Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – August 06, 2026

Transcript

AGI Dreams Podcast – August 06, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Demis
• 0:00
  Hassabis
• 0:00
  has
• 0:00
  described
• 0:01
  himself
• 0:01
  as
• 0:01
  standing
• 0:02
  in
• 0:02
  the
• 0:02
  foothills
• 0:02
  of
• 0:02
  the
• 0:02
  singularity,
• 0:03
  and
• 0:04
  now
• 0:04
  he
• 0:04
  wants
• 0:04
  to
• 0:04
  spend
• 0:04
  his
• 0:05
  time
• 0:05
  at
• 0:05
  the
• 0:05
  summit
• 0:05
  rather
• 0:06
  than
• 0:06
  in
• 0:06
  the
• 0:06
  base
• 0:06
  camp
• 0:07
  operations
• 0:07
  center.
• 0:08
  Google
• 0:08
  announced
• 0:09
  that
• 0:09
  Hassabis
• 0:09
  is
• 0:09
  stepping
• 0:10
  back
• 0:10
  from
• 0:10
  day-to-day
• 0:11
  operational
• 0:11
  leadership
• 0:12
  of
• 0:12
  Google
• 0:12
  DeepMind
• 0:12
  to
• 0:13
  become
• 0:13
  Chair
• 0:13
  of
• 0:13
  GDM
• 0:14
  and
• 0:14
  Chief
• 0:14
  Scientist
• 0:14
  of
• 0:15
  Alphabet,
• 0:16
  a
• 0:16
  move
• 0:16
  he
• 0:16
  and
• 0:16
  Sundar
• 0:16
  Pichai
• 0:17
  frame
• 0:17
  as
• 0:17
  freeing
• 0:18
  him
• 0:18
  to
• 0:18
  "actively
• 0:18
  shape
• 0:19
  the
• 0:19
  future
• 0:19
  of
• 0:19
  AGI."
• 0:20
  Koray
• 0:20
  Kavukcuoglu,
• 0:21
  GDM's
• 0:22
  CTO
• 0:22
  and
• 0:22
  Chief
• 0:23
  AI
• 0:23
  Architect
• 0:24
  since
• 0:24
  the
• 0:24
  early
• 0:24
  DeepMind
• 0:25
  days,
• 0:25
  steps
• 0:25
  up
• 0:26
  as
• 0:26
  SVP
• 0:26
  of
• 0:26
  Google
• 0:26
  DeepMind
• 0:27
  and
• 0:27
  takes
• 0:27
  over
• 0:27
  Gemini
• 0:28
  model
• 0:28
  development,
• 0:29
  frontier
• 0:29
  research,
• 0:30
  and
• 0:30
  the
• 0:30
  Gemini
• 0:30
  app
• 0:31
  and
• 0:31
  developer
• 0:31
  teams.
• 0:32
  The
• 0:32
  bigger
• 0:32
  surprise
• 0:33
  is
• 0:33
  Jeff
• 0:33
  Dean's
• 0:33
  departure
• 0:34
  after
• 0:34
  27
• 0:34
  years
• 0:35
  at
• 0:35
  Google.
• 0:36
  Dean
• 0:36
  and
• 0:36
  fellow
• 0:36
  Senior
• 0:37
  Fellow
• 0:37
  Sanjay
• 0:37
  Ghemawat
• 0:38
  are
• 0:38
  launching
• 0:38
  an
• 0:38
  independent
• 0:39
  public
• 0:39
  benefit
• 0:39
  corporation
• 0:40
  to
• 0:40
  accelerate
• 0:41
  discoveries
• 0:41
  in
• 0:41
  ML,
• 0:41
  science,
• 0:42
  and
• 0:42
  engineering.
• 0:43
  Google
• 0:43
  is
• 0:44
  backing
• 0:44
  it
• 0:44
  as
• 0:44
  a
• 0:44
  founding
• 0:45
  investor
• 0:45
  and
• 0:45
  Cloud
• 0:45
  partner.
• 0:46
  The
• 0:46
  pair
• 0:47
  built
• 0:47
  the
• 0:47
  infrastructure
• 0:47
  scaffolding
• 0:48
  --
• 0:49
  MapReduce,
• 0:49
  Bigtable,
• 0:50
  TensorFlow
• 0:50
  --
• 0:51
  that
• 0:51
  the
• 0:51
  modern
• 0:51
  AI
• 0:52
  era
• 0:52
  runs
• 0:52
  on.
• 0:53
  Their
• 0:53
  exit
• 0:54
  as
• 0:54
  a
• 0:54
  unit
• 0:54
  to
• 0:54
  start
• 0:55
  something
• 0:55
  mission-driven
• 0:55
  rather
• 0:56
  than
• 0:56
  retiring
• 0:56
  quietly
• 0:57
  signals
• 0:57
  they
• 0:57
  believe
• 0:58
  something
• 0:58
  more
• 0:58
  important
• 0:58
  can
• 0:59
  be
• 0:59
  built
• 0:59
  outside
• 0:59
  Google's
• 1:00
  orbit,
• 1:00
  not
• 1:00
  inside
• 1:01
  it.
• 1:01
  The
• 1:02
  operational
• 1:02
  read
• 1:02
  is
• 1:02
  straightforward:
• 1:04
  Pichai
• 1:04
  is
• 1:04
  consolidating
• 1:05
  direct
• 1:05
  reports.
• 1:06
  Kavukcuoglu
• 1:07
  now
• 1:07
  reports
• 1:07
  directly
• 1:08
  to
• 1:08
  him,
• 1:08
  shortening
• 1:08
  the
• 1:09
  chain
• 1:09
  between
• 1:09
  product
• 1:09
  decisions
• 1:10
  and
• 1:10
  AI
• 1:10
  research.
• 1:11
  Hassabis's
• 1:12
  pivot
• 1:12
  to
• 1:12
  Isomorphic
• 1:13
  Labs
• 1:13
  --
• 1:13
  where
• 1:13
  he
• 1:14
  wants
• 1:14
  AI
• 1:14
  to
• 1:14
  "help
• 1:14
  finally
• 1:15
  cure
• 1:15
  diseases
• 1:15
  like
• 1:16
  cancer"
• 1:16
  --
• 1:16
  suggests
• 1:17
  the
• 1:17
  best-case
• 1:17
  scenario
• 1:18
  for
• 1:18
  DeepMind's
• 1:18
  founder:
• 1:19
  the
• 1:19
  models
• 1:20
  are
• 1:20
  good
• 1:20
  enough
• 1:20
  that
• 1:20
  daily
• 1:21
  management
• 1:21
  is
• 1:21
  less
• 1:21
  interesting
• 1:22
  than
• 1:22
  applying
• 1:22
  them.
• 1:23
  The
• 1:23
  risk
• 1:23
  is
• 1:23
  more
• 1:24
  subtle.
• 1:25
  DeepMind's
• 1:25
  culture
• 1:25
  was
• 1:26
  built
• 1:26
  around
• 1:26
  Hassabis's
• 1:27
  personal
• 1:27
  credibility
• 1:27
  with
• 1:28
  researchers.
• 1:29
  Chair
• 1:29
  titles
• 1:29
  have
• 1:29
  a
• 1:30
  way
• 1:30
  of
• 1:30
  becoming
• 1:30
  ceremonial.
• 1:32
  Whether
• 1:32
  Kavukcuoglu
• 1:32
  can
• 1:33
  maintain
• 1:33
  that
• 1:33
  gravitational
• 1:34
  pull
• 1:34
  while
• 1:34
  simultaneously
• 1:35
  shipping
• 1:35
  Gemini
• 1:36
  4
• 1:36
  and
• 1:36
  managing
• 1:36
  950
• 1:37
  million
• 1:37
  monthly
• 1:38
  Gemini
• 1:38
  app
• 1:38
  users
• 1:39
  will
• 1:39
  determine
• 1:39
  whether
• 1:39
  this
• 1:39
  is
• 1:40
  a
• 1:40
  succession
• 1:40
  or
• 1:40
  a
• 1:41
  slow
• 1:41
  unwinding.
• 1:44
  The
• 1:44
  Grok
• 1:44
  Build
• 1:44
  CLI
• 1:45
  data
• 1:45
  exfiltration
• 1:46
  story
• 1:46
  broke
• 1:46
  in
• 1:46
  July,
• 1:47
  and
• 1:47
  now
• 1:47
  someone
• 1:47
  has
• 1:47
  built
• 1:48
  the
• 1:48
  forensic
• 1:48
  tool
• 1:48
  you
• 1:48
  actually
• 1:49
  need
• 1:49
  afterward.
• 1:50
  GrokPatrol
• 1:51
  is
• 1:51
  a
• 1:51
  single
• 1:51
  static
• 1:52
  Go
• 1:52
  binary
• 1:52
  --
• 1:53
  zero
• 1:53
  dependencies,
• 1:54
  no
• 1:54
  network
• 1:54
  calls
• 1:55
  (proven
• 1:55
  at
• 1:55
  the
• 1:55
  linker
• 1:55
  level
• 1:56
  by
• 1:56
  asserting
• 1:56
  that
• 1:56
  net,
• 1:56
  net/http,
• 1:58
  and
• 1:58
  crypto/tls
• 2:00
  are
• 2:01
  absent
• 2:01
  from
• 2:01
  the
• 2:01
  dependency
• 2:02
  tree),
• 2:02
  read-only
• 2:03
  (every
• 2:03
  file
• 2:03
  open
• 2:03
  is
• 2:03
  O
• 2:04
  RDONLY)
• 2:04
  --
• 2:05
  that
• 2:05
  answers
• 2:06
  the
• 2:06
  question:
• 2:06
  what
• 2:06
  did
• 2:07
  the
• 2:07
  Grok
• 2:07
  CLI
• 2:07
  take
• 2:07
  from
• 2:08
  this
• 2:08
  machine,
• 2:09
  and
• 2:09
  which
• 2:09
  secrets
• 2:09
  went
• 2:09
  with
• 2:09
  it?
