Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – July 20, 2026

Transcript

AGI Dreams Podcast – July 20, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  industry
• 0:00
  has
• 0:00
  been
• 0:00
  warning
• 0:01
  about
• 0:01
  autonomous
• 0:01
  AI-driven
• 0:02
  offensive
• 0:02
  tooling
• 0:03
  for
• 0:03
  over
• 0:03
  a
• 0:03
  year.
• 0:04
  This
• 0:04
  week,
• 0:05
  Hugging
• 0:05
  Face
• 0:05
  disclosed
• 0:06
  that
• 0:06
  the
• 0:06
  warning
• 0:06
  materialized
• 0:07
  —
• 0:08
  and
• 0:08
  it
• 0:08
  looked
• 0:08
  nothing
• 0:09
  like
• 0:09
  the
• 0:09
  drills.
• 0:09
  An
• 0:10
  autonomous
• 0:10
  agent
• 0:11
  framework,
• 0:11
  likely
• 0:12
  built
• 0:12
  on
• 0:12
  an
• 0:12
  agentic
• 0:13
  security-research
• 0:13
  harness
• 0:14
  (the
• 0:14
  specific
• 0:14
  LLM
• 0:15
  remains
• 0:16
  unidentified),
• 0:17
  breached
• 0:17
  Hugging
• 0:18
  Face's
• 0:18
  production
• 0:18
  infrastructure
• 0:19
  through
• 0:20
  two
• 0:20
  code-execution
• 0:21
  paths
• 0:21
  in
• 0:21
  the
• 0:21
  platform's
• 0:22
  dataset
• 0:22
  processing
• 0:23
  pipeline.
• 0:24
  From
• 0:24
  there,
• 0:25
  the
• 0:25
  agent
• 0:25
  escalated
• 0:26
  to
• 0:26
  node-level
• 0:26
  access,
• 0:27
  harvested
• 0:28
  cloud
• 0:28
  and
• 0:28
  cluster
• 0:28
  credentials,
• 0:30
  and
• 0:30
  moved
• 0:30
  laterally
• 0:31
  across
• 0:31
  internal
• 0:31
  clusters
• 0:32
  over
• 0:32
  a
• 0:32
  weekend
• 0:32
  —
• 0:33
  executing
• 0:34
  more
• 0:34
  than
• 0:34
  17,000
• 0:35
  individual
• 0:35
  actions
• 0:36
  across
• 0:36
  a
• 0:36
  swarm
• 0:37
  of
• 0:37
  short-lived
• 0:37
  sandboxes
• 0:38
  with
• 0:38
  self-migrating
• 0:39
  command-and-control
• 0:39
  staged
• 0:40
  on
• 0:40
  public
• 0:40
  services.
• 0:41
  The
• 0:41
  most
• 0:42
  operationally
• 0:42
  significant
• 0:43
  detail
• 0:43
  in
• 0:43
  the
• 0:43
  disclosure
• 0:44
  is
• 0:44
  the
• 0:44
  guardrail
• 0:45
  asymmetry
• 0:45
  problem.
• 0:46
  When
• 0:46
  Hugging
• 0:47
  Face's
• 0:47
  incident
• 0:48
  response
• 0:48
  team
• 0:48
  attempted
• 0:49
  to
• 0:49
  use
• 0:49
  frontier
• 0:49
  models
• 0:50
  behind
• 0:50
  commercial
• 0:51
  APIs
• 0:51
  to
• 0:51
  analyze
• 0:52
  the
• 0:52
  attack
• 0:52
  logs
• 0:52
  —
• 0:53
  real
• 0:53
  exploit
• 0:54
  payloads,
• 0:54
  C2
• 0:55
  artifacts,
• 0:56
  credential
• 0:56
  references
• 0:56
  —
• 0:57
  the
• 0:57
  providers'
• 0:58
  safety
• 0:58
  guardrails
• 0:58
  blocked
• 0:59
  the
• 0:59
  requests,
• 1:00
  unable
• 1:00
  to
• 1:00
  distinguish
• 1:01
  an
• 1:01
  incident
• 1:01
  responder
• 1:02
  from
• 1:02
  an
• 1:02
  attacker.
• 1:03
  The
• 1:03
  team
• 1:03
  pivoted
• 1:04
  to
• 1:04
  GLM
• 1:04
  5.2,
• 1:05
  an
• 1:05
  open-weight
• 1:06
  model
• 1:06
  running
• 1:07
  on
• 1:07
  their
• 1:07
  own
• 1:07
  infrastructure,
• 1:08
  which
• 1:08
  had
• 1:08
  the
• 1:08
  secondary
• 1:09
  benefit
• 1:09
  of
• 1:09
  keeping
• 1:10
  attacker
• 1:10
  data
• 1:10
  and
• 1:10
  compromised
• 1:11
  credentials
• 1:12
  from
• 1:12
  leaving
• 1:12
  the
• 1:12
  environment.
• 1:13
  The
• 1:13
  practical
• 1:14
  lesson
• 1:14
  Hugging
• 1:14
  Face
• 1:15
  draws
• 1:15
  explicitly:
• 1:16
  "have
• 1:16
  a
• 1:16
  capable
• 1:17
  model
• 1:17
  you
• 1:17
  can
• 1:17
  run
• 1:18
  on
• 1:18
  your
• 1:18
  own
• 1:18
  infrastructure
• 1:19
  vetted
• 1:19
  and
• 1:19
  ready
• 1:19
  before
• 1:20
  an
• 1:20
  incident."
• 1:21
  The
• 1:21
  attacker
• 1:21
  was
• 1:22
  bound
• 1:22
  by
• 1:22
  no
• 1:22
  usage
• 1:22
  policy.
• 1:24
  The
• 1:24
  defenders
• 1:24
  were
• 1:24
  blocked
• 1:25
  by
• 1:25
  their
• 1:25
  own
• 1:25
  tools.
• 1:26
  That
• 1:26
  asymmetry
• 1:27
  is
• 1:27
  now
• 1:27
  a
• 1:27
  documented
• 1:28
  operational
• 1:28
  failure
• 1:29
  mode,
• 1:29
  not
• 1:30
  a
• 1:30
  theoretical
• 1:30
  concern.
• 1:31
  The
• 1:31
  same
• 1:32
  week,
• 1:32
  a
• 1:33
  hacker
• 1:33
  wiped
• 1:33
  Romania's
• 1:33
  entire
• 1:34
  land
• 1:34
  registry
• 1:35
  database
• 1:35
  after
• 1:35
  a
• 1:35
  failed
• 1:36
  extortion
• 1:36
  attempt,
• 1:37
  bringing
• 1:37
  the
• 1:37
  country's
• 1:38
  real-estate
• 1:38
  market
• 1:38
  to
• 1:39
  a
• 1:39
  standstill.
• 1:40
  The
• 1:40
  National
• 1:41
  Agency
• 1:41
  for
• 1:41
  Cadastre
• 1:42
  and
• 1:42
  Real
• 1:42
  Estate
• 1:42
  Advertising
• 1:43
  has
• 1:43
  been
• 1:43
  offline
• 1:43
  for
• 1:44
  a
• 1:44
  week
• 1:44
  —
• 1:44
  notaries
• 1:45
  cannot
• 1:45
  record
• 1:46
  transactions,
• 1:47
  citizens
• 1:47
  cannot
• 1:48
  obtain
• 1:48
  proof
• 1:48
  of
• 1:48
  ownership.
• 1:49
  The
• 1:50
  attacker
• 1:50
  entered
• 1:50
  with
• 1:50
  valid
• 1:51
  credentials,
• 1:52
  mapped
• 1:52
  internal
• 1:52
  systems,
• 1:53
  and
• 1:53
  deleted
• 1:54
  both
• 1:54
  live
• 1:54
  data
• 1:54
  and
• 1:55
  backups.
• 1:56
  An
• 1:56
  offline
• 1:57
  copy
• 1:57
  appears
• 1:57
  to
• 1:57
  have
• 1:57
  survived,
• 1:58
  but
• 1:59
  Romania
• 1:59
  joins
• 1:59
  Greece,
• 2:00
  Montenegro,
• 2:01
  Thailand,
• 2:02
  Sri
• 2:02
  Lanka,
• 2:03
  Bermuda,
• 2:04
  and
• 2:04
  Antigua
• 2:05
  as
• 2:05
  countries
• 2:05
  whose
• 2:05
  land
• 2:06
  registries
• 2:06
  have
• 2:06
  been
• 2:07
  compromised
• 2:07
  in
• 2:07
  the
• 2:07
  past
• 2:08
  three
• 2:08
  years.
