Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 18, 2026

Transcript

AGI Dreams Podcast – March 18, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  AWS
• 0:00
  Bedrock
• 0:01
  AgentCore
• 0:01
  advertises
• 0:02
  a
• 0:03
  "Sandbox"
• 0:03
  network
• 0:04
  mode
• 0:04
  for
• 0:04
  its
• 0:04
  Code
• 0:04
  Interpreter
• 0:05
  service
• 0:05
  —
• 0:06
  the
• 0:06
  execution
• 0:07
  environment
• 0:07
  where
• 0:07
  AI
• 0:08
  agents
• 0:08
  run
• 0:08
  Python,
• 0:09
  JavaScript,
• 0:09
  or
• 0:09
  shell
• 0:10
  commands
• 0:10
  on
• 0:10
  behalf
• 0:10
  of
• 0:10
  users.
• 0:11
  According
• 0:12
  to
• 0:12
  BeyondTrust's
• 0:12
  Phantom
• 0:12
  Labs,
• 0:13
  that
• 0:13
  sandbox
• 0:14
  has
• 0:14
  a
• 0:14
  hole
• 0:14
  you
• 0:14
  could
• 0:14
  drive
• 0:15
  a
• 0:15
  DNS
• 0:15
  tunnel
• 0:15
  through.
• 0:17
  Researcher
• 0:17
  Kinnaird
• 0:17
  McQuade
• 0:18
  discovered
• 0:18
  that
• 0:18
  Sandbox
• 0:18
  mode
• 0:19
  permits
• 0:19
  public
• 0:19
  DNS
• 0:20
  resolution
• 0:20
  despite
• 0:21
  marketing
• 0:21
  language
• 0:21
  promising
• 0:22
  "complete
• 0:23
  isolation
• 0:23
  with
• 0:23
  no
• 0:23
  external
• 0:24
  access."
• 0:25
  Phantom
• 0:25
  Labs
• 0:25
  built
• 0:26
  a
• 0:26
  full
• 0:26
  proof-of-concept:
• 0:27
  a
• 0:27
  custom
• 0:27
  DNS
• 0:28
  command-and-control
• 0:28
  protocol
• 0:29
  that
• 0:29
  encodes
• 0:30
  commands
• 0:30
  as
• 0:30
  chunked
• 0:30
  ASCII
• 0:31
  in
• 0:31
  DNS
• 0:31
  A
• 0:31
  records
• 0:32
  and
• 0:32
  exfiltrates
• 0:33
  data
• 0:33
  via
• 0:33
  long
• 0:33
  subdomains.
• 0:35
  The
• 0:35
  result?
• 0:36
  A
• 0:36
  bidirectional
• 0:37
  interactive
• 0:37
  reverse
• 0:37
  shell
• 0:37
  inside
• 0:38
  what
• 0:38
  was
• 0:38
  supposed
• 0:38
  to
• 0:39
  be
• 0:39
  a
• 0:39
  network-isolated
• 0:39
  Firecracker
• 0:40
  microVM.
• 0:41
  The
• 0:41
  attack
• 0:42
  chain
• 0:42
  is
• 0:42
  worth
• 0:42
  understanding
• 0:43
  in
• 0:43
  detail.
• 0:44
  Commands
• 0:44
  are
• 0:44
  delivered
• 0:45
  through
• 0:45
  DNS
• 0:45
  A
• 0:45
  record
• 0:46
  responses
• 0:46
  where
• 0:46
  the
• 0:47
  first
• 0:47
  octet
• 0:47
  signals
• 0:48
  continuation
• 0:48
  (10)
• 0:48
  or
• 0:49
  termination
• 0:49
  (11),
• 0:50
  and
• 0:50
  octets
• 0:50
  2-4
• 0:51
  carry
• 0:51
  ASCII
• 0:51
  values
• 0:52
  of
• 0:52
  base64-encoded
• 0:53
  characters.
• 0:54
  Output
• 0:54
  exfiltration
• 0:55
  embeds
• 0:55
  base64
• 0:56
  data
• 0:57
  in
• 0:57
  DNS
• 0:57
  subdomain
• 0:58
  queries
• 0:58
  with
• 0:58
  cache-busting
• 0:59
  fields
• 1:00
  to
• 1:00
  ensure
• 1:00
  each
• 1:00
  query
• 1:01
  reaches
• 1:01
  the
• 1:01
  attacker's
• 1:01
  server.
• 1:02
  Phantom
• 1:03
  Labs
• 1:03
  demonstrated
• 1:03
  listing
• 1:04
  S3
• 1:04
  buckets,
• 1:05
  reading
• 1:05
  files
• 1:06
  containing
• 1:06
  PII
• 1:07
  and
• 1:07
  credentials,
• 1:08
  and
• 1:08
  accessing
• 1:08
  financial
• 1:08
  data
• 1:09
  —
• 1:09
  all
• 1:10
  through
• 1:10
  a
• 1:10
  Code
• 1:10
  Interpreter
• 1:10
  instance
• 1:11
  configured
• 1:11
  for
• 1:11
  "no
• 1:11
  network
• 1:12
  access."