• 2:10
  It
• 2:10
  scans
• 2:11
  Grok's
• 2:11
  logs
• 2:11
  (including
• 2:12
  rotated
• 2:12
  and
• 2:12
  gzipped),
• 2:13
  the
• 2:13
  upload
• 2:13
  queue
• 2:14
  at
• 2:14
  ~/.grok/upload
• 2:15
  queue/,
• 2:16
  staged
• 2:16
  archive
• 2:16
  manifests
• 2:17
  naming
• 2:18
  the
• 2:18
  gs://grok-code-session-traces/
• 2:20
  destination
• 2:21
  bucket,
• 2:21
  and
• 2:21
  critically,
• 2:22
  enumerates
• 2:23
  secrets
• 2:23
  in
• 2:23
  the
• 2:23
  uploaded
• 2:24
  object
• 2:24
  set
• 2:24
  by
• 2:24
  diffing
• 2:25
  git
• 2:25
  rev-list
• 2:25
  --objects
• 2:26
  HEAD
• 2:26
  against
• 2:26
  the
• 2:26
  working
• 2:26
  tree.
• 2:27
  The
• 2:27
  secrets
• 2:28
  it
• 2:28
  flags
• 2:28
  first
• 2:28
  are
• 2:28
  the
• 2:29
  ones
• 2:29
  you
• 2:29
  cannot
• 2:29
  see
• 2:29
  in
• 2:29
  your
• 2:29
  checkout
• 2:30
  anymore
• 2:30
  --
• 2:31
  the
• 2:31
  deleted
• 2:31
  .env,
• 2:32
  the
• 2:32
  rotated
• 2:32
  .pem
• 2:33
  --
• 2:33
  because
• 2:34
  those
• 2:34
  are
• 2:34
  the
• 2:34
  credentials
• 2:34
  that
• 2:34
  went
• 2:35
  out
• 2:35
  in
• 2:35
  git
• 2:35
  history
• 2:35
  and
• 2:36
  are
• 2:36
  hardest
• 2:36
  to
• 2:36
  remember
• 2:36
  to
• 2:36
  rotate.
• 2:37
  The
• 2:37
  design
• 2:38
  constraints
• 2:38
  are
• 2:38
  worth
• 2:38
  noting.
• 2:40
  GrokPatrol
• 2:41
  never
• 2:41
  executes
• 2:42
  the
• 2:42
  grok
• 2:42
  binary,
• 2:43
  not
• 2:43
  even
• 2:43
  for
• 2:44
  --version,
• 2:44
  because
• 2:45
  the
• 2:45
  Grok
• 2:45
  CLI
• 2:46
  carries
• 2:46
  a
• 2:46
  collector
• 2:46
  that
• 2:46
  runs
• 2:47
  outside
• 2:47
  its
• 2:47
  own
• 2:47
  permission
• 2:48
  system
• 2:48
  --
• 2:49
  launching
• 2:49
  it
• 2:49
  to
• 2:50
  ask
• 2:50
  a
• 2:50
  question
• 2:50
  could
• 2:50
  start
• 2:51
  a
• 2:51
  session.
• 2:52
  A
• 2:52
  degraded
• 2:53
  scan
• 2:53
  (macOS
• 2:54
  TCC
• 2:54
  blocked
• 2:54
  a
• 2:54
  directory)
• 2:55
  never
• 2:55
  reports
• 2:55
  CLEAN;
• 2:56
  it
• 2:56
  reports
• 2:57
  INDETERMINATE
• 2:58
  and
• 2:58
  prints
• 2:58
  what
• 2:58
  it
• 2:58
  could
• 2:58
  not
• 2:58
  see.
• 2:59
  Version
• 3:00
  detection
• 3:00
  covers
• 3:00
  0.1.212
• 3:03
  through
• 3:03
  the
• 3:03
  latest
• 3:03
  observed,
• 3:04
  with
• 3:04
  0.2.93
• 3:06
  confirmed
• 3:07
  affected
• 3:07
  and
• 3:07
  the
• 3:07
  collector
• 3:08
  reportedly
• 3:08
  present
• 3:09
  through
• 3:09
  at
• 3:09
  least
• 3:09
  0.2.99.
• 3:12
  This
• 3:12
  is
• 3:12
  what
• 3:12
  incident
• 3:13
  response
• 3:13
  tooling
• 3:13
  should
• 3:14
  look
• 3:14
  like:
• 3:14
  every
• 3:15
  finding
• 3:15
  cites
• 3:15
  something
• 3:16
  you
• 3:16
  can
• 3:16
  go
• 3:16
  look
• 3:16
  at,
• 3:17
  absence
• 3:17
  of
• 3:17
  evidence
• 3:18
  is
• 3:18
  explicitly
• 3:19
  not
• 3:19
  evidence
• 3:19
  of
• 3:19
  absence,
• 3:20
  and
• 3:20
  the
• 3:20
  tool
• 3:21
  itself
• 3:21
  is
• 3:21
  small
• 3:21
  enough
• 3:22
  to
• 3:22
  grep
• 3:22
  in
• 3:22
  an
• 3:22
  afternoon.
• 3:23
  On
• 3:23
  a
• 3:24
  different
• 3:24
  attack
• 3:24
  surface
• 3:24
  entirely,
• 3:25
  researchers
• 3:26
  at
• 3:26
  Mysk
• 3:26
  found
• 3:27
  three
• 3:27
  WebKit
• 3:27
  features
• 3:28
  --
• 3:28
  DNS
• 3:29
  prefetching,
• 3:30
  WebAuthn
• 3:30
  Related
• 3:31
  Origin
• 3:31
  Requests,
• 3:31
  and
• 3:32
  WebTransport
• 3:32
  --
• 3:33
  that
• 3:33
  bypass
• 3:34
  proxy
• 3:34
  configurations
• 3:35
  on
• 3:35
  iOS
• 3:36
  and
• 3:36
  macOS,
• 3:37
  leaking
• 3:37
  the
• 3:37
  user's
• 3:38
  real
• 3:38
  IP
• 3:38
  address
• 3:39
  and
• 3:39
  DNS
• 3:39
  queries
• 3:40
  even
• 3:40
  when
• 3:40
  traffic
• 3:40
  is
• 3:41
  routed
• 3:41
  through
• 3:41
  a
• 3:41
  proxy.
• 3:42
  The
• 3:42
  DNS
• 3:43
  prefetch
• 3:43
  leak
• 3:44
  is
• 3:44
  triggered
• 3:44
  by
• 3:44
  a
• 3:44
  simple
• 3:45
  HTML
• 3:45
  tag,
• 3:46
  works
• 3:46
  since
• 3:46
  iOS
• 3:47
  26.0,
• 3:48
  and
• 3:49
  affects
• 3:49
  Apple's
• 3:49
  iCloud
• 3:50
  Private
• 3:50
  Relay
• 3:50
  as
• 3:51
  well.
• 3:52
  WebAuthn's
• 3:52
  Related
• 3:53
  Origin
• 3:53
  Requests
• 3:54
  hand
• 3:54
  credential
• 3:54
  validation
• 3:55
  to
• 3:55
  the
• 3:55
  OS
• 3:55
  credential
• 3:56
  service,
• 3:57
  which
• 3:57
  fires
• 3:57
  an
• 3:57
  HTTPS
• 3:58
  fetch
• 3:59
  directly
• 3:59
  from
• 3:59
  the
• 3:59
  device,
• 4:00
  unaware
• 4:01
  of
• 4:01
  any
• 4:01
  proxy.
• 4:03
  WebTransport
• 4:03
  opens
• 4:04
  a
• 4:04
  direct
• 4:04
  QUIC
• 4:05
  connection
• 4:05
  that
• 4:06
  WebKit
• 4:06
  never
• 4:06
  offers
• 4:07
  to
• 4:07
  the
• 4:07
  proxy
• 4:07
  session.
• 4:08
  All
• 4:08
  three
• 4:09
  affect
• 4:09
  every
• 4:09
  iOS
• 4:10
  browser
• 4:10
  (since
• 4:10
  Apple
• 4:11
  requires
• 4:11
  WebKit),
• 4:12
  including
• 4:13
  iOS
• 4:13
  Tor
• 4:13
  browsers.
• 4:14
  VPNs
• 4:15
  are
• 4:15
  not
• 4:15
  affected
• 4:16
  because
• 4:16
  they
• 4:16
  tunnel
• 4:17
  at
• 4:17
  the
• 4:17
  system
• 4:17
  level.
• 4:18
  Meanwhile,
• 4:19
  Cloud.ru
• 4:20
  released
• 4:20
  guardrails-llm-filter-extproc,
• 4:23
  an
• 4:23
  Envoy
• 4:23
  external
• 4:24
  processor
• 4:24
  that
• 4:24
  masks
• 4:25
  PII
• 4:25
  and
• 4:26
  secrets
• 4:26
  in
• 4:26
  LLM-bound
• 4:27
  traffic
• 4:27
  using
• 4:28
  roughly
• 4:28
  260
• 4:29
  built-in
• 4:29
  regex
• 4:30
  rules
• 4:30
  covering
• 4:30
  credentials,
• 4:32
  API
• 4:32
  keys,
• 4:33
  access
• 4:33
  tokens,
• 4:34
  IP
• 4:35
  addresses,
• 4:35
  and
• 4:35
  Russian-specific
• 4:36
  personal
• 4:37
  data
• 4:37
  (SNILS,
• 4:37
  INN,
• 4:38
  OGRN
• 4:39
  with
• 4:39
  checksum
• 4:39
  validation).
