Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 05, 2026

Transcript

AGI Dreams Podcast – March 05, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  largest
• 0:00
  fully
• 0:01
  open-weight
• 0:01
  model
• 0:01
  just
• 0:02
  dropped,
• 0:02
  and
• 0:02
  it
• 0:02
  isn't
• 0:03
  from
• 0:03
  a
• 0:03
  Silicon
• 0:03
  Valley
• 0:04
  lab.
• 0:04
  YuanLabAI's
• 0:06
  Yuan3.0-Ultra
• 0:07
  is
• 0:07
  a
• 0:08
  1,010-billion-parameter
• 0:09
  Mixture-of-Experts
• 0:10
  model
• 0:11
  with
• 0:11
  68.8
• 0:12
  billion
• 0:12
  active
• 0:12
  parameters
• 0:13
  per
• 0:13
  forward
• 0:13
  pass
• 0:14
  —
• 0:14
  and
• 0:14
  every
• 0:15
  weight,
• 0:15
  the
• 0:15
  technical
• 0:16
  report,
• 0:16
  and
• 0:16
  training
• 0:17
  details
• 0:17
  are
• 0:17
  published
• 0:18
  under
• 0:18
  a
• 0:18
  permissive
• 0:18
  license.
• 0:20
  The
• 0:20
  headline
• 0:20
  innovation
• 0:21
  is
• 0:21
  LAEP
• 0:22
  (Layer-Adaptive
• 0:22
  Expert
• 0:23
  Pruning),
• 0:23
  which
• 0:24
  the
• 0:24
  team
• 0:24
  says
• 0:24
  cut
• 0:24
  the
• 0:24
  original
• 0:25
  1.515-trillion-parameter
• 0:27
  checkpoint
• 0:28
  by
• 0:28
  33%
• 0:29
  while
• 0:29
  boosting
• 0:30
  pre-training
• 0:30
  efficiency
• 0:31
  by
• 0:31
  49%.
• 0:33
  Whether
• 0:33
  you
• 0:33
  trust
• 0:33
  the
• 0:34
  benchmarks
• 0:34
  or
• 0:34
  not,
• 0:34
  the
• 0:35
  sheer
• 0:35
  audacity
• 0:35
  of
• 0:35
  releasing
• 0:36
  a
• 0:36
  trillion-parameter
• 0:37
  MoE
• 0:37
  with
• 0:37
  full
• 0:37
  16-bit
• 0:38
  and
• 0:38
  4-bit
• 0:38
  weights
• 0:39
  is
• 0:39
  a
• 0:39
  statement:
• 0:40
  the
• 0:40
  open-weight
• 0:41
  frontier
• 0:41
  is
• 0:41
  not
• 0:41
  slowing
• 0:42
  down.
• 0:43
  The
• 0:43
  model
• 0:43
  targets
• 0:44
  enterprise
• 0:44
  RAG,
• 0:45
  complex
• 0:45
  table
• 0:46
  understanding,
• 0:47
  and
• 0:47
  long-document
• 0:48
  analysis
• 0:48
  rather
• 0:49
  than
• 0:49
  the
• 0:49
  coding
• 0:49
  benchmarks
• 0:50
  everyone
• 0:50
  else
• 0:51
  chases,
• 0:51
  which
• 0:51
  is
• 0:52
  a
• 0:52
  refreshing
• 0:52
  change
• 0:52
  of
• 0:52
  pace.
• 0:53
  Community
• 0:54
  reactions
• 0:54
  range
• 0:55
  from
• 0:55
  excitement
• 0:55
  to
• 0:55
  sticker
• 0:56
  shock
• 0:56
  —
• 0:57
  running
• 0:57
  even
• 0:57
  the
• 0:57
  4-bit
• 0:58
  version
• 0:58
  requires
• 0:58
  infrastructure
• 0:59
  that
• 0:59
  makes
• 0:59
  consumer
• 1:00
  GPUs
• 1:00
  weep,
• 1:01
  with
• 1:01
  one
• 1:01
  commenter
• 1:02
  noting
• 1:02
  that
• 1:02
  16xA100
• 1:04
  rental
• 1:04
  runs
• 1:04
  $28.61/hour.
• 1:07
  But
• 1:07
  the
• 1:07
  LAEP
• 1:08
  pruning
• 1:08
  technique
• 1:08
  itself
• 1:09
  is
• 1:09
  worth
• 1:09
  watching:
• 1:10
  if
• 1:10
  you
• 1:11
  can
• 1:11
  train
• 1:11
  at
• 1:11
  1.5T
• 1:12
  and
• 1:12
  prune
• 1:13
  to
• 1:13
  1T
• 1:13
  with
• 1:13
  quality
• 1:14
  preservation,
• 1:15
  the
• 1:15
  open-weight
• 1:15
  community
• 1:16
  may
• 1:16
  have
• 1:16
  found
• 1:16
  a
• 1:17
  new
• 1:17
  playbook
• 1:17
  for
• 1:17
  training-time
• 1:18
  efficiency.
• 1:19
  On
• 1:20
  the
• 1:20
  inference
• 1:20
  side,
• 1:20
  llama.cpp
• 1:22
  is
• 1:22
  on
• 1:23
  the
• 1:23
  cusp
• 1:23
  of
• 1:23
  native
• 1:23
  NVFP4
• 1:26
  support
• 1:26
  via
• 1:26
  PR
• 1:27
  #19769.
• 1:29
  The
• 1:29
  initial
• 1:30
  merge
• 1:30
  adds
• 1:31
  the
• 1:31
  `GGML
• 1:32
  TYPE
• 1:32
  NVFP4`
• 1:34
  block
• 1:35
  struct,
• 1:35
  conversion
• 1:36
  logic
• 1:36
  for
• 1:36
  NVIDIA
• 1:37
  ModelOpt
• 1:37
  models,
• 1:38
  and
• 1:38
  reference
• 1:39
  CPU/ARM
• 1:41
  NEON
• 1:41
  implementations
• 1:41
  —
• 1:42
  but
• 1:43
  no
• 1:43
  CUDA
• 1:43
  kernels
• 1:43
  yet.
• 1:44
  That
• 1:44
  last
• 1:45
  detail
• 1:45
  matters:
• 1:46
  until
• 1:46
  someone
• 1:47
  writes
• 1:47
  the
• 1:47
  CUDA
• 1:47
  backend
• 1:48
  targeting
• 1:48
  Blackwell's
• 1:49
  native
• 1:49
  FP4
• 1:50
  Tensor
• 1:50
  Cores,
• 1:51
  this
• 1:51
  is
• 1:51
  CPU
• 1:52
  emulation,
• 1:53
  not
• 1:53
  the
• 1:53
  promised
• 1:54
  2.3x
• 1:55
  speed
• 1:55
  boost.
• 1:56
  The
• 1:56
  key
• 1:56
  distinction
• 1:57
  from
• 1:57
  standard
• 1:57
  Q4
• 1:58
  K
• 1:58
  M
• 1:58
  or
• 1:59
  IQ4
• 1:59
  XS
• 2:00
  quants
• 2:01
  is
• 2:01
  that
• 2:01
  NVFP4
• 2:03
  is
• 2:04
  designed
• 2:04
  for
• 2:04
  models
• 2:05
  trained
• 2:05
  in
• 2:05
  that
• 2:05
  format
• 2:06
  via
• 2:06
  ModelOpt,
• 2:07
  not
• 2:07
  for
• 2:07
  post-training
• 2:08
  quantization.
• 2:09
  When
• 2:10
  the
• 2:10
  CUDA
• 2:10
  path
• 2:10
  lands,
• 2:11
  Blackwell
• 2:12
  owners
• 2:12
  should
• 2:12
  see
• 2:12
  both
• 2:13
  quality
• 2:13
  and
• 2:13
  throughput
• 2:14
  advantages
• 2:14
  over
• 2:15
  traditional
• 2:15
  4-bit
• 2:16
  quants.