• 1:13
  The
• 1:13
  kicker:
• 1:13
  AWS's
• 1:14
  AgentCore
• 1:15
  Starter
• 1:15
  Toolkit
• 1:16
  Default
• 1:16
  Role
• 1:16
  grants
• 1:16
  full
• 1:17
  S3
• 1:17
  read
• 1:17
  access
• 1:18
  to
• 1:18
  all
• 1:18
  buckets,
• 1:19
  full
• 1:19
  DynamoDB
• 1:20
  access,
• 1:20
  and
• 1:20
  full
• 1:21
  Secrets
• 1:21
  Manager
• 1:21
  access.
• 1:22
  An
• 1:23
  overprivileged
• 1:23
  IAM
• 1:24
  role
• 1:24
  combined
• 1:24
  with
• 1:25
  a
• 1:25
  leaky
• 1:25
  sandbox
• 1:25
  is
• 1:25
  a
• 1:26
  data
• 1:26
  breach
• 1:26
  waiting
• 1:26
  to
• 1:26
  happen.
• 1:27
  AWS
• 1:28
  ultimately
• 1:28
  declined
• 1:29
  to
• 1:29
  patch
• 1:29
  the
• 1:29
  issue,
• 1:30
  instead
• 1:31
  updating
• 1:31
  documentation
• 1:32
  to
• 1:32
  note
• 1:32
  that
• 1:32
  DNS
• 1:33
  resolution
• 1:33
  is
• 1:33
  "enabled
• 1:34
  to
• 1:34
  support
• 1:34
  successful
• 1:35
  execution
• 1:35
  of
• 1:36
  S3
• 1:36
  operations."
• 1:37
  They
• 1:38
  awarded
• 1:38
  the
• 1:38
  researcher
• 1:38
  a
• 1:38
  $100
• 1:39
  gift
• 1:39
  card
• 1:40
  and
• 1:40
  a
• 1:40
  CVSS
• 1:41
  7.5
• 1:41
  score.
• 1:42
  The
• 1:42
  defensive
• 1:43
  posture
• 1:43
  here
• 1:43
  matters.
• 1:44
  Code
• 1:44
  interpreters
• 1:45
  are
• 1:45
  the
• 1:45
  execution
• 1:46
  endpoints
• 1:46
  for
• 1:46
  AI
• 1:47
  agents
• 1:47
  —
• 1:48
  prompt
• 1:48
  injection,
• 1:49
  supply
• 1:50
  chain
• 1:50
  compromises
• 1:50
  in
• 1:51
  the
• 1:51
  270+
• 1:52
  bundled
• 1:52
  Python
• 1:52
  packages,
• 1:53
  and
• 1:53
  adversarial
• 1:54
  code
• 1:54
  generation
• 1:55
  all
• 1:55
  provide
• 1:56
  realistic
• 1:56
  paths
• 1:56
  to
• 1:57
  attacker-controlled
• 1:57
  code
• 1:58
  execution.
• 1:59
  When
• 1:59
  the
• 1:59
  sandbox
• 1:59
  is
• 1:59
  the
• 2:00
  last
• 2:00
  line
• 2:00
  of
• 2:00
  defense
• 2:01
  and
• 2:01
  that
• 2:01
  sandbox
• 2:01
  leaks
• 2:01
  DNS,
• 2:02
  the
• 2:02
  isolation
• 2:03
  guarantee
• 2:03
  collapses.
• 2:05
  Organizations
• 2:05
  running
• 2:05
  AgentCore
• 2:06
  should
• 2:07
  inventory
• 2:07
  their
• 2:07
  Code
• 2:07
  Interpreter
• 2:08
  instances
• 2:09
  immediately,
• 2:10
  audit
• 2:10
  IAM
• 2:10
  roles
• 2:11
  for
• 2:11
  least
• 2:11
  privilege,
• 2:12
  and
• 2:12
  migrate
• 2:12
  sensitive
• 2:13
  workloads
• 2:13
  to
• 2:13
  VPC
• 2:14
  Mode
• 2:14
  with
• 2:14
  Route53
• 2:15
  Resolver
• 2:15
  DNS
• 2:16
  Firewall.
• 2:17
  Meanwhile,
• 2:17
  the
• 2:17
  offensive
• 2:18
  AI
• 2:18
  security
• 2:19
  market
• 2:19
  just
• 2:19
  got
• 2:19
  a
• 2:19
  billion-dollar
• 2:20
  vote
• 2:20
  of
• 2:20
  confidence:
• 2:21
  xBow
• 2:21
  raised
• 2:22
  $120
• 2:22
  million
• 2:23
  in
• 2:23
  Series
• 2:24
  C
• 2:24
  at
• 2:24
  a
• 2:24
  valuation
• 2:25
  exceeding
• 2:25
  $1
• 2:26
  billion.
• 2:27
  Founded
• 2:27
  by
• 2:27
  Oege
• 2:27
  de
• 2:28
  Moor
• 2:28
  —
• 2:28
  creator
• 2:29
  of
• 2:29
  GitHub
• 2:29
  Copilot
• 2:29
  and
• 2:30
  GitHub
• 2:30
  Advanced
• 2:30
  Security
• 2:30
  —
• 2:31
  xBow
• 2:32
  applies
• 2:32
  AI
• 2:32
  reasoning
• 2:33
  and
• 2:33
  adversarial
• 2:33
  workflows
• 2:34
  to
• 2:34
  find
• 2:34
  vulnerabilities
• 2:35
  at
• 2:35
  machine
• 2:35
  speed,
• 2:36
  claiming
• 2:36
  top
• 2:36
  placement
• 2:37
  on
• 2:37
  the
• 2:37
  HackerOne
• 2:37
  leaderboard.