• 4:41
  It
• 4:41
  replaces
• 4:41
  detected
• 4:42
  values
• 4:42
  with
• 4:42
  safe
• 4:43
  placeholders
• 4:43
  like
• 4:43
  1>
• 4:43
  on
• 4:44
  the
• 4:44
  way
• 4:44
  to
• 4:44
  the
• 4:44
  model
• 4:44
  and
• 4:44
  restores
• 4:45
  originals
• 4:45
  on
• 4:45
  the
• 4:46
  way
• 4:46
  back,
• 4:46
  including
• 4:47
  in
• 4:47
  streaming
• 4:47
  SSE
• 4:48
  responses
• 4:49
  and
• 4:49
  tool-call
• 4:49
  arguments.
• 4:50
  The
• 4:51
  provider
• 4:51
  never
• 4:51
  sees
• 4:52
  sensitive
• 4:52
  values;
• 4:53
  the
• 4:53
  client
• 4:53
  never
• 4:54
  sees
• 4:54
  placeholders.
• 4:55
  Storage
• 4:56
  backends
• 4:56
  include
• 4:56
  in-memory,
• 4:57
  Redis,
• 4:58
  or
• 4:58
  Postgres
• 4:59
  with
• 4:59
  optional
• 4:59
  AES-256-GCM
• 5:04
  encryption.
• 5:09
  Liquid
• 5:09
  AI
• 5:10
  released
• 5:10
  LFM2.5-2.6B,
• 5:13
  a
• 5:13
  2.69-billion-parameter
• 5:15
  model
• 5:15
  with
• 5:15
  128K
• 5:16
  context
• 5:17
  and
• 5:17
  tool
• 5:17
  calling,
• 5:18
  post-trained
• 5:18
  specifically
• 5:19
  for
• 5:19
  multi-step
• 5:19
  agent
• 5:20
  workflows.
• 5:21
  The
• 5:21
  Q4
• 5:21
  K
• 5:21
  M
• 5:21
  GGUF
• 5:22
  is
• 5:22
  1.67
• 5:23
  GB
• 5:23
  and
• 5:23
  already
• 5:24
  works
• 5:24
  with
• 5:24
  llama.cpp.
• 5:26
  Vendor-reported
• 5:26
  speeds:
• 5:27
  30
• 5:27
  tok/s
• 5:27
  on
• 5:28
  a
• 5:28
  phone,
• 5:28
  113
• 5:29
  tok/s
• 5:29
  on
• 5:30
  a
• 5:30
  Ryzen
• 5:30
  AI
• 5:30
  Max+
• 5:31
  395,
• 5:32
  220
• 5:32
  tok/s
• 5:33
  on
• 5:33
  an
• 5:34
  M5
• 5:34
  Max,
• 5:34
  under
• 5:35
  2.5
• 5:35
  GB
• 5:35
  memory.
• 5:36
  Benchmarks
• 5:36
  are
• 5:36
  competitive
• 5:37
  for
• 5:37
  the
• 5:37
  size
• 5:37
  --
• 5:38
  ToolSandbox
• 5:38
  77.83
• 5:40
  versus
• 5:40
  76.44
• 5:41
  for
• 5:41
  Qwen3.5-9B,
• 5:44
  IFBench
• 5:44
  59.17
• 5:45
  versus
• 5:46
  56.47
• 5:46
  --
• 5:47
  but
• 5:47
  coding
• 5:47
  and
• 5:48
  knowledge-heavy
• 5:48
  work
• 5:48
  remain
• 5:48
  weaknesses,
• 5:49
  and
• 5:49
  Liquid's
• 5:50
  own
• 5:50
  model
• 5:50
  card
• 5:50
  says
• 5:51
  it
• 5:51
  is
• 5:51
  not
• 5:51
  recommended
• 5:51
  for
• 5:51
  agentic
• 5:52
  coding.
• 5:53
  Community
• 5:53
  testing
• 5:54
  on
• 5:54
  a
• 5:54
  Snapdragon
• 5:54
  8
• 5:55
  Gen3
• 5:55
  showed
• 5:55
  20
• 5:56
  tok/s
• 5:56
  on
• 5:56
  a
• 5:56
  test
• 5:57
  prompt,
• 5:57
  falling
• 5:57
  to
• 5:57
  16
• 5:58
  tok/s
• 5:59
  on
• 5:59
  a
• 5:59
  2K-token
• 5:59
  response,
• 6:00
  and
• 6:00
  one
• 6:00
  user's
• 6:01
  extended
• 6:01
  tool-calling
• 6:02
  test
• 6:02
  produced
• 6:02
  a
• 6:02
  spectacular
• 6:03
  failure:
• 6:04
  the
• 6:04
  model
• 6:04
  hallucinated
• 6:05
  a
• 6:05
  date
• 6:05
  of
• 6:05
  March
• 6:06
  2026,
• 6:07
  then
• 6:07
  snowballed
• 6:07
  through
• 6:07
  pages
• 6:08
  of
• 6:08
  web
• 6:08
  searches
• 6:09
  all
• 6:09
  confirming
• 6:09
  that
• 6:09
  hallucinated
• 6:10
  timeframe,
• 6:11
  ultimately
• 6:11
  delivering
• 6:11
  a
• 6:12
  confidently
• 6:12
  wrong
• 6:12
  answer.
• 6:13
  Small
• 6:14
  models
• 6:14
  make
• 6:14
  sense
• 6:14
  as
• 6:15
  cheap
• 6:15
  worker
• 6:15
  agents
• 6:15
  doing
• 6:16
  extraction
• 6:16
  and
• 6:16
  repetitive
• 6:17
  tool
• 6:17
  calls,
• 6:18
  not
• 6:18
  as
• 6:18
  your
• 6:18
  smartest
• 6:18
  assistant.
• 6:19
  InclusionAI's
• 6:20
  Ling-3.0-flash
• 6:21
  weights
• 6:22
  landed
• 6:22
  on
• 6:22
  Hugging
• 6:22
  Face
• 6:23
  --
• 6:24
  MIT
• 6:24
  licensed,
• 6:25
  ungated,
• 6:25
  with
• 6:25
  BF16
• 6:26
  at
• 6:26
  roughly
• 6:27
  255
• 6:28
  GB
• 6:28
  and
• 6:28
  an
• 6:28
  official
• 6:29
  FP8
• 6:29
  at
• 6:29
  roughly
• 6:30
  128
• 6:31
  GB.
• 6:31
  The
• 6:31
  architecture
• 6:32
  is
• 6:32
  127.5B
• 6:34
  total
• 6:35
  with
• 6:35
  5.1B
• 6:36
  active
• 6:36
  per
• 6:36
  token,
• 6:37
  using
• 6:37
  512
• 6:38
  experts
• 6:38
  with
• 6:39
  8
• 6:39
  active,
• 6:39
  considerably
• 6:40
  finer-grained
• 6:40
  than
• 6:41
  most
• 6:41
  MoE
• 6:41
  releases.
• 6:42
  Thinking
• 6:42
  mode
• 6:43
  is
• 6:43
  a
• 6:43
  per-request
• 6:43
  toggle
• 6:44
  in
• 6:44
  the
• 6:44
  chat
• 6:44
  template
• 6:45
  rather
• 6:45
  than
• 6:45
  a
• 6:45
  separate
• 6:45
  SKU.
• 6:46
  The
• 6:47
  FP8
• 6:47
  at
• 6:47
  128
• 6:48
  GB
• 6:48
  is
• 6:48
  the
• 6:49
  practical
• 6:49
  number:
• 6:50
  it
• 6:50
  fits
• 6:50
  a
• 6:50
  big
• 6:50
  unified-memory
• 6:51
  box
• 6:51
  or
• 6:51
  a
• 6:51
  multi-GPU
• 6:52
  rig
• 6:53
  without
• 6:53
  community
• 6:53
  quanting.
• 6:54
  The
• 6:55
  open
• 6:55
  question
• 6:55
  is
• 6:55
  inference
• 6:56
  backend
• 6:56
  support
• 6:57
  --
• 6:58
  llama.cpp's
• 6:58
  handling
• 6:59
  of
• 6:59
  the
• 6:59
  bailing
• 6:59
  hybrid
• 7:00
  architecture
• 7:00
  remains
• 7:00
  unclear,
• 7:01
  potentially
• 7:02
  limiting
• 7:02
  this
• 7:02
  to
• 7:03
  vLLM
• 7:03
  and
• 7:03
  SGLang
• 7:04
  for
• 7:04
  now.
• 7:05
  DeepSeek
• 7:06
  is
• 7:06
  raising
• 7:06
  prices,
• 7:07
  and
• 7:07
  the
• 7:07
  community
• 7:08
  response
• 7:08
  on
• 7:08
  LocalLLaMA
• 7:09
  amounts
• 7:09
  to
• 7:09
  a
• 7:09
  collective
• 7:10
  shrug.
• 7:11
  The
• 7:11
  consensus:
• 7:12
  DeepSeek's
• 7:13
  first-party
• 7:13
  API
• 7:14
  was
• 7:14
  significantly
• 7:15
  cheaper
• 7:15
  than
• 7:15
  anyone
• 7:15
  else
• 7:15
  hosting
• 7:16
  DeepSeek
• 7:16
  V4,
• 7:17
  so
• 7:17
  the
• 7:17
  price
• 7:18
  increase
• 7:18
  likely
• 7:18
  brings
• 7:18
  it
• 7:19
  in
• 7:19
  line
• 7:19
  with
• 7:19
  third-party
• 7:19
  providers
• 7:20
  on
• 7:20
  OpenRouter
• 7:20
  --
• 7:22
  roughly
• 7:22
  a
• 7:22
  5x
• 7:22
  increase
• 7:23
  but
• 7:23
  still
• 7:23
  very
• 7:23
  cheap
• 7:24
  in
• 7:24
  absolute
• 7:24
  terms.