• 2:09
  The
• 2:09
  hacker,
• 2:10
  operating
• 2:10
  under
• 2:10
  the
• 2:10
  handle
• 2:11
  ByteToBreach,
• 2:12
  was
• 2:12
  subsequently
• 2:12
  doxxed
• 2:13
  as
• 2:13
  an
• 2:13
  individual
• 2:14
  from
• 2:14
  Oran,
• 2:14
  Algeria.
• 2:15
  On
• 2:16
  the
• 2:16
  defensive
• 2:16
  side,
• 2:17
  Capital
• 2:17
  One
• 2:17
  open-sourced
• 2:18
  VulnHunter,
• 2:19
  an
• 2:19
  agentic
• 2:19
  AI
• 2:20
  security
• 2:20
  tool
• 2:20
  that
• 2:21
  reasons
• 2:21
  like
• 2:21
  an
• 2:21
  adversary
• 2:22
  rather
• 2:22
  than
• 2:22
  scanning
• 2:23
  for
• 2:23
  patterns.
• 2:24
  Built
• 2:24
  on
• 2:24
  Claude
• 2:25
  Opus
• 2:25
  running
• 2:25
  inside
• 2:25
  Claude
• 2:26
  Code,
• 2:26
  VulnHunter
• 2:27
  flips
• 2:27
  the
• 2:27
  conventional
• 2:28
  SAST
• 2:28
  model:
• 2:29
  instead
• 2:29
  of
• 2:29
  searching
• 2:30
  backward
• 2:30
  from
• 2:30
  dangerous
• 2:31
  sinks,
• 2:31
  it
• 2:32
  starts
• 2:32
  at
• 2:32
  attacker-accessible
• 2:33
  entry
• 2:33
  points
• 2:34
  (APIs,
• 2:35
  file
• 2:35
  uploads,
• 2:36
  network
• 2:36
  messages)
• 2:37
  and
• 2:37
  reasons
• 2:37
  forward
• 2:38
  to
• 2:38
  evaluate
• 2:38
  whether
• 2:38
  an
• 2:38
  attacker
• 2:39
  can
• 2:39
  actually
• 2:39
  break
• 2:40
  through.
• 2:41
  After
• 2:41
  identifying
• 2:42
  a
• 2:42
  potential
• 2:42
  vulnerability,
• 2:43
  a
• 2:44
  falsification
• 2:44
  engine
• 2:45
  runs
• 2:45
  a
• 2:45
  structured
• 2:46
  workflow
• 2:46
  designed
• 2:47
  to
• 2:47
  disprove
• 2:47
  its
• 2:47
  own
• 2:47
  findings
• 2:48
  —
• 2:49
  searching
• 2:49
  for
• 2:49
  flawed
• 2:50
  assumptions,
• 2:50
  logic
• 2:51
  gaps,
• 2:51
  or
• 2:51
  security
• 2:52
  controls
• 2:52
  that
• 2:52
  would
• 2:52
  block
• 2:53
  the
• 2:53
  attack.
• 2:54
  What
• 2:54
  survives
• 2:54
  gets
• 2:55
  an
• 2:55
  executable
• 2:55
  exploit
• 2:56
  path
• 2:56
  and
• 2:56
  a
• 2:56
  proposed
• 2:57
  fix.
• 2:58
  The
• 2:58
  three-skill
• 2:59
  loop
• 2:59
  —
• 3:00
  /vulnhunt
• 3:00
  to
• 3:00
  find,
• 3:01
  /vulnhunter-fix
• 3:02
  to
• 3:02
  remediate
• 3:03
  with
• 3:03
  TDD,
• 3:04
  /vulnhunt-fix-verify
• 3:06
  to
• 3:06
  independently
• 3:07
  validate
• 3:07
  —
• 3:07
  is
• 3:08
  designed
• 3:08
  to
• 3:08
  run
• 3:08
  unattended
• 3:09
  at
• 3:09
  scale
• 3:10
  across
• 3:10
  multiple
• 3:10
  repositories.
• 3:12
  ClaudeBrain
• 3:12
  takes
• 3:13
  a
• 3:13
  complementary
• 3:13
  approach:
• 3:14
  a
• 3:14
  500-page
• 3:15
  offensive-security
• 3:16
  wiki
• 3:17
  indexed
• 3:17
  for
• 3:17
  semantic
• 3:18
  search,
• 3:18
  wired
• 3:19
  into
• 3:19
  Claude
• 3:19
  Code
• 3:20
  via
• 3:20
  MCP
• 3:21
  and
• 3:21
  deterministic
• 3:22
  hooks.
• 3:23
  Hunt
• 3:23
  skills
• 3:23
  fire
• 3:24
  per
• 3:24
  vulnerability
• 3:25
  class
• 3:25
  (XSS,
• 3:26
  SQLi,
• 3:27
  SSRF,
• 3:28
  IDOR,
• 3:29
  cloud
• 3:29
  misconfigs,
• 3:30
  Active
• 3:31
  Directory),
• 3:32
  each
• 3:32
  querying
• 3:32
  the
• 3:32
  wiki
• 3:33
  before
• 3:33
  attacking
• 3:34
  and
• 3:34
  emitting
• 3:34
  uniform
• 3:34
  findings.
• 3:36
  The
• 3:36
  state-first
• 3:36
  engagement
• 3:37
  model
• 3:37
  ensures
• 3:38
  nothing
• 3:38
  in
• 3:38
  scope
• 3:39
  gets
• 3:39
  skipped
• 3:40
  and
• 3:40
  no
• 3:40
  dead
• 3:40
  end
• 3:40
  gets
• 3:41
  retried
• 3:41
  —
• 3:42
  the
• 3:42
  kind
• 3:42
  of
• 3:42
  operational
• 3:43
  discipline
• 3:43
  that
• 3:43
  separates
• 3:44
  a
• 3:44
  structured
• 3:45
  assessment
• 3:45
  from
• 3:45
  an
• 3:45
  AI
• 3:46
  throwing
• 3:46
  payloads
• 3:47
  at
• 3:47
  a
• 3:47
  wall.
• 3:48
  The
• 3:48
  through-line
• 3:49
  is
• 3:49
  unmistakable:
• 3:50
  cybersecurity
• 3:51
  teams
• 3:51
  are
• 3:51
  asking
• 3:52
  sharper
• 3:52
  questions
• 3:53
  about
• 3:53
  AI
• 3:53
  reliability
• 3:54
  than
• 3:54
  most
• 3:55
  AI
• 3:55
  companies
• 3:56
  ask
• 3:56
  about
• 3:56
  themselves.
• 3:57
  As
• 3:58
  one
• 3:58
  practitioner
• 3:59
  building
• 3:59
  customer-facing
• 4:00
  AI
• 4:00
  agents
• 4:01
  for
• 4:01
  security
• 4:01
  firms
• 4:02
  noted,
• 4:03
  these
• 4:03
  teams
• 4:03
  "immediately
• 4:04
  start
• 4:04
  thinking
• 4:05
  about
• 4:05
  everything
• 4:05
  that
• 4:06
  can
• 4:06
  go
• 4:06
  wrong"
• 4:06
  —
• 4:07
  reliability,
• 4:08
  hallucinations,
• 4:10
  policy
• 4:10
  violations,
• 4:11
  behavioral
• 4:12
  drift
• 4:12
  —
• 4:13
  because
• 4:13
  they
• 4:13
  already
• 4:14
  operate
• 4:14
  in
• 4:14
  domains
• 4:15
  where
• 4:15
  failure
• 4:16
  has
• 4:16
  consequences.
• 4:20
  Xi
• 4:20
  Jinping
• 4:21
  spoke
• 4:21
  at
• 4:21
  the
• 4:21
  World
• 4:21
  AI
• 4:22
  Conference
• 4:22
  this
• 4:22
  week
• 4:23
  and
• 4:23
  reaffirmed
• 4:24
  China's
• 4:24
  commitment
• 4:25
  to
• 4:25
  open-source
• 4:25
  AI
• 4:26
  as
• 4:26
  a
• 4:26
  vehicle
• 4:26
  for
• 4:27
  "openness
• 4:27
  and
• 4:28
  win-win."