• 2:38
  The
• 2:39
  round
• 2:39
  was
• 2:39
  led
• 2:39
  by
• 2:39
  DFJ
• 2:40
  Growth
• 2:40
  and
• 2:40
  Northzone,
• 2:41
  with
• 2:41
  Sequoia,
• 2:42
  Altimeter,
• 2:43
  and
• 2:43
  others
• 2:43
  participating.
• 2:45
  AI
• 2:45
  Cyber
• 2:45
  Magazine's
• 2:46
  Winter
• 2:46
  2026
• 2:47
  issue
• 2:47
  provides
• 2:47
  broader
• 2:48
  context
• 2:49
  on
• 2:49
  the
• 2:49
  intersection
• 2:49
  of
• 2:49
  artificial
• 2:50
  intelligence
• 2:50
  and
• 2:50
  cybersecurity,
• 2:51
  covering
• 2:52
  in-depth
• 2:52
  analysis,
• 2:53
  technical
• 2:53
  how-tos,
• 2:54
  and
• 2:54
  thought
• 2:54
  leadership
• 2:55
  across
• 2:55
  the
• 2:55
  field.
• 2:58
  Stripe
• 2:59
  merges
• 2:59
  over
• 2:59
  1,300
• 3:00
  AI-written
• 3:01
  pull
• 3:01
  requests
• 3:01
  every
• 3:02
  single
• 3:02
  week
• 3:02
  —
• 3:02
  zero
• 3:03
  human-written
• 3:03
  code
• 3:03
  in
• 3:03
  those
• 3:04
  PRs.
• 3:04
  The
• 3:04
  architecture
• 3:05
  behind
• 3:05
  that
• 3:05
  velocity
• 3:06
  is
• 3:06
  what
• 3:06
  Stripe
• 3:06
  calls
• 3:06
  "blueprints":
• 3:07
  structured
• 3:07
  workflows
• 3:08
  that
• 3:08
  alternate
• 3:09
  between
• 3:09
  deterministic
• 3:09
  steps
• 3:10
  (linting,
• 3:10
  type
• 3:10
  checking,
• 3:11
  CI
• 3:11
  pipelines)
• 3:11
  and
• 3:12
  agentic
• 3:12
  steps
• 3:12
  (where
• 3:12
  the
• 3:12
  coding
• 3:13
  agent
• 3:13
  reasons
• 3:13
  and
• 3:13
  writes
• 3:14
  code).
• 3:14
  The
• 3:14
  interleaving
• 3:15
  is
• 3:15
  the
• 3:15
  key
• 3:15
  insight.
• 3:16
  Deterministic
• 3:16
  guardrails
• 3:17
  constrain
• 3:17
  AI
• 3:17
  output
• 3:18
  at
• 3:18
  each
• 3:18
  stage;
• 3:19
  the
• 3:19
  AI
• 3:19
  handles
• 3:19
  creative
• 3:20
  reasoning
• 3:20
  between
• 3:20
  those
• 3:21
  guardrails.
• 3:22
  This
• 3:22
  hybrid
• 3:22
  pattern
• 3:23
  has
• 3:23
  independently
• 3:23
  converged
• 3:24
  across
• 3:24
  multiple
• 3:24
  major
• 3:25
  engineering
• 3:25
  organizations:
• 3:26
  Shopify
• 3:27
  built
• 3:27
  an
• 3:27
  equivalent
• 3:27
  system
• 3:28
  called
• 3:28
  "Roast,"
• 3:29
  Amazon
• 3:29
  attributed
• 3:30
  savings
• 3:30
  of
• 3:30
  4,500
• 3:31
  developer-years
• 3:32
  to
• 3:32
  the
• 3:32
  same
• 3:32
  architecture,
• 3:33
  and
• 3:34
  Airbnb
• 3:34
  migrated
• 3:35
  3,500
• 3:35
  test
• 3:36
  files
• 3:36
  in
• 3:36
  six
• 3:36
  weeks
• 3:36
  using
• 3:37
  it.
• 3:37
  The
• 3:38
  lesson
• 3:38
  is
• 3:38
  not
• 3:38
  that
• 3:38
  AI
• 3:38
  writes
• 3:39
  good
• 3:39
  code
• 3:39
  on
• 3:39
  its
• 3:39
  own
• 3:39
  —
• 3:40
  it
• 3:40
  is
• 3:40
  that
• 3:40
  well-designed
• 3:41
  agent
• 3:41
  harnesses
• 3:42
  with
• 3:42
  deterministic
• 3:42
  validation
• 3:43
  checkpoints
• 3:43
  unlock
• 3:44
  reliable,
• 3:44
  scalable
• 3:45
  AI
• 3:45
  coding.
• 3:46
  If
• 3:46
  Stripe
• 3:46
  represents
• 3:46
  the
• 3:47
  volume
• 3:47
  play,
• 3:47
  Mistral's
• 3:48
  Leanstral
• 3:48
  represents
• 3:48
  the
• 3:49
  correctness
• 3:49
  play.
