Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 01, 2026

Transcript

AGI Dreams Podcast – June 01, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Anthropic
• 0:00
  shipped
• 0:00
  Claude
• 0:01
  Opus
• 0:01
  4.8
• 0:02
  last
• 0:02
  week
• 0:02
  alongside
• 0:03
  a
• 0:03
  $65
• 0:03
  billion
• 0:03
  Series
• 0:04
  H
• 0:04
  at
• 0:04
  a
• 0:04
  $965
• 0:05
  billion
• 0:05
  valuation,
• 0:07
  but
• 0:07
  the
• 0:07
  more
• 0:07
  revealing
• 0:07
  publication
• 0:08
  was
• 0:08
  the
• 0:08
  engineering
• 0:09
  blog
• 0:09
  post
• 0:09
  detailing
• 0:10
  how
• 0:10
  the
• 0:10
  company
• 0:10
  actually
• 0:11
  contains
• 0:11
  its
• 0:11
  own
• 0:12
  agents
• 0:12
  across
• 0:12
  products.
• 0:13
  The
• 0:13
  model
• 0:14
  itself
• 0:14
  is
• 0:14
  an
• 0:14
  incremental
• 0:15
  upgrade:
• 0:16
  better
• 0:16
  calibration
• 0:16
  (four
• 0:17
  times
• 0:17
  less
• 0:18
  likely
• 0:18
  to
• 0:18
  let
• 0:18
  flawed
• 0:18
  code
• 0:19
  pass
• 0:19
  unremarked),
• 0:20
  improved
• 0:20
  agentic
• 0:21
  judgment,
• 0:21
  and
• 0:22
  a
• 0:22
  new
• 0:22
  "dynamic
• 0:22
  workflows"
• 0:23
  feature
• 0:23
  that
• 0:23
  lets
• 0:23
  Claude
• 0:24
  Code
• 0:24
  run
• 0:24
  hundreds
• 0:25
  of
• 0:25
  parallel
• 0:25
  subagents
• 0:26
  in
• 0:26
  a
• 0:26
  single
• 0:26
  session.
• 0:27
  Early
• 0:28
  testers
• 0:28
  call
• 0:28
  it
• 0:28
  "the
• 0:28
  first
• 0:28
  model
• 0:29
  to
• 0:29
  complete
• 0:29
  every
• 0:30
  case
• 0:30
  end-to-end"
• 0:31
  on
• 0:31
  Super-Agent
• 0:31
  benchmarks,
• 0:32
  beating
• 0:32
  prior
• 0:33
  Opus
• 0:33
  models
• 0:33
  and
• 0:34
  GPT-5.5
• 0:35
  at
• 0:35
  cost
• 0:35
  parity.
• 0:36
  Effort
• 0:37
  controls
• 0:37
  now
• 0:37
  let
• 0:38
  users
• 0:38
  choose
• 0:38
  how
• 0:38
  hard
• 0:39
  Claude
• 0:39
  thinks,
• 0:39
  from
• 0:40
  low
• 0:40
  (faster,
• 0:41
  cheaper)
• 0:41
  to
• 0:41
  max
• 0:41
  (deep
• 0:42
  reasoning).
• 0:43
  Fast
• 0:43
  mode
• 0:43
  for
• 0:43
  Opus
• 0:44
  4.8
• 0:44
  is
• 0:44
  three
• 0:45
  times
• 0:45
  cheaper
• 0:45
  than
• 0:45
  for
• 0:45
  previous
• 0:46
  models.
• 0:47
  Pricing
• 0:47
  holds
• 0:47
  at
• 0:47
  $5/$25
• 0:49
  per
• 0:49
  million
• 0:50
  input/output
• 0:50
  tokens.
• 0:51
  The
• 0:51
  containment
• 0:52
  post,
• 0:52
  however,
• 0:53
  is
• 0:53
  the
• 0:53
  piece
• 0:53
  worth
• 0:53
  reading
• 0:54
  carefully.
• 0:55
  Anthropic
• 0:55
  describes
• 0:56
  three
• 0:56
  isolation
• 0:57
  patterns
• 0:57
  —
• 0:58
  ephemeral
• 0:58
  containers
• 0:59
  for
• 0:59
  claude.ai,
• 1:00
  an
• 1:00
  OS-level
• 1:01
  sandbox
• 1:02
  (Seatbelt
• 1:02
  on
• 1:02
  macOS,
• 1:03
  bubblewrap
• 1:04
  on
• 1:04
  Linux)
• 1:05
  for
• 1:05
  Claude
• 1:05
  Code,
• 1:06
  and
• 1:06
  a
• 1:06
  full
• 1:06
  VM
• 1:06
  via
• 1:07
  Apple's
• 1:07
  Virtualization
• 1:08
  framework
• 1:08
  for
• 1:08
  Claude
• 1:08
  Cowork
• 1:09
  —
• 1:09
  and
• 1:10
  then
• 1:10
  candidly
• 1:10
  catalogues
• 1:11
  where
• 1:11
  each
• 1:11
  one
• 1:11
  broke.
• 1:13
  Between
• 1:13
  mid-2025
• 1:14
  and
• 1:14
  January
• 1:14
  2026,
• 1:16
  three
• 1:16
  Claude
• 1:16
  Code
• 1:16
  vulnerabilities
• 1:17
  exploited
• 1:18
  code
• 1:18
  that
• 1:18
  executed
• 1:19
  before
• 1:19
  the
• 1:19
  user
• 1:19
  even
• 1:20
  saw
• 1:20
  the
• 1:20
  trust
• 1:20
  dialog:
• 1:21
  a
• 1:21
  malicious
• 1:23
  .claude/settings.json
• 1:25
  hook
• 1:25
  in
• 1:25
  a
• 1:25
  cloned
• 1:25
  repo
• 1:26
  ran
• 1:26
  automatically
• 1:26
  at
• 1:27
  startup.
• 1:28
  In
• 1:28
  February
• 1:28
  2026,
• 1:29
  an
• 1:29
  internal
• 1:30
  red
• 1:30
  team
• 1:30
  phished
• 1:31
  an
• 1:31
  Anthropic
• 1:31
  employee
• 1:32
  into
• 1:32
  pasting
• 1:32
  a
• 1:32
  prompt
• 1:33
  that
• 1:33
  exfiltrated
• 1:34
  ~/.aws/credentials
• 1:35
  —
• 1:37
  successfully,
• 1:38
  24
• 1:38
  out
• 1:38
  of
• 1:38
  25
• 1:39
  attempts.
• 1:40
  The
• 1:40
  model
• 1:41
  layer
• 1:41
  had
• 1:41
  nothing
• 1:41
  anomalous
• 1:42
  to
• 1:42
  catch
• 1:42
  because
• 1:42
  the
• 1:42
  user
• 1:43
  was
• 1:43
  the
• 1:43
  injection
• 1:44
  vector.
• 1:45
  A
• 1:45
  third-party
• 1:45
  disclosure
• 1:46
  then
• 1:46
  showed
• 1:46
  that
• 1:46
  Claude
• 1:47
  Cowork's
• 1:47
  egress
• 1:47
  allowlist
• 1:48
  for
• 1:48
  api.anthropic.com
• 1:52
  could
• 1:52
  be
• 1:52
  weaponized:
• 1:54
  a
• 1:54
  malicious
• 1:55
  file
• 1:55
  in
• 1:55
  a
• 1:55
  mounted
• 1:55
  workspace
• 1:56
  instructed
• 1:57
  Claude
• 1:57
  to
• 1:57
  upload
• 1:57
  other
• 1:58
  workspace
• 1:58
  files
• 1:59
  using
• 1:59
  an
• 1:59
  attacker-controlled
• 2:00
  API
• 2:00
  key,
• 2:01
  and
• 2:01
  the
• 2:01
  proxy
• 2:02
  let
• 2:02
  it
• 2:02
  through
• 2:02
  because
• 2:03
  the
• 2:03
  destination
• 2:03
  was
• 2:04
  on
• 2:04
  the
• 2:04
  allowlist.