• 4:29
  The
• 4:29
  statement
• 4:29
  lands
• 4:30
  at
• 4:30
  a
• 4:30
  time
• 4:30
  when
• 4:30
  multiple
• 4:31
  signals
• 4:31
  pointed
• 4:32
  in
• 4:32
  the
• 4:32
  opposite
• 4:32
  direction
• 4:32
  —
• 4:34
  earlier
• 4:34
  reports
• 4:35
  suggested
• 4:35
  Beijing
• 4:36
  was
• 4:36
  exploring
• 4:36
  restrictions
• 4:37
  on
• 4:37
  overseas
• 4:37
  access
• 4:38
  to
• 4:38
  China's
• 4:38
  most
• 4:39
  advanced
• 4:39
  models,
• 4:40
  with
• 4:40
  the
• 4:40
  Ministry
• 4:40
  of
• 4:40
  Commerce
• 4:41
  holding
• 4:41
  meetings
• 4:41
  with
• 4:42
  Alibaba
• 4:42
  and
• 4:42
  ByteDance
• 4:43
  about
• 4:43
  potential
• 4:43
  limits.
• 4:45
  Xi's
• 4:45
  public
• 4:45
  posture
• 4:46
  resolves
• 4:46
  that
• 4:46
  ambiguity,
• 4:47
  at
• 4:47
  least
• 4:47
  rhetorically,
• 4:48
  toward
• 4:48
  openness.
• 4:50
  The
• 4:50
  community
• 4:51
  reaction
• 4:51
  reads
• 4:51
  the
• 4:51
  move
• 4:52
  as
• 4:52
  strategic:
• 4:53
  Chinese
• 4:54
  AI
• 4:54
  labs
• 4:54
  are
• 4:54
  training
• 4:55
  frontier
• 4:55
  models
• 4:55
  at
• 4:56
  a
• 4:56
  fraction
• 4:56
  of
• 4:56
  US
• 4:57
  costs,
• 4:58
  and
• 4:58
  coupling
• 4:58
  that
• 4:58
  with
• 4:59
  a
• 4:59
  policy
• 4:59
  pushing
• 4:59
  models
• 5:00
  to
• 5:00
  be
• 5:00
  trained
• 5:00
  and
• 5:00
  served
• 5:01
  on
• 5:01
  Chinese-made
• 5:02
  chips
• 5:02
  creates
• 5:03
  a
• 5:03
  potential
• 5:03
  path
• 5:03
  to
• 5:04
  selling
• 5:04
  "all
• 5:04
  the
• 5:04
  shovels
• 5:05
  in
• 5:05
  this
• 5:05
  AI
• 5:05
  rush"
• 5:05
  —
• 5:06
  compute
• 5:07
  hardware,
• 5:07
  power
• 5:08
  infrastructure,
• 5:08
  and
• 5:08
  the
• 5:08
  models
• 5:09
  to
• 5:09
  run
• 5:09
  on
• 5:09
  them
• 5:09
  —
• 5:10
  to
• 5:10
  any
• 5:10
  country
• 5:11
  without
• 5:11
  export
• 5:11
  restrictions.
• 5:12
  Case
• 5:13
  in
• 5:13
  point:
• 5:14
  Moonshot
• 5:14
  AI
• 5:15
  unveiled
• 5:15
  Kimi
• 5:15
  K3,
• 5:16
  an
• 5:17
  open-weight
• 5:17
  model
• 5:17
  the
• 5:17
  company
• 5:18
  claims
• 5:18
  rivals
• 5:18
  frontier
• 5:19
  offerings
• 5:19
  from
• 5:19
  Anthropic
• 5:20
  and
• 5:20
  OpenAI
• 5:20
  on
• 5:20
  agentic
• 5:21
  and
• 5:21
  visual
• 5:21
  benchmarks.
• 5:22
  The
• 5:23
  community
• 5:23
  response
• 5:23
  is
• 5:24
  appropriately
• 5:24
  divided
• 5:24
  between
• 5:25
  excitement
• 5:25
  and
• 5:25
  skepticism.
• 5:27
  At
• 5:27
  an
• 5:27
  estimated
• 5:27
  3
• 5:28
  trillion
• 5:28
  parameters,
• 5:29
  K3's
• 5:29
  practical
• 5:30
  reach
• 5:30
  is
• 5:30
  limited
• 5:31
  —
• 5:31
  Q8
• 5:32
  quantization
• 5:32
  would
• 5:32
  demand
• 5:33
  hardware
• 5:33
  well
• 5:33
  beyond
• 5:33
  any
• 5:34
  hobbyist's
• 5:34
  setup,
• 5:35
  and
• 5:35
  the
• 5:35
  model
• 5:35
  has
• 5:35
  only
• 5:35
  been
• 5:36
  available
• 5:36
  for
• 5:36
  hours
• 5:37
  on
• 5:37
  a
• 5:37
  first-party
• 5:37
  provider
• 5:38
  that
• 5:38
  few
• 5:38
  people
• 5:38
  actually
• 5:39
  use.
• 5:40
  The
• 5:40
  benchmarks
• 5:40
  are
• 5:40
  promising,
• 5:41
  but
• 5:41
  the
• 5:41
  community
• 5:42
  rightly
• 5:42
  notes
• 5:42
  that
• 5:42
  real-world
• 5:43
  performance
• 5:43
  on
• 5:43
  sustained
• 5:44
  coding
• 5:44
  or
• 5:44
  complex
• 5:45
  reasoning
• 5:45
  tasks
• 5:46
  rarely
• 5:46
  matches
• 5:46
  chart-topping
• 5:47
  scores.
• 5:48
  The
• 5:48
  trend
• 5:48
  is
• 5:48
  clear,
• 5:49
  though:
• 5:49
  the
• 5:49
  gap
• 5:50
  between
• 5:50
  open
• 5:50
  and
• 5:50
  closed
• 5:50
  models
• 5:51
  is
• 5:51
  no
• 5:51
  longer
• 5:51
  measured
• 5:52
  in
• 5:52
  years,
• 5:52
  or
• 5:52
  even
• 5:53
  months.
• 5:54
  At
• 5:54
  the
• 5:54
  other
• 5:54
  end
• 5:54
  of
• 5:54
  the
• 5:54
  scale
• 5:55
  spectrum,
• 5:55
  Apple
• 5:56
  is
• 5:56
  reportedly
• 5:56
  in
• 5:56
  talks
• 5:57
  with
• 5:57
  PrismML,
• 5:58
  a
• 5:58
  Caltech
• 5:58
  spinoff
• 5:59
  that
• 5:59
  claims
• 5:59
  to
• 5:59
  compress
• 6:00
  models
• 6:00
  like
• 6:00
  Qwen
• 6:00
  27B
• 6:01
  down
• 6:01
  to
• 6:01
  ternary
• 6:02
  quantization
• 6:02
  (1.58-bit)
• 6:03
  —
• 6:04
  small
• 6:04
  enough
• 6:05
  to
• 6:05
  fit
• 6:05
  on
• 6:05
  an
• 6:05
  iPhone.
• 6:06
  The
• 6:06
  discussions
• 6:07
  are
• 6:07
  characterized
• 6:07
  as
• 6:07
  "very
• 6:08
  early,"
• 6:09
  and
• 6:09
  the
• 6:09
  CEO's
• 6:09
  own
• 6:09
  description
• 6:10
  —
• 6:11
  "things
• 6:11
  are
• 6:11
  progressing
• 6:11
  nicely"
• 6:12
  —
• 6:13
  is
• 6:13
  the
• 6:13
  kind
• 6:13
  of
• 6:13
  phrasing
• 6:14
  that
• 6:14
  could
• 6:14
  describe
• 6:14
  anything
• 6:15
  from
• 6:15
  a
• 6:15
  signed
• 6:15
  partnership
• 6:16
  to
• 6:16
  a
• 6:16
  polite
• 6:16
  email
• 6:16
  reply.