• 3:50
  Leanstral
• 3:50
  is
• 3:50
  the
• 3:50
  first
• 3:51
  open-source
• 3:51
  coding
• 3:52
  agent
• 3:52
  designed
• 3:52
  specifically
• 3:53
  for
• 3:53
  Lean
• 3:53
  4,
• 3:54
  the
• 3:54
  proof
• 3:54
  assistant
• 3:55
  capable
• 3:55
  of
• 3:55
  expressing
• 3:56
  complex
• 3:56
  mathematical
• 3:56
  objects
• 3:57
  and
• 3:57
  software
• 3:57
  specifications.
• 3:59
  With
• 3:59
  just
• 3:59
  6
• 4:00
  billion
• 4:00
  active
• 4:00
  parameters
• 4:01
  (via
• 4:01
  a
• 4:01
  highly
• 4:01
  sparse
• 4:02
  mixture-of-experts
• 4:03
  architecture),
• 4:04
  Leanstral
• 4:04
  outperforms
• 4:05
  models
• 4:05
  10-60x
• 4:06
  its
• 4:06
  size
• 4:06
  on
• 4:06
  FLTEval,
• 4:08
  a
• 4:08
  new
• 4:08
  benchmark
• 4:08
  that
• 4:09
  evaluates
• 4:09
  formal
• 4:09
  proofs
• 4:10
  and
• 4:10
  correct
• 4:10
  mathematical
• 4:11
  definitions
• 4:12
  in
• 4:12
  realistic
• 4:12
  repository
• 4:13
  PRs
• 4:13
  rather
• 4:14
  than
• 4:14
  isolated
• 4:14
  competition
• 4:15
  problems.
• 4:16
  At
• 4:16
  pass@2,
• 4:17
  Leanstral
• 4:17
  scores
• 4:17
  26.3
• 4:18
  —
• 4:19
  beating
• 4:20
  Claude
• 4:20
  Sonnet
• 4:20
  4.6
• 4:21
  by
• 4:21
  2.6
• 4:22
  points
• 4:22
  while
• 4:22
  costing
• 4:23
  $36
• 4:23
  versus
• 4:24
  Sonnet's
• 4:24
  $549.
• 4:26
  Even
• 4:27
  at
• 4:27
  pass@16,
• 4:28
  where
• 4:28
  Leanstral
• 4:28
  reaches
• 4:29
  31.9,
• 4:30
  it
• 4:30
  costs
• 4:30
  92x
• 4:31
  less
• 4:31
  than
• 4:31
  Claude
• 4:32
  Opus
• 4:32
  4.6.
• 4:33
  The
• 4:33
  model
• 4:34
  supports
• 4:34
  MCP
• 4:35
  (Model
• 4:35
  Context
• 4:35
  Protocol)
• 4:36
  through
• 4:36
  Mistral
• 4:36
  Vibe,
• 4:37
  was
• 4:37
  specifically
• 4:38
  trained
• 4:38
  to
• 4:38
  work
• 4:38
  with
• 4:38
  lean-lsp-mcp,
• 4:40
  and
• 4:40
  ships
• 4:40
  under
• 4:40
  Apache
• 4:41
  2.0.
• 4:43
  The
• 4:43
  practitioner
• 4:43
  perspective
• 4:44
  comes
• 4:44
  from
• 4:44
  Hamilton
• 4:45
  Carter,
• 4:45
  who
• 4:45
  has
• 4:45
  been
• 4:45
  building
• 4:46
  with
• 4:46
  Gas
• 4:46
  Town
• 4:46
  —
• 4:47
  an
• 4:47
  agentic
• 4:48
  coding
• 4:48
  platform
• 4:49
  where
• 4:49
  Claude
• 4:49
  agents
• 4:50
  carry
• 4:50
  out
• 4:50
  development
• 4:50
  tasks
• 4:51
  autonomously.
• 4:52
  Carter
• 4:52
  describes
• 4:53
  how
• 4:53
  he
• 4:53
  and
• 4:53
  his
• 4:53
  kid
• 4:54
  built
• 4:54
  a
• 4:54
  Morse
• 4:54
  code
• 4:54
  game
• 4:55
  (CW
• 4:55
  Simon)
• 4:56
  by
• 4:56
  writing
• 4:56
  a
• 4:56
  software
• 4:57
  requirements
• 4:57
  specification
• 4:58
  and
• 4:58
  having
• 4:58
  "polecats"
• 4:59
  (Gas
• 5:00
  Town's
• 5:00
  term
• 5:00
  for
• 5:00
  agents)
• 5:01
  execute
• 5:01
  the
• 5:01
  tasks.
• 5:02
  The
• 5:02
  workflow
• 5:03
  is
• 5:03
  not
• 5:03
  frictionless:
• 5:04
  Carter
• 5:05
  encountered
• 5:05
  "zombie
• 5:06
  polecats"
• 5:06
  —
• 5:07
  agents
• 5:07
  coerced
• 5:08
  into
• 5:08
  existence
• 5:08
  with
• 5:08
  no
• 5:09
  work
• 5:09
  to
• 5:09
  do
• 5:09
  —
• 5:10
  when
• 5:10
  the
• 5:10
  system's
• 5:11
  mayor
• 5:11
  agent
• 5:11
  misreported
• 5:12
  project
• 5:12
  status.
