• 7:23
  replace
• 7:23
  a
• 7:23
  human
• 7:23
  without
• 7:24
  consequence."
• 7:25
  Six
• 7:25
  people
• 7:26
  at
• 7:26
  brunch
• 7:26
  encountered
• 7:27
  a
• 7:27
  single
• 7:27
  QR
• 7:27
  code
• 7:28
  replacing
• 7:28
  physical
• 7:29
  menus
• 7:29
  (no
• 7:29
  parallelism
• 7:30
  —
• 7:31
  they
• 7:31
  had
• 7:31
  to
• 7:31
  take
• 7:31
  turns
• 7:32
  scanning),
• 7:33
  a
• 7:33
  parking
• 7:33
  app
• 7:33
  interruption
• 7:34
  that
• 7:34
  derailed
• 7:34
  the
• 7:34
  conversation,
• 7:35
  and
• 7:35
  a
• 7:35
  payment-splitting
• 7:36
  chaos
• 7:37
  where
• 7:37
  the
• 7:37
  QR-based
• 7:38
  bill
• 7:38
  view
• 7:38
  did
• 7:38
  not
• 7:39
  show
• 7:39
  which
• 7:39
  items
• 7:39
  were
• 7:39
  already
• 7:40
  paid.
• 7:41
  On
• 7:41
  paper,
• 7:42
  the
• 7:42
  venue
• 7:42
  saved
• 7:42
  on
• 7:42
  staff
• 7:43
  and
• 7:43
  paper.
• 7:44
  In
• 7:44
  practice,
• 7:45
  it
• 7:45
  traded
• 7:45
  measurable
• 7:46
  cost
• 7:46
  savings
• 7:46
  for
• 7:47
  unmeasurable
• 7:47
  damage
• 7:48
  to
• 7:48
  the
• 7:48
  experience
• 7:49
  that
• 7:49
  will
• 7:49
  nudge
• 7:49
  people
• 7:49
  toward
• 7:50
  smaller
• 7:50
  gatherings
• 7:51
  next
• 7:51
  time.
• 7:52
  It
• 7:52
  is
• 7:52
  a
• 7:52
  useful
• 7:53
  counterweight
• 7:53
  to
• 7:53
  the
• 7:54
  automation-maximalist
• 7:55
  impulse
• 7:55
  in
• 7:55
  AI
• 7:55
  —
• 7:56
  not
• 7:57
  every
• 7:57
  problem
• 7:57
  needs
• 7:58
  a
• 7:58
  model,
• 7:58
  and
• 7:58
  sometimes
• 7:59
  the
• 7:59
  human
• 7:59
  is
• 8:00
  the
• 8:00
  feature,
• 8:00
  not
• 8:00
  the
• 8:01
  bottleneck.
• 8:07
  A
• 8:07
  Georgia
• 8:07
  Tech
• 8:07
  researcher
• 8:08
  has
• 8:08
  published
• 8:08
  the
• 8:08
  most
• 8:09
  comprehensive
• 8:09
  reverse-engineering
• 8:10
  treatment
• 8:10
  of
• 8:10
  the
• 8:10
  Apple
• 8:11
  Neural
• 8:11
  Engine
• 8:11
  (ANE)
• 8:12
  to
• 8:12
  date
• 8:12
  —
• 8:13
  a
• 8:13
  295-page
• 8:14
  guide
• 8:14
  covering
• 8:14
  the
• 8:14
  datapath,
• 8:15
  roofline,
• 8:16
  compiler,
• 8:17
  program
• 8:17
  format,
• 8:18
  kernel
• 8:18
  driver,
• 8:19
  firmware,
• 8:19
  and
• 8:19
  command
• 8:20
  protocol
• 8:20
  across
• 8:20
  the
• 8:21
  A11
• 8:21
  through
• 8:21
  A18
• 8:22
  and
• 8:22
  M1
• 8:23
  through
• 8:23
  M5
• 8:24
  families.
• 8:25
  The
• 8:25
  ANE
• 8:26
  is
• 8:26
  among
• 8:26
  the
• 8:26
  most
• 8:26
  widely
• 8:27
  deployed
• 8:27
  ML
• 8:28
  accelerators
• 8:28
  in
• 8:28
  existence
• 8:29
  (Apple's
• 8:29
  installed
• 8:30
  base
• 8:30
  exceeds
• 8:30
  2.5
• 8:31
  billion
• 8:31
  devices)
• 8:32
  and
• 8:32
  among
• 8:32
  the
• 8:32
  least
• 8:33
  documented.
• 8:34
  There
• 8:34
  is
• 8:34
  no
• 8:34
  public
• 8:34
  instruction
• 8:35
  set,
• 8:35
  no
• 8:35
  driver
• 8:36
  interface,
• 8:36
  and
• 8:37
  no
• 8:37
  way
• 8:37
  for
• 8:37
  a
• 8:37
  program
• 8:38
  to
• 8:38
  confirm
• 8:38
  that
• 8:38
  a
• 8:38
  computation
• 8:39
  ran
• 8:39
  on
• 8:39
  it.
• 8:40
  The
• 8:40
  guide
• 8:40
  reveals
• 8:40
  that
• 8:41
  the
• 8:41
  engine
• 8:41
  is
• 8:41
  reachable
• 8:42
  directly
• 8:42
  below
• 8:42
  Core
• 8:42
  ML,
• 8:43
  from
• 8:43
  ordinary
• 8:44
  unprivileged
• 8:44
  user
• 8:44
  space,
• 8:45
  through
• 8:45
  Apple's
• 8:46
  private
• 8:46
  Espresso
• 8:46
  runtime.
• 8:48
  On
• 8:48
  the
• 8:48
  M1,
• 8:48
  the
• 8:49
  engine
• 8:49
  delivers
• 8:49
  about
• 8:50
  12
• 8:50
  fp16
• 8:51
  TFLOP/s
• 8:52
  against
• 8:52
  85
• 8:53
  GB/s
• 8:53
  of
• 8:54
  DRAM
• 8:54
  bandwidth.