• 7:25
  The
• 7:25
  real
• 7:26
  question
• 7:26
  for
• 7:26
  local-first
• 7:27
  users
• 7:27
  is
• 7:27
  whether
• 7:27
  this
• 7:28
  changes
• 7:28
  the
• 7:28
  break-even
• 7:28
  math
• 7:29
  on
• 7:29
  buying
• 7:29
  hardware.
• 7:30
  Several
• 7:31
  commenters
• 7:31
  noted
• 7:31
  they
• 7:32
  were
• 7:32
  already
• 7:32
  considering
• 7:33
  upgrades,
• 7:34
  and
• 7:34
  the
• 7:34
  open-weight
• 7:34
  nature
• 7:35
  of
• 7:35
  the
• 7:35
  models
• 7:35
  means
• 7:35
  there
• 7:35
  is
• 7:36
  a
• 7:36
  price
• 7:36
  ceiling:
• 7:37
  anyone
• 7:37
  can
• 7:37
  host
• 7:37
  it
• 7:38
  if
• 7:38
  they
• 7:38
  find
• 7:38
  cheaper
• 7:38
  compute.
• 7:39
  The
• 7:39
  more
• 7:40
  substantive
• 7:40
  DeepSeek
• 7:41
  story
• 7:41
  is
• 7:41
  that
• 7:42
  V4
• 7:42
  Flash
• 7:42
  0731
• 7:44
  still
• 7:44
  struggles
• 7:44
  with
• 7:44
  instruction
• 7:45
  following
• 7:45
  and
• 7:45
  system
• 7:46
  prompt
• 7:46
  adherence.
• 7:47
  One
• 7:47
  practitioner
• 7:48
  running
• 7:48
  it
• 7:48
  locally
• 7:49
  at
• 7:49
  full
• 7:49
  precision
• 7:50
  reports
• 7:50
  that
• 7:50
  no
• 7:50
  matter
• 7:50
  what
• 7:51
  they
• 7:51
  try
• 7:51
  --
• 7:52
  first
• 7:52
  person,
• 7:52
  second
• 7:53
  person,
• 7:53
  Chinese,
• 7:53
  English
• 7:54
  --
• 7:54
  the
• 7:55
  model
• 7:55
  ignores
• 7:55
  rules,
• 7:56
  prompts,
• 7:56
  and
• 7:56
  skills.
• 7:57
  The
• 7:58
  technical
• 7:58
  explanation
• 7:58
  is
• 7:59
  revealing:
• 8:00
  DeepSeek
• 8:00
  V4
• 8:00
  stores
• 8:01
  rules
• 8:01
  and
• 8:01
  skills
• 8:02
  as
• 8:02
  compressed
• 8:02
  summaries,
• 8:03
  not
• 8:03
  raw
• 8:03
  text.
• 8:04
  Of
• 8:04
  43
• 8:05
  layers,
• 8:06
  20
• 8:06
  see
• 8:06
  the
• 8:06
  entire
• 8:07
  context
• 8:07
  squeezed
• 8:08
  into
• 8:08
  128-token
• 8:09
  compressed
• 8:09
  entries,
• 8:10
  21
• 8:11
  see
• 8:11
  it
• 8:11
  at
• 8:11
  4:1
• 8:12
  compression,
• 8:12
  and
• 8:13
  only
• 8:13
  two
• 8:13
  are
• 8:13
  fully
• 8:14
  dense.
• 8:15
  The
• 8:15
  system
• 8:15
  prompt
• 8:16
  "survives"
• 8:17
  but
• 8:17
  the
• 8:17
  exact
• 8:17
  wording
• 8:18
  does
• 8:18
  not.
• 8:19
  This
• 8:19
  is
• 8:19
  the
• 8:19
  same
• 8:20
  sparse-attention
• 8:20
  compatibility
• 8:21
  cliff
• 8:21
  observed
• 8:22
  with
• 8:22
  V3.2
• 8:23
  --
• 8:25
  running
• 8:25
  these
• 8:25
  models
• 8:26
  without
• 8:26
  native
• 8:26
  sparse-attention
• 8:27
  support
• 8:27
  in
• 8:27
  the
• 8:27
  inference
• 8:28
  backend
• 8:28
  concretely
• 8:29
  degrades
• 8:29
  quality.
• 8:31
  The
• 8:31
  practical
• 8:31
  advice
• 8:31
  from
• 8:32
  experienced
• 8:32
  users:
• 8:33
  write
• 8:33
  firm
• 8:34
  positive
• 8:34
  instructions,
• 8:35
  list
• 8:35
  workflow
• 8:36
  steps
• 8:36
  and
• 8:36
  their
• 8:36
  order,
• 8:37
  avoid
• 8:37
  hedging
• 8:38
  language,
• 8:38
  and
• 8:38
  accept
• 8:39
  that
• 8:39
  DeepSeek
• 8:39
  models
• 8:40
  need
• 8:40
  to
• 8:40
  work
• 8:40
  in
• 8:40
  their
• 8:40
  style
• 8:41
  rather
• 8:41
  than
• 8:41
  conforming
• 8:42
  to
• 8:42
  harness
• 8:42
  conventions
• 8:43
  designed
• 8:43
  for
• 8:43
  Claude
• 8:44
  or
• 8:44
  GPT.
• 8:48
  NVIDIA
• 8:48
  released
• 8:49
  OO-Agents
• 8:50
  (NOOA),
• 8:50
  a
• 8:50
  model-agnostic
• 8:51
  Python
• 8:51
  framework
• 8:52
  that
• 8:52
  takes
• 8:52
  a
• 8:52
  different
• 8:53
  approach
• 8:53
  to
• 8:53
  agent
• 8:53
  development:
• 8:54
  agents
• 8:55
  are
• 8:55
  Python
• 8:55
  objects,
• 8:56
  methods
• 8:56
  with
• 8:56
  ...
• 8:56
  bodies
• 8:57
  become
• 8:57
  LLM-driven
• 8:58
  generation
• 8:58
  methods
• 8:58
  at
• 8:59
  runtime,
• 8:59
  type
• 9:00
  annotations
• 9:00
  serve
• 9:00
  as
• 9:01
  contracts,
• 9:01
  and
• 9:02
  docstrings
• 9:02
  serve
• 9:02
  as
• 9:02
  prompts.
• 9:03
  The
• 9:03
  model
• 9:04
  acts
• 9:04
  by
• 9:04
  writing
• 9:05
  Python
• 9:05
  in
• 9:05
  a
• 9:05
  Jupyter-style
• 9:06
  REPL
• 9:07
  with
• 9:07
  access
• 9:07
  to
• 9:07
  self,
• 9:08
  imports,
• 9:08
  and
• 9:09
  helpers,
• 9:09
  which
• 9:09
  means
• 9:10
  Python
• 9:10
  methods
• 9:10
  and
• 9:11
  type
• 9:11
  annotations
• 9:12
  supply
• 9:12
  the
• 9:12
  callable
• 9:12
  interfaces
• 9:13
  instead
• 9:13
  of
• 9:13
  separate
• 9:14
  tool-schema
• 9:14
  definitions.
• 9:16
  The
• 9:16
  framework
• 9:16
  supports
• 9:17
  typed
• 9:17
  I/O
• 9:17
  with
• 9:18
  auto-retry,
• 9:18
  live-object
• 9:19
  arguments
• 9:19
  passed
• 9:20
  by
• 9:20
  reference,
• 9:21
  and
• 9:21
  model-callable
• 9:21
  context
• 9:22
  and
• 9:22
  event
• 9:23
  APIs.
• 9:24
  It
• 9:24
  comes
• 9:24
  with
• 9:24
  an
• 9:24
  important
• 9:25
  safety
• 9:25
  caveat:
• 9:26
  the
• 9:26
  AST-level
• 9:27
  code
• 9:27
  validation
• 9:28
  and
• 9:28
  module
• 9:28
  deny-lists
• 9:29
  are
• 9:29
  "defense-in-depth
• 9:30
  guardrails,
• 9:31
  not
• 9:31
  a
• 9:31
  containment
• 9:31
  boundary."
• 9:32
  NVIDIA
• 9:33
  explicitly
• 9:33
  states
• 9:33
  that
• 9:34
  open()
• 9:34
  gives
• 9:34
  arbitrary
• 9:35
  file
• 9:35
  access,
• 9:36
  importlib
• 9:36
  can
• 9:36
  load
• 9:36
  modules
• 9:37
  from
• 9:37
  a
• 9:37
  path,
• 9:38
  and
• 9:38
  reflection
• 9:38
  reaches
• 9:39
  the
• 9:39
  rest
• 9:39
  --
• 9:40
  the
• 9:40
  containment
• 9:40
  boundary
• 9:40
  must
• 9:41
  be
• 9:41
  OS-level
• 9:41
  isolation
• 9:42
  like
• 9:42
  a
• 9:42
  container,
• 9:43
  VM,
• 9:43
  or
• 9:44
  NVIDIA's
• 9:44
  OpenShell.
• 9:45
  Cua,
• 9:46
  which
• 9:46
  has
• 9:46
  evolved
• 9:46
  from
• 9:46
  its
• 9:47
  original
• 9:47
  pitch
• 9:47
  as
• 9:47
  "Docker
• 9:48
  for
• 9:48
  computer-use
• 9:48
  agents"
• 9:49
  to
• 9:49
  a
• 9:49
  cross-OS
• 9:50
  agent
• 9:50
  sandbox
• 9:51
  with
• 9:51
  benchmark-leading
• 9:52
  performance,
• 9:52
  now
• 9:53
  includes
• 9:53
  Cua
• 9:53
  Drivers
• 9:53
  --
• 9:54
  background
• 9:55
  computer-use
• 9:55
  on
• 9:55
  macOS,
• 9:56
  Windows,
• 9:56
  and
• 9:56
  Linux
• 9:57
  that
• 9:57
  lets
• 9:57
  agents
• 9:57
  click,
• 9:58
  type,
• 9:58
  and
• 9:58
  verify
• 9:58
  without
• 9:59
  stealing
• 9:59
  the
• 9:59
  cursor
• 9:59
  or
• 9:59
  focus.