• 5:14
  The
• 5:14
  fix
• 5:14
  was
• 5:14
  pragmatic:
• 5:15
  never
• 5:15
  ask
• 5:16
  yes/no
• 5:16
  questions,
• 5:17
  always
• 5:17
  request
• 5:18
  actual
• 5:18
  CLI
• 5:19
  output.
• 5:20
  On
• 5:20
  the
• 5:20
  final
• 5:20
  bug
• 5:20
  fix,
• 5:21
  Carter
• 5:21
  asked
• 5:22
  the
• 5:22
  mayor
• 5:22
  for
• 5:22
  rig
• 5:22
  status
• 5:23
  and
• 5:23
  discovered
• 5:23
  the
• 5:23
  work
• 5:24
  had
• 5:24
  already
• 5:24
  been
• 5:24
  completed,
• 5:25
  merged,
• 5:25
  and
• 5:25
  pushed
• 5:26
  while
• 5:26
  he
• 5:26
  was
• 5:26
  typing
• 5:27
  his
• 5:27
  query.
• 5:28
  The
• 5:28
  pattern
• 5:28
  is
• 5:28
  consistent
• 5:29
  with
• 5:29
  what
• 5:29
  the
• 5:29
  industry
• 5:29
  is
• 5:30
  converging
• 5:30
  on:
• 5:31
  agents
• 5:31
  as
• 5:31
  productivity
• 5:32
  multipliers,
• 5:33
  with
• 5:33
  human
• 5:33
  judgment
• 5:33
  applied
• 5:34
  at
• 5:34
  the
• 5:34
  specification
• 5:35
  and
• 5:35
  verification
• 5:35
  boundaries
• 5:36
  rather
• 5:36
  than
• 5:36
  in
• 5:37
  the
• 5:37
  coding
• 5:37
  loop
• 5:37
  itself.
• 5:40
  The
• 5:40
  infrastructure
• 5:41
  layer
• 5:41
  beneath
• 5:41
  coding
• 5:42
  agents
• 5:42
  is
• 5:42
  filling
• 5:43
  in
• 5:43
  fast.
• 5:44
  RTK
• 5:44
  (Rust
• 5:44
  Token
• 5:44
  Killer)
• 5:45
  is
• 5:45
  a
• 5:45
  single-binary
• 5:46
  CLI
• 5:46
  proxy
• 5:47
  that
• 5:47
  filters
• 5:47
  and
• 5:47
  compresses
• 5:48
  command
• 5:48
  outputs
• 5:49
  before
• 5:49
  they
• 5:49
  reach
• 5:49
  LLM
• 5:50
  context.
• 5:51
  The
• 5:51
  savings
• 5:51
  are
• 5:51
  substantial:
• 5:52
  in
• 5:52
  a
• 5:52
  30-minute
• 5:53
  Claude
• 5:53
  Code
• 5:53
  session,
• 5:54
  RTK
• 5:54
  reduces
• 5:55
  approximately
• 5:55
  118,000
• 5:56
  tokens
• 5:57
  to
• 5:57
  23,900
• 5:58
  —
• 5:59
  an
• 5:59
  80%
• 6:00
  cut.
• 6:01
  It
• 6:01
  works
• 6:01
  by
• 6:01
  transparently
• 6:02
  hooking
• 6:02
  into
• 6:02
  Claude
• 6:03
  Code's
• 6:03
  Bash
• 6:03
  tool
• 6:03
  calls,
• 6:04
  rewriting
• 6:05
  commands
• 6:05
  like
• 6:05
  git
• 6:05
  status
• 6:06
  to
• 6:06
  rtk
• 6:07
  git
• 6:07
  status
• 6:07
  and
• 6:07
  applying
• 6:08
  four
• 6:08
  strategies:
• 6:09
  smart
• 6:10
  filtering
• 6:10
  (removing
• 6:11
  boilerplate),
• 6:11
  grouping
• 6:11
  (aggregating
• 6:12
  files
• 6:12
  by
• 6:12
  directory),
• 6:13
  truncation
• 6:14
  (keeping
• 6:14
  relevant
• 6:15
  context),
• 6:16
  and
• 6:16
  deduplication
• 6:16
  (collapsing
• 6:17
  repeated
• 6:17
  log
• 6:17
  lines
• 6:18
  with
• 6:18
  counts).
• 6:19
  A
• 6:19
  git
• 6:19
  push
• 6:20
  that
• 6:20
  normally
• 6:20
  consumes
• 6:21
  200
• 6:21
  tokens
• 6:21
  becomes
• 6:22
  a
• 6:22
  single
• 6:22
  line:
• 6:23
  "ok
• 6:23
  main"
• 6:23
  at
• 6:24
  10
• 6:24
  tokens.
• 6:25
  Test
• 6:25
  output
• 6:25
  drops
• 6:26
  90%.
• 6:27
  The
• 6:27
  tool
• 6:27
  is
• 6:27
  MIT-licensed,
• 6:28
  Homebrew-installable,
• 6:29
  and
• 6:29
  written
• 6:30
  in
• 6:30
  Rust
• 6:30
  with
• 6:30
  sub-10ms
• 6:31
  overhead.
• 6:32
  One
• 6:32
  layer
• 6:33
  up
• 6:33
  from
• 6:33
  token
• 6:33
  economics
• 6:34
  sits
• 6:34
  agent
• 6:34
  packaging.