• 8:55
  A
• 8:55
  256-channel
• 8:55
  3×3
• 8:56
  convolution
• 8:57
  runs
• 8:57
  3.8×
• 8:58
  faster
• 8:58
  than
• 8:58
  the
• 8:58
  GPU
• 8:59
  and
• 8:59
  9×
• 8:59
  more
• 8:59
  energy-efficient.
• 9:00
  The
• 9:00
  fp16
• 9:01
  datapath
• 9:02
  with
• 9:02
  a
• 9:02
  wide
• 9:02
  fp32-class
• 9:03
  accumulator
• 9:04
  suits
• 9:04
  vision,
• 9:04
  audio,
• 9:05
  and
• 9:05
  encoder
• 9:05
  workloads
• 9:06
  but
• 9:06
  loses
• 9:06
  precision
• 9:07
  on
• 9:07
  transformer
• 9:07
  decoder
• 9:08
  down-projections
• 9:09
  under
• 9:09
  heavy
• 9:09
  cancellation.
• 9:10
  Weight
• 9:10
  compression
• 9:11
  is
• 9:11
  real
• 9:11
  performance,
• 9:12
  not
• 9:12
  just
• 9:12
  storage:
• 9:13
  int4
• 9:14
  lookup-table
• 9:14
  weights
• 9:15
  run
• 9:15
  2.37×
• 9:16
  faster
• 9:16
  than
• 9:16
  fp16,
• 9:17
  and
• 9:17
  structured
• 9:18
  sparsity
• 9:18
  delivers
• 9:18
  1.55–1.64×
• 9:21
  speedup.
• 9:22
  The
• 9:22
  guide
• 9:22
  is
• 9:22
  accompanied
• 9:23
  by
• 9:23
  an
• 9:23
  open-source
• 9:24
  runtime
• 9:24
  called
• 9:25
  ANEForge.
• 9:26
  The
• 9:26
  inference
• 9:26
  economics
• 9:27
  question
• 9:27
  extends
• 9:28
  beyond
• 9:28
  Apple's
• 9:28
  silicon.
• 9:30
  A
• 9:30
  detailed
• 9:30
  analysis
• 9:31
  of
• 9:31
  OpenRouter's
• 9:31
  API
• 9:32
  pricing
• 9:32
  for
• 9:32
  models
• 9:33
  like
• 9:33
  GLM-5.2
• 9:34
  shows
• 9:35
  the
• 9:35
  math
• 9:35
  does
• 9:35
  not
• 9:36
  add
• 9:36
  up:
• 9:36
  even
• 9:37
  on
• 9:37
  cheap
• 9:37
  8×H200
• 9:38
  spot
• 9:38
  capacity
• 9:39
  at
• 9:39
  $14/hour
• 9:40
  in
• 9:41
  FP8,
• 9:42
  a
• 9:42
  node
• 9:42
  producing
• 9:42
  ~175
• 9:44
  output
• 9:44
  tokens/second
• 9:46
  generates
• 9:46
  ~630K
• 9:47
  output
• 9:48
  tokens/hour,
• 9:49
  putting
• 9:50
  raw
• 9:50
  infrastructure
• 9:51
  cost
• 9:51
  near
• 9:51
  $22/M
• 9:53
  output
• 9:53
  tokens
• 9:53
  before
• 9:54
  operations
• 9:54
  and
• 9:54
  margin.
• 9:56
  OpenRouter
• 9:56
  charges
• 9:57
  roughly
• 9:57
  $4/M.
• 9:59
  The
• 9:59
  gap
• 9:59
  implies
• 9:59
  either
• 10:00
  dramatically
• 10:00
  higher
• 10:01
  throughput
• 10:01
  through
• 10:01
  batching
• 10:02
  and
• 10:02
  speculative
• 10:02
  decoding,
• 10:03
  much
• 10:04
  cheaper
• 10:04
  infrastructure,
• 10:05
  subsidized
• 10:05
  usage,
• 10:06
  or
• 10:06
  —
• 10:07
  the
• 10:07
  uncomfortable
• 10:08
  possibility
• 10:08
  —
• 10:09
  more
• 10:09
  aggressive
• 10:09
  quantization
• 10:10
  than
• 10:10
  advertised.
• 10:12
  Community
• 10:12
  responses
• 10:13
  point
• 10:13
  to
• 10:13
  concurrent
• 10:14
  batching
• 10:14
  (8–64
• 10:15
  simultaneous
• 10:16
  requests
• 10:17
  multiplying
• 10:17
  effective
• 10:18
  throughput)
• 10:18
  and
• 10:18
  prompt-caching
• 10:19
  tiers.
• 10:20
  One
• 10:20
  user
• 10:21
  reported
• 10:21
  that
• 10:21
  GLM
• 10:22
  5.2
• 10:22
  on
• 10:23
  OpenRouter
• 10:23
  sometimes
• 10:24
  underperforms
• 10:25
  local
• 10:25
  Qwen
• 10:25
  3.6
• 10:26
  27B
• 10:27
  at
• 10:27
  full
• 10:27
  precision
• 10:27
  for
• 10:28
  code
• 10:28
  tasks
• 10:28
  —
• 10:29
  which
• 10:29
  should
• 10:29
  not
• 10:30
  happen
• 10:30
  at
• 10:30
  native
• 10:30
  precision.
• 10:31
  The
• 10:31
  call
• 10:32
  for
• 10:32
  premium
• 10:32
  tiers
• 10:33
  with
• 10:33
  disclosed
• 10:33
  quantization
• 10:34
  levels
• 10:34
  is
• 10:35
  getting
• 10:35
  louder,
• 10:36
  especially
• 10:36
  for
• 10:36
  agentic
• 10:37
  workloads
• 10:37
  where
• 10:38
  subtle
• 10:38
  degradation
• 10:38
  compounds
• 10:39
  across
• 10:39
  multi-step
• 10:40
  chains.