• 2:17
  StepFun
• 2:17
  continues
• 2:18
  its
• 2:18
  quiet
• 2:18
  streak
• 2:18
  of
• 2:19
  open
• 2:19
  releases
• 2:19
  with
• 2:19
  Step-3.5-Flash-Base
• 2:21
  and
• 2:21
  a
• 2:21
  midtrain
• 2:22
  checkpoint,
• 2:23
  plus
• 2:23
  training
• 2:23
  code
• 2:24
  and
• 2:24
  a
• 2:24
  promise
• 2:24
  of
• 2:25
  SFT
• 2:25
  data
• 2:25
  soon
• 2:25
  —
• 2:26
  the
• 2:27
  kind
• 2:27
  of
• 2:27
  transparent
• 2:27
  release
• 2:28
  cadence
• 2:28
  that
• 2:28
  earns
• 2:29
  community
• 2:29
  trust
• 2:29
  .
• 2:30
  Meanwhile,
• 2:31
  the
• 2:31
  uncensored-model
• 2:32
  cottage
• 2:32
  industry
• 2:33
  keeps
• 2:33
  producing:
• 2:34
  a
• 2:34
  Qwen3.5-9B
• 2:38
  "Aggressive"
• 2:38
  GGUF
• 2:39
  hit
• 2:39
  zero
• 2:39
  refusals
• 2:40
  across
• 2:40
  465
• 2:41
  test
• 2:42
  prompts
• 2:42
  with
• 2:42
  zero
• 2:42
  reported
• 2:43
  capability
• 2:43
  loss,
• 2:45
  leveraging
• 2:45
  the
• 2:45
  new
• 2:45
  hybrid
• 2:46
  Gated
• 2:46
  DeltaNet
• 2:47
  +
• 2:47
  softmax
• 2:48
  architecture
• 2:48
  with
• 2:48
  262K
• 2:50
  native
• 2:50
  context
• 2:50
  and
• 2:51
  multimodal
• 2:51
  support.
• 2:53
  The
• 2:53
  27B
• 2:53
  and
• 2:54
  35B
• 2:54
  variants
• 2:55
  are
• 2:55
  reportedly
• 2:55
  in
• 2:55
  progress
• 2:56
  .
• 2:58
  If
• 2:58
  AI
• 2:59
  agents
• 2:59
  are
• 2:59
  going
• 2:59
  to
• 2:59
  call
• 2:59
  APIs,
• 3:00
  manage
• 3:01
  credentials,
• 3:02
  and
• 3:02
  act
• 3:02
  on
• 3:02
  behalf
• 3:02
  of
• 3:02
  users,
• 3:03
  someone
• 3:03
  eventually
• 3:04
  has
• 3:04
  to
• 3:04
  formalize
• 3:04
  how
• 3:04
  they
• 3:05
  authenticate.
• 3:06
  An
• 3:06
  IETF
• 3:06
  Internet-Draft
• 3:07
  published
• 3:07
  March
• 3:08
  2,
• 3:08
  2026
• 3:09
  by
• 3:09
  authors
• 3:10
  from
• 3:10
  Defakto
• 3:10
  Security,
• 3:11
  AWS,
• 3:11
  Zscaler,
• 3:12
  and
• 3:12
  Ping
• 3:12
  Identity
• 3:13
  does
• 3:13
  exactly
• 3:13
  that.
• 3:14
  Draft
• 3:15
  `draft-klrc-aiagent-auth-00`
• 3:16
  proposes
• 3:17
  treating
• 3:17
  AI
• 3:17
  agents
• 3:18
  as
• 3:18
  workloads
• 3:18
  under
• 3:18
  the
• 3:19
  existing
• 3:19
  WIMSE
• 3:19
  (Workload
• 3:20
  Identity
• 3:20
  in
• 3:20
  Multi-System
• 3:21
  Environments)
• 3:21
  architecture
• 3:22
  and
• 3:22
  composing
• 3:22
  OAuth
• 3:23
  2.0,
• 3:24
  Transaction
• 3:24
  Tokens,
• 3:25
  and
• 3:25
  HTTP
• 3:26
  Message
• 3:26
  Signatures
• 3:27
  into
• 3:27
  a
• 3:27
  coherent
• 3:27
  agent
• 3:28
  auth
• 3:28
  framework.
• 3:29
  The
• 3:29
  document
• 3:30
  explicitly
• 3:30
  avoids
• 3:31
  inventing
• 3:31
  new
• 3:31
  protocols
• 3:32
  —
• 3:32
  instead
• 3:33
  mapping
• 3:33
  the
• 3:33
  agent
• 3:34
  lifecycle
• 3:34
  (identity
• 3:34
  provisioning,
• 3:35
  credential
• 3:36
  attestation,
• 3:37
  transport-layer
• 3:37
  and
• 3:38
  application-layer
• 3:38
  authentication,
• 3:39
  delegated
• 3:40
  authorization,
• 3:41
  human-in-the-loop
• 3:41
  gating,
• 3:42
  cross-domain
• 3:43
  access,
• 3:43
  and
• 3:43
  monitoring/remediation)
• 3:45
  onto
• 3:45
  standards
• 3:45
  that
• 3:45
  already
• 3:46
  exist.
• 3:47
  The
• 3:47
  gaps
• 3:47
  it
• 3:47
  identifies
• 3:48
  —
• 3:48
  like
• 3:49
  how
• 3:49
  to
• 3:49
  handle
• 3:49
  tool-to-service
• 3:50
  authorization
• 3:50
  when
• 3:51
  an
• 3:51
  agent
• 3:51
  chains
• 3:51
  MCP
• 3:52
  calls
• 3:52
  across
• 3:52
  trust
• 3:53
  boundaries
• 3:53
  —
• 3:54
  are
• 3:54
  precisely
• 3:54
  the
• 3:55
  unsolved
• 3:55
  problems
• 3:55
  that
• 3:56
  matter.
• 3:57
  This
• 3:57
  is
• 3:57
  the
• 3:57
  kind
• 3:57
  of
• 3:57
  boring,
• 3:58
  essential
• 3:58
  infrastructure
• 3:59
  work
• 3:59
  that
• 3:59
  prevents
• 4:00
  the
• 4:00
  agent
• 4:00
  ecosystem
• 4:00
  from
• 4:01
  becoming
• 4:01
  an
• 4:01
  authentication
• 4:02
  free-for-all.
• 4:03
  On
• 4:03
  the
• 4:03
  defensive
• 4:04
  tooling
• 4:04
  side,
• 4:05
  a
• 4:05
  detailed
• 4:05
  walkthrough
• 4:05
  of
• 4:06
  running
• 4:06
  Claude
• 4:06
  Code
• 4:06
  inside
• 4:07
  `nono`,
• 4:07
  an
• 4:07
  open-source
• 4:08
  kernel-level
• 4:09
  sandbox,
• 4:09
  on
• 4:09
  Red
• 4:10
  Hat
• 4:10
  OpenShift
• 4:10
  demonstrates
• 4:11
  both
• 4:11
  the
• 4:11
  promise
• 4:12
  and
• 4:12
  rough
• 4:12
  edges.
• 4:13
  Nono
• 4:13
  uses
• 4:14
  Linux
• 4:14
  Landlock
• 4:14
  (and
• 4:15
  macOS
• 4:15
  Seatbelt)
• 4:16
  to
• 4:16
  enforce
• 4:16
  irreversible
• 4:17
  filesystem
• 4:17
  restrictions
• 4:18
  at
• 4:18
  the
• 4:18
  syscall
• 4:18
  level
• 4:19
  —
• 4:19
  no
• 4:19
  escape
• 4:20
  hatch,
• 4:20
  no
• 4:20
  API
• 4:21
  to
• 4:21
  widen
• 4:21
  permissions
• 4:22
  after
• 4:22
  they're
• 4:22
  set.