• 2:05
  The
• 2:05
  fix
• 2:05
  required
• 2:06
  a
• 2:06
  defensive
• 2:06
  man-in-the-middle
• 2:07
  proxy
• 2:08
  inside
• 2:08
  the
• 2:08
  VM
• 2:08
  that
• 2:08
  rejects
• 2:09
  any
• 2:09
  request
• 2:10
  not
• 2:10
  carrying
• 2:10
  the
• 2:10
  VM's
• 2:11
  own
• 2:11
  provisioned
• 2:11
  session
• 2:12
  token.
• 2:13
  The
• 2:13
  recurring
• 2:13
  lesson:
• 2:14
  battle-tested
• 2:15
  hypervisors
• 2:15
  held;
• 2:16
  every
• 2:16
  failure
• 2:17
  came
• 2:17
  from
• 2:17
  custom
• 2:18
  code
• 2:18
  the
• 2:18
  team
• 2:18
  built
• 2:18
  around
• 2:19
  them.
• 2:21
  OpenAI's
• 2:22
  ChatGPT
• 2:23
  for
• 2:23
  Google
• 2:23
  Sheets
• 2:24
  extension,
• 2:24
  which
• 2:25
  accumulated
• 2:25
  over
• 2:25
  185,000
• 2:27
  downloads
• 2:27
  in
• 2:27
  its
• 2:28
  first
• 2:28
  month,
• 2:28
  turns
• 2:29
  out
• 2:29
  to
• 2:29
  be
• 2:29
  a
• 2:29
  textbook
• 2:30
  case
• 2:30
  of
• 2:30
  the
• 2:30
  trust-inversion
• 2:31
  problem.
• 2:32
  PromptArmor
• 2:32
  demonstrated
• 2:33
  that
• 2:33
  a
• 2:33
  single
• 2:33
  indirect
• 2:34
  prompt
• 2:34
  injection
• 2:34
  —
• 2:35
  hidden
• 2:35
  in
• 2:35
  white
• 2:36
  text
• 2:36
  in
• 2:36
  an
• 2:36
  imported
• 2:37
  spreadsheet
• 2:37
  —
• 2:38
  can
• 2:38
  trigger
• 2:38
  ChatGPT
• 2:39
  to
• 2:39
  run
• 2:39
  an
• 2:39
  attacker-controlled
• 2:40
  Google
• 2:40
  Apps
• 2:40
  Script
• 2:41
  that
• 2:41
  exfiltrates
• 2:41
  workbooks
• 2:42
  across
• 2:42
  the
• 2:42
  victim's
• 2:42
  entire
• 2:43
  account,
• 2:44
  displays
• 2:44
  a
• 2:44
  phishing
• 2:44
  overlay
• 2:45
  replacing
• 2:45
  the
• 2:45
  ChatGPT
• 2:46
  sidebar,
• 2:47
  and
• 2:47
  edits
• 2:47
  sheets
• 2:48
  with
• 2:48
  attacker-controlled
• 2:48
  content.
• 2:49
  The
• 2:50
  kicker:
• 2:50
  this
• 2:51
  works
• 2:51
  even
• 2:51
  when
• 2:51
  the
• 2:51
  user
• 2:52
  has
• 2:52
  explicitly
• 2:52
  disabled
• 2:53
  automatic
• 2:53
  edits.
• 2:54
  The
• 2:54
  malicious
• 2:55
  script
• 2:55
  crawls
• 2:55
  links
• 2:56
  between
• 2:56
  workbooks,
• 2:57
  exfiltrating
• 2:58
  12
• 2:58
  in
• 2:58
  the
• 2:58
  demonstration.
• 2:59
  OpenAI's
• 3:00
  response,
• 3:01
  after
• 3:01
  three
• 3:01
  weeks
• 3:01
  of
• 3:01
  silence
• 3:02
  following
• 3:02
  responsible
• 3:03
  disclosure,
• 3:03
  was
• 3:04
  to
• 3:04
  remove
• 3:04
  the
• 3:04
  model's
• 3:05
  ability
• 3:05
  to
• 3:05
  generate
• 3:06
  Apps
• 3:06
  Script
• 3:06
  code
• 3:06
  entirely.
• 3:07
  That
• 3:08
  incident
• 3:08
  lands
• 3:08
  alongside
• 3:09
  PIArena,
• 3:10
  a
• 3:10
  new
• 3:10
  unified
• 3:11
  platform
• 3:11
  for
• 3:11
  prompt
• 3:12
  injection
• 3:12
  evaluation
• 3:13
  out
• 3:13
  of
• 3:13
  Penn
• 3:14
  State
• 3:14
  that
• 3:14
  systematically
• 3:15
  demonstrates
• 3:15
  how
• 3:16
  badly
• 3:16
  current
• 3:16
  defenses
• 3:17
  generalize.
• 3:18
  The
• 3:18
  researchers
• 3:19
  integrated
• 3:19
  eight
• 3:19
  state-of-the-art
• 3:20
  defenses
• 3:20
  —
• 3:21
  both
• 3:21
  prevention-based
• 3:22
  (PISanitizer,
• 3:23
  SecAlign++,
• 3:24
  DataFilter,
• 3:25
  PromptArmor)
• 3:26
  and
• 3:26
  detection-based
• 3:27
  (DataSentinel,
• 3:28
  PromptGuard,
• 3:29
  AttentionTracker,
• 3:30
  PIGuard)
• 3:30
  —
• 3:31
  and
• 3:31
  evaluated
• 3:31
  them
• 3:31
  across
• 3:32
  diverse
• 3:32
  benchmarks
• 3:33
  spanning
• 3:33
  QA,
• 3:34
  summarization,
• 3:35
  RAG,
• 3:35
  and
• 3:35
  code
• 3:35
  generation.
• 3:37
  The
• 3:37
  results
• 3:37
  are
• 3:37
  sobering.
• 3:38
  A
• 3:38
  novel
• 3:39
  strategy-based
• 3:39
  adaptive
• 3:40
  attack
• 3:40
  that
• 3:40
  iteratively
• 3:41
  rewrites
• 3:41
  injected
• 3:42
  prompts
• 3:42
  based
• 3:43
  on
• 3:43
  defense
• 3:43
  feedback
• 3:44
  achieves
• 3:44
  99%
• 3:45
  attack
• 3:45
  success
• 3:45
  rate
• 3:46
  without
• 3:46
  any
• 3:46
  defense,
• 3:47
  and
• 3:47
  86%
• 3:48
  against
• 3:48
  PISanitizer
• 3:49
  —
• 3:50
  versus
• 3:50
  4%
• 3:51
  for
• 3:51
  static
• 3:51
  combined
• 3:52
  attacks
• 3:52
  against
• 3:52
  the
• 3:52
  same
• 3:53
  defense.
• 3:54
  Even
• 3:54
  GPT-5,
• 3:55
  deployed
• 3:56
  with
• 3:56
  a
• 3:56
  multilayered
• 3:56
  defense
• 3:57
  stack,
• 3:57
  exhibits
• 3:58
  70%
• 3:59
  ASR.
• 4:00
  The
• 4:00
  most
• 4:00
  uncomfortable
• 4:01
  finding:
• 4:01
  when
• 4:01
  the
• 4:02
  injected
• 4:02
  task
• 4:02
  aligns
• 4:03
  with
• 4:03
  the
• 4:03
  target
• 4:03
  task
• 4:04
  (like
• 4:04
  corrupting
• 4:05
  answers
• 4:05
  in
• 4:05
  a
• 4:05
  QA
• 4:06
  pipeline),
• 4:06
  the
• 4:07
  attack
• 4:07
  reduces
• 4:07
  to
• 4:08
  disinformation,
• 4:09
  and
• 4:09
  no
• 4:09
  existing
• 4:09
  defense
• 4:10
  can
• 4:10
  handle
• 4:10
  it
• 4:11
  because
• 4:11
  there
• 4:11
  are
• 4:11
  no
• 4:11
  malicious
• 4:12
  instructions
• 4:12
  to
• 4:13
  detect.