• 6:35
  Agent
• 6:36
  Clip
• 6:36
  is
• 6:36
  a
• 6:36
  Go-based
• 6:36
  AI
• 6:37
  agent
• 6:37
  built
• 6:37
  on
• 6:37
  the
• 6:37
  Pinix
• 6:38
  platform
• 6:39
  that
• 6:39
  packages
• 6:39
  an
• 6:39
  entire
• 6:40
  agentic
• 6:40
  loop
• 6:40
  —
• 6:41
  LLM
• 6:41
  calls,
• 6:42
  tool
• 6:42
  execution,
• 6:43
  semantic
• 6:44
  memory
• 6:44
  search,
• 6:45
  browser
• 6:45
  vision,
• 6:46
  and
• 6:46
  clip-to-clip
• 6:46
  invocation
• 6:47
  —
• 6:47
  into
• 6:48
  a
• 6:48
  deployable
• 6:48
  .clip
• 6:49
  archive.
• 6:50
  The
• 6:50
  architecture
• 6:50
  uses
• 6:51
  a
• 6:51
  three-layer
• 6:51
  model:
• 6:52
  workspace
• 6:52
  (source
• 6:53
  code),
• 6:53
  package
• 6:54
  (immutable
• 6:54
  ZIP),
• 6:55
  and
• 6:55
  instance
• 6:55
  (runtime
• 6:56
  with
• 6:56
  mutable
• 6:56
  data
• 6:56
  directory).
• 6:58
  The
• 6:58
  agent
• 6:58
  supports
• 6:58
  topics
• 6:59
  (named
• 6:59
  conversation
• 7:00
  namespaces),
• 7:01
  runs
• 7:01
  (agentic
• 7:02
  loop
• 7:02
  cycles
• 7:02
  with
• 7:03
  multiple
• 7:03
  LLM
• 7:03
  calls
• 7:04
  and
• 7:04
  tool
• 7:04
  executions),
• 7:05
  and
• 7:05
  three-tier
• 7:06
  memory
• 7:06
  (persistent
• 7:07
  facts,
• 7:07
  LLM-generated
• 7:08
  summaries,
• 7:09
  and
• 7:09
  cosine
• 7:10
  similarity
• 7:10
  search
• 7:10
  over
• 7:11
  embeddings).
• 7:12
  Commands
• 7:12
  pipe
• 7:12
  together
• 7:13
  (cmd1
• 7:13
  |
• 7:14
  cmd2
• 7:15
  &&
• 7:16
  cmd3),
• 7:17
  browser
• 7:17
  screenshots
• 7:18
  auto-attach
• 7:18
  as
• 7:18
  vision
• 7:19
  content,
• 7:19
  and
• 7:20
  cross-clip
• 7:20
  file
• 7:20
  transfer
• 7:21
  enables
• 7:21
  multi-agent
• 7:22
  collaboration.
• 7:23
  At
• 7:23
  the
• 7:23
  application
• 7:24
  layer,
• 7:24
  DreamLab
• 7:25
  OctoBot
• 7:25
  demonstrates
• 7:26
  what
• 7:26
  a
• 7:26
  fully
• 7:26
  autonomous
• 7:27
  local
• 7:27
  agent
• 7:28
  looks
• 7:28
  like
• 7:28
  when
• 7:28
  pointed
• 7:29
  at
• 7:29
  creative
• 7:29
  work
• 7:29
  instead
• 7:30
  of
• 7:30
  code.
• 7:31
  Running
• 7:31
  entirely
• 7:31
  on
• 7:32
  Ollama
• 7:32
  (default
• 7:32
  model:
• 7:33
  gemma3:4b),
• 7:36
  OctoBot
• 7:36
  executes
• 7:37
  a
• 7:37
  continuous
• 7:37
  loop:
• 7:38
  pick
• 7:38
  an
• 7:38
  idea
• 7:39
  domain,
• 7:40
  invent
• 7:40
  something,
• 7:41
  save
• 7:41
  it
• 7:41
  to
• 7:41
  a
• 7:41
  markdown
• 7:42
  vault,
• 7:42
  fuse
• 7:43
  past
• 7:43
  ideas
• 7:43
  into
• 7:44
  hybrids,
• 7:44
  and
• 7:44
  follow
• 7:44
  idea
• 7:45
  chains
• 7:45
  deeper.
• 7:46
  Leave
• 7:47
  it
• 7:47
  running
• 7:47
  overnight
• 7:48
  and
• 7:48
  you
• 7:48
  wake
• 7:48
  up
• 7:48
  to
• 7:48
  hundreds
• 7:49
  of
• 7:49
  fully-written
• 7:49
  idea
• 7:49
  pitches.
• 7:51
  The
• 7:51
  pixel-art
• 7:51
  game
• 7:51
  interface
• 7:52
  tracks
• 7:52
  15
• 7:53
  environments,
• 7:53
  achievements,
• 7:54
  and
• 7:54
  inventor
• 7:55
  levels.
• 7:56
  The
• 7:56
  privacy
• 7:56
  pitch
• 7:56
  is
• 7:57
  real
• 7:57
  —
• 7:58
  everything
• 7:58
  stays
• 7:58
  local,
• 7:59
  no
• 7:59
  cloud,
• 8:00
  no
• 8:00
  API
• 8:00
  calls
• 8:01
  unless
• 8:01
  you
• 8:01
  explicitly
• 8:02
  configure
• 8:02
  an
• 8:02
  alternative
• 8:03
  backend.