• 10:01
  The
• 10:01
  same
• 10:01
  CLI
• 10:01
  and
• 10:02
  MCP
• 10:02
  server
• 10:02
  works
• 10:03
  across
• 10:03
  all
• 10:03
  three
• 10:04
  operating
• 10:04
  systems
• 10:04
  from
• 10:05
  Claude
• 10:05
  Code,
• 10:05
  Cursor,
• 10:06
  Codex,
• 10:07
  and
• 10:07
  custom
• 10:07
  clients.
• 10:08
  The
• 10:08
  broader
• 10:09
  Cua
• 10:09
  ecosystem
• 10:10
  includes
• 10:10
  agent-ready
• 10:11
  sandboxes
• 10:11
  (one
• 10:11
  API
• 10:12
  for
• 10:12
  any
• 10:12
  VM
• 10:13
  or
• 10:13
  container
• 10:13
  image),
• 10:14
  Cua-Bench
• 10:14
  for
• 10:15
  evaluating
• 10:15
  computer-use
• 10:16
  agents
• 10:16
  on
• 10:16
  OSWorld
• 10:17
  and
• 10:17
  other
• 10:17
  benchmarks,
• 10:18
  and
• 10:18
  Lume
• 10:18
  for
• 10:18
  macOS
• 10:19
  virtualization
• 10:20
  on
• 10:20
  Apple
• 10:20
  Silicon.
• 10:21
  Google's
• 10:21
  Gemini
• 10:22
  Spark
• 10:22
  can
• 10:22
  now
• 10:22
  tap
• 10:23
  into
• 10:23
  Chrome's
• 10:23
  auto-browse
• 10:24
  feature
• 10:24
  to
• 10:25
  handle
• 10:25
  complex
• 10:25
  online
• 10:26
  errands
• 10:26
  using
• 10:27
  the
• 10:27
  user's
• 10:27
  logged-in
• 10:28
  accounts
• 10:28
  --
• 10:29
  scheduling
• 10:29
  apartment
• 10:30
  viewings,
• 10:30
  researching
• 10:31
  flights,
• 10:31
  starting
• 10:31
  booking
• 10:32
  processes.
• 10:33
  The
• 10:33
  permission
• 10:34
  model
• 10:34
  is
• 10:34
  explicit:
• 10:35
  Gemini
• 10:36
  tells
• 10:36
  you
• 10:36
  it
• 10:36
  can
• 10:36
  use
• 10:36
  your
• 10:36
  money
• 10:37
  before
• 10:37
  it
• 10:37
  does
• 10:37
  anything.
• 10:38
  Whether
• 10:39
  this
• 10:39
  runs
• 10:39
  in
• 10:39
  the
• 10:39
  user's
• 10:40
  actual
• 10:40
  Chrome
• 10:40
  session
• 10:41
  or
• 10:41
  a
• 10:41
  sandboxed
• 10:41
  browser
• 10:42
  appears
• 10:42
  to
• 10:42
  vary
• 10:43
  by
• 10:43
  region,
• 10:43
  with
• 10:44
  some
• 10:44
  users
• 10:44
  reporting
• 10:44
  sandboxed-only
• 10:45
  access.
• 10:47
  The
• 10:47
  more
• 10:47
  interesting
• 10:47
  design
• 10:48
  question
• 10:48
  comes
• 10:48
  from
• 10:49
  a
• 10:49
  discussion
• 10:49
  about
• 10:49
  Meta's
• 10:50
  Muse
• 10:50
  Code
• 10:50
  beta,
• 10:51
  which
• 10:51
  can
• 10:51
  keep
• 10:52
  asynchronous
• 10:52
  subagents
• 10:53
  running
• 10:53
  across
• 10:53
  restarts,
• 10:54
  record
• 10:55
  work
• 10:55
  in
• 10:55
  an
• 10:55
  append-only
• 10:56
  local
• 10:56
  event
• 10:56
  log,
• 10:57
  and
• 10:57
  sustain
• 10:57
  1,000+
• 10:59
  tool
• 10:59
  calls
• 11:00
  for
• 11:00
  up
• 11:00
  to
• 11:00
  24
• 11:00
  hours.
• 11:01
  The
• 11:02
  thesis:
• 11:03
  coding
• 11:03
  agents
• 11:03
  often
• 11:04
  fail
• 11:04
  not
• 11:04
  because
• 11:05
  they
• 11:05
  cannot
• 11:05
  write
• 11:05
  a
• 11:05
  function
• 11:06
  but
• 11:06
  because
• 11:07
  they
• 11:07
  lose
• 11:07
  state,
• 11:08
  repeat
• 11:08
  work,
• 11:09
  or
• 11:09
  report
• 11:09
  success
• 11:10
  before
• 11:10
  the
• 11:10
  task
• 11:11
  is
• 11:11
  verified.
• 11:12
  Persistence
• 11:13
  attacks
• 11:13
  that
• 11:13
  bottleneck
• 11:14
  directly.
• 11:15
  The
• 11:15
  counterargument
• 11:16
  is
• 11:16
  equally
• 11:16
  sharp:
• 11:18
  a
• 11:18
  durable
• 11:18
  agent
• 11:18
  can
• 11:19
  also
• 11:19
  preserve
• 11:19
  a
• 11:19
  bad
• 11:20
  assumption
• 11:20
  for
• 11:20
  12
• 11:20
  hours,
• 11:21
  burn
• 11:22
  through
• 11:22
  tokens,
• 11:23
  and
• 11:23
  create
• 11:23
  coordination
• 11:24
  bugs
• 11:24
  harder
• 11:24
  to
• 11:24
  audit
• 11:25
  than
• 11:25
  a
• 11:25
  single
• 11:25
  model's
• 11:26
  mistake.
• 11:27
  As
• 11:27
  one
• 11:28
  practitioner
• 11:28
  put
• 11:28
  it,
• 11:29
  "if
• 11:29
  it
• 11:29
  cant
• 11:29
  check
• 11:30
  its
• 11:30
  own
• 11:30
  work,
• 11:30
  run
• 11:31
  the
• 11:31
  tests,
• 11:31
  actually
• 11:32
  look
• 11:32
  at
• 11:32
  the
• 11:32
  diff
• 11:32
  --
• 11:33
  then
• 11:34
  all
• 11:34
  persistence
• 11:34
  does
• 11:35
  is
• 11:35
  make
• 11:35
  the
• 11:35
  wrong
• 11:35
  answer
• 11:36
  bigger."
• 11:39
  A
• 11:39
  research
• 11:39
  paper
• 11:39
  from
• 11:40
  Multiverse
• 11:40
  Computing
• 11:41
  introduces
• 11:41
  ProvenanceGuard,
• 11:42
  a
• 11:42
  source-aware
• 11:43
  verifier
• 11:43
  for
• 11:43
  answers
• 11:44
  produced
• 11:44
  by
• 11:44
  LLM
• 11:45
  agents
• 11:45
  using
• 11:45
  Model
• 11:46
  Context
• 11:46
  Protocol
• 11:47
  (MCP).
• 11:48
  The
• 11:48
  core
• 11:48
  insight
• 11:48
  is
• 11:49
  that
• 11:49
  existing
• 11:49
  factuality
• 11:50
  metrics
• 11:50
  evaluate
• 11:50
  whether
• 11:51
  an
• 11:51
  answer
• 11:51
  is
• 11:51
  supported
• 11:52
  by
• 11:52
  available
• 11:52
  context
• 11:53
  after
• 11:53
  evidence
• 11:53
  has
• 11:53
  been
• 11:54
  pooled
• 11:54
  --
• 11:54
  but
• 11:55
  they
• 11:55
  miss
• 11:55
  a
• 11:55
  critical
• 11:55
  failure
• 11:56
  mode
• 11:56
  the
• 11:56
  authors
• 11:56
  call
• 11:56
  cross-source
• 11:57
  conflation:
• 11:58
  a
• 11:58
  claim
• 11:59
  may
• 11:59
  be
• 11:59
  supported
• 11:59
  somewhere
• 11:59
  in
• 12:00
  the
• 12:00
  evidence
• 12:00
  while
• 12:00
  being
• 12:00
  attributed
• 12:01
  to
• 12:01
  the
• 12:01
  wrong
• 12:01
  source.
• 12:02
  Consider
• 12:03
  a
• 12:03
  medical
• 12:03
  agent
• 12:03
  that
• 12:03
  states
• 12:04
  "According
• 12:04
  to
• 12:04
  the
• 12:04
  patient's
• 12:05
  chart,
• 12:05
  empagliflozin
• 12:06
  reduced
• 12:07
  a
• 12:07
  mortality
• 12:07
  endpoint."
• 12:08
  The
• 12:08
  mortality
• 12:09
  claim
• 12:09
  might
• 12:09
  be
• 12:09
  supported
• 12:10
  by
• 12:10
  a
• 12:10
  clinical-trial
• 12:11
  abstract,
• 12:11
  but
• 12:12
  not
• 12:12
  by
• 12:12
  the
• 12:12
  patient's
• 12:12
  chart.
• 12:14
  A
• 12:14
  source-blind
• 12:14
  verifier
• 12:15
  sees
• 12:15
  support
• 12:15
  in
• 12:16
  pooled
• 12:16
  evidence
• 12:16
  and
• 12:17
  passes
• 12:17
  it.