• 4:14
  The
• 4:14
  researchers
• 4:15
  note
• 4:15
  this
• 4:15
  aligns
• 4:15
  with
• 4:15
  OpenAI's
• 4:16
  own
• 4:16
  recent
• 4:17
  observation
• 4:17
  that
• 4:18
  "the
• 4:18
  most
• 4:18
  effective
• 4:18
  real-world
• 4:19
  prompt
• 4:19
  injection
• 4:20
  attacks
• 4:20
  increasingly
• 4:21
  resemble
• 4:21
  social
• 4:22
  engineering."
• 4:23
  On
• 4:23
  the
• 4:23
  offensive
• 4:24
  side
• 4:24
  of
• 4:24
  the
• 4:24
  ledger,
• 4:25
  the
• 4:25
  CAI
• 4:26
  (Cybersecurity
• 4:26
  AI)
• 4:27
  framework
• 4:28
  from
• 4:28
  Alias
• 4:28
  Robotics
• 4:29
  continues
• 4:29
  to
• 4:29
  mature
• 4:30
  as
• 4:30
  the
• 4:30
  de
• 4:30
  facto
• 4:30
  open-source
• 4:31
  toolkit
• 4:31
  for
• 4:32
  AI-powered
• 4:32
  security
• 4:33
  testing.
• 4:34
  The
• 4:34
  framework
• 4:34
  supports
• 4:35
  over
• 4:35
  300
• 4:36
  AI
• 4:36
  models,
• 4:37
  includes
• 4:37
  built-in
• 4:37
  tools
• 4:38
  organized
• 4:38
  along
• 4:39
  the
• 4:39
  security
• 4:39
  kill
• 4:39
  chain,
• 4:40
  and
• 4:40
  has
• 4:40
  demonstrated
• 4:41
  a
• 4:41
  claimed
• 4:41
  3,600x
• 4:43
  performance
• 4:44
  improvement
• 4:44
  over
• 4:44
  human
• 4:45
  pentesters
• 4:45
  in
• 4:45
  standardized
• 4:46
  CTF
• 4:46
  benchmarks.
• 4:48
  Its
• 4:48
  alias1
• 4:49
  model
• 4:49
  reached
• 4:49
  Rank
• 4:49
  1
• 4:50
  during
• 4:50
  hours
• 4:50
  7-8
• 4:51
  of
• 4:51
  the
• 4:51
  Dragos
• 4:51
  OT
• 4:52
  CTF
• 4:52
  2025,
• 4:54
  completing
• 4:54
  32
• 4:55
  of
• 4:55
  34
• 4:56
  challenges.
• 4:57
  Real-world
• 4:57
  case
• 4:57
  studies
• 4:58
  include
• 4:58
  finding
• 4:59
  critical
• 4:59
  vulnerabilities
• 5:00
  in
• 5:00
  Unitree
• 5:00
  G1
• 5:01
  humanoid
• 5:01
  robots
• 5:02
  (unauthorized
• 5:03
  telemetry
• 5:04
  to
• 5:04
  China-related
• 5:04
  servers,
• 5:05
  exposed
• 5:06
  RSA
• 5:06
  keys)
• 5:07
  and
• 5:07
  Ecoforest
• 5:08
  heat
• 5:08
  pumps
• 5:09
  (remote
• 5:09
  access
• 5:10
  with
• 5:10
  DES
• 5:10
  encryption
• 5:11
  weaknesses).
• 5:12
  The
• 5:12
  framework
• 5:13
  includes
• 5:13
  guardrails
• 5:14
  against
• 5:14
  prompt
• 5:14
  injection
• 5:15
  of
• 5:15
  the
• 5:15
  security
• 5:15
  agents
• 5:16
  themselves
• 5:16
  and
• 5:16
  supports
• 5:17
  MCP
• 5:17
  integration
• 5:18
  —
• 5:18
  a
• 5:19
  reminder
• 5:19
  that
• 5:19
  the
• 5:19
  same
• 5:20
  protocol
• 5:20
  designed
• 5:20
  for
• 5:21
  helpful
• 5:21
  tool
• 5:21
  use
• 5:21
  is
• 5:22
  also
• 5:22
  an
• 5:22
  attack
• 5:22
  surface.
• 5:26
  A
• 5:26
  studio
• 5:27
  called
• 5:27
  Firespawn
• 5:27
  ran
• 5:27
  25
• 5:28
  LLM
• 5:28
  agents
• 5:29
  across
• 5:29
  eight
• 5:29
  open-weight
• 5:30
  models
• 5:30
  —
• 5:31
  Qwen3
• 5:31
  235B,
• 5:32
  Qwen3
• 5:33
  32B,
• 5:33
  Nemotron
• 5:34
  3
• 5:34
  Nano
• 5:34
  30B,
• 5:35
  Ministral
• 5:35
  14B
• 5:36
  and
• 5:36
  8B,
• 5:37
  Gemma
• 5:37
  3
• 5:37
  12B,
• 5:38
  and
• 5:38
  others
• 5:38
  —
• 5:38
  in
• 5:38
  a
• 5:38
  persistent
• 5:39
  text-based
• 5:39
  MMO
• 5:40
  for
• 5:40
  10
• 5:40
  days,
• 5:41
  logging
• 5:41
  93,000
• 5:42
  events
• 5:42
  with
• 5:42
  reasoning
• 5:43
  traces.
• 5:43
  The
• 5:44
  dataset,
• 5:44
  released
• 5:44
  under
• 5:44
  CC-BY-4.0,
• 5:46
  reveals
• 5:47
  patterns
• 5:47
  that
• 5:47
  static
• 5:48
  benchmarks
• 5:48
  cannot.
• 5:49
  Qwen3
• 5:50
  235B
• 5:50
  independently
• 5:51
  invented
• 5:51
  arbitrage:
• 5:52
  nobody
• 5:52
  told
• 5:53
  it
• 5:53
  to
• 5:53
  be
• 5:53
  a
• 5:53
  pacifist
• 5:53
  merchant,
• 5:54
  but
• 5:54
  it
• 5:55
  examined
• 5:55
  the
• 5:55
  JSON
• 5:55
  state,
• 5:56
  calculated
• 5:57
  risk/reward,
• 5:58
  and
• 5:58
  adopted
• 5:58
  a
• 5:58
  buy-low-relist-high
• 5:59
  strategy,
• 6:00
  hoarding
• 6:01
  a
• 6:01
  third
• 6:01
  of
• 6:01
  all
• 6:01
  shard
• 6:02
  wealth
• 6:02
  while
• 6:02
  engaging
• 6:02
  in
• 6:02
  combat
• 6:03
  only
• 6:03
  8%
• 6:04
  of
• 6:04
  the
• 6:04
  time.
• 6:04
  Ministral
• 6:05
  8B
• 6:05
  and
• 6:05
  14B
• 6:06
  held
• 6:06
  their
• 6:06
  own
• 6:06
  on
• 6:06
  long-term
• 6:07
  state
• 6:07
  awareness
• 6:07
  despite
• 6:08
  their
• 6:08
  size.
• 6:09
  One
• 6:09
  Nemotron
• 6:10
  agent
• 6:10
  died
• 6:10
  over
• 6:10
  300
• 6:11
  times
• 6:11
  because
• 6:11
  its
• 6:11
  directive
• 6:12
  was
• 6:12
  simply
• 6:12
  "gather"
• 6:13
  with
• 6:13
  no
• 6:13
  survival
• 6:14
  cost
• 6:14
  —
• 6:15
  dying
• 6:15
  and
• 6:15
  respawning
• 6:16
  was
• 6:16
  the
• 6:16
  optimal
• 6:16
  policy
• 6:17
  under
• 6:17
  the
• 6:17
  reward
• 6:17
  function
• 6:18
  as
• 6:18
  specified.