• 8:04
  It
• 8:04
  is
• 8:04
  whimsical
• 8:05
  but
• 8:05
  structurally
• 8:05
  interesting
• 8:06
  as
• 8:06
  a
• 8:06
  reference
• 8:06
  implementation
• 8:07
  for
• 8:08
  autonomous
• 8:08
  creative
• 8:08
  agents
• 8:09
  with
• 8:09
  persistent
• 8:09
  memory
• 8:10
  and
• 8:10
  knowledge
• 8:10
  ingestion.
• 8:13
  Every
• 8:14
  AI
• 8:14
  coding
• 8:14
  session
• 8:15
  starts
• 8:15
  from
• 8:15
  zero.
• 8:16
  You
• 8:16
  re-explain
• 8:17
  your
• 8:17
  project
• 8:17
  structure,
• 8:18
  your
• 8:18
  naming
• 8:18
  conventions,
• 8:19
  your
• 8:19
  opinions
• 8:20
  about
• 8:20
  barrel
• 8:20
  exports.
• 8:21
  The
• 8:21
  model
• 8:21
  nods
• 8:21
  along,
• 8:22
  produces
• 8:23
  reasonable
• 8:23
  code,
• 8:23
  and
• 8:24
  forgets
• 8:24
  everything
• 8:24
  when
• 8:24
  the
• 8:24
  session
• 8:25
  ends.
• 8:26
  Most
• 8:26
  MCP
• 8:26
  memory
• 8:27
  servers
• 8:27
  solve
• 8:27
  basic
• 8:28
  session
• 8:28
  persistence.
• 8:29
  Pi,
• 8:30
  the
• 8:30
  collective
• 8:30
  intelligence
• 8:31
  system
• 8:31
  built
• 8:31
  on
• 8:31
  the
• 8:31
  RuVector
• 8:32
  platform,
• 8:32
  tries
• 8:32
  to
• 8:33
  solve
• 8:33
  a
• 8:33
  harder
• 8:33
  problem:
• 8:34
  what
• 8:34
  if
• 8:34
  knowledge
• 8:34
  compounded
• 8:35
  across
• 8:35
  sessions,
• 8:36
  across
• 8:36
  developers,
• 8:36
  across
• 8:37
  projects?
• 8:38
  Every
• 8:38
  piece
• 8:38
  of
• 8:38
  knowledge
• 8:38
  entering
• 8:39
  Pi
• 8:39
  gets
• 8:39
  packaged
• 8:40
  into
• 8:40
  a
• 8:40
  cognitive
• 8:40
  container
• 8:41
  using
• 8:41
  the
• 8:41
  RVF
• 8:41
  binary
• 8:42
  format,
• 8:42
  carrying
• 8:43
  vector
• 8:43
  embeddings,
• 8:43
  witness
• 8:44
  chains
• 8:44
  tracking
• 8:44
  who
• 8:44
  contributed
• 8:45
  what,
• 8:46
  Ed25519
• 8:47
  digital
• 8:47
  signatures,
• 8:48
  and
• 8:48
  SHAKE-256
• 8:50
  integrity
• 8:50
  hashes.
• 8:51
  The
• 8:51
  search
• 8:51
  layer
• 8:52
  runs
• 8:52
  HNSW
• 8:53
  indexing
• 8:53
  over
• 8:54
  a
• 8:54
  graph
• 8:54
  neural
• 8:54
  network
• 8:55
  where
• 8:55
  memories
• 8:55
  are
• 8:55
  nodes
• 8:56
  and
• 8:56
  similarities
• 8:56
  are
• 8:56
  edges,
• 8:57
  with
• 8:57
  MinCut-based
• 8:58
  self-organization
• 8:59
  so
• 8:59
  you
• 8:59
  never
• 8:59
  have
• 8:59
  to
• 8:59
  decide
• 9:00
  whether
• 9:00
  "JWT
• 9:01
  refresh
• 9:01
  token
• 9:01
  rotation"
• 9:02
  belongs
• 9:03
  under
• 9:03
  security
• 9:03
  or
• 9:03
  authentication.
• 9:05
  The
• 9:05
  security
• 9:05
  model
• 9:05
  is
• 9:05
  what
• 9:06
  distinguishes
• 9:06
  Pi
• 9:06
  from
• 9:07
  simpler
• 9:07
  alternatives.
• 9:08
  It
• 9:08
  assumes
• 9:08
  every
• 9:09
  input
• 9:09
  is
• 9:09
  adversarial,
• 9:10
  running
• 9:10
  seven
• 9:11
  defense
• 9:11
  layers:
• 9:12
  input
• 9:12
  sanitization
• 9:13
  with
• 9:13
  PII
• 9:13
  stripping,
• 9:14
  cryptographic
• 9:15
  verification
• 9:15
  via
• 9:15
  SHAKE-256
• 9:18
  and
• 9:19
  Ed25519,
• 9:20
  embedding
• 9:21
  space
• 9:21
  validation
• 9:21
  rejecting
• 9:22
  NaN/Inf/magnitude
• 9:23
  outliers,
• 9:24
  rate
• 9:24
  limiting
• 9:24
  with
• 9:25
  single-use
• 9:25
  nonces,
• 9:26
  2-sigma
• 9:26
  statistical
• 9:27
  exclusion
• 9:27
  for
• 9:27
  outlier
• 9:28
  contributions,
• 9:29
  reputation
• 9:29
  scoring
• 9:30
  weighted
• 9:30
  by
• 9:30
  accuracy-squared
• 9:31
  times
• 9:31
  uptime
• 9:31
  times
• 9:31
  stake,
• 9:32
  and
• 9:32
  anomaly
• 9:32
  detection
• 9:33
  via
• 9:33
  spiking
• 9:34
  neural
• 9:34
  networks.