• 4:23
  The
• 4:23
  walkthrough
• 4:23
  reveals
• 4:24
  real
• 4:24
  friction:
• 4:25
  Claude
• 4:25
  Code
• 4:25
  (built
• 4:25
  on
• 4:26
  Bun)
• 4:26
  segfaulted
• 4:26
  because
• 4:27
  the
• 4:27
  default
• 4:27
  claude-code
• 4:27
  profile
• 4:28
  didn't
• 4:28
  whitelist
• 4:28
  `/dev/urandom`,
• 4:31
  `/dev/null`,
• 4:34
  and
• 4:34
  `/dev/tty`
• 4:37
  that
• 4:37
  Bun's
• 4:37
  JSC
• 4:38
  engine
• 4:38
  needs
• 4:38
  at
• 4:38
  startup.
• 4:39
  Once
• 4:40
  patched
• 4:40
  with
• 4:40
  explicit
• 4:41
  `--allow-read`
• 4:42
  flags
• 4:42
  for
• 4:42
  those
• 4:42
  pseudodevices,
• 4:43
  the
• 4:43
  sandbox
• 4:44
  successfully
• 4:44
  blocked
• 4:45
  `pkexec
• 4:45
  ls
• 4:46
  /root`
• 4:46
  attempts,
• 4:47
  caught
• 4:47
  an
• 4:47
  incorrect
• 4:47
  `KUBECONFIG`
• 4:50
  path,
• 4:50
  and
• 4:50
  let
• 4:51
  Claude
• 4:51
  deploy
• 4:51
  a
• 4:51
  Flask
• 4:52
  app
• 4:52
  to
• 4:52
  OpenShift
• 4:52
  with
• 4:52
  proper
• 4:53
  auditing
• 4:53
  and
• 4:53
  atomic
• 4:54
  rollback
• 4:54
  support.
• 4:55
  The
• 4:55
  immutable
• 4:56
  audit
• 4:56
  trail
• 4:56
  with
• 4:56
  Merkle-root
• 4:57
  verification
• 4:58
  is
• 4:58
  particularly
• 4:58
  compelling
• 4:59
  for
• 4:59
  enterprise
• 4:59
  compliance.
• 5:01
  Without
• 5:01
  the
• 5:01
  sandbox,
• 5:02
  Claude
• 5:02
  happily
• 5:03
  ran
• 5:03
  `pkexec
• 5:03
  ls
• 5:04
  /root`
• 5:04
  after
• 5:04
  prompting
• 5:05
  for
• 5:05
  authentication
• 5:05
  —
• 5:06
  a
• 5:06
  reminder
• 5:07
  that
• 5:07
  the
• 5:07
  agent
• 5:08
  permission
• 5:08
  model
• 5:08
  today
• 5:09
  is
• 5:09
  essentially
• 5:10
  "whatever
• 5:10
  the
• 5:10
  user
• 5:11
  approves
• 5:11
  in
• 5:11
  the
• 5:11
  moment,"
• 5:12
  which
• 5:12
  is
• 5:12
  not
• 5:13
  a
• 5:13
  security
• 5:13
  model
• 5:13
  at
• 5:13
  all.
• 5:14
  Zero
• 5:15
  Day
• 5:15
  Clock
• 5:15
  tracks
• 5:15
  the
• 5:16
  accelerating
• 5:16
  collapse
• 5:16
  of
• 5:17
  time-to-exploit
• 5:17
  —
• 5:18
  the
• 5:18
  gap
• 5:18
  between
• 5:19
  vulnerability
• 5:19
  disclosure
• 5:20
  and
• 5:20
  first
• 5:20
  confirmed
• 5:21
  exploitation
• 5:21
  —
• 5:22
  using
• 5:23
  3,500+
• 5:24
  CVE-exploit
• 5:25
  pairs
• 5:25
  from
• 5:25
  CISA
• 5:25
  KEV
• 5:26
  and
• 5:26
  VulnCheck.
• 5:27
  The
• 5:27
  median
• 5:28
  TTE
• 5:28
  fell
• 5:28
  from
• 5:29
  771
• 5:30
  days
• 5:30
  in
• 5:30
  2018
• 5:31
  to
• 5:31
  4
• 5:31
  hours
• 5:32
  in
• 5:32
  2024;
• 5:33
  by
• 5:33
  2026,
• 5:35
  67%
• 5:36
  of
• 5:36
  exploited
• 5:36
  CVEs
• 5:37
  are
• 5:37
  zero-days.
• 5:38
  The
• 5:38
  site
• 5:39
  pairs
• 5:39
  this
• 5:39
  data
• 5:39
  with
• 5:40
  a
• 5:40
  historical
• 5:40
  timeline
• 5:41
  of
• 5:41
  how
• 5:41
  AI
• 5:41
  accelerated
• 5:42
  the
• 5:42
  collapse,
• 5:43
  a
• 5:43
  searchable
• 5:44
  CVE
• 5:44
  explorer,
• 5:45
  and
• 5:45
  a
• 5:45
  ten-point
• 5:46
  call
• 5:46
  to
• 5:46
  action
• 5:46
  —
• 5:47
  from
• 5:47
  vendor
• 5:47
  liability
• 5:48
  and
• 5:48
  memory-safe
• 5:49
  mandates
• 5:49
  to
• 5:49
  open-source
• 5:50
  defense
• 5:50
  and
• 5:50
  machine-speed
• 5:51
  regulation
• 5:51
  —
• 5:52
  backed
• 5:53
  by
• 5:53
  40+
• 5:53
  signatories
• 5:54
  including
• 5:54
  Schneier,
• 5:55
  Adkins,
• 5:55
  Jeff
• 5:56
  Moss,
• 5:56
  and
• 5:56
  Halvar
• 5:56
  Flake.
• 5:57
  (more:
• 5:58
  zerodayclock.com)
• 6:02
  A
• 6:02
  provocatively
• 6:02
  titled
• 6:03
  essay
• 6:03
  argues
• 6:03
  that
• 6:03
  the
• 6:03
  word
• 6:04
  missing
• 6:04
  from
• 6:04
  the
• 6:04
  LLM
• 6:05
  discourse
• 6:05
  is
• 6:05
  forgery.
• 6:06
  The
• 6:06
  core
• 6:06
  claim:
• 6:08
  LLMs
• 6:08
  allow
• 6:08
  individuals
• 6:09
  to
• 6:09
  produce
• 6:09
  forgeries
• 6:09
  of
• 6:10
  their
• 6:10
  own
• 6:10
  potential
• 6:10
  output
• 6:11
  faster
• 6:11
  than
• 6:11
  they
• 6:11
  could
• 6:11
  make
• 6:12
  it
• 6:12
  themselves,
• 6:13
  and
• 6:13
  the
• 6:13
  only
• 6:13
  way
• 6:13
  to
• 6:13
  separate
• 6:14
  gold
• 6:14
  from
• 6:14
  slop
• 6:14
  is
• 6:15
  source
• 6:15
  attribution
• 6:15
  during
• 6:16
  inference
• 6:16
  —
• 6:16
  a
• 6:16
  technical
• 6:17
  impossibility
• 6:17
  with
• 6:18
  current
• 6:18
  architectures.
• 6:19
  The
• 6:19
  piece
• 6:19
  draws
• 6:19
  a
• 6:19
  line
• 6:19
  from
• 6:20
  controlled-appellation
• 6:20
  cheese
• 6:21
  (where
• 6:21
  geographic
• 6:21
  origin
• 6:22
  protects
• 6:22
  quality
• 6:23
  and
• 6:23
  expertise)
• 6:23
  to
• 6:23
  open-source
• 6:24
  code
• 6:24
  maintenance
• 6:25
  (where
• 6:25
  slop
• 6:25
  PRs
• 6:25
  from
• 6:25
  resume-padding
• 6:26
  vibe-coders
• 6:27
  are
• 6:27
  forcing
• 6:27
  projects
• 6:27
  to
• 6:27
  ban
• 6:28
  or
• 6:28
  ignore
• 6:28
  AI-generated
• 6:29
  contributions).