• 6:18
  The
• 6:18
  compression
• 6:18
  technique
• 6:18
  itself
• 6:19
  faces
• 6:19
  legitimate
• 6:20
  scrutiny:
• 6:21
  nobody
• 6:21
  in
• 6:21
  the
• 6:21
  community
• 6:22
  has
• 6:22
  demonstrated
• 6:22
  that
• 6:23
  ternary
• 6:23
  quantization
• 6:24
  at
• 6:24
  this
• 6:24
  ratio
• 6:25
  preserves
• 6:25
  the
• 6:25
  capabilities
• 6:26
  that
• 6:26
  make
• 6:26
  models
• 6:26
  useful
• 6:27
  beyond
• 6:27
  benchmarks,
• 6:28
  particularly
• 6:29
  long-context
• 6:29
  reasoning
• 6:30
  and
• 6:30
  coherent
• 6:30
  multi-turn
• 6:31
  interaction.
• 6:35
  IBM
• 6:35
  Research
• 6:36
  published
• 6:36
  a
• 6:36
  refreshingly
• 6:36
  honest
• 6:37
  account
• 6:37
  of
• 6:37
  why
• 6:37
  model
• 6:38
  routing
• 6:38
  is
• 6:38
  harder
• 6:39
  than
• 6:39
  it
• 6:39
  looks.
• 6:40
  The
• 6:40
  core
• 6:40
  finding:
• 6:41
  across
• 6:41
  417
• 6:42
  tasks
• 6:42
  on
• 6:42
  the
• 6:43
  AppWorld
• 6:43
  Test
• 6:43
  Challenge,
• 6:44
  Claude
• 6:44
  Sonnet
• 6:44
  4.6
• 6:45
  cost
• 6:45
  $79
• 6:46
  total
• 6:46
  ($0.19/task)
• 6:48
  while
• 6:48
  GPT-4.1
• 6:49
  cost
• 6:50
  $155
• 6:51
  ($0.37/task)
• 6:52
  —
• 6:54
  nearly
• 6:54
  double
• 6:54
  —
• 6:55
  despite
• 6:55
  GPT-4.1
• 6:56
  having
• 6:56
  lower
• 6:56
  token
• 6:57
  pricing
• 6:57
  on
• 6:57
  both
• 6:57
  input
• 6:57
  and
• 6:58
  output.
• 6:59
  The
• 6:59
  explanation
• 6:59
  is
• 6:59
  caching,
• 7:00
  which
• 7:00
  most
• 7:00
  routing
• 7:01
  discussions
• 7:01
  ignore.
• 7:02
  Agent
• 7:02
  workloads
• 7:03
  reuse
• 7:03
  large
• 7:03
  chunks
• 7:03
  of
• 7:04
  context
• 7:04
  across
• 7:04
  steps,
• 7:05
  and
• 7:05
  Sonnet's
• 7:05
  lower
• 7:06
  cache-read
• 7:06
  pricing
• 7:06
  meant
• 7:07
  it
• 7:07
  benefited
• 7:07
  disproportionately
• 7:08
  from
• 7:08
  this
• 7:08
  pattern,
• 7:09
  enough
• 7:09
  to
• 7:09
  overcome
• 7:10
  both
• 7:10
  its
• 7:10
  higher
• 7:10
  base
• 7:11
  pricing
• 7:11
  and
• 7:11
  its
• 7:11
  longer
• 7:11
  reasoning
• 7:12
  trajectories.
• 7:13
  The
• 7:13
  team's
• 7:13
  broader
• 7:14
  lesson:
• 7:15
  routing
• 7:15
  is
• 7:15
  not
• 7:15
  a
• 7:15
  classification
• 7:16
  problem
• 7:16
  ("which
• 7:16
  model
• 7:17
  is
• 7:17
  best
• 7:17
  for
• 7:17
  this
• 7:17
  task?")
• 7:18
  but
• 7:18
  a
• 7:18
  systems
• 7:19
  optimization
• 7:19
  problem
• 7:20
  across
• 7:20
  cost,
• 7:20
  quality,
• 7:21
  and
• 7:21
  latency
• 7:21
  simultaneously.
• 7:23
  Their
• 7:23
  lightweight
• 7:23
  optimizer
• 7:24
  (6ms,
• 7:24
  2KB
• 7:25
  memory
• 7:25
  per
• 7:25
  decision)
• 7:26
  traced
• 7:26
  out
• 7:26
  a
• 7:26
  cost-accuracy
• 7:27
  frontier
• 7:27
  —
• 7:28
  one
• 7:28
  configuration
• 7:29
  delivered
• 7:29
  84%
• 7:30
  accuracy
• 7:30
  at
• 7:30
  $93
• 7:31
  with
• 7:31
  a
• 7:31
  21%
• 7:32
  cost
• 7:32
  reduction
• 7:32
  versus
• 7:33
  running
• 7:33
  Opus
• 7:33
  alone.
• 7:34
  A
• 7:34
  standard
• 7:35
  difficulty-based
• 7:35
  router
• 7:36
  landed
• 7:36
  at
• 7:36
  similar
• 7:36
  accuracy
• 7:37
  but
• 7:37
  higher
• 7:37
  cost.
• 7:38
  OmniRoute
• 7:38
  is
• 7:39
  one
• 7:39
  practical
• 7:39
  answer
• 7:39
  to
• 7:40
  this
• 7:40
  problem
• 7:40
  at
• 7:40
  scale.
• 7:41
  The
• 7:41
  open-source
• 7:42
  gateway
• 7:42
  connects
• 7:42
  250
• 7:43
  AI
• 7:44
  providers
• 7:44
  —
• 7:45
  90+
• 7:45
  with
• 7:45
  free
• 7:46
  tiers
• 7:46
  —
• 7:47
  through
• 7:47
  a
• 7:47
  single
• 7:47
  endpoint,
• 7:48
  with
• 7:48
  18
• 7:49
  routing
• 7:49
  strategies
• 7:49
  ranging
• 7:50
  from
• 7:50
  simple
• 7:50
  priority
• 7:51
  lists
• 7:51
  to
• 7:51
  cost-optimized
• 7:52
  selection,
• 7:53
  context-relay,
• 7:53
  and
• 7:54
  a
• 7:54
  fusion
• 7:54
  strategy
• 7:55
  that
• 7:55
  fans
• 7:55
  out
• 7:55
  to
• 7:55
  a
• 7:56
  panel
• 7:56
  of
• 7:56
  models
• 7:56
  and
• 7:56
  synthesizes
• 7:57
  via
• 7:57
  a
• 7:58
  judge.
• 7:58
  Its
• 7:59
  compression
• 7:59
  pipeline
• 8:00
  stacks
• 8:00
  10
• 8:00
  composable
• 8:01
  engines
• 8:01
  (RTK
• 8:02
  tool-result
• 8:02
  filtering,
• 8:03
  Caveman
• 8:03
  prose
• 8:04
  compression,
• 8:04
  LLMLingua-2
• 8:05
  ML
• 8:06
  pruning)
• 8:06
  to
• 8:06
  save
• 8:06
  15–95%
• 8:08
  of
• 8:08
  tokens
• 8:09
  transparently.
• 8:10
  The
• 8:10
  architecture
• 8:10
  includes
• 8:11
  circuit
• 8:11
  breakers,
• 8:12
  connection
• 8:12
  cooldowns,
• 8:13
  and
• 8:13
  per-model
• 8:13
  lockouts
• 8:14
  —
• 8:15
  three
• 8:15
  independent
• 8:15
  resilience
• 8:16
  layers
• 8:16
  that
• 8:16
  keep
• 8:16
  requests
• 8:17
  flowing
• 8:17
  when
• 8:18
  individual
• 8:18
  providers
• 8:19
  fail.
• 8:20
  It
• 8:20
  ships
• 8:20
  as
• 8:20
  npm,
• 8:21
  Docker,
• 8:22
  desktop
• 8:22
  Electron,
• 8:23
  or
• 8:23
  even
• 8:23
  on
• 8:23
  Android
• 8:24
  via
• 8:24
  Termux.