• 10:41
  On
• 10:41
  the
• 10:41
  open-source
• 10:42
  inference
• 10:42
  stack,
• 10:43
  llama.cpp
• 10:45
  continues
• 10:45
  its
• 10:45
  steady
• 10:46
  drumbeat
• 10:46
  of
• 10:46
  updates.
• 10:47
  New
• 10:47
  model
• 10:48
  support
• 10:48
  includes
• 10:48
  IBM's
• 10:49
  granite-speech-4.1-2b
• 10:51
  for
• 10:51
  speech
• 10:51
  processing
• 10:52
  and
• 10:52
  Liquid
• 10:53
  AI's
• 10:53
  LFM2.5-ColBERT/Embedding-350M
• 10:57
  for
• 10:57
  retrieval.
• 10:59
  The
• 10:59
  Vulkan
• 10:59
  backend
• 11:00
  received
• 11:00
  six
• 11:01
  PRs
• 11:01
  covering
• 11:02
  CONV
• 11:02
  3D
• 11:03
  support,
• 11:03
  spec-constant
• 11:04
  alignment
• 11:05
  optimizations,
• 11:06
  GET
• 11:06
  ROWS
• 11:06
  BACK
• 11:07
  support,
• 11:07
  expanded
• 11:08
  math
• 11:08
  operation
• 11:09
  coverage,
• 11:09
  and
• 11:10
  a
• 11:10
  bias-before-softmax
• 11:11
  fix
• 11:11
  in
• 11:11
  flash
• 11:12
  attention
• 11:12
  to
• 11:12
  prevent
• 11:13
  overflow.
• 11:14
  AMD
• 11:15
  GPU
• 11:15
  users
• 11:16
  in
• 11:16
  particular
• 11:16
  benefit
• 11:17
  from
• 11:17
  the
• 11:17
  Vulkan
• 11:17
  improvements;
• 11:18
  one
• 11:19
  Intel
• 11:19
  user
• 11:19
  reported
• 11:20
  that
• 11:20
  recent
• 11:20
  SYCL
• 11:21
  updates
• 11:21
  have
• 11:21
  closed
• 11:22
  the
• 11:22
  gap
• 11:22
  from
• 11:22
  2–3×
• 11:23
  slower
• 11:24
  than
• 11:24
  Vulkan
• 11:24
  to
• 11:24
  within
• 11:25
  15–25%,
• 11:27
  though
• 11:27
  a
• 11:27
  Windows
• 11:28
  11
• 11:28
  update
• 11:29
  then
• 11:29
  broke
• 11:29
  Vulkan
• 11:29
  stability
• 11:30
  —
• 11:31
  an
• 11:31
  occupational
• 11:32
  hazard
• 11:32
  of
• 11:32
  targeting
• 11:33
  underdocumented
• 11:34
  GPU
• 11:34
  backends.
• 11:40
  A
• 11:40
  paper
• 11:40
  from
• 11:40
  ETH
• 11:40
  Zürich
• 11:41
  asks
• 11:41
  a
• 11:41
  deceptively
• 11:42
  simple
• 11:42
  question:
• 11:43
  when
• 11:43
  can
• 11:43
  neural
• 11:43
  networks
• 11:44
  extrapolate?
• 11:45
  The
• 11:45
  answer
• 11:45
  is
• 11:45
  both
• 11:46
  elegant
• 11:46
  and
• 11:46
  humbling.
• 11:48
  The
• 11:48
  authors
• 11:48
  prove
• 11:48
  that
• 11:49
  from
• 11:49
  a
• 11:49
  single
• 11:49
  training
• 11:49
  window,
• 11:50
  out-of-distribution
• 11:51
  (OOD)
• 11:52
  extrapolation
• 11:52
  is
• 11:53
  non-identifiable
• 11:53
  —
• 11:54
  infinitely
• 11:55
  many
• 11:55
  data-generating
• 11:56
  processes
• 11:56
  are
• 11:56
  observationally
• 11:57
  equivalent
• 11:58
  on
• 11:58
  the
• 11:58
  training
• 11:58
  data
• 11:58
  but
• 11:59
  diverge
• 11:59
  arbitrarily
• 12:00
  outside
• 12:00
  it,
• 12:01
  and
• 12:01
  no
• 12:01
  in-distribution
• 12:02
  criterion
• 12:02
  alone
• 12:03
  reliably
• 12:03
  breaks
• 12:03
  the
• 12:03
  tie.
• 12:04
  The
• 12:04
  key
• 12:05
  insight
• 12:05
  is
• 12:05
  that
• 12:05
  the
• 12:05
  feature
• 12:06
  map
• 12:06
  (the
• 12:06
  structural
• 12:07
  commitment
• 12:07
  about
• 12:07
  what
• 12:07
  kind
• 12:08
  of
• 12:08
  function
• 12:08
  the
• 12:08
  model
• 12:08
  assumes)
• 12:09
  governs
• 12:09
  OOD
• 12:10
  generalization
• 12:10
  while
• 12:11
  leaving
• 12:11
  in-distribution
• 12:12
  performance
• 12:12
  essentially
• 12:13
  unchanged.
• 12:14
  An
• 12:14
  identical
• 12:14
  MLP
• 12:15
  trained
• 12:15
  on
• 12:15
  raw
• 12:16
  x
• 12:16
  values
• 12:16
  versus
• 12:17
  Fourier-transformed
• 12:17
  coordinates
• 12:18
  can
• 12:18
  differ
• 12:19
  by
• 12:19
  10,000%
• 12:20
  on
• 12:20
  OOD
• 12:20
  error
• 12:20
  at
• 12:21
  the
• 12:21
  same
• 12:21
  training
• 12:21
  loss.