• 6:30
  The
• 6:30
  observation
• 6:31
  that
• 6:31
  experienced
• 6:32
  engineers
• 6:32
  who
• 6:32
  adopt
• 6:32
  AI
• 6:33
  coding
• 6:33
  still
• 6:34
  produce
• 6:34
  "highly
• 6:35
  embarrassing
• 6:35
  goofs"
• 6:35
  rings
• 6:36
  true
• 6:36
  for
• 6:36
  anyone
• 6:36
  who's
• 6:36
  reviewed
• 6:37
  a
• 6:37
  PR
• 6:37
  where
• 6:37
  the
• 6:37
  thought
• 6:38
  process
• 6:38
  was
• 6:38
  clearly
• 6:38
  "none
• 6:39
  at
• 6:39
  all."
• 6:40
  The
• 6:40
  proposed
• 6:40
  fix
• 6:40
  —
• 6:41
  making
• 6:41
  backpropagation
• 6:42
  auditable
• 6:42
  and
• 6:43
  weights
• 6:43
  attributable
• 6:43
  —
• 6:44
  is
• 6:44
  a
• 6:44
  research
• 6:44
  moonshot,
• 6:45
  but
• 6:45
  the
• 6:45
  diagnosis
• 6:45
  is
• 6:46
  sharp:
• 6:46
  a
• 6:46
  technology
• 6:46
  that
• 6:47
  cannot
• 6:47
  tell
• 6:47
  you
• 6:47
  where
• 6:47
  its
• 6:48
  information
• 6:48
  comes
• 6:48
  from
• 6:48
  is,
• 6:48
  by
• 6:49
  design,
• 6:49
  sloppy.
• 6:50
  That
• 6:50
  diagnosis
• 6:51
  gets
• 6:51
  empirical
• 6:51
  support
• 6:52
  from
• 6:52
  a
• 6:52
  candid
• 6:52
  founders'
• 6:53
  conversation
• 6:54
  about
• 6:54
  the
• 6:54
  "massive
• 6:54
  discrepancy
• 6:55
  between
• 6:55
  people's
• 6:55
  AI
• 6:56
  world
• 6:56
  views
• 6:56
  on
• 6:56
  software
• 6:56
  engineering."
• 6:57
  One
• 6:58
  founder
• 6:58
  describes
• 6:58
  seeing
• 6:58
  diametrically
• 6:59
  opposed
• 6:59
  LinkedIn
• 7:00
  posts
• 7:00
  —
• 7:00
  one
• 7:01
  bragging
• 7:01
  about
• 7:01
  10x
• 7:02
  commit
• 7:02
  rates,
• 7:02
  another
• 7:02
  calling
• 7:03
  AI
• 7:03
  output
• 7:03
  a
• 7:03
  mountain
• 7:03
  of
• 7:04
  junk
• 7:04
  —
• 7:04
  and
• 7:04
  not
• 7:05
  knowing
• 7:05
  which
• 7:05
  universe
• 7:05
  is
• 7:05
  real.
• 7:06
  The
• 7:06
  answer,
• 7:07
  they
• 7:07
  conclude,
• 7:07
  is
• 7:07
  both:
• 7:08
  AI
• 7:08
  produces
• 7:09
  genius-level
• 7:09
  fixes
• 7:10
  on
• 7:10
  prototypical
• 7:10
  problems
• 7:11
  but
• 7:11
  hallucinates
• 7:11
  entire
• 7:12
  background
• 7:12
  processes
• 7:13
  for
• 7:13
  anything
• 7:13
  with
• 7:13
  actual
• 7:13
  stateful
• 7:14
  logic.
• 7:15
  One
• 7:15
  participant
• 7:16
  recounts
• 7:16
  an
• 7:16
  AI
• 7:16
  that
• 7:16
  concluded
• 7:16
  —
• 7:17
  entirely
• 7:17
  from
• 7:17
  thin
• 7:18
  air
• 7:18
  —
• 7:18
  that
• 7:18
  a
• 7:18
  "get
• 7:19
  transitions"
• 7:19
  function
• 7:20
  only
• 7:20
  returns
• 7:20
  automated
• 7:21
  transitions,
• 7:21
  then
• 7:21
  started
• 7:22
  tearing
• 7:22
  the
• 7:22
  codebase
• 7:22
  apart
• 7:23
  based
• 7:23
  on
• 7:23
  that
• 7:23
  hallucination.
• 7:25
  The
• 7:25
  session's
• 7:25
  verdict:
• 7:26
  "Software
• 7:26
  engineering
• 7:26
  is
• 7:26
  too
• 7:27
  large
• 7:27
  to
• 7:27
  fit
• 7:27
  into
• 7:27
  the
• 7:27
  context
• 7:28
  window
• 7:28
  of
• 7:28
  an
• 7:28
  AI.
• 7:29
  The
• 7:29
  AI
• 7:29
  is
• 7:29
  always
• 7:30
  running
• 7:30
  with
• 7:30
  tunnel
• 7:30
  vision,
• 7:31
  like
• 7:31
  a
• 7:31
  guy
• 7:31
  driving
• 7:32
  a
• 7:32
  tank."
• 7:32
  A
• 7:33
  detailed
• 7:33
  video
• 7:33
  analysis
• 7:34
  puts
• 7:34
  numbers
• 7:34
  to
• 7:34
  the
• 7:34
  adoption
• 7:35
  gap.
• 7:36
  While
• 7:36
  90%
• 7:36
  of
• 7:36
  engineering
• 7:37
  teams
• 7:37
  nominally
• 7:38
  "use
• 7:38
  AI
• 7:39
  in
• 7:39
  their
• 7:39
  workflow,"
• 7:40
  only
• 7:40
  51%
• 7:41
  of
• 7:41
  professional
• 7:41
  developers
• 7:42
  use
• 7:42
  AI
• 7:42
  tools
• 7:43
  daily
• 7:43
  per
• 7:43
  Stack
• 7:43
  Overflow.
• 7:44
  A
• 7:44
  full
• 7:45
  76%
• 7:45
  of
• 7:45
  executives
• 7:46
  believe
• 7:46
  their
• 7:46
  teams
• 7:47
  have
• 7:47
  embraced
• 7:47
  AI
• 7:48
  while
• 7:48
  only
• 7:48
  52%
• 7:49
  of
• 7:49
  engineers
• 7:50
  agree.
• 7:51
  And
• 7:51
  21%
• 7:51
  of
• 7:51
  AI
• 7:52
  coding
• 7:52
  licenses
• 7:52
  remain
• 7:53
  underutilized.
• 7:54
  Enterprise
• 7:55
  adoption
• 7:55
  faces
• 7:56
  corporate
• 7:56
  red
• 7:56
  tape,
• 7:57
  procurement
• 7:57
  processes,
• 7:58
  security
• 7:59
  reviews,
• 7:59
  and
• 7:59
  senior
• 8:00
  engineers
• 8:00
  who
• 8:00
  rightly
• 8:01
  distrust
• 8:01
  hallucination-prone
• 8:02
  tools
• 8:02
  for
• 8:02
  production-grade
• 8:03
  systems.
• 8:04
  The
• 8:04
  evolution
• 8:05
  from
• 8:05
  prompt
• 8:05
  engineering
• 8:06
  to
• 8:06
  context
• 8:06
  engineering
• 8:07
  to
• 8:07
  "intent
• 8:07
  engineering"
• 8:07
  —
• 8:08
  where
• 8:09
  developers
• 8:09
  define
• 8:09
  architecture,
• 8:10
  success
• 8:11
  criteria,
• 8:11
  and
• 8:12
  validation
• 8:12
  strategies
• 8:13
  while
• 8:13
  delegating
• 8:13
  only
• 8:14
  the
• 8:14
  typing
• 8:14
  —
• 8:15
  looks
• 8:15
  less
• 8:15
  like
• 8:16
  engineering
• 8:16
  extinction
• 8:16
  and
• 8:17
  more
• 8:17
  like
• 8:17
  engineering
• 8:18
  with
• 8:18
  a
• 8:18
  very
• 8:18
  demanding
• 8:19
  junior
• 8:19
  developer.