• 8:25
  A
• 8:25
  Quesma
• 8:26
  engineer
• 8:26
  learned
• 8:26
  the
• 8:26
  token
• 8:27
  economics
• 8:27
  lesson
• 8:27
  the
• 8:27
  hard
• 8:28
  way:
• 8:29
  the
• 8:29
  first
• 8:29
  run
• 8:29
  of
• 8:29
  a
• 8:29
  deep
• 8:30
  research
• 8:30
  pipeline
• 8:30
  burned
• 8:31
  through
• 8:31
  an
• 8:31
  entire
• 8:31
  Claude
• 8:32
  Max
• 8:32
  5x
• 8:32
  subscription
• 8:33
  limit
• 8:33
  in
• 8:33
  30
• 8:34
  minutes
• 8:34
  —
• 8:35
  111
• 8:35
  agents
• 8:35
  launched,
• 8:36
  123
• 8:37
  claims
• 8:37
  queued,
• 8:38
  25
• 8:38
  verified,
• 8:39
  zero
• 8:39
  final
• 8:40
  output.
• 8:41
  The
• 8:41
  fix
• 8:41
  was
• 8:41
  architectural:
• 8:42
  use
• 8:43
  Claude
• 8:43
  Code
• 8:43
  as
• 8:43
  the
• 8:43
  main
• 8:44
  harness,
• 8:44
  Codex
• 8:45
  and
• 8:45
  Antigravity
• 8:46
  as
• 8:46
  headless
• 8:46
  subagents
• 8:47
  via
• 8:47
  a
• 8:47
  Bash
• 8:47
  wrapper,
• 8:48
  all
• 8:48
  sharing
• 8:49
  memory
• 8:49
  through
• 8:49
  a
• 8:49
  locally
• 8:50
  extended
• 8:50
  plugin.
• 8:51
  Model
• 8:52
  routing
• 8:52
  by
• 8:52
  role
• 8:52
  —
• 8:53
  Sonnet
• 8:53
  for
• 8:54
  search,
• 8:54
  Opus
• 8:55
  for
• 8:55
  verification,
• 8:56
  Fable
• 8:56
  only
• 8:56
  for
• 8:57
  planning
• 8:57
  and
• 8:57
  dispute
• 8:57
  resolution
• 8:58
  —
• 8:59
  extended
• 8:59
  continuous
• 9:00
  research
• 9:00
  time
• 9:00
  from
• 9:00
  30
• 9:01
  minutes
• 9:01
  to
• 9:01
  hours
• 9:01
  without
• 9:02
  additional
• 9:02
  spend.
• 9:03
  The
• 9:03
  most
• 9:03
  practical
• 9:04
  insight:
• 9:04
  "context
• 9:05
  compaction
• 9:05
  can
• 9:05
  double
• 9:06
  your
• 9:06
  bill
• 9:06
  instead
• 9:06
  of
• 9:06
  saving
• 9:07
  it"
• 9:07
  when
• 9:07
  the
• 9:07
  summarization
• 9:08
  call
• 9:08
  consumes
• 9:09
  tokens
• 9:09
  and
• 9:09
  evicts
• 9:09
  files
• 9:10
  the
• 9:10
  agent
• 9:10
  still
• 9:10
  needs,
• 9:11
  triggering
• 9:12
  a
• 9:12
  re-read-recompact
• 9:13
  loop
• 9:13
  that
• 9:13
  inflated
• 9:14
  one
• 9:14
  practitioner's
• 9:15
  usage
• 9:15
  from
• 9:15
  89
• 9:16
  million
• 9:16
  to
• 9:16
  185
• 9:17
  million
• 9:17
  tokens
• 9:18
  on
• 9:18
  identical
• 9:18
  tasks.
• 9:20
  Underneath
• 9:20
  the
• 9:20
  routing
• 9:20
  layer,
• 9:21
  the
• 9:21
  wire
• 9:21
  format
• 9:22
  matters.
• 9:22
  GCF
• 9:23
  (GPU
• 9:23
  Compute
• 9:23
  Format)
• 9:24
  is
• 9:24
  a
• 9:24
  Rust-implemented
• 9:25
  alternative
• 9:25
  to
• 9:25
  JSON
• 9:26
  that
• 9:26
  claims
• 9:26
  50–92%
• 9:28
  smaller
• 9:28
  payloads
• 9:29
  with
• 9:29
  100%
• 9:30
  comprehension
• 9:30
  across
• 9:31
  every
• 9:31
  frontier
• 9:31
  model,
• 9:32
  zero
• 9:32
  training
• 9:33
  required.
• 9:34
  The
• 9:34
  key
• 9:34
  differentiator
• 9:35
  over
• 9:35
  prior
• 9:35
  contenders
• 9:36
  like
• 9:36
  TOON
• 9:36
  is
• 9:36
  multi-turn
• 9:37
  session
• 9:37
  deduplication
• 9:38
  and
• 9:38
  delta
• 9:38
  encoding:
• 9:39
  by
• 9:39
  the
• 9:39
  fifth
• 9:39
  overlapping
• 9:40
  call
• 9:40
  in
• 9:40
  a
• 9:40
  session,
• 9:41
  GCF
• 9:41
  delivers
• 9:42
  99%
• 9:43
  fewer
• 9:43
  tokens
• 9:43
  than
• 9:43
  re-sending
• 9:44
  JSON
• 9:44
  each
• 9:44
  turn.
• 9:45
  Adoption
• 9:46
  already
• 9:46
  includes
• 9:46
  Chrome
• 9:47
  DevTools
• 9:47
  MCP
• 9:48
  (47K
• 9:49
  stars),
• 9:49
  Speakeasy,
• 9:50
  and
• 9:50
  OmniRoute
• 9:51
  itself.
• 9:54
  Command
• 9:54
  Code
• 9:55
  raised
• 9:55
  a
• 9:55
  $5M
• 9:56
  seed
• 9:56
  for
• 9:56
  what
• 9:56
  it
• 9:56
  calls
• 9:57
  "the
• 9:57
  first
• 9:57
  frontier
• 9:57
  coding
• 9:58
  agent
• 9:58
  that
• 9:58
  continuously
• 9:59
  learns
• 9:59
  your
• 9:59
  coding
• 9:59
  taste."
• 10:00
  The
• 10:00
  pitch:
• 10:01
  every
• 10:02
  accept,
• 10:02
  reject,
• 10:03
  and
• 10:03
  edit
• 10:03
  is
• 10:04
  a
• 10:04
  signal
• 10:04
  that
• 10:04
  auto-generates
• 10:05
  into
• 10:05
  project-level
• 10:06
  skills
• 10:06
  —
• 10:06
  no
• 10:07
  rules
• 10:07
  to
• 10:07
  write,
• 10:08
  no
• 10:08
  prompts
• 10:08
  to
• 10:08
  maintain.
• 10:09
  The
• 10:10
  taste-1
• 10:10
  system
• 10:10
  uses
• 10:11
  continuous
• 10:11
  reinforcement
• 10:12
  learning
• 10:12
  to
• 10:12
  distill
• 10:13
  developer
• 10:13
  preferences,
• 10:14
  and
• 10:14
  the
• 10:14
  results
• 10:15
  are
• 10:15
  sharable
• 10:15
  across
• 10:15
  teams
• 10:16
  via
• 10:16
  npx
• 10:16
  taste
• 10:17
  push/pull.
• 10:18
  The
• 10:18
  tool
• 10:19
  ships
• 10:19
  with
• 10:19
  interactive
• 10:19
  CLI,
• 10:20
  headless
• 10:20
  -p
• 10:21
  mode,
• 10:21
  and
• 10:21
  background
• 10:22
  sandbox,
• 10:22
  supporting
• 10:23
  open
• 10:23
  models
• 10:23
  from
• 10:24
  DeepSeek,
• 10:24
  MiniMax,
• 10:25
  Kimi,
• 10:26
  GLM,
• 10:26
  and
• 10:26
  others
• 10:27
  with
• 10:27
  $10–40
• 10:29
  in
• 10:29
  free
• 10:29
  credits.
• 10:30
  Tom
• 10:30
  Preston-Werner,
• 10:31
  GitHub's
• 10:32
  co-founder,
• 10:33
  offered
• 10:33
  the
• 10:33
  endorsement:
• 10:35
  "If
• 10:35
  I
• 10:35
  were
• 10:35
  building
• 10:35
  GitHub
• 10:36
  for
• 10:36
  software
• 10:36
  built
• 10:36
  with
• 10:36
  AI,
• 10:37
  Command
• 10:38
  is
• 10:38
  what
• 10:38
  it
• 10:38
  would
• 10:38
  look
• 10:38
  like."