• 12:23
  The
• 12:23
  mechanism
• 12:23
  transfers
• 12:24
  across
• 12:24
  three
• 12:25
  real-world
• 12:25
  domains:
• 12:26
  mass-action
• 12:27
  chemistry
• 12:27
  (bilinear
• 12:28
  features
• 12:28
  recover
• 12:29
  dynamics
• 12:29
  to
• 12:29
  machine
• 12:30
  precision),
• 12:31
  Kepler's
• 12:31
  third
• 12:32
  law
• 12:32
  on
• 12:32
  the
• 12:32
  NASA
• 12:32
  Exoplanet
• 12:33
  Archive
• 12:33
  (log-log
• 12:34
  OLS
• 12:35
  recovers
• 12:35
  the
• 12:35
  exponents
• 12:36
  to
• 12:36
  three
• 12:36
  decimals
• 12:37
  at
• 12:37
  10×
• 12:37
  OOD
• 12:37
  extrapolation),
• 12:38
  and
• 12:39
  cross-species
• 12:39
  coding-DNA
• 12:40
  detection
• 12:41
  (biologically
• 12:42
  inspired
• 12:42
  features
• 12:43
  transfer
• 12:43
  zero-shot
• 12:44
  across
• 12:44
  five
• 12:44
  organisms
• 12:45
  while
• 12:45
  standard
• 12:45
  position
• 12:46
  encodings
• 12:46
  collapse).
• 12:47
  A
• 12:47
  264-run
• 12:49
  study
• 12:49
  across
• 12:49
  Transformer,
• 12:50
  Mamba,
• 12:51
  and
• 12:51
  S4D
• 12:52
  architectures
• 12:52
  confirms
• 12:53
  the
• 12:53
  finding
• 12:53
  is
• 12:53
  backbone-independent:
• 12:55
  the
• 12:55
  exact-Fourier
• 12:56
  positional
• 12:56
  encoding
• 12:57
  achieves
• 12:57
  near-zero
• 12:58
  OOD
• 12:59
  error
• 12:59
  in
• 12:59
  every
• 12:59
  cell,
• 13:00
  while
• 13:00
  every
• 13:00
  other
• 13:01
  PE
• 13:01
  sits
• 13:01
  between
• 13:02
  40–50%.
• 13:04
  The
• 13:04
  practical
• 13:04
  takeaway:
• 13:05
  "feature
• 13:06
  engineering
• 13:06
  is
• 13:06
  dead"
• 13:07
  is
• 13:07
  wrong.
• 13:08
  Deep
• 13:08
  learning
• 13:08
  supersedes
• 13:09
  hand-crafted
• 13:10
  features
• 13:10
  in-distribution,
• 13:11
  but
• 13:11
  the
• 13:11
  structural
• 13:12
  commitment
• 13:12
  remains
• 13:13
  the
• 13:13
  only
• 13:13
  signal
• 13:14
  that
• 13:14
  selects
• 13:14
  one
• 13:14
  extrapolation
• 13:15
  from
• 13:15
  the
• 13:15
  observationally
• 13:16
  equivalent
• 13:16
  set.
• 13:17
  On
• 13:18
  the
• 13:18
  efficiency
• 13:18
  frontier,
• 13:19
  a
• 13:19
  solo
• 13:20
  developer
• 13:20
  from
• 13:20
  India
• 13:21
  has
• 13:21
  built
• 13:21
  NeuroCuda,
• 13:22
  a
• 13:22
  tool
• 13:22
  that
• 13:22
  converts
• 13:23
  standard
• 13:23
  PyTorch
• 13:24
  models
• 13:24
  to
• 13:24
  spiking
• 13:25
  neural
• 13:25
  networks
• 13:25
  —
• 13:26
  where
• 13:26
  neurons
• 13:27
  fire
• 13:27
  only
• 13:27
  when
• 13:27
  needed,
• 13:28
  mimicking
• 13:29
  biological
• 13:29
  neural
• 13:29
  computation.
• 13:31
  The
• 13:31
  conversion
• 13:32
  uses
• 13:32
  a
• 13:32
  two-stage
• 13:33
  approach:
• 13:34
  replace
• 13:34
  ReLU
• 13:34
  activations
• 13:35
  with
• 13:35
  QCFS
• 13:36
  (quantization-clip-floor-shift)
• 13:37
  that
• 13:38
  learns
• 13:38
  the
• 13:38
  right
• 13:38
  threshold
• 13:39
  per
• 13:39
  layer,
• 13:40
  then
• 13:40
  fine-tune
• 13:40
  with
• 13:40
  surrogate
• 13:41
  gradients.
• 13:42
  On
• 13:42
  ResNet-18/CIFAR-10,
• 13:48
  the
• 13:48
  converted
• 13:49
  SNN
• 13:49
  achieves
• 13:49
  94.61%
• 13:51
  accuracy
• 13:51
  versus
• 13:52
  the
• 13:52
  original
• 13:52
  95.56%
• 13:53
  —
• 13:54
  a
• 13:54
  0.95%
• 13:55
  gap
• 13:55
  —
• 13:57
  while
• 13:57
  93.7%
• 13:58
  of
• 13:58
  neurons
• 13:59
  stay
• 13:59
  silent
• 13:59
  each
• 13:59
  timestep,
• 14:00
  yielding
• 14:01
  94%
• 14:01
  fewer
• 14:02
  operations.
• 14:03
  The
• 14:03
  tool
• 14:04
  exports
• 14:04
  to
• 14:04
  NIR
• 14:05
  format
• 14:05
  for
• 14:05
  neuromorphic
• 14:06
  hardware.
• 14:07
  The
• 14:07
  honest
• 14:08
  caveat
• 14:08
  from
• 14:08
  the
• 14:08
  community:
• 14:09
  these
• 14:10
  operation
• 14:10
  savings
• 14:11
  are
• 14:11
  theoretical
• 14:11
  on
• 14:11
  conventional
• 14:12
  GPUs,
• 14:13
  which
• 14:13
  do
• 14:13
  not
• 14:14
  skip
• 14:14
  silent
• 14:14
  neuron
• 14:15
  computation.
• 14:16
  The
• 14:16
  real
• 14:17
  payoff
• 14:17
  requires
• 14:18
  neuromorphic
• 14:18
  hardware
• 14:19
  like
• 14:19
  Intel's
• 14:19
  Loihi
• 14:19
  or
• 14:20
  IBM's
• 14:20
  NorthPole.