• 8:20
  Anthropic's
• 8:21
  own
• 8:21
  CEO
• 8:21
  says
• 8:21
  software
• 8:22
  engineers
• 8:22
  could
• 8:22
  "go
• 8:23
  extinct"
• 8:23
  in
• 8:23
  2026;
• 8:25
  the
• 8:25
  creator
• 8:25
  of
• 8:25
  Claude
• 8:26
  Code
• 8:26
  says
• 8:26
  "great
• 8:26
  engineers
• 8:27
  are
• 8:27
  more
• 8:27
  important
• 8:28
  than
• 8:28
  ever."
• 8:29
  That
• 8:29
  contradiction
• 8:29
  from
• 8:30
  within
• 8:30
  the
• 8:30
  same
• 8:30
  company
• 8:31
  tells
• 8:31
  you
• 8:31
  everything
• 8:31
  about
• 8:31
  the
• 8:32
  incentive
• 8:32
  structures
• 8:32
  at
• 8:33
  play.
• 8:35
  Abnormal
• 8:36
  AI's
• 8:36
  engineering
• 8:36
  team
• 8:37
  has
• 8:37
  gone
• 8:37
  further
• 8:38
  than
• 8:38
  most
• 8:38
  in
• 8:38
  operationalizing
• 8:39
  spec-driven
• 8:39
  development.
• 8:41
  Their
• 8:41
  system
• 8:41
  starts
• 8:41
  with
• 8:42
  four
• 8:42
  markdown
• 8:42
  files
• 8:43
  in
• 8:43
  the
• 8:43
  repo
• 8:44
  —
• 8:45
  `ARCHITECTURE.md`,
• 8:46
  `LEGAL.md`,
• 8:47
  `SECURITY.md`,
• 8:48
  and
• 8:48
  a
• 8:48
  spec
• 8:49
  template
• 8:49
  —
• 8:49
  that
• 8:49
  encode
• 8:50
  organizational
• 8:50
  constraints.
• 8:52
  A
• 8:52
  CLI
• 8:52
  built
• 8:53
  on
• 8:53
  the
• 8:53
  Claude
• 8:53
  Code
• 8:53
  SDK
• 8:54
  ingests
• 8:54
  anything
• 8:55
  from
• 8:55
  a
• 8:55
  full
• 8:56
  PRD
• 8:56
  to
• 8:56
  a
• 8:56
  few
• 8:56
  bullet
• 8:57
  points
• 8:57
  and
• 8:57
  generates
• 8:57
  a
• 8:57
  two-audience
• 8:58
  technical
• 8:58
  plan:
• 8:59
  skimmable
• 8:59
  architecture
• 9:00
  decisions
• 9:00
  for
• 9:00
  human
• 9:01
  reviewers
• 9:01
  up
• 9:01
  front,
• 9:02
  detailed
• 9:02
  function
• 9:03
  signatures
• 9:03
  and
• 9:03
  verification
• 9:04
  checklists
• 9:04
  for
• 9:04
  agents
• 9:05
  in
• 9:05
  the
• 9:05
  back.
• 9:06
  The
• 9:06
  clever
• 9:06
  bit
• 9:07
  is
• 9:07
  the
• 9:07
  feedback
• 9:07
  flywheel:
• 9:08
  a
• 9:08
  Zoom
• 9:09
  bot
• 9:09
  records
• 9:09
  every
• 9:10
  design
• 9:10
  review,
• 9:11
  an
• 9:11
  agent
• 9:11
  processes
• 9:12
  recordings
• 9:12
  plus
• 9:13
  Slack
• 9:13
  threads
• 9:13
  and
• 9:13
  PR
• 9:14
  comments
• 9:14
  weekly,
• 9:15
  and
• 9:15
  generates
• 9:15
  a
• 9:15
  PR
• 9:16
  updating
• 9:16
  the
• 9:16
  system
• 9:17
  files.
• 9:18
  Every
• 9:18
  review
• 9:18
  trains
• 9:19
  the
• 9:19
  system,
• 9:19
  not
• 9:19
  just
• 9:20
  evaluates
• 9:20
  a
• 9:20
  single
• 9:21
  plan.
• 9:21
  Non-engineers
• 9:22
  —
• 9:23
  PMs,
• 9:24
  data
• 9:24
  analysts
• 9:24
  —
• 9:25
  are
• 9:25
  now
• 9:25
  self-serving
• 9:26
  production
• 9:26
  changes
• 9:26
  through
• 9:27
  this
• 9:27
  pipeline.
• 9:28
  They're
• 9:28
  open-sourcing
• 9:28
  the
• 9:28
  template
• 9:29
  as
• 9:29
  a
• 9:29
  Claude
• 9:29
  skill.
• 9:30
  The
• 9:30
  "context
• 9:31
  compilation"
• 9:31
  approach
• 9:32
  takes
• 9:32
  a
• 9:32
  different
• 9:32
  path
• 9:33
  to
• 9:33
  the
• 9:33
  same
• 9:33
  goal.
• 9:35
  Instead
• 9:35
  of
• 9:35
  stuffing
• 9:35
  a
• 9:35
  million
• 9:36
  tokens
• 9:36
  into
• 9:36
  a
• 9:36
  context
• 9:37
  window,
• 9:37
  one
• 9:38
  practitioner
• 9:38
  built
• 9:39
  a
• 9:39
  Recursive
• 9:39
  Language
• 9:40
  Model
• 9:40
  Gateway
• 9:41
  that
• 9:41
  loads
• 9:41
  an
• 9:41
  entire
• 9:42
  repo
• 9:42
  into
• 9:42
  a
• 9:42
  REPL
• 9:43
  workspace,
• 9:43
  writes
• 9:44
  programmatic
• 9:44
  code
• 9:45
  to
• 9:45
  walk,
• 9:45
  slice,
• 9:46
  and
• 9:46
  search
• 9:46
  it,
• 9:47
  then
• 9:47
  hands
• 9:47
  only
• 9:47
  a
• 9:47
  tiny
• 9:48
  compiled
• 9:48
  context
• 9:49
  pack
• 9:49
  to
• 9:49
  the
• 9:49
  downstream
• 9:50
  model
• 9:50
  —
• 9:50
  treating
• 9:51
  tokens
• 9:51
  as
• 9:51
  CPU,
• 9:52
  not
• 9:53
  storage.
• 9:54
  The
• 9:54
  result:
• 9:55
  Claude
• 9:55
  and
• 9:56
  MiniMax
• 9:56
  "suddenly
• 9:57
  act
• 9:57
  like
• 9:57
  they've
• 9:57
  been
• 9:57
  on
• 9:57
  the
• 9:57
  project
• 9:58
  for
• 9:58
  2
• 9:58
  years"
• 9:59
  without
• 9:59
  the
• 9:59
  architecture
• 10:00
  drift
• 10:00
  that
• 10:01
  comes
• 10:01
  from
• 10:01
  trying
• 10:01
  to
• 10:01
  grep
• 10:01
  across
• 10:02
  an
• 10:02
  entire
• 10:02
  codebase
• 10:03
  inside
• 10:03
  a
• 10:03
  prompt.