• 12:18
  ProvenanceGuard
• 12:19
  rejects
• 12:19
  it
• 12:19
  because
• 12:19
  the
• 12:19
  stated
• 12:20
  attribution
• 12:20
  is
• 12:20
  wrong.
• 12:21
  The
• 12:21
  system
• 12:21
  consumes
• 12:21
  captured
• 12:22
  MCP
• 12:22
  traces
• 12:23
  with
• 12:23
  stable
• 12:24
  tool
• 12:24
  IDs
• 12:24
  and
• 12:24
  source
• 12:25
  IDs,
• 12:26
  decomposes
• 12:26
  answers
• 12:27
  into
• 12:27
  atomic
• 12:27
  claims,
• 12:28
  routes
• 12:28
  claims
• 12:29
  to
• 12:29
  source-specific
• 12:30
  evidence
• 12:30
  using
• 12:30
  cosine
• 12:31
  similarity
• 12:31
  over
• 12:31
  source
• 12:32
  embeddings,
• 12:33
  checks
• 12:33
  support
• 12:33
  with
• 12:33
  NLI
• 12:34
  and
• 12:34
  an
• 12:34
  attention-derived
• 12:35
  token-alignment
• 12:35
  proxy,
• 12:36
  and
• 12:36
  separately
• 12:37
  compares
• 12:37
  the
• 12:37
  claim's
• 12:37
  stated
• 12:38
  attribution
• 12:38
  with
• 12:38
  the
• 12:38
  routed
• 12:39
  source.
• 12:40
  On
• 12:40
  a
• 12:40
  held-out
• 12:40
  split
• 12:40
  of
• 12:41
  40
• 12:41
  traces
• 12:41
  and
• 12:41
  361
• 12:42
  claims
• 12:42
  from
• 12:43
  a
• 12:43
  medical-domain
• 12:43
  corpus,
• 12:44
  ProvenanceGuard
• 12:45
  reaches
• 12:45
  block
• 12:45
  F1
• 12:46
  of
• 12:46
  0.802
• 12:47
  and
• 12:47
  source
• 12:47
  accuracy
• 12:47
  of
• 12:48
  0.858
• 12:49
  over
• 12:49
  260
• 12:50
  source-eligible
• 12:50
  claims.
• 12:52
  Source-blind
• 12:52
  baselines
• 12:52
  on
• 12:53
  the
• 12:53
  same
• 12:53
  data:
• 12:54
  MiniCheck
• 12:54
  0.783,
• 12:55
  RAGAS
• 12:56
  Faithfulness
• 12:56
  0.758,
• 12:58
  AlignScore
• 12:58
  0.662,
• 12:59
  SummaC-ZS
• 13:00
  0.436
• 13:01
  --
• 13:02
  but
• 13:02
  none
• 13:02
  of
• 13:02
  them
• 13:02
  emit
• 13:02
  claim-to-source
• 13:03
  IDs.
• 13:04
  On
• 13:05
  a
• 13:05
  harder
• 13:05
  multi-source
• 13:05
  benchmark
• 13:06
  with
• 13:06
  many
• 13:06
  semantically
• 13:07
  close
• 13:07
  candidate
• 13:07
  sources,
• 13:09
  block
• 13:09
  F1
• 13:09
  holds
• 13:09
  at
• 13:10
  0.846
• 13:11
  but
• 13:11
  source-plus-relation
• 13:12
  accuracy
• 13:12
  drops
• 13:12
  to
• 13:12
  0.229,
• 13:14
  confirming
• 13:14
  that
• 13:15
  exact
• 13:15
  source
• 13:15
  ownership
• 13:16
  remains
• 13:16
  genuinely
• 13:16
  difficult
• 13:17
  when
• 13:17
  wrong
• 13:17
  sources
• 13:18
  are
• 13:18
  topically
• 13:18
  plausible.
• 13:19
  The
• 13:19
  paper
• 13:20
  is
• 13:20
  honest
• 13:20
  about
• 13:20
  this
• 13:20
  limitation:
• 13:22
  a
• 13:22
  conservative
• 13:22
  blocker
• 13:23
  can
• 13:23
  still
• 13:23
  fail
• 13:23
  to
• 13:23
  identify
• 13:24
  the
• 13:24
  exact
• 13:24
  supporting
• 13:25
  provenance
• 13:25
  object.
• 13:26
  On
• 13:26
  the
• 13:27
  practical
• 13:27
  MCP
• 13:27
  tooling
• 13:28
  side,
• 13:28
  a
• 13:28
  developer
• 13:29
  released
• 13:29
  quillpdf-mcp,
• 13:31
  an
• 13:31
  MIT-licensed
• 13:31
  MCP
• 13:32
  server
• 13:32
  and
• 13:32
  CLI
• 13:33
  for
• 13:33
  PDF
• 13:33
  operations
• 13:34
  --
• 13:34
  merge,
• 13:35
  split,
• 13:35
  rotate,
• 13:36
  watermark,
• 13:37
  Bates
• 13:37
  numbering,
• 13:37
  metadata
• 13:38
  cleaning
• 13:38
  --
• 13:39
  that
• 13:39
  runs
• 13:39
  entirely
• 13:39
  on
• 13:40
  the
• 13:40
  local
• 13:40
  filesystem
• 13:40
  with
• 13:41
  no
• 13:41
  network
• 13:41
  calls.
• 13:42
  The
• 13:42
  motivation:
• 13:43
  every
• 13:43
  PDF
• 13:44
  API
• 13:44
  wants
• 13:45
  you
• 13:45
  to
• 13:45
  upload
• 13:45
  the
• 13:45
  file,
• 13:46
  and
• 13:46
  if
• 13:46
  your
• 13:46
  agent
• 13:47
  handles
• 13:47
  contracts,
• 13:48
  discovery
• 13:48
  docs,
• 13:49
  or
• 13:49
  medical
• 13:49
  records,
• 13:50
  "the
• 13:50
  upload
• 13:51
  is
• 13:51
  the
• 13:51
  problem."
• 13:52
  Built
• 13:52
  on
• 13:52
  pdf-lib
• 13:53
  with
• 13:53
  stdio
• 13:53
  transport
• 13:54
  only,
• 13:55
  small
• 13:55
  enough
• 13:55
  to
• 13:55
  audit
• 13:56
  in
• 13:56
  an
• 13:56
  afternoon.
• 13:57
  AWS
• 13:57
  open-sourced
• 13:58
  the
• 13:58
  Context
• 13:59
  Ontology
• 13:59
  Accelerator,
• 14:00
  a
• 14:00
  semantic
• 14:01
  context
• 14:01
  layer
• 14:02
  that
• 14:02
  combines
• 14:02
  knowledge
• 14:02
  graphs,
• 14:03
  formal
• 14:03
  ontologies
• 14:04
  (with
• 14:05
  HermiT/ELK
• 14:05
  reasoning),
• 14:06
  and
• 14:06
  rule-based
• 14:06
  systems
• 14:07
  with
• 14:07
  AI
• 14:07
  agents.
• 14:08
  The
• 14:08
  architecture
• 14:09
  follows
• 14:09
  a
• 14:09
  Scan-Model-Serve
• 14:10
  workflow:
• 14:11
  connect
• 14:11
  data
• 14:11
  sources
• 14:12
  and
• 14:12
  discover
• 14:12
  schemas,
• 14:13
  induce
• 14:14
  and
• 14:14
  manage
• 14:14
  ontologies
• 14:15
  with
• 14:15
  a
• 14:15
  unified
• 14:15
  semantic
• 14:16
  graph,
• 14:16
  then
• 14:17
  serve
• 14:17
  context
• 14:17
  to
• 14:17
  agents
• 14:18
  via
• 14:18
  SPARQL
• 14:21
  federation
• 14:21
  and
• 14:21
  MCP
• 14:22
  tools.
• 14:23
  Access
• 14:23
  is
• 14:23
  governed
• 14:24
  by
• 14:24
  namespace
• 14:24
  isolation
• 14:25
  and
• 14:25
  role-based
• 14:25
  access
• 14:26
  control.
• 14:27
  It
• 14:27
  is
• 14:27
  published
• 14:27
  as
• 14:27
  a
• 14:27
  read-only
• 14:28
  mirror.
• 14:32
  Prime
• 14:32
  Intellect
• 14:32
  has
• 14:33
  assembled
• 14:33
  a
• 14:33
  full-stack
• 14:33
  platform
• 14:34
  for
• 14:34
  training,
• 14:34
  deploying,
• 14:35
  and
• 14:35
  continuously
• 14:36
  improving
• 14:36
  custom
• 14:37
  models,
• 14:37
  anchored
• 14:38
  around
• 14:38
  reinforcement
• 14:38
  learning
• 14:39
  for
• 14:39
  agentic
• 14:39
  workflows.
• 14:40
  The
• 14:40
  platform
• 14:41
  includes
• 14:41
  the
• 14:41
  open-source
• 14:42
  Verifiers
• 14:42
  library
• 14:43
  for
• 14:43
  creating
• 14:43
  RL
• 14:43
  environments,
• 14:44
  a
• 14:44
  CLI
• 14:45
  loop
• 14:45
  (init,
• 14:45
  develop,
• 14:46
  eval,
• 14:47
  push)
• 14:47
  for
• 14:47
  iterating
• 14:47
  on
• 14:47
  RL-trained
• 14:48
  agents,
• 14:49
  2,500+
• 14:50
  community
• 14:50
  environments
• 14:50
  on
• 14:50
  their
• 14:51
  Hub,
• 14:51
  managed
• 14:52
  training
• 14:52
  workflows,
• 14:52
  and
• 14:53
  dedicated
• 14:53
  or
• 14:53
  serverless
• 14:54
  inference
• 14:54
  with
• 14:54
  native
• 14:55
  LoRA
• 14:55
  support.