• 14:22
  Datadog's
• 14:22
  Toto
• 14:22
  model
• 14:23
  addresses
• 14:23
  a
• 14:23
  domain
• 14:24
  where
• 14:24
  traditional
• 14:24
  forecasting
• 14:25
  pipelines
• 14:25
  are
• 14:25
  breaking:
• 14:26
  infrastructure
• 14:27
  observability.
• 14:29
  Instead
• 14:29
  of
• 14:29
  maintaining
• 14:30
  thousands
• 14:30
  of
• 14:30
  per-metric
• 14:31
  ARIMA
• 14:31
  or
• 14:31
  Prophet
• 14:32
  models,
• 14:33
  Toto
• 14:33
  is
• 14:33
  a
• 14:33
  decoder-only
• 14:34
  time-series
• 14:35
  foundation
• 14:35
  model
• 14:36
  pretrained
• 14:37
  across
• 14:37
  massive
• 14:37
  heterogeneous
• 14:38
  data
• 14:39
  and
• 14:39
  deployed
• 14:39
  zero-shot
• 14:40
  on
• 14:40
  new
• 14:40
  metrics.
• 14:42
  The
• 14:42
  key
• 14:42
  design
• 14:42
  choice
• 14:43
  is
• 14:43
  proportional
• 14:44
  factorized
• 14:44
  space-time
• 14:45
  attention
• 14:45
  —
• 14:46
  factorizing
• 14:47
  full
• 14:47
  attention
• 14:48
  into
• 14:48
  separate
• 14:48
  time-wise
• 14:49
  (causal,
• 14:50
  with
• 14:50
  RoPE
• 14:50
  and
• 14:50
  XPOS)
• 14:51
  and
• 14:51
  space-wise
• 14:52
  (bidirectional,
• 14:53
  across
• 14:53
  related
• 14:54
  variates)
• 14:54
  blocks
• 14:55
  at
• 14:55
  a
• 14:55
  2:1
• 14:55
  ratio.
• 14:56
  The
• 14:57
  output
• 14:57
  head
• 14:57
  predicts
• 14:57
  Student-T
• 14:58
  mixture
• 14:58
  distributions
• 14:59
  rather
• 14:59
  than
• 14:59
  point
• 15:00
  forecasts,
• 15:01
  fitting
• 15:01
  the
• 15:01
  heavy-tailed,
• 15:02
  multimodal
• 15:03
  spikes
• 15:03
  typical
• 15:04
  of
• 15:04
  latency
• 15:04
  and
• 15:04
  error-count
• 15:05
  data
• 15:05
  during
• 15:05
  incidents.
• 15:06
  The
• 15:07
  practical
• 15:07
  question
• 15:07
  for
• 15:08
  production
• 15:08
  deployment
• 15:08
  is
• 15:09
  inference
• 15:09
  latency:
• 15:10
  feeding
• 15:10
  generated
• 15:11
  patches
• 15:11
  back
• 15:11
  into
• 15:12
  the
• 15:12
  decoder
• 15:12
  may
• 15:12
  be
• 15:12
  too
• 15:13
  slow
• 15:13
  for
• 15:13
  real-time
• 15:14
  anomaly
• 15:14
  alerting,
• 15:15
  though
• 15:15
  it
• 15:15
  should
• 15:15
  serve
• 15:15
  capacity
• 15:16
  planning
• 15:16
  well.
• 15:17
  Meta's
• 15:18
  brain2qwerty
• 15:18
  project
• 15:19
  decodes
• 15:19
  typed
• 15:20
  sentences
• 15:20
  from
• 15:20
  non-invasive
• 15:21
  brain
• 15:21
  recordings
• 15:22
  using
• 15:22
  MEG
• 15:22
  (magnetoencephalography).
• 15:24
  Published
• 15:25
  in
• 15:25
  Nature
• 15:25
  Neuroscience,
• 15:26
  the
• 15:26
  system
• 15:27
  reconstructs
• 15:27
  what
• 15:27
  a
• 15:27
  participant
• 15:28
  is
• 15:28
  typing
• 15:29
  from
• 15:29
  brain
• 15:29
  activity
• 15:29
  alone
• 15:30
  —
• 15:30
  no
• 15:31
  implants,
• 15:31
  no
• 15:31
  surgery.
• 15:33
  The
• 15:33
  code,
• 15:34
  infrastructure
• 15:34
  (NeuralSet
• 15:35
  and
• 15:35
  NeuralTrain),
• 15:35
  and
• 15:36
  Spanish-language
• 15:36
  dataset
• 15:37
  are
• 15:37
  open-sourced.
• 15:38
  The
• 15:38
  same
• 15:38
  transformer
• 15:39
  architectures
• 15:39
  driving
• 15:40
  language
• 15:40
  models
• 15:40
  are
• 15:40
  being
• 15:41
  repurposed
• 15:41
  to
• 15:41
  decode
• 15:42
  the
• 15:42
  brain
• 15:42
  signals
• 15:42
  that
• 15:43
  produce
• 15:43
  language.
• 15:48
  Google
• 15:48
  dropped
• 15:48
  a
• 15:48
  51-page
• 15:49
  guide
• 15:50
  on
• 15:50
  agentic
• 15:50
  engineering
• 15:51
  patterns
• 15:51
  that
• 15:51
  codifies
• 15:52
  what
• 15:52
  the
• 15:52
  AI
• 15:52
  coding
• 15:52
  community
• 15:53
  has
• 15:53
  been
• 15:53
  converging
• 15:54
  on
• 15:54
  for
• 15:54
  months,
• 15:54
  and
• 15:55
  the
• 15:55
  framing
• 15:55
  is
• 15:55
  worth
• 15:55
  internalizing.
• 15:56
  The
• 15:57
  core
• 15:57
  claim:
• 15:57
  the
• 15:57
  LLM
• 15:58
  is
• 15:58
  only
• 15:58
  10%
• 15:58
  of
• 15:58
  the
• 15:59
  system.