• 10:04
  Seine,
• 10:05
  an
• 10:05
  open-source
• 10:06
  agentic
• 10:06
  deep-search
• 10:07
  orchestrator,
• 10:08
  attacks
• 10:08
  the
• 10:08
  context
• 10:08
  quality
• 10:09
  problem
• 10:09
  with
• 10:09
  20
• 10:10
  specialized
• 10:11
  agents
• 10:11
  in
• 10:11
  a
• 10:11
  3-phase
• 10:12
  gated
• 10:12
  pipeline:
• 10:13
  Phase
• 10:14
  A
• 10:14
  discovers
• 10:14
  evidence
• 10:15
  across
• 10:15
  9
• 10:15
  domains,
• 10:16
  Phase
• 10:16
  B
• 10:16
  deploys
• 10:17
  a
• 10:17
  dedicated
• 10:18
  Skeptic
• 10:18
  agent
• 10:18
  to
• 10:18
  challenge
• 10:19
  every
• 10:19
  claim
• 10:19
  while
• 10:20
  a
• 10:20
  Referee
• 10:20
  resolves
• 10:20
  conflicts,
• 10:21
  and
• 10:22
  Phase
• 10:22
  C
• 10:22
  runs
• 10:22
  adversarial
• 10:23
  red-team
• 10:24
  analysis
• 10:24
  with
• 10:24
  calibrated
• 10:25
  confidence
• 10:25
  scores.
• 10:26
  Validators
• 10:27
  at
• 10:27
  each
• 10:27
  gate
• 10:28
  push
• 10:28
  back
• 10:28
  for
• 10:28
  more
• 10:28
  evidence
• 10:29
  or
• 10:29
  stop
• 10:29
  the
• 10:29
  pipeline
• 10:30
  with
• 10:30
  a
• 10:30
  partial
• 10:30
  result
• 10:31
  and
• 10:31
  a
• 10:31
  clear
• 10:31
  label.
• 10:32
  The
• 10:32
  adversarial
• 10:33
  architecture
• 10:34
  is
• 10:34
  the
• 10:34
  right
• 10:34
  instinct
• 10:34
  —
• 10:35
  if
• 10:35
  context
• 10:36
  is
• 10:36
  king,
• 10:36
  bad
• 10:37
  context
• 10:37
  is
• 10:37
  poison
• 10:38
  .
• 10:39
  On
• 10:39
  a
• 10:39
  complementary
• 10:40
  note,
• 10:40
  the
• 10:40
  Semantic
• 10:41
  Anchors
• 10:41
  project
• 10:42
  catalogs
• 10:43
  46
• 10:43
  well-defined
• 10:44
  terms
• 10:44
  and
• 10:44
  frameworks
• 10:44
  —
• 10:45
  from
• 10:46
  "TDD,
• 10:47
  London
• 10:47
  School"
• 10:47
  to
• 10:48
  "Clean
• 10:48
  Architecture"
• 10:48
  —
• 10:49
  that
• 10:49
  serve
• 10:50
  as
• 10:50
  shared
• 10:50
  vocabulary
• 10:51
  between
• 10:51
  humans
• 10:51
  and
• 10:52
  LLMs,
• 10:53
  compressing
• 10:53
  complex
• 10:54
  context
• 10:54
  into
• 10:54
  tokens
• 10:55
  that
• 10:55
  reliably
• 10:56
  activate
• 10:56
  the
• 10:56
  right
• 10:56
  knowledge
• 10:57
  domains
• 10:57
  .
• 11:01
  Simon
• 11:02
  Willison's
• 11:02
  new
• 11:02
  guide
• 11:02
  on
• 11:03
  agentic
• 11:03
  engineering
• 11:03
  patterns
• 11:04
  distills
• 11:04
  hard-won
• 11:05
  lessons
• 11:05
  into
• 11:05
  a
• 11:05
  compact
• 11:06
  reference
• 11:06
  for
• 11:06
  getting
• 11:06
  the
• 11:06
  best
• 11:07
  results
• 11:07
  from
• 11:07
  Claude
• 11:07
  Code
• 11:08
  and
• 11:08
  OpenAI
• 11:08
  Codex.
• 11:09
  The
• 11:09
  framing
• 11:10
  —
• 11:11
  "writing
• 11:11
  code
• 11:11
  is
• 11:12
  cheap
• 11:12
  now"
• 11:12
  and
• 11:13
  "hoard
• 11:13
  things
• 11:13
  you
• 11:13
  know
• 11:13
  how
• 11:13
  to
• 11:14
  do"
• 11:14
  —
• 11:15
  captures
• 11:15
  the
• 11:15
  shift
• 11:15
  from
• 11:16
  treating
• 11:16
  code
• 11:16
  as
• 11:16
  precious
• 11:17
  output
• 11:17
  to
• 11:17
  treating
• 11:17
  it
• 11:17
  as
• 11:17
  disposable
• 11:18
  intermediate
• 11:18
  artifact.
• 11:19
  The
• 11:19
  guide
• 11:20
  joins
• 11:20
  a
• 11:20
  growing
• 11:20
  library
• 11:21
  of
• 11:21
  practitioner-driven
• 11:21
  pattern
• 11:22
  catalogs
• 11:22
  that
• 11:22
  are
• 11:22
  quietly
• 11:23
  becoming
• 11:23
  more
• 11:23
  useful
• 11:24
  than
• 11:24
  vendor
• 11:24
  documentation.
• 11:25
  Anthropic's
• 11:25
  new
• 11:25
  Playground
• 11:26
  plugin
• 11:26
  for
• 11:26
  Claude
• 11:26
  Code
• 11:27
  lets
• 11:27
  users
• 11:27
  build
• 11:27
  custom
• 11:28
  interactive
• 11:28
  HTML
• 11:29
  tools
• 11:29
  —
• 11:29
  sliders,
• 11:30
  color
• 11:30
  pickers,
• 11:31
  click
• 11:31
  regions,
• 11:31
  approve/reject
• 11:32
  buttons
• 11:32
  —
• 11:33
  that
• 11:33
  generate
• 11:33
  natural-language
• 11:34
  prompts
• 11:34
  at
• 11:34
  the
• 11:34
  bottom.
• 11:35
  Draw
• 11:35
  rectangles
• 11:36
  on
• 11:36
  a
• 11:36
  UI
• 11:36
  screenshot
• 11:37
  to
• 11:37
  create
• 11:37
  location-aware
• 11:38
  instructions;
• 11:39
  paste
• 11:39
  a
• 11:39
  document
• 11:40
  for
• 11:40
  inline
• 11:40
  critique
• 11:41
  with
• 11:41
  approve/reject
• 11:42
  per
• 11:42
  suggestion;
• 11:43
  dial
• 11:43
  in
• 11:43
  spacing
• 11:44
  and
• 11:44
  typography
• 11:45
  with
• 11:45
  live
• 11:45
  sliders
• 11:45
  and
• 11:45
  copy
• 11:46
  the
• 11:46
  resulting
• 11:46
  spec.
• 11:47
  Six
• 11:47
  templates
• 11:47
  ship
• 11:48
  built-in,
• 11:49
  and
• 11:49
  someone
• 11:49
  has
• 11:49
  already
• 11:49
  wired
• 11:50
  one
• 11:50
  up
• 11:50
  to
• 11:50
  send
• 11:50
  feedback
• 11:51
  directly
• 11:51
  back
• 11:51
  to
• 11:51
  Claude
• 11:52
  via
• 11:52
  an
• 11:52
  MCP
• 11:52
  server,
• 11:53
  closing
• 11:53
  the
• 11:53
  loop
• 11:54
  entirely.
• 11:55
  Open
• 11:55
  WebUI
• 11:55
  v0.8.6
• 11:58
  ships
• 11:59
  terminal
• 11:59
  integration
• 12:00
  that
• 12:00
  goes
• 12:00
  beyond
• 12:00
  tool
• 12:00
  calling
• 12:00
  —
• 12:01
  it's
• 12:01
  a
• 12:01
  full
• 12:02
  Linux+Python
• 12:03
  sandbox
• 12:03
  Docker
• 12:03
  container
• 12:04
  with
• 12:04
  file
• 12:04
  explorer,
• 12:05
  upload/download,
• 12:06
  and
• 12:06
  inline
• 12:07
  editing
• 12:07
  in
• 12:07
  the
• 12:07
  sidebar.