• 10:40
  Kimchi
• 10:40
  takes
• 10:40
  a
• 10:40
  different
• 10:40
  angle
• 10:41
  on
• 10:41
  the
• 10:41
  same
• 10:41
  problem.
• 10:42
  Rather
• 10:43
  than
• 10:43
  learning
• 10:43
  taste
• 10:43
  implicitly,
• 10:44
  it
• 10:44
  orchestrates
• 10:45
  models
• 10:45
  explicitly:
• 10:46
  a
• 10:46
  task
• 10:47
  classifier
• 10:47
  routes
• 10:47
  each
• 10:48
  request
• 10:48
  to
• 10:48
  the
• 10:48
  best
• 10:48
  model
• 10:49
  for
• 10:49
  the
• 10:49
  job,
• 10:49
  with
• 10:50
  a
• 10:50
  high-reasoning
• 10:50
  model
• 10:50
  scoring
• 10:51
  output
• 10:51
  against
• 10:51
  original
• 10:52
  intent
• 10:52
  after
• 10:52
  each
• 10:52
  milestone.
• 10:54
  The
• 10:54
  spend
• 10:54
  console
• 10:55
  provides
• 10:55
  coding-specific
• 10:56
  cost
• 10:56
  breakdowns
• 10:57
  by
• 10:57
  user,
• 10:57
  tag,
• 10:58
  and
• 10:58
  API
• 10:58
  key
• 10:59
  —
• 11:00
  the
• 11:00
  kind
• 11:00
  of
• 11:00
  financial
• 11:00
  visibility
• 11:01
  that
• 11:01
  enterprise
• 11:01
  teams
• 11:02
  actually
• 11:02
  need.
• 11:03
  Migration
• 11:04
  from
• 11:04
  Claude
• 11:04
  Code
• 11:04
  is
• 11:05
  one-prompt:
• 11:06
  Kimchi
• 11:06
  detects
• 11:07
  existing
• 11:07
  installations
• 11:08
  and
• 11:08
  offers
• 11:08
  to
• 11:08
  import
• 11:09
  MCP
• 11:09
  servers
• 11:10
  and
• 11:10
  skills.
• 11:11
  Kastor
• 11:11
  addresses
• 11:12
  a
• 11:12
  layer
• 11:12
  neither
• 11:12
  touches:
• 11:13
  the
• 11:13
  declarative
• 11:14
  specification
• 11:15
  of
• 11:15
  agents
• 11:15
  themselves.
• 11:16
  Written
• 11:16
  in
• 11:16
  HCL
• 11:17
  (the
• 11:17
  same
• 11:17
  language
• 11:18
  as
• 11:18
  Terraform),
• 11:19
  a
• 11:19
  Kastor
• 11:19
  module
• 11:19
  defines
• 11:20
  agents,
• 11:20
  tools,
• 11:21
  prompts,
• 11:22
  models,
• 11:22
  and
• 11:22
  targets
• 11:23
  as
• 11:23
  a
• 11:23
  versionable,
• 11:24
  reviewable
• 11:24
  contract.
• 11:25
  kastor
• 11:26
  validate
• 11:26
  checks
• 11:27
  references,
• 11:28
  kastor
• 11:28
  build
• 11:28
  compiles
• 11:29
  to
• 11:29
  runnable
• 11:29
  LangGraph
• 11:30
  projects,
• 11:31
  and
• 11:31
  kastor
• 11:31
  plan
• 11:31
  /
• 11:32
  kastor
• 11:32
  apply
• 11:32
  reconciles
• 11:33
  hosted
• 11:33
  agents
• 11:33
  with
• 11:34
  Terraform-style
• 11:34
  diffs
• 11:35
  and
• 11:35
  drift
• 11:35
  detection.
• 11:36
  The
• 11:36
  philosophy
• 11:37
  is
• 11:37
  that
• 11:37
  agents
• 11:38
  need
• 11:38
  a
• 11:38
  source
• 11:38
  of
• 11:38
  truth
• 11:39
  before
• 11:39
  they
• 11:39
  become
• 11:39
  serious
• 11:40
  software
• 11:40
  —
• 11:41
  the
• 11:41
  generated
• 11:41
  code
• 11:42
  is
• 11:42
  an
• 11:42
  output,
• 11:43
  not
• 11:43
  the
• 11:43
  contract.
• 11:44
  It
• 11:44
  is
• 11:44
  early
• 11:45
  (proof
• 11:45
  of
• 11:45
  concept,
• 11:46
  LangGraph
• 11:46
  is
• 11:46
  the
• 11:47
  only
• 11:47
  codegen
• 11:47
  target),
• 11:48
  but
• 11:48
  the
• 11:48
  bet
• 11:48
  that
• 11:48
  agent
• 11:49
  specifications
• 11:50
  should
• 11:50
  be
• 11:50
  reviewable
• 11:50
  and
• 11:50
  diffable
• 11:51
  mirrors
• 11:51
  what
• 11:51
  happened
• 11:52
  to
• 11:52
  infrastructure
• 11:52
  a
• 11:52
  decade
• 11:53
  ago.
• 11:56
  Cole
• 11:56
  Medin's
• 11:57
  analysis
• 11:57
  of
• 11:57
  why
• 11:57
  the
• 11:58
  Karpathy
• 11:58
  LLM
• 11:58
  Wiki
• 11:59
  pattern
• 11:59
  does
• 11:59
  not
• 11:59
  scale
• 12:00
  beyond
• 12:00
  personal
• 12:00
  use
• 12:01
  cuts
• 12:01
  to
• 12:01
  a
• 12:01
  structural
• 12:02
  limitation.
• 12:02
  For
• 12:03
  a
• 12:03
  personal
• 12:03
  agent
• 12:03
  —
• 12:04
  Claude
• 12:04
  Code,
• 12:04
  Hermes,
• 12:05
  or
• 12:05
  any
• 12:05
  coding
• 12:05
  SDK
• 12:06
  managing
• 12:06
  an
• 12:06
  Obsidian
• 12:06
  vault
• 12:07
  —
• 12:07
  markdown
• 12:07
  is
• 12:08
  ideal:
• 12:08
  simple,
• 12:08
  flexible,
• 12:09
  no
• 12:09
  governance
• 12:09
  overhead.
• 12:10
  But
• 12:10
  as
• 12:10
  soon
• 12:11
  as
• 12:11
  other
• 12:11
  users
• 12:11
  interact
• 12:12
  with
• 12:12
  the
• 12:12
  agent,
• 12:12
  the
• 12:12
  model
• 12:12
  breaks.
• 12:13
  Markdown
• 12:14
  cannot
• 12:14
  handle
• 12:14
  concurrent
• 12:15
  access,
• 12:15
  organization
• 12:16
  degrades
• 12:17
  as
• 12:17
  the
• 12:17
  knowledge
• 12:17
  base
• 12:18
  grows,
• 12:18
  and
• 12:18
  the
• 12:18
  subscription-based
• 12:19
  coding
• 12:19
  SDKs
• 12:20
  are
• 12:20
  both
• 12:20
  too
• 12:21
  expensive
• 12:21
  and
• 12:21
  too
• 12:22
  slow
• 12:22
  for
• 12:22
  production
• 12:22
  traffic.
• 12:23
  The
• 12:23
  architecture
• 12:24
  that
• 12:24
  replaces
• 12:24
  it
• 12:25
  has
• 12:25
  two
• 12:25
  components:
• 12:26
  a
• 12:26
  context
• 12:26
  retriever
• 12:27
  (structured
• 12:27
  MCP
• 12:27
  access
• 12:28
  to
• 12:28
  business
• 12:28
  data
• 12:28
  with
• 12:29
  auto-generated
• 12:29
  filter
• 12:30
  and
• 12:30
  search
• 12:30
  tools)
• 12:30
  and
• 12:30
  agent
• 12:31
  memory
• 12:31
  (short-term
• 12:31
  conversation
• 12:32
  state
• 12:33
  automatically
• 12:33
  promoted
• 12:33
  to
• 12:34
  long-term
• 12:34
  vector-indexed
• 12:34
  memories
• 12:35
  by
• 12:35
  a
• 12:35
  background
• 12:35
  extraction
• 12:36
  process).