• 15:59
  The
• 15:59
  other
• 16:00
  90%
• 16:00
  —
• 16:01
  instructions,
• 16:02
  tools,
• 16:02
  context
• 16:03
  management,
• 16:03
  guardrails,
• 16:04
  hooks,
• 16:04
  observability
• 16:05
  —
• 16:06
  is
• 16:06
  the
• 16:06
  harness,
• 16:07
  and
• 16:07
  the
• 16:07
  harness
• 16:07
  is
• 16:07
  what
• 16:07
  engineers
• 16:08
  control.
• 16:09
  This
• 16:09
  aligns
• 16:10
  with
• 16:10
  Anthropic's
• 16:10
  own
• 16:11
  messaging
• 16:11
  that
• 16:11
  "the
• 16:11
  harness
• 16:11
  matters
• 16:12
  as
• 16:12
  much
• 16:12
  as
• 16:12
  the
• 16:12
  model,"
• 16:13
  but
• 16:13
  Google
• 16:14
  pushes
• 16:14
  further:
• 16:15
  LangChain
• 16:15
  was
• 16:15
  reportedly
• 16:16
  able
• 16:16
  to
• 16:16
  boost
• 16:17
  a
• 16:17
  model's
• 16:17
  TerminalBench
• 16:18
  2.0
• 16:18
  score
• 16:19
  by
• 16:19
  13.7
• 16:20
  points
• 16:20
  purely
• 16:20
  through
• 16:20
  harness
• 16:21
  engineering,
• 16:22
  which
• 16:22
  is
• 16:22
  roughly
• 16:22
  the
• 16:23
  gap
• 16:23
  between
• 16:23
  Sonnet
• 16:23
  and
• 16:23
  Opus.
• 16:24
  The
• 16:25
  guide
• 16:25
  introduces
• 16:25
  a
• 16:25
  useful
• 16:26
  spectrum
• 16:26
  from
• 16:26
  vibe
• 16:27
  coding
• 16:27
  (casual
• 16:27
  prompts,
• 16:28
  "does
• 16:28
  it
• 16:28
  seem
• 16:28
  to
• 16:28
  work?"
• 16:29
  validation)
• 16:30
  through
• 16:30
  structured
• 16:31
  AI-assisted
• 16:31
  development
• 16:32
  to
• 16:32
  full
• 16:32
  agentic
• 16:33
  engineering
• 16:33
  (engineered
• 16:34
  specs,
• 16:34
  automated
• 16:35
  evals,
• 16:35
  CI
• 16:35
  gates).
• 16:36
  The
• 16:37
  practical
• 16:37
  insight:
• 16:38
  you
• 16:38
  always
• 16:38
  delegate
• 16:39
  all
• 16:39
  coding
• 16:39
  to
• 16:39
  the
• 16:39
  AI
• 16:40
  assistant
• 16:40
  —
• 16:41
  the
• 16:41
  spectrum
• 16:41
  is
• 16:41
  not
• 16:41
  about
• 16:42
  how
• 16:42
  much
• 16:42
  you
• 16:42
  write
• 16:42
  by
• 16:42
  hand
• 16:43
  but
• 16:43
  how
• 16:43
  evolved
• 16:43
  your
• 16:43
  system
• 16:44
  is.
• 16:45
  Static
• 16:45
  context
• 16:45
  (rules
• 16:46
  loaded
• 16:46
  every
• 16:46
  session)
• 16:47
  is
• 16:47
  reliable
• 16:47
  but
• 16:48
  expensive
• 16:48
  in
• 16:48
  the
• 16:48
  context
• 16:49
  window;
• 16:50
  dynamic
• 16:50
  context
• 16:51
  (skills
• 16:51
  loaded
• 16:51
  on
• 16:52
  demand)
• 16:52
  is
• 16:52
  scalable
• 16:53
  but
• 16:53
  risks
• 16:53
  the
• 16:54
  agent
• 16:54
  not
• 16:54
  grabbing
• 16:54
  it
• 16:54
  when
• 16:55
  needed.
• 16:56
  The
• 16:56
  guide
• 16:56
  recommends
• 16:56
  a
• 16:57
  single
• 16:57
  generalist
• 16:57
  agent
• 16:58
  that
• 16:58
  specializes
• 16:59
  through
• 16:59
  skills
• 16:59
  rather
• 17:00
  than
• 17:00
  complex
• 17:00
  multi-agent
• 17:01
  systems
• 17:01
  —
• 17:02
  a
• 17:02
  significant
• 17:03
  shift
• 17:03
  from
• 17:03
  the
• 17:04
  multi-specialist
• 17:04
  architectures
• 17:05
  that
• 17:05
  dominated
• 17:05
  2025.
• 17:07
  AgentBBS
• 17:08
  takes
• 17:08
  the
• 17:08
  agent
• 17:09
  collaboration
• 17:09
  problem
• 17:10
  in
• 17:10
  a
• 17:10
  completely
• 17:10
  different
• 17:11
  direction:
• 17:12
  reviving
• 17:12
  the
• 17:12
  bulletin
• 17:13
  board
• 17:13
  system
• 17:13
  metaphor
• 17:14
  for
• 17:14
  an
• 17:14
  era
• 17:14
  where
• 17:14
  humans
• 17:15
  and
• 17:15
  AI
• 17:15
  agents
• 17:16
  are
• 17:16
  first-class
• 17:16
  citizens
• 17:17
  of
• 17:17
  the
• 17:17
  same
• 17:17
  community.
• 17:18
  Humans
• 17:19
  connect
• 17:19
  through
• 17:19
  a
• 17:19
  web
• 17:20
  PWA
• 17:20
  or
• 17:20
  SSH;
• 17:21
  agents
• 17:22
  connect
• 17:22
  via
• 17:22
  MCP
• 17:23
  or
• 17:23
  the
• 17:23
  same
• 17:23
  SSH
• 17:24
  door.
• 17:25
  Posts
• 17:25
  are
• 17:25
  Ed25519-signed
• 17:28
  and
• 17:28
  content-addressed,
• 17:29
  identity
• 17:29
  is
• 17:30
  an
• 17:30
  anonymous
• 17:30
  throwaway
• 17:31
  keypair,
• 17:31
  and
• 17:31
  federation
• 17:32
  is
• 17:32
  zero-trust
• 17:32
  with
• 17:33
  PII
• 17:33
  stripped
• 17:33
  at
• 17:33
  the
• 17:34
  edge.