• 12:08
  The
• 12:08
  release
• 12:09
  also
• 12:09
  addresses
• 12:09
  what
• 12:10
  users
• 12:10
  had
• 12:10
  been
• 12:10
  calling
• 12:11
  "a
• 12:11
  bloated
• 12:11
  mess"
• 12:11
  with
• 12:12
  substantial
• 12:12
  frontend
• 12:12
  and
• 12:13
  backend
• 12:13
  performance
• 12:13
  improvements:
• 12:14
  eliminated
• 12:15
  CPU/memory
• 12:16
  hogging,
• 12:17
  fixed
• 12:17
  memory
• 12:18
  leaks,
• 12:18
  and
• 12:18
  smoothed
• 12:19
  token
• 12:19
  streaming.
• 12:20
  The
• 12:20
  comment
• 12:20
  that
• 12:21
  Open
• 12:21
  WebUI
• 12:21
  should
• 12:21
  "find
• 12:22
  courage
• 12:22
  to
• 12:22
  remove
• 12:23
  legacy
• 12:23
  parts
• 12:23
  like
• 12:23
  pipelines/pipes/filters
• 12:25
  and
• 12:26
  focus
• 12:26
  more
• 12:26
  on
• 12:26
  Skills"
• 12:27
  suggests
• 12:28
  the
• 12:28
  project
• 12:28
  is
• 12:28
  at
• 12:28
  an
• 12:28
  architectural
• 12:29
  inflection
• 12:29
  point.
• 12:30
  World
• 12:31
  Intelligence
• 12:31
  MCP
• 12:32
  Server
• 12:32
  pushes
• 12:32
  the
• 12:33
  MCP
• 12:33
  ecosystem
• 12:34
  into
• 12:34
  ambitious
• 12:35
  territory:
• 12:35
  68
• 12:36
  tools
• 12:36
  across
• 12:36
  27
• 12:37
  intelligence
• 12:37
  domains
• 12:38
  —
• 12:38
  financial
• 12:39
  markets,
• 12:39
  conflict
• 12:40
  tracking,
• 12:40
  military
• 12:41
  flight
• 12:41
  monitoring,
• 12:42
  cyber
• 12:42
  threats,
• 12:43
  space
• 12:43
  weather,
• 12:44
  disease
• 12:44
  outbreaks,
• 12:45
  shipping
• 12:45
  stress
• 12:45
  indices,
• 12:46
  and
• 12:46
  more
• 12:46
  —
• 12:47
  all
• 12:47
  from
• 12:47
  free
• 12:47
  public
• 12:48
  APIs
• 12:48
  with
• 12:48
  per-source
• 12:49
  circuit
• 12:49
  breakers
• 12:49
  and
• 12:50
  a
• 12:50
  live
• 12:50
  ops-center
• 12:50
  dashboard.
• 12:52
  It
• 12:52
  evolved
• 12:52
  from
• 12:52
  a
• 12:52
  focused
• 12:53
  threat-intel
• 12:54
  MCP
• 12:54
  server
• 12:55
  into
• 12:55
  a
• 12:55
  comprehensive
• 12:55
  OSINT
• 12:56
  platform,
• 12:57
  which
• 12:57
  is
• 12:57
  exactly
• 12:57
  the
• 12:58
  trajectory
• 12:58
  the
• 12:58
  MCP
• 12:59
  ecosystem
• 13:00
  needs:
• 13:00
  specialized
• 13:01
  servers
• 13:02
  that
• 13:02
  do
• 13:02
  one
• 13:02
  vertical
• 13:03
  extremely
• 13:03
  well
• 13:03
  .
• 13:04
  For
• 13:04
  developers
• 13:05
  building
• 13:05
  their
• 13:05
  own
• 13:06
  integrations,
• 13:07
  the
• 13:07
  public-apis
• 13:07
  repository
• 13:08
  remains
• 13:09
  the
• 13:09
  canonical
• 13:09
  directory
• 13:10
  of
• 13:10
  free
• 13:10
  APIs
• 13:11
  across
• 13:11
  40+
• 13:11
  categories
• 13:12
  .
• 13:14
  A
• 13:14
  Hacker
• 13:15
  News
• 13:15
  Show
• 13:15
  HN
• 13:15
  details
• 13:16
  building
• 13:16
  a
• 13:16
  sub-500ms
• 13:18
  latency
• 13:18
  voice
• 13:18
  agent
• 13:18
  from
• 13:19
  scratch
• 13:19
  —
• 13:19
  and
• 13:20
  beating
• 13:20
  Vapi's
• 13:21
  equivalent
• 13:21
  setup
• 13:21
  by
• 13:21
  2x.
• 13:22
  The
• 13:22
  architecture
• 13:23
  is
• 13:23
  deceptively
• 13:23
  simple:
• 13:24
  Twilio
• 13:24
  streams
• 13:25
  audio
• 13:25
  to
• 13:25
  Deepgram's
• 13:26
  Flux
• 13:26
  (which
• 13:26
  combines
• 13:27
  transcription
• 13:27
  and
• 13:28
  turn
• 13:28
  detection
• 13:28
  in
• 13:28
  a
• 13:28
  single
• 13:28
  model),
• 13:29
  end-of-turn
• 13:30
  triggers
• 13:30
  an
• 13:30
  LLM
• 13:30
  →
• 13:31
  TTS
• 13:31
  streaming
• 13:31
  pipeline,
• 13:32
  and
• 13:32
  barge-ins
• 13:33
  propagate
• 13:33
  cancellation
• 13:34
  to
• 13:34
  all
• 13:34
  components
• 13:35
  simultaneously.
• 13:36
  Three
• 13:36
  optimizations
• 13:37
  made
• 13:37
  the
• 13:37
  biggest
• 13:37
  difference.
• 13:38
  First,
• 13:39
  keeping
• 13:39
  TTS
• 13:39
  WebSocket
• 13:40
  connections
• 13:40
  warm
• 13:41
  (pre-connected
• 13:41
  pool)
• 13:41
  shaved
• 13:42
  ~300ms.
• 13:43
  Second,
• 13:44
  deploying
• 13:44
  the
• 13:44
  orchestration
• 13:45
  layer
• 13:45
  in
• 13:45
  the
• 13:45
  EU
• 13:45
  region
• 13:46
  near
• 13:46
  the
• 13:46
  API
• 13:46
  endpoints
• 13:47
  cut
• 13:47
  latency
• 13:47
  in
• 13:47
  half
• 13:48
  —
• 13:48
  geography
• 13:49
  is
• 13:49
  a
• 13:49
  first-class
• 13:49
  design
• 13:50
  parameter.
• 13:51
  Third,
• 13:51
  swapping
• 13:52
  GPT-4o-mini
• 13:53
  for
• 13:53
  Groq's
• 13:53
  Llama-3.3-70b
• 13:55
  dropped
• 13:55
  first-token
• 13:56
  latency
• 13:56
  to
• 13:56
  ~80ms
• 13:57
  (faster
• 13:57
  than
• 13:57
  a
• 13:57
  human
• 13:58
  blink),
• 13:58
  bringing
• 13:59
  end-to-end
• 13:59
  response
• 13:59
  time
• 14:00
  to
• 14:00
  ~400ms.
• 14:01
  The
• 14:01
  author's
• 14:02
  latency
• 14:02
  measurements
• 14:03
  are
• 14:03
  worth
• 14:03
  internalizing:
• 14:05
  Groq's
• 14:05
  Llama-3.3-70b
• 14:06
  delivered
• 14:07
  ~80ms
• 14:08
  time-to-first-token,
• 14:09
  which
• 14:09
  accounts
• 14:10
  for
• 14:10
  more
• 14:10
  than
• 14:10
  half
• 14:10
  of
• 14:10
  total
• 14:11
  latency.
• 14:12
  First-token
• 14:12
  latency
• 14:13
  is
• 14:13
  the
• 14:13
  metric
• 14:13
  that
• 14:13
  actually
• 14:14
  matters
• 14:14
  in
• 14:14
  voice
• 14:14
  —
• 14:15
  everything
• 14:15
  downstream
• 14:16
  (TTS
• 14:16
  synthesis,
• 14:17
  audio
• 14:17
  buffering)
• 14:17
  can
• 14:18
  be
• 14:18
  pipelined,
• 14:18
  but
• 14:19
  nothing
• 14:19
  moves
• 14:19
  until
• 14:19
  that
• 14:20
  first
• 14:20
  token
• 14:20
  arrives.