• 6:19
  The
• 6:19
  most
• 6:19
  universal
• 6:20
  failure
• 6:20
  was
• 6:20
  the
• 6:20
  "cooldown
• 6:21
  paradox":
• 6:22
  a
• 6:22
  resource
• 6:22
  node
• 6:23
  showed
• 6:23
  node
• 6:23
  available:
• 6:24
  true
• 6:24
  while
• 6:24
  the
• 6:25
  agent's
• 6:25
  personal
• 6:25
  cooldown
• 6:26
  was
• 6:26
  still
• 6:26
  active.
• 6:27
  Every
• 6:27
  model
• 6:28
  from
• 6:28
  8B
• 6:28
  to
• 6:28
  235B
• 6:30
  failed
• 6:30
  identically,
• 6:31
  retrying
• 6:31
  until
• 6:32
  a
• 6:32
  single-line
• 6:32
  state
• 6:33
  clarification
• 6:33
  fixed
• 6:33
  the
• 6:34
  problem.
• 6:35
  The
• 6:35
  takeaway:
• 6:36
  much
• 6:36
  of
• 6:36
  what
• 6:36
  looks
• 6:36
  like
• 6:36
  reasoning
• 6:37
  failure
• 6:37
  may
• 6:37
  actually
• 6:38
  be
• 6:38
  ambiguous
• 6:38
  state
• 6:39
  signaling.
• 6:40
  That
• 6:40
  ambiguity
• 6:41
  problem
• 6:41
  scales
• 6:41
  poorly
• 6:42
  in
• 6:42
  multi-agent
• 6:43
  architectures.
• 6:44
  A
• 6:44
  practitioner
• 6:45
  running
• 6:45
  Qwen3.6-35B-A3B
• 6:51
  as
• 6:51
  a
• 6:51
  sub-agent
• 6:51
  on
• 6:51
  a
• 6:51
  single
• 6:52
  4090
• 6:53
  reports
• 6:53
  a
• 6:53
  specific
• 6:53
  failure
• 6:54
  pattern:
• 6:55
  the
• 6:55
  model
• 6:55
  processes
• 6:56
  tasks
• 6:56
  in
• 6:56
  thinking
• 6:57
  mode,
• 6:57
  produces
• 6:58
  structurally
• 6:58
  correct
• 6:59
  but
• 6:59
  semantically
• 7:00
  wrong
• 7:00
  output,
• 7:01
  and
• 7:01
  the
• 7:01
  orchestrator
• 7:02
  accepts
• 7:02
  it
• 7:02
  because
• 7:02
  format
• 7:03
  validation
• 7:03
  passes.
• 7:05
  With
• 7:05
  MoE
• 7:05
  (Mixture
• 7:06
  of
• 7:06
  Experts)
• 7:07
  architecture,
• 7:08
  certain
• 7:08
  task
• 7:08
  types
• 7:09
  hit
• 7:09
  cold
• 7:09
  experts
• 7:10
  and
• 7:10
  performance
• 7:10
  drops
• 7:10
  with
• 7:11
  no
• 7:11
  signal
• 7:11
  that
• 7:11
  it
• 7:11
  happened.
• 7:12
  The
• 7:13
  fix
• 7:13
  is
• 7:13
  validation
• 7:14
  at
• 7:14
  the
• 7:14
  boundary
• 7:14
  —
• 7:15
  schema
• 7:15
  checks
• 7:15
  catch
• 7:16
  "wrong
• 7:16
  shape"
• 7:17
  but
• 7:17
  not
• 7:17
  "wrong
• 7:17
  claim,"
• 7:18
  so
• 7:19
  each
• 7:19
  sub-agent
• 7:19
  handoff
• 7:20
  needs
• 7:20
  a
• 7:20
  task-specific
• 7:21
  verifier
• 7:22
  beyond
• 7:22
  format
• 7:22
  compliance.
• 7:24
  For
• 7:24
  those
• 7:24
  giving
• 7:24
  agents
• 7:25
  more
• 7:25
  autonomy,
• 7:26
  a
• 7:26
  developer
• 7:26
  built
• 7:27
  an
• 7:27
  LXC-based
• 7:29
  sandbox
• 7:29
  framework
• 7:30
  for
• 7:30
  Codex
• 7:30
  on
• 7:30
  headless
• 7:31
  Linux
• 7:31
  with
• 7:31
  GPU
• 7:32
  passthrough,
• 7:33
  persistent
• 7:33
  environments,
• 7:34
  and
• 7:34
  multiple
• 7:35
  parallel
• 7:35
  browser
• 7:35
  sessions
• 7:36
  —
• 7:37
  the
• 7:37
  goal
• 7:37
  being
• 7:37
  "let
• 7:37
  the
• 7:38
  agent
• 7:38
  run
• 7:38
  free
• 7:38
  while
• 7:39
  limiting
• 7:39
  the
• 7:39
  damage
• 7:39
  it
• 7:40
  can
• 7:40
  do."
• 7:41
  The
• 7:41
  design
• 7:41
  uses
• 7:42
  LXC
• 7:42
  containers
• 7:43
  rather
• 7:43
  than
• 7:43
  full
• 7:44
  VMs
• 7:44
  so
• 7:44
  multiple
• 7:45
  instances
• 7:45
  can
• 7:46
  share
• 7:46
  a
• 7:46
  GPU,
• 7:47
  with
• 7:47
  git
• 7:48
  push
• 7:48
  hooks
• 7:48
  preventing
• 7:49
  history
• 7:49
  rewrites.
• 7:50
  Meanwhile,
• 7:51
  SDAR
• 7:52
  (Self-Distilled
• 7:52
  Agentic
• 7:53
  Reinforcement
• 7:53
  Learning)
• 7:54
  from
• 7:54
  Zhejiang
• 7:55
  University
• 7:55
  tackles
• 7:56
  agent
• 7:56
  training
• 7:56
  itself,
• 7:57
  using
• 7:57
  self-distillation
• 7:58
  to
• 7:59
  improve
• 7:59
  agentic
• 7:59
  RL
• 8:00
  on
• 8:00
  environments
• 8:01
  like
• 8:01
  ALFWorld,
• 8:02
  WebShop,
• 8:02
  and
• 8:03
  Search-QA
• 8:03
  —
• 8:04
  achieving
• 8:04
  substantial
• 8:05
  improvements
• 8:05
  over
• 8:06
  standard
• 8:06
  GRPO
• 8:07
  baselines
• 8:08
  by
• 8:08
  having
• 8:08
  the
• 8:08
  model
• 8:08
  learn
• 8:09
  from
• 8:09
  its
• 8:09
  own
• 8:09
  successful
• 8:10
  trajectories.
• 8:13
  PrismML's
• 8:14
  Bonsai
• 8:14
  Image
• 8:14
  4B
• 8:15
  is
• 8:15
  the
• 8:15
  first
• 8:16
  image-generation
• 8:16
  model
• 8:17
  in
• 8:17
  its
• 8:17
  parameter
• 8:18
  class
• 8:18
  to
• 8:18
  run
• 8:18
  on
• 8:18
  an
• 8:18
  iPhone.
• 8:19
  Built
• 8:19
  from
• 8:20
  FLUX.2
• 8:20
  Klein
• 8:20
  4B,
• 8:21
  it
• 8:21
  keeps
• 8:21
  the
• 8:22
  architecture
• 8:22
  intact
• 8:22
  but
• 8:23
  moves
• 8:23
  transformer
• 8:23
  weights
• 8:24
  to
• 8:24
  binary
• 8:24
  ({-1,
• 8:25
  +1},
• 8:26
  1.125
• 8:27
  effective
• 8:27
  bits)
• 8:28
  or
• 8:28
  ternary
• 8:29
  ({-1,
• 8:29
  0,
• 8:30
  +1},
• 8:31
  1.71
• 8:32
  effective
• 8:32
  bits)
• 8:32
  form.