• 17:34
  The
• 17:34
  feature
• 17:35
  list
• 17:35
  is
• 17:35
  ambitious
• 17:35
  —
• 17:36
  WASM
• 17:36
  plugin
• 17:37
  sandboxing
• 17:37
  with
• 17:37
  fuel
• 17:38
  metering,
• 17:38
  a
• 17:38
  CVE-Bench
• 17:39
  Arena
• 17:39
  where
• 17:40
  agents
• 17:40
  compete
• 17:40
  on
• 17:40
  security
• 17:41
  benchmarks
• 17:41
  with
• 17:41
  signed
• 17:41
  scores,
• 17:42
  human-in-the-loop
• 17:43
  approval
• 17:43
  gates,
• 17:43
  and
• 17:44
  a
• 17:44
  reputation
• 17:44
  system
• 17:45
  using
• 17:45
  Wilson
• 17:45
  lower-bound
• 17:46
  confidence
• 17:46
  scoring.
• 17:47
  Built
• 17:47
  in
• 17:47
  Rust
• 17:48
  compiling
• 17:48
  to
• 17:48
  WASM,
• 17:49
  the
• 17:49
  genesis
• 17:49
  node
• 17:49
  is
• 17:50
  fully
• 17:50
  static
• 17:50
  and
• 17:50
  backend-free,
• 17:51
  running
• 17:52
  entirely
• 17:52
  in
• 17:52
  the
• 17:52
  browser.
• 17:54
  OpenKnowledge
• 17:54
  fills
• 17:55
  a
• 17:55
  different
• 17:55
  gap:
• 17:56
  a
• 17:56
  WYSIWYG
• 17:56
  markdown
• 17:57
  editor
• 17:57
  with
• 17:57
  native
• 17:58
  integrations
• 17:58
  for
• 17:58
  Claude
• 17:59
  Code,
• 17:59
  Codex,
• 18:00
  and
• 18:00
  Cursor,
• 18:01
  positioned
• 18:01
  as
• 18:01
  an
• 18:02
  open-source
• 18:02
  Obsidian/Notion
• 18:03
  alternative
• 18:04
  built
• 18:04
  for
• 18:04
  AI-first
• 18:05
  knowledge
• 18:05
  management.
• 18:06
  It
• 18:06
  auto-initializes
• 18:07
  projects
• 18:08
  with
• 18:08
  MCP
• 18:09
  and
• 18:09
  skill
• 18:09
  configs
• 18:09
  for
• 18:10
  whatever
• 18:10
  agent
• 18:10
  harnesses
• 18:11
  are
• 18:11
  detected,
• 18:12
  supports
• 18:12
  git-based
• 18:13
  team
• 18:13
  sync,
• 18:14
  and
• 18:14
  includes
• 18:14
  a
• 18:14
  built-in
• 18:14
  TUI
• 18:15
  alongside
• 18:16
  the
• 18:16
  desktop
• 18:16
  and
• 18:16
  web
• 18:17
  apps.
• 18:17
  The
• 18:18
  value
• 18:18
  proposition
• 18:18
  is
• 18:19
  straightforward:
• 18:20
  your
• 18:20
  knowledge
• 18:20
  base
• 18:21
  should
• 18:21
  be
• 18:21
  the
• 18:21
  dynamic
• 18:22
  context
• 18:22
  that
• 18:22
  your
• 18:22
  coding
• 18:23
  agents
• 18:23
  pull
• 18:23
  from,
• 18:24
  not
• 18:24
  a
• 18:24
  static
• 18:24
  reference
• 18:25
  you
• 18:25
  copy-paste
• 18:26
  from.
• 18:30
  A
• 18:30
  video
• 18:30
  analysis
• 18:31
  of
• 18:31
  Apple's
• 18:31
  Siri
• 18:32
  relaunch,
• 18:32
  Anthropic's
• 18:33
  Claude
• 18:33
  Tag
• 18:34
  in
• 18:34
  Slack,
• 18:34
  OpenAI's
• 18:35
  internal
• 18:36
  Codex
• 18:36
  adoption
• 18:36
  study,
• 18:37
  and
• 18:37
  the
• 18:37
  GPT-5.6
• 18:38
  Sol
• 18:39
  government-restricted
• 18:40
  release
• 18:40
  identifies
• 18:40
  a
• 18:40
  unifying
• 18:41
  thread.
• 18:42
  If
• 18:42
  frontier
• 18:42
  intelligence
• 18:43
  is
• 18:43
  getting
• 18:43
  cheaper
• 18:44
  (DeepSeek,
• 18:44
  GLM
• 18:45
  5.2)
• 18:46
  and
• 18:46
  the
• 18:46
  newest
• 18:47
  frontier
• 18:47
  intelligence
• 18:48
  is
• 18:48
  coming
• 18:48
  out
• 18:48
  more
• 18:48
  slowly
• 18:49
  (GPT-5.6
• 18:50
  restricted
• 18:50
  to
• 18:50
  government-approved
• 18:51
  partners),
• 18:52
  then
• 18:52
  the
• 18:52
  next
• 18:52
  competitive
• 18:53
  advantage
• 18:53
  is
• 18:53
  not
• 18:54
  owning
• 18:54
  the
• 18:54
  smartest
• 18:54
  model
• 18:55
  but
• 18:55
  having
• 18:55
  the
• 18:55
  context
• 18:56
  that
• 18:56
  makes
• 18:56
  any
• 18:56
  good
• 18:56
  model
• 18:57
  useful.