• 14:56
  Their
• 14:56
  case
• 14:57
  studies
• 14:57
  are
• 14:57
  specific:
• 14:58
  Ramp
• 14:58
  trained
• 14:59
  a
• 14:59
  small
• 14:59
  RL-trained
• 15:00
  subagent
• 15:00
  called
• 15:01
  Fast
• 15:01
  Ask
• 15:01
  that
• 15:02
  beat
• 15:02
  frontier
• 15:02
  models
• 15:03
  on
• 15:03
  accuracy
• 15:03
  at
• 15:03
  spreadsheet
• 15:04
  search
• 15:04
  while
• 15:04
  running
• 15:05
  faster
• 15:05
  and
• 15:05
  cheaper,
• 15:06
  and
• 15:06
  Zapier
• 15:07
  uses
• 15:07
  their
• 15:07
  eval
• 15:07
  infrastructure
• 15:08
  as
• 15:08
  a
• 15:08
  continuous
• 15:09
  agent
• 15:09
  improvement
• 15:09
  loop.
• 15:10
  The
• 15:10
  Prime
• 15:11
  Agent
• 15:11
  harness,
• 15:11
  built
• 15:12
  around
• 15:12
  their
• 15:12
  Recursive
• 15:13
  Language
• 15:13
  Model,
• 15:13
  reaches
• 15:14
  95.5%
• 15:15
  on
• 15:15
  ARC-AGI-3
• 15:18
  with
• 15:18
  Opus
• 15:18
  5.
• 15:19
  Their
• 15:19
  INTELLECT-3
• 15:20
  release
• 15:20
  --
• 15:21
  a
• 15:21
  100B+
• 15:22
  MoE
• 15:22
  trained
• 15:22
  with
• 15:23
  large-scale
• 15:23
  RL,
• 15:24
  12B
• 15:25
  active
• 15:25
  per
• 15:25
  forward
• 15:25
  pass,
• 15:26
  MIT/Apache
• 15:26
  licensed
• 15:27
  --
• 15:28
  remains
• 15:29
  one
• 15:29
  of
• 15:29
  the
• 15:29
  most
• 15:29
  thoroughly
• 15:30
  open
• 15:30
  frontier
• 15:30
  model
• 15:31
  releases
• 15:31
  to
• 15:31
  date.
• 15:32
  Superlinked
• 15:32
  open-sourced
• 15:33
  its
• 15:33
  inference
• 15:33
  engine
• 15:34
  (SIE),
• 15:35
  a
• 15:35
  Kubernetes-native
• 15:36
  cluster
• 15:36
  designed
• 15:36
  to
• 15:37
  run
• 15:37
  the
• 15:37
  full
• 15:37
  stack
• 15:37
  of
• 15:37
  models
• 15:38
  behind
• 15:38
  agents:
• 15:39
  LLM,
• 15:40
  OCR,
• 15:41
  vision,
• 15:42
  embeddings,
• 15:43
  reranking,
• 15:43
  and
• 15:43
  content
• 15:44
  policy,
• 15:45
  all
• 15:45
  in
• 15:45
  one
• 15:45
  deployment.
• 15:46
  The
• 15:46
  differentiator
• 15:47
  is
• 15:47
  a
• 15:47
  cluster-wide
• 15:48
  queue
• 15:48
  that
• 15:48
  maximizes
• 15:49
  GPU
• 15:49
  utilization
• 15:50
  by
• 15:50
  letting
• 15:51
  worker
• 15:51
  pods
• 15:51
  pull
• 15:51
  from
• 15:52
  a
• 15:52
  shared
• 15:52
  queue
• 15:52
  rather
• 15:53
  than
• 15:53
  having
• 15:53
  each
• 15:53
  worker
• 15:53
  manage
• 15:54
  its
• 15:54
  own.
• 15:54
  SIE
• 15:55
  wraps
• 15:55
  whichever
• 15:55
  model
• 15:56
  server
• 15:56
  proves
• 15:56
  best
• 15:56
  for
• 15:57
  a
• 15:57
  given
• 15:57
  model
• 15:57
  --
• 15:58
  vLLM,
• 15:58
  SGLang,
• 15:59
  TensorRT-LLM,
• 16:00
  TEI
• 16:01
  --
• 16:02
  and
• 16:02
  handles
• 16:03
  model-profile
• 16:03
  hot
• 16:03
  reloads
• 16:04
  without
• 16:04
  restarts.
• 16:05
  It
• 16:05
  runs
• 16:05
  on
• 16:05
  AWS
• 16:06
  EKS,
• 16:07
  GCP
• 16:07
  GKE,
• 16:08
  or
• 16:08
  Azure
• 16:09
  AKS,
• 16:10
  and
• 16:10
  offers
• 16:10
  air-gapped
• 16:11
  installs
• 16:11
  from
• 16:11
  mirrored
• 16:12
  model
• 16:12
  snapshots.
• 16:13
  They
• 16:13
  also
• 16:13
  offer
• 16:13
  free
• 16:14
  inference
• 16:14
  grants
• 16:14
  for
• 16:14
  selected
• 16:15
  projects.
• 16:16
  Repowise
• 16:17
  bills
• 16:17
  itself
• 16:17
  as
• 16:17
  "codebase
• 16:18
  intelligence
• 16:18
  for
• 16:19
  the
• 16:19
  AI
• 16:19
  era"
• 16:19
  --
• 16:20
  an
• 16:20
  open-source
• 16:21
  (AGPL-3.0)
• 16:23
  tool
• 16:23
  that
• 16:24
  indexes
• 16:24
  a
• 16:24
  repository
• 16:25
  once
• 16:25
  and
• 16:25
  serves
• 16:25
  both
• 16:26
  AI
• 16:26
  agents
• 16:26
  (via
• 16:26
  ten
• 16:27
  MCP
• 16:27
  tools
• 16:28
  and
• 16:28
  a
• 16:28
  generated
• 16:29
  CLAUDE.md)
• 16:30
  and
• 16:30
  humans
• 16:30
  (a
• 16:30
  verified
• 16:31
  wiki,
• 16:31
  defect-validated
• 16:32
  code-health
• 16:33
  score,
• 16:33
  and
• 16:33
  architectural
• 16:34
  decision
• 16:34
  records).
• 16:35
  The
• 16:35
  headline
• 16:35
  benchmark:
• 16:36
  in
• 16:36
  a
• 16:36
  real
• 16:36
  agent
• 16:37
  loop
• 16:37
  with
• 16:37
  Codex
• 16:37
  on
• 16:37
  django/django
• 16:39
  (n=43,
• 16:41
  p
• 16:43
  Repowise
• 16:44
  cut
• 16:44
  output
• 16:44
  tokens
• 16:45
  31.6%
• 16:46
  and
• 16:47
  reached
• 16:47
  answers
• 16:47
  in
• 16:47
  3.8
• 16:48
  tool
• 16:48
  calls
• 16:48
  instead
• 16:48
  of
• 16:49
  7.2.
• 16:50
  The
• 16:50
  code-health
• 16:50
  score
• 16:51
  --
• 16:51
  one
• 16:52
  1-10
• 16:52
  number
• 16:52
  from
• 16:52
  21
• 16:53
  deterministic
• 16:53
  signals
• 16:53
  --
• 16:54
  achieves
• 16:55
  ROC
• 16:55
  AUC
• 16:56
  of
• 16:56
  0.74
• 16:57
  across
• 16:57
  21
• 16:57
  repos
• 16:58
  and
• 16:58
  9
• 16:58
  languages
• 16:59
  at
• 16:59
  predicting
• 16:59
  which
• 16:59
  files
• 17:00
  will
• 17:00
  have
• 17:00
  bugs,
• 17:00
  matching
• 17:01
  or
• 17:01
  beating
• 17:01
  published
• 17:01
  academic
• 17:02
  models.
• 17:03
  The
• 17:03
  PR
• 17:03
  bot
• 17:04
  posts
• 17:04
  deterministic
• 17:05
  reviews
• 17:05
  with
• 17:05
  zero
• 17:06
  LLM
• 17:06
  calls,
• 17:07
  so
• 17:07
  the
• 17:07
  same
• 17:08
  diff
• 17:08
  always
• 17:08
  produces
• 17:09
  the
• 17:09
  same
• 17:09
  analysis.
• 17:10
  Soup,
• 17:11
  the
• 17:11
  fine-tuning
• 17:12
  CLI
• 17:12
  that
• 17:12
  turns
• 17:13
  LLM
• 17:13
  training
• 17:13
  into
• 17:13
  a
• 17:13
  three-command
• 17:14
  workflow,
• 17:15
  shipped
• 17:15
  v0.72.4
• 17:18
  with
• 17:19
  preference-learning
• 17:19
  losses
• 17:20
  (DPO,
• 17:20
  ORPO,
• 17:21
  SimPO,
• 17:22
  KTO)
• 17:22
  now
• 17:23
  running
• 17:23
  over
• 17:23
  layer
• 17:23
  streaming.
• 17:24
  The
• 17:24
  clever
• 17:25
  bit:
• 17:25
  DPO
• 17:26
  needs
• 17:26
  a
• 17:26
  reference
• 17:27
  model
• 17:27
  to
• 17:27
  compare
• 17:27
  against,
• 17:28
  and
• 17:28
  a
• 17:28
  second
• 17:28
  copy
• 17:29
  would
• 17:29
  double
• 17:29
  memory,
• 17:30
  defeating
• 17:30
  the
• 17:30
  purpose.