• 8:33
  The
• 8:33
  1-bit
• 8:34
  variant
• 8:34
  compresses
• 8:34
  the
• 8:34
  diffusion
• 8:35
  transformer
• 8:35
  from
• 8:35
  7.75
• 8:36
  GB
• 8:37
  to
• 8:37
  0.94
• 8:38
  GB
• 8:38
  —
• 8:39
  an
• 8:39
  8.3x
• 8:40
  reduction.
• 8:41
  The
• 8:41
  ternary
• 8:41
  variant
• 8:42
  lands
• 8:42
  at
• 8:42
  1.21
• 8:43
  GB,
• 8:44
  retaining
• 8:44
  95%
• 8:45
  of
• 8:45
  FLUX.2
• 8:46
  Klein
• 8:46
  4B
• 8:46
  accuracy
• 8:47
  across
• 8:47
  GenEval,
• 8:48
  HPSv3,
• 8:49
  and
• 8:49
  DPG-Bench.
• 8:50
  On
• 8:50
  iPhone
• 8:50
  17
• 8:51
  Pro
• 8:51
  Max,
• 8:52
  Bonsai
• 8:52
  generates
• 8:52
  a
• 8:52
  512x512
• 8:54
  image
• 8:54
  in
• 8:54
  9.4
• 8:55
  seconds;
• 8:56
  on
• 8:56
  Mac
• 8:56
  M4
• 8:57
  Pro,
• 8:57
  about
• 8:57
  6
• 8:57
  seconds
• 8:58
  and
• 8:58
  up
• 8:58
  to
• 8:58
  5.6x
• 8:59
  faster
• 8:59
  than
• 8:59
  the
• 8:59
  stock
• 8:59
  full-precision
• 9:00
  MFLUX
• 9:01
  pipeline.
• 9:01
  Mean
• 9:02
  active
• 9:02
  memory
• 9:02
  during
• 9:02
  1024x1024
• 9:04
  generation
• 9:04
  is
• 9:04
  1.95
• 9:05
  GB
• 9:06
  for
• 9:06
  the
• 9:06
  binary
• 9:06
  model
• 9:07
  versus
• 9:07
  14.39
• 9:08
  GB
• 9:08
  for
• 9:08
  full
• 9:08
  precision.
• 9:09
  Open
• 9:10
  weights
• 9:10
  under
• 9:10
  Apache
• 9:10
  license.
• 9:11
  Not
• 9:11
  everyone
• 9:12
  needs
• 9:12
  new
• 9:12
  hardware.
• 9:13
  A
• 9:13
  developer
• 9:14
  dusted
• 9:14
  off
• 9:14
  a
• 9:14
  2013
• 9:15
  "trash
• 9:15
  can"
• 9:16
  Mac
• 9:16
  Pro
• 9:16
  —
• 9:17
  originally
• 9:17
  £10,000,
• 9:18
  with
• 9:18
  dual
• 9:18
  D700
• 9:19
  GPUs
• 9:20
  (Southern
• 9:20
  Islands
• 9:20
  architecture)
• 9:21
  —
• 9:22
  and
• 9:22
  discovered
• 9:22
  that
• 9:22
  new
• 9:23
  Vulkan
• 9:23
  driver
• 9:23
  support
• 9:23
  in
• 9:24
  recent
• 9:24
  Linux
• 9:24
  kernels
• 9:25
  brought
• 9:25
  its
• 9:25
  12
• 9:25
  GB
• 9:26
  of
• 9:26
  VRAM
• 9:26
  back
• 9:26
  to
• 9:27
  life
• 9:27
  for
• 9:27
  llama.cpp
• 9:28
  inference.
• 9:30
  Benchmarks:
• 9:31
  Qwen
• 9:31
  3.5
• 9:31
  9B
• 9:32
  Q4
• 9:32
  at
• 9:32
  11
• 9:33
  t/s,
• 9:34
  Qwen
• 9:34
  2.5
• 9:34
  Coder
• 9:35
  Q4
• 9:35
  at
• 9:35
  22
• 9:36
  t/s,
• 9:37
  both
• 9:37
  at
• 9:37
  70K
• 9:37
  context.
• 9:38
  Not
• 9:39
  fast,
• 9:39
  but
• 9:39
  usable
• 9:40
  for
• 9:40
  planning
• 9:40
  tasks
• 9:41
  where
• 9:41
  you
• 9:41
  set
• 9:41
  it
• 9:41
  and
• 9:41
  forget
• 9:41
  it.
• 9:42
  On
• 9:42
  the
• 9:42
  opposite
• 9:43
  end
• 9:43
  of
• 9:43
  the
• 9:43
  hardware
• 9:43
  spectrum,
• 9:44
  a
• 9:44
  heterogeneous
• 9:45
  GPU
• 9:45
  optimization
• 9:46
  for
• 9:46
  Ollama
• 9:46
  reverses
• 9:46
  the
• 9:47
  default
• 9:47
  layer-splitting
• 9:47
  algorithm
• 9:48
  so
• 9:48
  the
• 9:48
  strongest
• 9:49
  GPU
• 9:49
  fills
• 9:49
  first
• 9:50
  instead
• 9:50
  of
• 9:50
  last.
• 9:51
  The
• 9:51
  single
• 9:52
  most
• 9:52
  impactful
• 9:52
  change:
• 9:53
  greedyFit
• 9:53
  iteration
• 9:54
  direction.
• 9:55
  Combined
• 9:55
  with
• 9:55
  compute-power
• 9:56
  weighting
• 9:56
  (SM
• 9:56
  count
• 9:57
  ×
• 9:57
  clock
• 9:58
  MHz)
• 9:59
  and
• 9:59
  forced
• 9:59
  output-layer
• 10:00
  placement
• 10:00
  on
• 10:00
  the
• 10:00
  fastest
• 10:01
  card,
• 10:01
  an
• 10:02
  RTX
• 10:02
  5090
• 10:03
  +
• 10:03
  RTX
• 10:04
  3090
• 10:05
  combo
• 10:05
  now
• 10:06
  outperforms
• 10:06
  the
• 10:06
  5090
• 10:07
  alone,
• 10:08
  where
• 10:08
  previously
• 10:09
  the
• 10:09
  5090
• 10:10
  idled
• 10:10
  while
• 10:10
  the
• 10:11
  3090
• 10:11
  bottlenecked.
• 10:13
  Intel's
• 10:13
  Arrow
• 10:13
  Lake
• 10:14
  NPU,
• 10:14
  often
• 10:15
  dismissed
• 10:15
  as
• 10:15
  a
• 10:15
  marketing
• 10:16
  gimmick
• 10:16
  for
• 10:16
  LLM
• 10:16
  workloads,
• 10:17
  turns
• 10:17
  out
• 10:17
  to
• 10:18
  be
• 10:18
  genuinely
• 10:18
  useful
• 10:18
  for
• 10:19
  automatic
• 10:19
  speech
• 10:19
  recognition.
• 10:20
  Running
• 10:21
  ONNX-compiled
• 10:22
  Whisper
• 10:23
  on
• 10:23
  the
• 10:23
  13
• 10:23
  TOPS
• 10:24
  NPU
• 10:24
  via
• 10:25
  OpenVINO
• 10:25
  delivers
• 10:25
  4.8–6.1x
• 10:28
  faster
• 10:28
  transcription
• 10:29
  than
• 10:29
  INT8
• 10:29
  CPU
• 10:30
  inference,
• 10:31
  but
• 10:31
  the
• 10:31
  energy
• 10:31
  numbers
• 10:32
  are
• 10:32
  the
• 10:32
  real
• 10:32
  story:
• 10:33
  10.7–21.6x
• 10:36
  less
• 10:36
  energy
• 10:36
  per
• 10:36
  transcription.