• 14:21
  The
• 14:21
  lesson:
• 14:22
  voice
• 14:22
  agents
• 14:23
  are
• 14:23
  an
• 14:23
  orchestration
• 14:24
  problem,
• 14:25
  and
• 14:25
  understanding
• 14:26
  the
• 14:26
  pipeline
• 14:26
  end-to-end
• 14:27
  lets
• 14:27
  you
• 14:27
  outperform
• 14:28
  abstraction
• 14:28
  layers
• 14:29
  that
• 14:29
  hide
• 14:29
  the
• 14:29
  complexity.
• 14:31
  For
• 14:31
  teams
• 14:31
  evaluating
• 14:32
  Vapi,
• 14:32
  ElevenLabs,
• 14:33
  or
• 14:33
  rolling
• 14:34
  their
• 14:34
  own,
• 14:35
  the
• 14:35
  takeaway
• 14:35
  is
• 14:35
  clear:
• 14:36
  measure
• 14:37
  TTFT,
• 14:38
  co-locate
• 14:38
  your
• 14:38
  services,
• 14:39
  and
• 14:39
  keep
• 14:39
  your
• 14:40
  WebSocket
• 14:40
  connections
• 14:40
  warm.
• 14:44
  A
• 14:44
  military
• 14:44
  training
• 14:45
  presentation
• 14:45
  —
• 14:45
  marked
• 14:46
  "UNCLASSIFIED
• 14:47
  //
• 14:48
  FOR
• 14:48
  TRAINING
• 14:48
  USE
• 14:49
  ONLY"
• 14:49
  —
• 14:49
  lays
• 14:50
  out
• 14:50
  how
• 14:50
  US
• 14:50
  Special
• 14:51
  Operations
• 14:51
  Forces
• 14:51
  are
• 14:51
  framing
• 14:52
  agentic
• 14:53
  AI.
• 14:53
  The
• 14:53
  document
• 14:54
  maps
• 14:54
  the
• 14:54
  OODA
• 14:55
  loop
• 14:55
  (Observe,
• 14:56
  Orient,
• 14:56
  Decide,
• 14:57
  Act)
• 14:57
  directly
• 14:57
  onto
• 14:58
  agent
• 14:58
  architectures:
• 14:59
  the
• 14:59
  agent
• 15:00
  receives
• 15:00
  a
• 15:00
  goal,
• 15:01
  gathers
• 15:01
  information,
• 15:02
  reasons
• 15:02
  about
• 15:02
  the
• 15:02
  best
• 15:03
  action
• 15:03
  sequence,
• 15:03
  executes,
• 15:04
  observes
• 15:05
  results,
• 15:05
  and
• 15:05
  loops.
• 15:06
  SOF
• 15:06
  applications
• 15:07
  include
• 15:07
  autonomous
• 15:08
  target
• 15:08
  package
• 15:08
  generation
• 15:09
  from
• 15:09
  a
• 15:09
  single
• 15:10
  tasking,
• 15:10
  same-night
• 15:11
  follow-on
• 15:11
  raids
• 15:12
  driven
• 15:12
  by
• 15:12
  AI-processed
• 15:13
  captured
• 15:13
  media,
• 15:14
  and
• 15:14
  multi-stream
• 15:15
  pattern-of-life
• 15:15
  monitoring.
• 15:16
  The
• 15:16
  presentation
• 15:17
  cites
• 15:17
  three
• 15:17
  real
• 15:17
  examples:
• 15:18
  Project
• 15:18
  Maven
• 15:19
  automating
• 15:19
  ISR
• 15:20
  video
• 15:20
  analysis
• 15:20
  (processing
• 15:21
  years
• 15:21
  of
• 15:21
  drone
• 15:22
  footage
• 15:22
  backlog
• 15:22
  in
• 15:22
  seconds),
• 15:23
  SOCOM's
• 15:24
  push
• 15:24
  to
• 15:24
  collapse
• 15:24
  sensitive-site
• 15:25
  exploitation
• 15:26
  from
• 15:26
  months
• 15:26
  to
• 15:26
  minutes
• 15:27
  using
• 15:27
  AI
• 15:27
  at
• 15:28
  point-of-capture,
• 15:28
  and
• 15:29
  Ukrainian
• 15:29
  AI-guided
• 15:30
  FPV
• 15:30
  drones
• 15:31
  boosting
• 15:31
  strike
• 15:31
  accuracy
• 15:31
  from
• 15:32
  30-50%
• 15:34
  to
• 15:34
  approximately
• 15:35
  80%.
• 15:36
  The
• 15:36
  critical
• 15:37
  guardrail:
• 15:37
  "Agents
• 15:38
  are
• 15:38
  force
• 15:38
  multipliers
• 15:39
  for
• 15:39
  planning
• 15:39
  and
• 15:39
  analysis.
• 15:40
  They
• 15:40
  are
• 15:40
  not
• 15:41
  authorized
• 15:41
  to
• 15:41
  make
• 15:41
  targeting
• 15:42
  decisions,
• 15:43
  transmit
• 15:43
  operational
• 15:43
  information,
• 15:44
  or
• 15:45
  take
• 15:45
  actions
• 15:45
  with
• 15:45
  irreversible
• 15:46
  consequences
• 15:46
  without
• 15:47
  explicit
• 15:47
  human
• 15:47
  approval
• 15:48
  at
• 15:48
  each
• 15:48
  step."
• 15:49
  The
• 15:49
  SOFAI
• 15:49
  (Synergies
• 15:50
  of
• 15:50
  Fast
• 15:50
  and
• 15:50
  Slow
• 15:50
  AI)
• 15:51
  Workshop
• 15:51
  explores
• 15:51
  the
• 15:51
  theoretical
• 15:52
  underpinning:
• 15:53
  System
• 15:53
  1
• 15:53
  (fast,
• 15:54
  intuitive)
• 15:54
  and
• 15:54
  System
• 15:55
  2
• 15:55
  (slow,
• 15:55
  deliberative)
• 15:55
  cognitive
• 15:56
  architectures
• 15:56
  applied
• 15:56
  to
• 15:56
  AI
• 15:57
  agents.
• 15:58
  The
• 15:58
  dual-process
• 15:59
  framework
• 15:59
  maps
• 15:59
  neatly
• 16:00
  onto
• 16:00
  the
• 16:00
  pattern
• 16:00
  emerging
• 16:01
  in
• 16:01
  practice
• 16:01
  —
• 16:02
  fast
• 16:02
  reactive
• 16:03
  loops
• 16:03
  (voice
• 16:03
  turn
• 16:03
  detection,
• 16:04
  drone
• 16:05
  guidance)
• 16:05
  paired
• 16:05
  with
• 16:05
  deliberative
• 16:06
  planning
• 16:06
  (target
• 16:06
  package
• 16:07
  assembly,
• 16:07
  spec
• 16:08
  generation).
• 16:09
  Whether
• 16:09
  the
• 16:09
  cognitive
• 16:10
  science
• 16:10
  analogy
• 16:10
  holds
• 16:11
  up
• 16:11
  under
• 16:11
  scrutiny
• 16:12
  is
• 16:12
  debatable,
• 16:13
  but
• 16:13
  the
• 16:13
  engineering
• 16:14
  pattern
• 16:14
  of
• 16:14
  layered
• 16:14
  speed-vs-depth
• 16:15
  processing
• 16:16
  is
• 16:16
  showing
• 16:17
  up
• 16:17
  everywhere
• 16:17
  from
• 16:17
  voice
• 16:17
  pipelines
• 16:18
  to
• 16:18
  military
• 16:18
  kill
• 16:19
  chains
• 16:19
  .