• 17:31
  Soup
• 17:32
  uses
• 17:32
  the
• 17:32
  same
• 17:32
  streamed
• 17:33
  base
• 17:33
  with
• 17:33
  its
• 17:33
  adapters
• 17:34
  switched
• 17:34
  off
• 17:34
  --
• 17:35
  one
• 17:35
  set
• 17:35
  of
• 17:35
  weights,
• 17:36
  one
• 17:36
  stream.
• 17:37
  Measured
• 17:38
  on
• 17:38
  an
• 17:38
  RTX
• 17:38
  3050
• 17:39
  4
• 17:41
  GB:
• 17:41
  streamed
• 17:42
  DPO
• 17:42
  peaked
• 17:42
  at
• 17:43
  0.914x
• 17:44
  the
• 17:45
  supervised-fine-tuning
• 17:46
  peak
• 17:46
  memory.
• 17:47
  Forcing
• 17:47
  a
• 17:47
  real
• 17:47
  second
• 17:48
  model
• 17:48
  cost
• 17:48
  +730
• 17:49
  MB.
• 17:50
  The
• 17:50
  honest
• 17:50
  trade-off:
• 17:51
  the
• 17:51
  reference
• 17:51
  is
• 17:52
  free
• 17:52
  in
• 17:52
  memory,
• 17:53
  not
• 17:53
  in
• 17:53
  time
• 17:53
  --
• 17:54
  DPO
• 17:55
  reads
• 17:55
  the
• 17:55
  layer
• 17:55
  stack
• 17:55
  1.52x
• 17:57
  as
• 17:57
  often
• 17:57
  per
• 17:57
  step.
• 17:58
  GRPO
• 17:59
  and
• 17:59
  PPO
• 17:59
  are
• 17:59
  explicitly
• 18:00
  excluded
• 18:00
  because
• 18:01
  generation
• 18:01
  re-reads
• 18:02
  every
• 18:02
  layer
• 18:02
  per
• 18:02
  token,
• 18:03
  which
• 18:03
  streaming
• 18:03
  cannot
• 18:04
  amortize.
• 18:08
  Seedance
• 18:08
  2.5
• 18:09
  and
• 18:09
  Minimax
• 18:10
  H3
• 18:10
  (open-weight)
• 18:11
  went
• 18:11
  head-to-head
• 18:11
  in
• 18:11
  a
• 18:11
  video
• 18:12
  generation
• 18:12
  comparison
• 18:13
  that
• 18:13
  has
• 18:13
  the
• 18:13
  LocalLLaMA
• 18:14
  community
• 18:14
  cautiously
• 18:15
  optimistic
• 18:16
  about
• 18:16
  the
• 18:16
  state
• 18:16
  of
• 18:16
  AI
• 18:17
  filmmaking.
• 18:18
  The
• 18:18
  outputs
• 18:18
  look
• 18:18
  impressive
• 18:19
  enough
• 18:19
  to
• 18:19
  prompt
• 18:19
  one
• 18:20
  commenter
• 18:20
  to
• 18:20
  hope
• 18:20
  for
• 18:21
  AI-generated
• 18:21
  seasons
• 18:22
  of
• 18:22
  Firefly,
• 18:23
  though
• 18:23
  another
• 18:23
  captured
• 18:24
  the
• 18:24
  current
• 18:24
  aesthetic
• 18:24
  limitation
• 18:25
  well:
• 18:26
  "it
• 18:26
  has
• 18:26
  taken
• 18:26
  Hollywood's
• 18:27
  worst
• 18:27
  quick
• 18:27
  cut
• 18:27
  spectacle
• 18:28
  without
• 18:28
  essence
• 18:28
  to
• 18:28
  the
• 18:28
  maximum."
• 18:30
  The
• 18:30
  technical
• 18:30
  quality
• 18:31
  is
• 18:31
  advancing
• 18:31
  faster
• 18:32
  than
• 18:32
  the
• 18:32
  artistic
• 18:32
  control,
• 18:33
  but
• 18:33
  the
• 18:33
  R&D
• 18:34
  behind
• 18:34
  these
• 18:34
  models
• 18:35
  continues
• 18:35
  to
• 18:35
  push
• 18:36
  boundaries
• 18:36
  that
• 18:36
  may
• 18:36
  yield
• 18:37
  breakthroughs
• 18:37
  in
• 18:37
  other
• 18:37
  domains.
• 18:39
  On
• 18:39
  the
• 18:39
  audio
• 18:39
  side,
• 18:40
  srt2speech
• 18:41
  is
• 18:41
  an
• 18:41
  open-source
• 18:42
  tool
• 18:42
  for
• 18:42
  converting
• 18:43
  SRT
• 18:43
  subtitles
• 18:44
  to
• 18:44
  narrated
• 18:44
  speech
• 18:44
  with
• 18:45
  automatic
• 18:45
  duration
• 18:45
  matching
• 18:46
  --
• 18:47
  the
• 18:47
  genuinely
• 18:47
  hard
• 18:47
  problem
• 18:48
  that
• 18:48
  most
• 18:48
  TTS
• 18:49
  tools
• 18:49
  ignore.
• 18:50
  It
• 18:50
  uses
• 18:50
  pitch-corrected
• 18:51
  speed
• 18:51
  adjustment,
• 18:52
  automatic
• 18:53
  regeneration,
• 18:54
  pause
• 18:54
  modification
• 18:55
  between
• 18:55
  words,
• 18:56
  and
• 18:56
  exact
• 18:56
  placement
• 18:56
  of
• 18:57
  silence
• 18:57
  between
• 18:57
  subtitle
• 18:58
  cues.
• 18:59
  Multi-speaker
• 18:59
  support
• 19:00
  works
• 19:00
  via
• 19:00
  simple
• 19:01
  templating
• 19:01
  ({{speaker
• 19:02
  name}}
• 19:02
  in
• 19:02
  the
• 19:02
  subtitle),
• 19:03
  and
• 19:03
  voice
• 19:03
  cloning
• 19:04
  is
• 19:04
  supported.
• 19:05
  Dependencies
• 19:06
  are
• 19:06
  minimal:
• 19:07
  Python,
• 19:07
  NumPy,
• 19:08
  llama.cpp,
• 19:10
  and
• 19:10
  GGUF
• 19:11
  speech
• 19:11
  models.
• 19:12
  Performance
• 19:13
  on
• 19:13
  an
• 19:13
  RTX
• 19:13
  4080
• 19:14
  laptop
• 19:15
  GPU
• 19:15
  runs
• 19:16
  12-13x
• 19:17
  real
• 19:17
  time;
• 19:17
  CPU-only
• 19:18
  manages
• 19:19
  1.5-2x.
• 19:21
  While
• 19:21
  waiting
• 19:21
  for
• 19:21
  Qwen3.8
• 19:23
  27B,
• 19:25
  one
• 19:25
  hobbyist
• 19:25
  repurposed
• 19:26
  a
• 19:26
  kid's
• 19:26
  old
• 19:26
  32x32
• 19:28
  pixel
• 19:28
  Divoom
• 19:28
  display
• 19:29
  as
• 19:29
  a
• 19:29
  GPU
• 19:30
  cluster
• 19:30
  status
• 19:31
  monitor
• 19:31
  for
• 19:31
  three
• 19:31
  RTX
• 19:32
  A6000s
• 19:34
  running
• 19:34
  Qwen3.5
• 19:36
  122B
• 19:38
  on
• 19:38
  a
• 19:38
  vLLM
• 19:38
  Ray
• 19:39
  cluster.
• 19:40
  The
• 19:40
  build
• 19:40
  --
• 19:40
  including
• 19:41
  Ray
• 19:41
  Dashboard
• 19:42
  API
• 19:42
  integration,
• 19:43
  vLLM
• 19:43
  metrics
• 19:44
  endpoint,
• 19:44
  and
• 19:44
  nvidia-smi
• 19:45
  calls
• 19:45
  over
• 19:46
  SSH
• 19:46
  --
• 19:47
  took
• 19:47
  GLM
• 19:47
  5.2
• 19:48
  about
• 19:48
  20
• 19:49
  minutes
• 19:49
  to
• 19:49
  build
• 19:49
  and
• 19:49
  another
• 19:50
  5
• 19:50
  minutes
• 19:50
  of
• 19:50
  human
• 19:50
  polish.
• 19:51
  The
• 19:51
  hardware
• 19:52
  setup
• 19:52
  (three
• 19:52
  Dell
• 19:53
  Precision
• 19:53
  7960
• 19:55
  workstations,
• 19:56
  each
• 19:56
  with
• 19:56
  a
• 19:56
  64GB
• 19:57
  RTX
• 19:58
  A6000,
• 20:00
  plus
• 20:00
  a
• 20:00
  Dell
• 20:00
  Pro
• 20:01
  Max
• 20:01
  GB10
• 20:02
  for
• 20:02
  Bluetooth
• 20:02
  and
• 20:02
  secondary
• 20:03
  inference)
• 20:03
  represents
• 20:04
  the
• 20:04
  kind
• 20:04
  of
• 20:04
  local-first
• 20:05
  infrastructure
• 20:06
  that
• 20:06
  has
• 20:06
  steadily
• 20:06
  drifted
• 20:06
  from
• 20:07
  hobbyist
• 20:07
  roots
• 20:07
  toward
• 20:08
  the
• 20:08
  $40K-workstation
• 20:09
  territory.
• 20:10
  But
• 20:10
  the
• 20:10
  pixel-art
• 20:11
  dashboard
• 20:11
  is
• 20:11
  a
• 20:12
  reminder
• 20:12
  that
• 20:12
  the
• 20:12
  tinkering
• 20:13
  spirit
• 20:13
  persists
• 20:14
  even
• 20:14
  as
• 20:14
  the
• 20:14
  hardware
• 20:15
  floor
• 20:15
  rises.