• 10:37
  A
• 10:37
  60-second
• 10:38
  audio
• 10:38
  clip
• 10:39
  takes
• 10:39
  818ms
• 10:40
  on
• 10:40
  the
• 10:40
  NPU
• 10:41
  at
• 10:41
  13.4W
• 10:42
  above
• 10:42
  idle,
• 10:43
  versus
• 10:43
  5,011ms
• 10:44
  on
• 10:44
  CPU
• 10:45
  at
• 10:45
  47.4W.
• 10:47
  For
• 10:47
  smart
• 10:47
  home
• 10:47
  voice
• 10:47
  commands,
• 10:48
  the
• 10:48
  NPU's
• 10:49
  instant
• 10:49
  wake-up
• 10:50
  from
• 10:50
  dormancy
• 10:51
  actually
• 10:51
  beats
• 10:51
  an
• 10:51
  RTX
• 10:52
  3060
• 10:53
  eGPU
• 10:54
  on
• 10:54
  short
• 10:54
  utterances.
• 10:55
  MOSS-TTS
• 10:57
  v1.5
• 10:58
  rounds
• 10:58
  out
• 10:59
  the
• 10:59
  local
• 10:59
  audio
• 10:59
  story
• 11:00
  with
• 11:00
  improved
• 11:00
  multilingual
• 11:01
  synthesis
• 11:01
  across
• 11:02
  31
• 11:02
  languages,
• 11:03
  more
• 11:03
  stable
• 11:04
  voice
• 11:04
  cloning,
• 11:05
  explicit
• 11:05
  pause
• 11:05
  control
• 11:06
  via
• 11:06
  inline
• 11:06
  markers
• 11:07
  like
• 11:07
  [pause
• 11:07
  3.2s],
• 11:07
  and
• 11:08
  a
• 11:08
  companion
• 11:08
  MOSS-SoundEffect
• 11:10
  v2.0
• 11:11
  model
• 11:12
  for
• 11:12
  48
• 11:12
  kHz
• 11:13
  bilingual
• 11:14
  sound
• 11:14
  effects
• 11:14
  up
• 11:14
  to
• 11:15
  30
• 11:15
  seconds.
• 11:18
  A
• 11:18
  Brian
• 11:18
  Krebs
• 11:19
  LinkedIn
• 11:19
  post
• 11:20
  captured
• 11:20
  a
• 11:20
  WSJ
• 11:21
  report
• 11:21
  that
• 11:21
  enterprises
• 11:22
  are
• 11:22
  hitting
• 11:22
  their
• 11:22
  annual
• 11:23
  AI
• 11:23
  token
• 11:24
  budgets
• 11:24
  in
• 11:24
  three
• 11:24
  months,
• 11:25
  with
• 11:25
  spending
• 11:26
  doubling
• 11:26
  or
• 11:26
  tripling
• 11:27
  as
• 11:27
  providers
• 11:27
  stop
• 11:27
  subsidizing
• 11:28
  prices.
• 11:29
  Companies
• 11:29
  are
• 11:30
  scrambling
• 11:30
  to
• 11:30
  ration
• 11:31
  AI
• 11:31
  use,
• 11:31
  steer
• 11:32
  employees
• 11:32
  toward
• 11:33
  cheaper
• 11:33
  internal
• 11:33
  tools,
• 11:34
  and
• 11:34
  hone
• 11:34
  skills
• 11:34
  to
• 11:34
  improve
• 11:35
  returns.
• 11:36
  The
• 11:36
  community
• 11:36
  response
• 11:37
  was
• 11:37
  pointed:
• 11:38
  one
• 11:38
  practitioner
• 11:39
  running
• 11:39
  three
• 11:39
  concurrent
• 11:40
  research
• 11:40
  projects,
• 11:41
  eight
• 11:41
  apps,
• 11:42
  and
• 11:42
  four
• 11:42
  harnesses
• 11:43
  reports
• 11:43
  spending
• 11:43
  $300/month
• 11:45
  total
• 11:46
  across
• 11:46
  Claude
• 11:46
  and
• 11:47
  GPT
• 11:47
  subscriptions,
• 11:48
  arguing
• 11:49
  that
• 11:49
  the
• 11:49
  cost
• 11:49
  blowouts
• 11:50
  come
• 11:50
  from
• 11:50
  handing
• 11:50
  tools
• 11:51
  to
• 11:51
  people
• 11:51
  who
• 11:51
  don't
• 11:52
  understand
• 11:52
  model
• 11:52
  tiering
• 11:53
  ("Opus
• 11:54
  is
• 11:54
  not
• 11:54
  the
• 11:54
  tool
• 11:54
  to
• 11:54
  use
• 11:55
  in
• 11:55
  a
• 11:55
  business
• 11:55
  situation
• 11:56
  99%
• 11:56
  of
• 11:56
  the
• 11:57
  time")
• 11:57
  or
• 11:57
  context
• 11:58
  hygiene
• 11:58
  ("/compact
• 11:59
  and
• 11:59
  /clear").
• 12:01
  The
• 12:01
  pattern
• 12:01
  echoes
• 12:02
  the
• 12:02
  cloud
• 12:02
  migration
• 12:02
  regret
• 12:03
  cycle:
• 12:04
  low
• 12:04
  prices
• 12:04
  drove
• 12:04
  adoption,
• 12:05
  then
• 12:05
  the
• 12:05
  bill
• 12:05
  arrived,
• 12:06
  and
• 12:06
  the
• 12:06
  switching
• 12:07
  costs
• 12:07
  were
• 12:07
  already
• 12:08
  locked
• 12:08
  in.
• 12:09
  Reuven
• 12:09
  Cohen
• 12:09
  quantified
• 12:10
  the
• 12:10
  professional
• 12:10
  end
• 12:10
  of
• 12:11
  this
• 12:11
  curve:
• 12:11
  a
• 12:12
  single
• 12:12
  developer
• 12:12
  running
• 12:12
  Claude
• 12:13
  Code
• 12:13
  in
• 12:13
  swarm-style
• 12:14
  development
• 12:14
  burns
• 12:14
  roughly
• 12:15
  $2,500/day
• 12:16
  or
• 12:16
  $75,000/month
• 12:18
  via
• 12:18
  the
• 12:18
  Anthropic
• 12:18
  enterprise
• 12:19
  API.
• 12:20
  His
• 12:20
  argument
• 12:20
  is
• 12:20
  that
• 12:21
  the
• 12:21
  biggest
• 12:21
  cost
• 12:21
  isn't
• 12:22
  the
• 12:22
  model
• 12:22
  —
• 12:23
  it's
• 12:23
  "the
• 12:23
  constant
• 12:23
  replay
• 12:24
  of
• 12:24
  context."
• 12:25
  Architecture
• 12:25
  documents,
• 12:26
  ADRs,
• 12:27
  source
• 12:27
  files,
• 12:28
  and
• 12:28
  conversation
• 12:29
  history
• 12:29
  get
• 12:29
  resent
• 12:30
  every
• 12:30
  loop
• 12:30
  iteration.
• 12:31
  His
• 12:31
  company's
• 12:32
  approach
• 12:32
  operates
• 12:33
  at
• 12:33
  the
• 12:33
  AST
• 12:33
  and
• 12:33
  semantic
• 12:34
  level
• 12:34
  rather
• 12:35
  than
• 12:35
  treating
• 12:35
  code
• 12:35
  as
• 12:35
  raw
• 12:36
  text,
• 12:36
  routing
• 12:37
  only
• 12:37
  the
• 12:37
  pieces
• 12:37
  relevant
• 12:38
  to
• 12:38
  the
• 12:38
  current
• 12:38
  task.
• 12:39
  Combined
• 12:40
  with
• 12:40
  prompt
• 12:40
  caching,
• 12:41
  selective
• 12:41
  model
• 12:42
  escalation,
• 12:42
  and
• 12:42
  retrieval,
• 12:43
  he
• 12:43
  claims
• 12:44
  multi-multiple
• 12:44
  cost
• 12:45
  reductions
• 12:45
  while
• 12:45
  maintaining
• 12:46
  output
• 12:46
  quality.