• 12:36
  In
• 12:37
  Medin's
• 12:37
  Redis
• 12:37
  Iris
• 12:37
  demonstration,
• 12:38
  a
• 12:38
  single
• 12:39
  MCP
• 12:39
  tool
• 12:39
  call
• 12:39
  replaces
• 12:40
  what
• 12:40
  would
• 12:40
  have
• 12:40
  been
• 12:40
  reading
• 12:41
  an
• 12:41
  index
• 12:41
  document,
• 12:42
  navigating
• 12:42
  to
• 12:42
  sub-documents,
• 12:43
  and
• 12:43
  parsing
• 12:43
  unstructured
• 12:44
  text
• 12:44
  —
• 12:45
  the
• 12:45
  query
• 12:45
  for
• 12:45
  delayed
• 12:45
  orders
• 12:46
  returns
• 12:46
  in
• 12:46
  one
• 12:46
  call
• 12:47
  regardless
• 12:47
  of
• 12:47
  database
• 12:47
  size.
• 12:48
  The
• 12:49
  Incident-to-Eval
• 12:49
  Synthesis
• 12:50
  pattern
• 12:50
  from
• 12:50
  the
• 12:50
  AI
• 12:50
  Pattern
• 12:51
  Book
• 12:51
  addresses
• 12:51
  the
• 12:51
  other
• 12:51
  production
• 12:52
  gap:
• 12:52
  ensuring
• 12:53
  agents
• 12:53
  do
• 12:53
  not
• 12:54
  regress.
• 12:55
  The
• 12:55
  discipline
• 12:55
  is
• 12:55
  straightforward
• 12:56
  but
• 12:56
  hard
• 12:56
  to
• 12:56
  maintain
• 12:56
  —
• 12:57
  every
• 12:57
  production
• 12:58
  incident
• 12:58
  becomes
• 12:58
  a
• 12:58
  two-part
• 12:59
  fix:
• 12:59
  repair
• 13:00
  the
• 13:00
  underlying
• 13:00
  pattern,
• 13:01
  then
• 13:01
  encode
• 13:01
  the
• 13:01
  failure
• 13:02
  as
• 13:02
  a
• 13:02
  permanent
• 13:02
  eval
• 13:03
  that
• 13:03
  reproduces
• 13:03
  it.
• 13:04
  The
• 13:04
  eval
• 13:04
  lands
• 13:05
  in
• 13:05
  the
• 13:05
  same
• 13:05
  commit
• 13:05
  as
• 13:05
  the
• 13:05
  fix,
• 13:06
  the
• 13:06
  CI
• 13:07
  gate
• 13:07
  runs
• 13:07
  it
• 13:07
  on
• 13:07
  every
• 13:07
  relevant
• 13:08
  change,
• 13:08
  and
• 13:08
  recurrence
• 13:09
  becomes
• 13:09
  structurally
• 13:10
  hard
• 13:10
  rather
• 13:10
  than
• 13:10
  routinely
• 13:10
  surprising.
• 13:11
  A
• 13:11
  support
• 13:12
  agent
• 13:12
  that
• 13:12
  invented
• 13:13
  a
• 13:13
  nonexistent
• 13:13
  refund
• 13:13
  policy,
• 13:14
  a
• 13:14
  coding
• 13:14
  agent
• 13:15
  that
• 13:15
  deleted
• 13:15
  a
• 13:15
  migration
• 13:16
  file
• 13:16
  it
• 13:16
  judged
• 13:17
  "unused"
• 13:17
  —
• 13:18
  each
• 13:18
  becomes
• 13:18
  a
• 13:18
  golden
• 13:18
  test
• 13:19
  case
• 13:19
  that
• 13:19
  no
• 13:19
  up-front
• 13:20
  test
• 13:20
  designer
• 13:20
  would
• 13:20
  have
• 13:20
  thought
• 13:20
  to
• 13:21
  write.
• 13:21
  The
• 13:22
  maturity
• 13:22
  signal
• 13:22
  is
• 13:22
  the
• 13:22
  ratio
• 13:23
  of
• 13:23
  incident-derived
• 13:24
  evals
• 13:24
  to
• 13:24
  hand-designed
• 13:25
  ones:
• 13:25
  a
• 13:25
  young
• 13:26
  suite
• 13:26
  is
• 13:26
  mostly
• 13:26
  imagination;
• 13:28
  a
• 13:28
  mature
• 13:28
  suite
• 13:28
  has
• 13:29
  absorbed
• 13:29
  enough
• 13:29
  production
• 13:30
  reality
• 13:30
  to
• 13:30
  carry
• 13:30
  incident
• 13:31
  numbers
• 13:31
  on
• 13:31
  a
• 13:31
  growing
• 13:31
  share
• 13:32
  of
• 13:32
  its
• 13:32
  cases.
• 13:35
  The
• 13:35
  CMP
• 13:35
  170HX
• 13:38
  is
• 13:38
  a
• 13:38
  cryptocurrency
• 13:39
  mining
• 13:39
  GPU
• 13:39
  that
• 13:40
  NVIDIA
• 13:40
  shipped
• 13:40
  with
• 13:40
  its
• 13:41
  memory
• 13:41
  artificially
• 13:41
  capped
• 13:42
  at
• 13:42
  8GB.
• 13:43
  A
• 13:43
  patched
• 13:43
  set
• 13:44
  of
• 13:44
  open
• 13:44
  kernel
• 13:44
  modules
• 13:45
  for
• 13:45
  driver
• 13:45
  610.43.03
• 13:48
  unlocks
• 13:48
  the
• 13:48
  full
• 13:48
  64GB
• 13:49
  of
• 13:49
  HBM2e
• 13:50
  and
• 13:50
  173
• 13:51
  TFLOPS
• 13:53
  of
• 13:53
  BF16
• 13:53
  compute.
• 13:54
  The
• 13:54
  installer
• 13:55
  builds
• 13:55
  from
• 13:55
  source,
• 13:56
  blacklists
• 13:56
  nouveau,
• 13:57
  rebuilds
• 13:57
  initramfs,
• 13:57
  and
• 13:58
  requires
• 13:58
  a
• 13:58
  cold
• 13:58
  reboot.
• 13:59
  The
• 13:59
  catch
• 13:59
  is
• 14:00
  PCIe:
• 14:01
  the
• 14:01
  card
• 14:01
  is
• 14:01
  physically
• 14:02
  wired
• 14:02
  as
• 14:02
  x4
• 14:02
  (missing
• 14:03
  AC
• 14:03
  coupling
• 14:04
  capacitors
• 14:04
  on
• 14:04
  lanes
• 14:04
  4–15),
• 14:06
  and
• 14:06
  the
• 14:06
  current
• 14:06
  unlock
• 14:07
  only
• 14:07
  reaches
• 14:07
  Gen2
• 14:08
  —
• 14:09
  about
• 14:09
  2GB/s
• 14:10
  of
• 14:10
  bandwidth.
• 14:11
  That
• 14:12
  wrecks
• 14:12
  multi-GPU
• 14:13
  inference
• 14:13
  and
• 14:13
  limits
• 14:13
  host-device
• 14:14
  transfer
• 14:14
  speed,
• 14:15
  but
• 14:15
  for
• 14:15
  single-card
• 14:16
  image
• 14:16
  generation
• 14:17
  or
• 14:17
  batched
• 14:17
  local
• 14:17
  inference
• 14:18
  where
• 14:18
  the
• 14:18
  model
• 14:18
  fits
• 14:19
  entirely
• 14:19
  in
• 14:19
  HBM,
• 14:20
  64GB
• 14:21
  at
• 14:21
  this
• 14:21
  price
• 14:21
  point
• 14:22
  is
• 14:22
  extraordinary.
• 14:23
  Some
• 14:23
  community
• 14:23
  members
• 14:24
  report
• 14:24
  renting
• 14:24
  unlocked
• 14:25
  CMPs
• 14:25
  for
• 14:25
  half
• 14:26
  the
• 14:26
  cost
• 14:26
  of
• 14:26
  A100s.
• 14:27
  The
• 14:28
  GGML
• 14:28
  runtime
• 14:29
  continues
• 14:29
  expanding
• 14:29
  beyond
• 14:30
  text.