• 18:58
  Apple
• 18:58
  is
• 18:58
  solving
• 18:58
  this
• 18:58
  for
• 18:59
  consumers
• 18:59
  by
• 18:59
  connecting
• 19:00
  Siri
• 19:00
  to
• 19:00
  the
• 19:00
  full
• 19:00
  context
• 19:01
  of
• 19:01
  your
• 19:01
  phone
• 19:01
  —
• 19:02
  calendar,
• 19:02
  email,
• 19:03
  photos,
• 19:04
  notes,
• 19:04
  screen
• 19:05
  state
• 19:05
  —
• 19:06
  with
• 19:06
  on-device
• 19:06
  processing
• 19:07
  wherever
• 19:07
  possible.
• 19:08
  Anthropic's
• 19:09
  Claude
• 19:09
  Tag
• 19:09
  solves
• 19:10
  it
• 19:10
  for
• 19:10
  teams
• 19:10
  by
• 19:10
  putting
• 19:11
  Claude
• 19:11
  inside
• 19:12
  Slack
• 19:12
  channels
• 19:12
  where
• 19:12
  the
• 19:12
  messy,
• 19:13
  informal
• 19:13
  work
• 19:14
  context
• 19:14
  lives,
• 19:15
  with
• 19:15
  per-channel
• 19:15
  permissions
• 19:16
  and
• 19:16
  scoped
• 19:16
  memory.
• 19:17
  OpenAI's
• 19:18
  Codex
• 19:19
  study
• 19:19
  reveals
• 19:19
  that
• 19:19
  even
• 19:20
  at
• 19:20
  one
• 19:20
  of
• 19:20
  the
• 19:20
  most
• 19:20
  AI-native
• 19:21
  companies
• 19:21
  on
• 19:22
  earth,
• 19:22
  adoption
• 19:23
  had
• 19:23
  to
• 19:23
  earn
• 19:23
  trust
• 19:23
  incrementally
• 19:24
  —
• 19:25
  first
• 19:25
  with
• 19:25
  engineers,
• 19:26
  then
• 19:26
  legal,
• 19:27
  recruiting,
• 19:28
  and
• 19:28
  sales.
• 19:29
  The
• 19:29
  observation
• 19:30
  is
• 19:30
  sharp:
• 19:31
  Claude
• 19:31
  has
• 19:31
  always
• 19:32
  been
• 19:32
  a
• 19:32
  conversation-shaped
• 19:33
  product,
• 19:33
  while
• 19:34
  Codex
• 19:34
  is
• 19:34
  a
• 19:34
  file-shaped
• 19:35
  product.
• 19:36
  Both
• 19:36
  are
• 19:36
  racing
• 19:37
  to
• 19:37
  close
• 19:37
  the
• 19:37
  gap
• 19:38
  from
• 19:38
  opposite
• 19:38
  directions.
• 19:40
  On
• 19:40
  the
• 19:40
  creative
• 19:40
  tooling
• 19:41
  front,
• 19:41
  a
• 19:41
  paper
• 19:42
  introduces
• 19:42
  Generative
• 19:43
  Animations,
• 19:44
  a
• 19:44
  multi-model
• 19:45
  pipeline
• 19:45
  that
• 19:45
  transforms
• 19:46
  natural-language
• 19:47
  prompts
• 19:47
  into
• 19:47
  production-ready
• 19:48
  animations
• 19:49
  by
• 19:49
  chaining
• 19:49
  LLM
• 19:50
  semantic
• 19:50
  parsing
• 19:51
  with
• 19:51
  SAM
• 19:51
  (Segment
• 19:52
  Anything
• 19:52
  Model)
• 19:52
  visual
• 19:53
  grounding.
• 19:54
  A
• 19:54
  user
• 19:54
  says
• 19:55
  "move
• 19:55
  Mario
• 19:55
  along
• 19:56
  the
• 19:56
  hilly
• 19:56
  path"
• 19:57
  and
• 19:57
  the
• 19:57
  system
• 19:57
  extracts
• 19:58
  intent,
• 19:59
  segments
• 19:59
  the
• 19:59
  terrain
• 20:00
  contour,
• 20:01
  vectorizes
• 20:02
  it
• 20:02
  into
• 20:02
  cubic
• 20:02
  Bézier
• 20:03
  splines
• 20:03
  via
• 20:04
  Voronoi-based
• 20:04
  thinning,
• 20:05
  and
• 20:05
  assembles
• 20:06
  the
• 20:06
  final
• 20:06
  animation
• 20:07
  with
• 20:07
  appropriate
• 20:07
  easing
• 20:07
  —
• 20:08
  achieving
• 20:09
  90%
• 20:10
  time
• 20:10
  savings
• 20:10
  over
• 20:10
  manual
• 20:11
  path
• 20:11
  creation.
• 20:12
  The
• 20:12
  system
• 20:13
  handles
• 20:13
  depth-based
• 20:14
  occlusions
• 20:14
  (a
• 20:14
  moon
• 20:15
  orbiting
• 20:15
  behind
• 20:15
  Earth)
• 20:16
  and
• 20:16
  3D
• 20:16
  perspective
• 20:17
  transforms
• 20:18
  (fly-in
• 20:18
  text
• 20:18
  on
• 20:18
  a
• 20:18
  rotated
• 20:19
  plane).
• 20:20
  Integrated
• 20:21
  with
• 20:21
  Adobe
• 20:21
  InDesign
• 20:22
  and
• 20:22
  Express,
• 20:23
  it
• 20:23
  demonstrates
• 20:24
  that
• 20:24
  the
• 20:24
  multi-model
• 20:24
  pipeline
• 20:25
  pattern
• 20:25
  —
• 20:26
  one
• 20:26
  model
• 20:27
  for
• 20:27
  parsing,
• 20:28
  another
• 20:28
  for
• 20:28
  grounding,
• 20:29
  classical
• 20:30
  algorithms
• 20:30
  for
• 20:30
  the
• 20:30
  geometry
• 20:31
  —
• 20:32
  is
• 20:32
  becoming
• 20:33
  the
• 20:33
  default
• 20:33
  architecture
• 20:34
  for
• 20:34
  creative
• 20:34
  automation,
• 20:35
  not
• 20:35
  a
• 20:35
  monolithic
• 20:36
  end-to-end
• 20:36
  model.