• 12:47
  "Every
• 12:48
  token
• 12:48
  you
• 12:48
  don't
• 12:48
  send
• 12:49
  is
• 12:49
  a
• 12:49
  token
• 12:49
  you
• 12:49
  don't
• 12:50
  pay
• 12:50
  for."
• 12:51
  The
• 12:51
  workflow
• 12:51
  evolution
• 12:52
  is
• 12:52
  visible
• 12:52
  in
• 12:52
  practice.
• 12:54
  A
• 12:54
  content
• 12:54
  creator
• 12:55
  describes
• 12:55
  assembling
• 12:56
  context
• 12:56
  windows
• 12:56
  manually
• 12:57
  —
• 12:58
  telling
• 12:58
  Codex
• 12:58
  to
• 12:58
  find
• 12:59
  files
• 12:59
  by
• 12:59
  natural-language
• 13:00
  description,
• 13:01
  copy
• 13:01
  them
• 13:01
  into
• 13:01
  a
• 13:01
  clean
• 13:02
  working
• 13:02
  folder,
• 13:03
  then
• 13:03
  open
• 13:03
  a
• 13:03
  new
• 13:03
  chat
• 13:04
  pointed
• 13:04
  at
• 13:04
  that
• 13:04
  folder
• 13:05
  with
• 13:05
  explicit
• 13:05
  instructions.
• 13:07
  The
• 13:07
  result:
• 13:07
  30,000–50,000-word
• 13:11
  document
• 13:12
  work,
• 13:12
  complex
• 13:13
  spreadsheet
• 13:13
  operations,
• 13:14
  and
• 13:14
  multi-threaded
• 13:15
  idea
• 13:15
  incubation
• 13:16
  that
• 13:16
  "just
• 13:16
  works"
• 13:17
  because
• 13:18
  Codex
• 13:18
  treats
• 13:18
  text
• 13:19
  files
• 13:19
  the
• 13:19
  same
• 13:19
  as
• 13:19
  code
• 13:20
  files
• 13:20
  in
• 13:20
  a
• 13:20
  repo.
• 13:21
  The
• 13:21
  shift
• 13:22
  from
• 13:22
  prompt
• 13:22
  engineering
• 13:23
  to
• 13:23
  collaborative
• 13:23
  task-shaping
• 13:24
  —
• 13:25
  defining
• 13:25
  the
• 13:25
  shape
• 13:25
  of
• 13:26
  the
• 13:26
  work
• 13:26
  with
• 13:26
  the
• 13:26
  model
• 13:26
  before
• 13:27
  executing
• 13:27
  agentically
• 13:28
  —
• 13:29
  marks
• 13:29
  a
• 13:29
  maturity
• 13:29
  inflection.
• 13:31
  A
• 13:31
  developer
• 13:31
  reflecting
• 13:32
  on
• 13:32
  eight
• 13:32
  years
• 13:32
  at
• 13:32
  Atlassian,
• 13:33
  where
• 13:33
  they
• 13:34
  built
• 13:34
  an
• 13:34
  Envoy
• 13:34
  control
• 13:35
  plane
• 13:35
  called
• 13:35
  "Sovereign"
• 13:36
  serving
• 13:36
  2,000
• 13:37
  proxies
• 13:37
  across
• 13:38
  13
• 13:38
  AWS
• 13:39
  regions
• 13:40
  plus
• 13:40
  an
• 13:40
  authentication
• 13:40
  sidecar
• 13:41
  in
• 13:41
  Rust,
• 13:42
  offered
• 13:42
  a
• 13:42
  counterpoint
• 13:43
  to
• 13:43
  the
• 13:43
  AI-everything
• 13:43
  narrative:
• 13:44
  "Building
• 13:45
  something
• 13:45
  is
• 13:45
  easy.
• 13:46
  Changing
• 13:47
  it
• 13:47
  and
• 13:47
  making
• 13:47
  sure
• 13:47
  that
• 13:47
  you
• 13:47
  can
• 13:48
  still
• 13:48
  change
• 13:48
  it
• 13:48
  over
• 13:48
  time
• 13:49
  is
• 13:49
  difficult."
• 13:50
  The
• 13:50
  maintenance
• 13:51
  burden
• 13:51
  —
• 13:52
  onboarding
• 13:52
  cycles,
• 13:53
  codebase
• 13:53
  churn,
• 13:54
  coupled
• 13:54
  changes
• 13:55
  —
• 13:56
  doesn't
• 13:56
  appear
• 13:56
  at
• 13:56
  the
• 13:56
  start.
• 13:57
  Whether
• 13:58
  AI-assisted
• 13:58
  "vibe
• 13:59
  coded"
• 13:59
  apps
• 14:00
  can
• 14:00
  handle
• 14:00
  that
• 14:00
  long
• 14:01
  tail
• 14:01
  remains
• 14:01
  an
• 14:01
  open
• 14:02
  question.
• 14:04
  Deja
• 14:05
  Vu
• 14:05
  is
• 14:05
  a
• 14:05
  local-first
• 14:06
  AI
• 14:06
  memory
• 14:06
  layer
• 14:07
  that
• 14:07
  stores
• 14:07
  context
• 14:08
  in
• 14:08
  SQLite
• 14:08
  at
• 14:08
  ~/.dejavu
• 14:09
  and
• 14:09
  exposes
• 14:10
  it
• 14:10
  via
• 14:10
  Python
• 14:10
  SDK,
• 14:11
  CLI,
• 14:12
  REST
• 14:12
  API,
• 14:13
  and
• 14:13
  MCP
• 14:14
  server
• 14:14
  —
• 14:14
  meaning
• 14:15
  a
• 14:15
  preference
• 14:15
  added
• 14:15
  from
• 14:16
  the
• 14:16
  terminal
• 14:16
  is
• 14:16
  immediately
• 14:17
  available
• 14:17
  in
• 14:17
  Claude
• 14:18
  Desktop
• 14:18
  or
• 14:18
  a
• 14:19
  Python
• 14:19
  agent,
• 14:19
  same
• 14:20
  database,
• 14:20
  no
• 14:20
  sync,
• 14:21
  no
• 14:21
  account.
• 14:22
  LLM
• 14:23
  calls
• 14:23
  route
• 14:23
  through
• 14:23
  Venice's
• 14:24
  privacy-focused
• 14:24
  API;
• 14:25
  everything
• 14:26
  else
• 14:26
  runs
• 14:26
  locally.
• 14:27
  It's
• 14:27
  the
• 14:27
  latest
• 14:28
  in
• 14:28
  a
• 14:28
  crowded
• 14:28
  field
• 14:28
  of
• 14:28
  local
• 14:29
  memory
• 14:29
  tools,
• 14:30
  but
• 14:30
  the
• 14:30
  multi-interface
• 14:31
  approach
• 14:31
  (one
• 14:31
  store,
• 14:32
  every
• 14:32
  tool)
• 14:33
  and
• 14:33
  the
• 14:33
  MCP
• 14:33
  integration
• 14:34
  give
• 14:34
  it
• 14:34
  practical
• 14:35
  stickiness.
• 14:36
  UniMM-Trainer
• 14:38
  fills
• 14:38
  a
• 14:38
  gap
• 14:38
  that
• 14:39
  more
• 14:39
  prominent
• 14:39
  projects
• 14:40
  have
• 14:40
  stepped
• 14:40
  around:
• 14:41
  a
• 14:41
  small,
• 14:42
  opinionated
• 14:42
  library
• 14:43
  for
• 14:43
  training
• 14:43
  multimodal
• 14:44
  models
• 14:44
  that
• 14:44
  combine
• 14:44
  at
• 14:45
  least
• 14:45
  two
• 14:45
  of
• 14:45
  text,
• 14:45
  vision,
• 14:46
  and
• 14:46
  audio.