• 14:31
  Trellis.cpp,
• 14:33
  a
• 14:33
  GGML
• 14:34
  port
• 14:34
  of
• 14:34
  TRELLIS.2,
• 14:35
  now
• 14:35
  produces
• 14:36
  3D
• 14:36
  assets
• 14:36
  on
• 14:36
  par
• 14:37
  with
• 14:37
  the
• 14:37
  reference
• 14:37
  implementation
• 14:38
  after
• 14:38
  a
• 14:38
  significant
• 14:39
  debugging
• 14:39
  effort.
• 14:40
  This
• 14:40
  means
• 14:41
  top-tier
• 14:41
  open-source
• 14:42
  image-to-3D
• 14:42
  generation
• 14:43
  runs
• 14:43
  on
• 14:43
  any
• 14:44
  GPU
• 14:44
  with
• 14:44
  enough
• 14:44
  memory
• 14:45
  —
• 14:45
  or
• 14:45
  even
• 14:46
  on
• 14:46
  CPU
• 14:46
  for
• 14:46
  the
• 14:46
  patient
• 14:47
  —
• 14:47
  without
• 14:48
  requiring
• 14:48
  CUDA.
• 14:49
  The
• 14:49
  project
• 14:50
  connects
• 14:50
  to
• 14:50
  Lemonade
• 14:51
  for
• 14:51
  an
• 14:51
  integrated
• 14:51
  experience
• 14:52
  with
• 14:52
  optional
• 14:52
  text-to-3D
• 14:53
  cascading.
• 14:54
  Topology
• 14:54
  remains
• 14:55
  rough
• 14:55
  (useful
• 14:55
  for
• 14:55
  "rough
• 14:56
  drafts"
• 14:56
  that
• 14:56
  need
• 14:56
  retopologizing
• 14:57
  for
• 14:57
  rigging),
• 14:58
  but
• 14:58
  the
• 14:58
  output
• 14:59
  quality
• 14:59
  in
• 14:59
  the
• 14:59
  screenshots
• 15:00
  is
• 15:00
  a
• 15:00
  visible
• 15:00
  step
• 15:00
  change
• 15:01
  from
• 15:01
  earlier
• 15:01
  ports.
• 15:02
  The
• 15:02
  full
• 15:03
  creative
• 15:03
  pipeline
• 15:03
  is
• 15:03
  converging
• 15:04
  on
• 15:04
  local
• 15:04
  hardware.
• 15:05
  One
• 15:05
  developer
• 15:06
  produced
• 15:06
  a
• 15:06
  4-minute
• 15:07
  animated
• 15:07
  film
• 15:07
  overnight
• 15:08
  on
• 15:08
  an
• 15:08
  M5
• 15:08
  Max
• 15:08
  MacBook
• 15:09
  (128GB,
• 15:11
  wifi
• 15:11
  off):
• 15:12
  Flux
• 15:12
  draws
• 15:12
  the
• 15:13
  stills,
• 15:13
  Wan
• 15:13
  2.2
• 15:14
  and
• 15:14
  LTX
• 15:14
  animate
• 15:15
  them,
• 15:15
  Piper
• 15:16
  narrates,
• 15:16
  Ace-Step
• 15:17
  writes
• 15:17
  and
• 15:17
  performs
• 15:18
  the
• 15:18
  score,
• 15:18
  and
• 15:19
  ffmpeg
• 15:19
  stitches
• 15:19
  everything
• 15:20
  into
• 15:20
  a
• 15:20
  finished
• 15:20
  film
• 15:21
  with
• 15:21
  titles
• 15:21
  and
• 15:21
  credits.
• 15:22
  Characters
• 15:23
  drift
• 15:23
  between
• 15:23
  shots,
• 15:24
  motion
• 15:24
  comes
• 15:24
  in
• 15:25
  short
• 15:25
  clips,
• 15:25
  and
• 15:25
  longer
• 15:26
  films
• 15:26
  take
• 15:26
  about
• 15:26
  12
• 15:27
  hours
• 15:27
  of
• 15:27
  compute.
• 15:28
  But
• 15:28
  it
• 15:28
  is
• 15:28
  a
• 15:28
  complete
• 15:29
  movie
• 15:29
  from
• 15:29
  one
• 15:29
  pipeline
• 15:30
  on
• 15:30
  one
• 15:30
  laptop
• 15:30
  —
• 15:31
  an
• 15:31
  outcome
• 15:32
  that
• 15:32
  would
• 15:32
  have
• 15:32
  been
• 15:32
  unthinkable
• 15:33
  a
• 15:33
  year
• 15:33
  ago.
• 15:34
  The
• 15:34
  pipeline
• 15:34
  is
• 15:34
  open-source.
• 15:35
  Ghost
• 15:36
  in
• 15:36
  the
• 15:36
  Droid
• 15:36
  gives
• 15:36
  AI
• 15:37
  agents
• 15:37
  a
• 15:37
  physical
• 15:37
  phone
• 15:38
  as
• 15:38
  a
• 15:38
  tool
• 15:38
  surface
• 15:38
  via
• 15:39
  Model
• 15:39
  Context
• 15:39
  Protocol.
• 15:40
  The
• 15:40
  agent
• 15:41
  observes
• 15:41
  a
• 15:41
  screenshot
• 15:42
  plus
• 15:42
  a
• 15:42
  compacted
• 15:43
  accessibility
• 15:43
  tree
• 15:44
  and
• 15:44
  acts
• 15:44
  through
• 15:44
  62
• 15:45
  tools
• 15:45
  —
• 15:46
  tap,
• 15:46
  swipe,
• 15:47
  type,
• 15:48
  launch,
• 15:48
  batch
• 15:49
  flows
• 15:49
  —
• 15:50
  over
• 15:50
  ADB
• 15:50
  with
• 15:51
  no
• 15:51
  app
• 15:51
  on
• 15:51
  the
• 15:51
  phone
• 15:51
  and
• 15:52
  no
• 15:52
  root
• 15:52
  required.
• 15:53
  iPhone
• 15:53
  support
• 15:54
  arrived
• 15:54
  this
• 15:54
  release
• 15:54
  via
• 15:55
  WebDriverAgent.
• 15:56
  Because
• 15:56
  Ghost
• 15:56
  is
• 15:56
  just
• 15:57
  the
• 15:57
  body,
• 15:57
  any
• 15:58
  model
• 15:58
  —
• 15:58
  cloud
• 15:59
  or
• 15:59
  local
• 15:59
  —
• 16:00
  can
• 16:00
  serve
• 16:00
  as
• 16:00
  the
• 16:00
  brain.
• 16:01
  Meanwhile,
• 16:02
  Grok's
• 16:02
  video
• 16:03
  generation
• 16:03
  is
• 16:03
  drawing
• 16:04
  reactions
• 16:04
  like
• 16:04
  "this
• 16:05
  doesn't
• 16:05
  even
• 16:05
  look
• 16:05
  AI-generated
• 16:06
  anymore."
• 16:07
  The
• 16:07
  community
• 16:08
  is
• 16:08
  more
• 16:08
  measured
• 16:08
  —
• 16:09
  eyes
• 16:09
  that
• 16:10
  do
• 16:10
  not
• 16:10
  blink,
• 16:11
  mouths
• 16:11
  that
• 16:11
  open
• 16:11
  too
• 16:11
  wide,
• 16:12
  the
• 16:12
  uncanny-valley
• 16:13
  artifacts
• 16:13
  that
• 16:14
  read
• 16:14
  as
• 16:14
  "70s
• 16:14
  TV
• 16:15
  show"
• 16:15
  to
• 16:16
  a
• 16:16
  critical
• 16:16
  viewer
• 16:16
  —
• 16:17
  but
• 16:17
  the
• 16:17
  trajectory
• 16:18
  is
• 16:18
  undeniable.
• 16:19
  The
• 16:19
  gap
• 16:20
  between
• 16:20
  AI-generated
• 16:21
  video
• 16:21
  and
• 16:21
  conventional
• 16:22
  production
• 16:22
  narrows
• 16:23
  with
• 16:23
  every
• 16:23
  release,
• 16:24
  even
• 16:25
  if
• 16:25
  the
• 16:25
  last
• 16:25
  mile
• 16:25
  of
• 16:25
  realism
• 16:26
  remains
• 16:26
  stubborn.