• 14:47
  Plug
• 14:48
  a
• 14:48
  frozen
• 14:48
  encoder
• 14:48
  (Whisper,
• 14:49
  CLIP,
• 14:50
  SigLIP,
• 14:50
  DINOv2)
• 14:51
  into
• 14:51
  a
• 14:51
  language
• 14:52
  backbone
• 14:52
  (Llama,
• 14:53
  Qwen,
• 14:54
  Mistral),
• 14:54
  train
• 14:55
  only
• 14:55
  the
• 14:55
  projection
• 14:55
  adapter
• 14:56
  (linear,
• 14:57
  Q-Former,
• 14:58
  perceiver-resampler),
• 14:58
  and
• 14:59
  get
• 14:59
  modality-balanced
• 15:00
  loss
• 15:00
  reporting,
• 15:00
  frozen-encoder
• 15:01
  feature
• 15:01
  caching,
• 15:02
  and
• 15:02
  sensible
• 15:03
  LR
• 15:03
  ratio
• 15:04
  defaults.
• 15:04
  It
• 15:04
  doesn't
• 15:05
  try
• 15:05
  to
• 15:05
  be
• 15:05
  a
• 15:05
  foundation
• 15:06
  model
• 15:06
  release
• 15:06
  or
• 15:06
  a
• 15:06
  serving
• 15:07
  framework
• 15:07
  —
• 15:08
  just
• 15:08
  a
• 15:08
  correct
• 15:08
  training
• 15:09
  loop
• 15:09
  that
• 15:09
  saves
• 15:09
  you
• 15:09
  from
• 15:10
  forking
• 15:10
  BLIP-2's
• 15:11
  5,000-line
• 15:12
  codebase.
• 15:13
  The
• 15:14
  Cognitum
• 15:14
  One
• 15:14
  Seed
• 15:14
  pushes
• 15:15
  further
• 15:15
  into
• 15:15
  ambient
• 15:16
  intelligence:
• 15:17
  a
• 15:17
  self-contained
• 15:18
  AI
• 15:18
  computer
• 15:18
  that
• 15:19
  fuses
• 15:19
  mmWave
• 15:20
  radar
• 15:20
  (24
• 15:21
  GHz
• 15:21
  for
• 15:21
  2D
• 15:21
  position
• 15:22
  tracking,
• 15:23
  60
• 15:23
  GHz
• 15:24
  for
• 15:24
  contactless
• 15:24
  heart
• 15:25
  rate
• 15:25
  and
• 15:25
  breathing),
• 15:26
  WiFi
• 15:26
  CSI
• 15:27
  from
• 15:27
  ESP32
• 15:29
  node
• 15:29
  arrays
• 15:29
  for
• 15:29
  through-wall
• 15:30
  presence
• 15:30
  detection,
• 15:31
  and
• 15:31
  IMU
• 15:32
  data
• 15:32
  —
• 15:33
  all
• 15:33
  processed
• 15:33
  on-device
• 15:34
  with
• 15:34
  vector
• 15:34
  memory,
• 15:35
  no
• 15:35
  cloud
• 15:35
  required.
• 15:36
  The
• 15:36
  critical
• 15:37
  constraint
• 15:37
  most
• 15:37
  guides
• 15:38
  skip:
• 15:39
  WiFi
• 15:39
  CSI
• 15:40
  needs
• 15:40
  a
• 15:40
  minimum
• 15:41
  of
• 15:41
  three
• 15:41
  nodes
• 15:41
  for
• 15:41
  basic
• 15:42
  triangulation,
• 15:43
  with
• 15:43
  seven
• 15:43
  recommended.
• 15:45
  The
• 15:45
  RuView
• 15:45
  software
• 15:45
  layer
• 15:46
  produces
• 15:46
  a
• 15:46
  128-dimensional
• 15:48
  anonymous
• 15:48
  identity
• 15:48
  fingerprint
• 15:49
  from
• 15:49
  CSI
• 15:50
  data,
• 15:50
  enabling
• 15:51
  re-identification
• 15:52
  across
• 15:52
  sessions
• 15:52
  without
• 15:53
  cameras
• 15:53
  or
• 15:53
  wearables
• 15:53
  —
• 15:54
  though
• 15:55
  it
• 15:55
  currently
• 15:55
  ships
• 15:55
  without
• 15:56
  pretrained
• 15:56
  weights.
• 15:57
  On
• 15:58
  the
• 15:58
  Cognitive
• 15:58
  Revolution
• 15:58
  podcast,
• 15:59
  Nathan
• 16:00
  Labenz
• 16:00
  detailed
• 16:00
  his
• 16:01
  personal
• 16:01
  AI
• 16:01
  infrastructure:
• 16:03
  a
• 16:03
  1
• 16:03
  GB
• 16:03
  SQLite
• 16:04
  database
• 16:04
  containing
• 16:05
  five
• 16:05
  years
• 16:05
  of
• 16:05
  digital
• 16:06
  history
• 16:06
  (emails,
• 16:07
  calls,
• 16:08
  podcasts,
• 16:08
  DMs,
• 16:09
  social
• 16:10
  media)
• 16:10
  with
• 16:11
  monthly
• 16:11
  and
• 16:11
  annual
• 16:11
  summarization
• 16:12
  layers
• 16:12
  plus
• 16:13
  a
• 16:13
  500-article
• 16:14
  wiki
• 16:15
  of
• 16:15
  profiled
• 16:15
  individuals
• 16:16
  and
• 16:16
  organizations,
• 16:17
  all
• 16:17
  searchable
• 16:18
  by
• 16:18
  Claude
• 16:18
  Code
• 16:19
  on
• 16:19
  his
• 16:19
  laptop.
• 16:20
  His
• 16:20
  experimental
• 16:21
  second
• 16:21
  layer:
• 16:22
  two
• 16:22
  autonomous
• 16:23
  AI
• 16:23
  employees
• 16:23
  named
• 16:24
  Aid
• 16:24
  and
• 16:24
  Clay,
• 16:25
  running
• 16:25
  on
• 16:25
  a
• 16:25
  dedicated
• 16:26
  Mac
• 16:26
  Mini
• 16:27
  via
• 16:27
  Tailscale,
• 16:28
  with
• 16:28
  their
• 16:28
  own
• 16:28
  Gmail
• 16:29
  accounts,
• 16:29
  GitHub
• 16:30
  access,
• 16:31
  and
• 16:31
  restricted
• 16:31
  virtual
• 16:32
  credit
• 16:32
  cards.
• 16:33
  As
• 16:33
  a
• 16:33
  proof
• 16:33
  of
• 16:34
  concept,
• 16:34
  Aid
• 16:35
  autonomously
• 16:35
  booked
• 16:36
  a
• 16:36
  full
• 16:36
  week
• 16:36
  of
• 16:36
  podcast
• 16:37
  guests
• 16:37
  from
• 16:37
  a
• 16:37
  25-person
• 16:38
  candidate
• 16:39
  list,
• 16:39
  managing
• 16:40
  communications
• 16:41
  without
• 16:41
  most
• 16:41
  people
• 16:42
  realizing
• 16:42
  they
• 16:42
  were
• 16:42
  interacting
• 16:43
  with
• 16:43
  an
• 16:43
  AI.
• 16:44
  Security
• 16:45
  researcher
• 16:45
  Daniel
• 16:45
  Mesler
• 16:46
  emphasized
• 16:46
  that
• 16:47
  a
• 16:47
  clear
• 16:47
  hierarchy
• 16:47
  among
• 16:48
  AI
• 16:48
  agents
• 16:48
  still
• 16:49
  beats
• 16:49
  emergent
• 16:49
  teamwork,
• 16:50
  and
• 16:50
  both
• 16:51
  agreed
• 16:51
  on
• 16:51
  always
• 16:51
  preserving
• 16:52
  raw
• 16:52
  data
• 16:53
  so
• 16:53
  systems
• 16:53
  can
• 16:53
  be
• 16:53
  rebuilt
• 16:54
  as
• 16:54
  capabilities
• 16:55
  improve
• 16:55
  —
• 16:56
  "bitter
• 16:56
  lesson
• 16:56
  engineering."