• 9:35
  The
• 9:35
  Byzantine
• 9:35
  fault
• 9:36
  tolerance
• 9:36
  layer
• 9:37
  uses
• 9:37
  FedAvg
• 9:38
  with
• 9:38
  BFT
• 9:38
  for
• 9:38
  knowledge
• 9:39
  aggregation
• 9:39
  —
• 9:39
  no
• 9:40
  single
• 9:40
  contributor
• 9:40
  can
• 9:41
  shift
• 9:41
  what
• 9:41
  the
• 9:41
  system
• 9:41
  considers
• 9:42
  true.
• 9:43
  For
• 9:43
  a
• 9:43
  solo
• 9:43
  developer,
• 9:44
  this
• 9:44
  is
• 9:44
  overkill.
• 9:45
  For
• 9:45
  teams
• 9:46
  or
• 9:46
  open-source
• 9:46
  communities
• 9:47
  sharing
• 9:47
  patterns
• 9:47
  across
• 9:48
  projects,
• 9:48
  it
• 9:48
  is
• 9:49
  load-bearing
• 9:49
  infrastructure.
• 9:50
  On
• 9:50
  the
• 9:50
  forensic
• 9:51
  side
• 9:51
  of
• 9:51
  memory
• 9:51
  persistence,
• 9:52
  wechat-decrypt
• 9:53
  is
• 9:53
  a
• 9:53
  cross-platform
• 9:53
  tool
• 9:54
  that
• 9:54
  extracts
• 9:54
  encryption
• 9:55
  keys
• 9:55
  from
• 9:55
  running
• 9:55
  WeChat
• 9:56
  4.x
• 9:56
  processes
• 9:57
  and
• 9:57
  decrypts
• 9:58
  all
• 9:58
  local
• 9:59
  SQLCipher
• 9:59
  4
• 9:59
  databases.
• 10:00
  WeChat
• 10:01
  4.0
• 10:01
  uses
• 10:02
  AES-256-CBC
• 10:06
  with
• 10:06
  HMAC-SHA512,
• 10:11
  PBKDF2-HMAC-SHA512
• 10:17
  at
• 10:17
  256,000
• 10:18
  iterations,
• 10:19
  and
• 10:19
  independent
• 10:20
  salt
• 10:20
  and
• 10:20
  key
• 10:20
  per
• 10:21
  database.
• 10:22
  The
• 10:22
  tool
• 10:22
  scans
• 10:23
  process
• 10:23
  memory
• 10:23
  for
• 10:23
  cached
• 10:24
  raw
• 10:24
  keys
• 10:25
  (a
• 10:26
  64-hex
• 10:26
  encryption
• 10:27
  key
• 10:27
  plus
• 10:27
  32-hex
• 10:28
  salt
• 10:28
  pattern),
• 10:29
  validates
• 10:30
  them
• 10:30
  against
• 10:30
  page
• 10:30
  1
• 10:31
  HMAC,
• 10:31
  and
• 10:32
  decrypts
• 10:32
  everything
• 10:32
  —
• 10:33
  session
• 10:34
  lists,
• 10:34
  chat
• 10:35
  history,
• 10:35
  contacts,
• 10:36
  media
• 10:36
  indexes.
• 10:37
  It
• 10:38
  includes
• 10:38
  a
• 10:38
  real-time
• 10:38
  Web
• 10:39
  UI
• 10:39
  with
• 10:39
  30ms
• 10:40
  WAL
• 10:40
  polling,
• 10:41
  SSE
• 10:41
  push,
• 10:42
  and
• 10:42
  ~100ms
• 10:43
  total
• 10:43
  latency.
• 10:45
  An
• 10:45
  MCP
• 10:46
  server
• 10:46
  integration
• 10:47
  lets
• 10:47
  Claude
• 10:47
  Code
• 10:47
  query
• 10:48
  WeChat
• 10:48
  data
• 10:48
  directly.
• 10:50
  The
• 10:50
  tool
• 10:50
  works
• 10:50
  on
• 10:51
  Windows,
• 10:51
  Linux,
• 10:51
  and
• 10:51
  macOS
• 10:52
  (via
• 10:52
  Mach
• 10:53
  VM
• 10:53
  API),
• 10:54
  covering
• 10:54
  the
• 10:54
  platforms
• 10:55
  where
• 10:55
  forensic
• 10:55
  access
• 10:56
  to
• 10:56
  encrypted
• 10:56
  messaging
• 10:57
  databases
• 10:57
  matters
• 10:58
  most.
• 11:00
  Mistral
• 11:00
  released
• 11:01
  Mistral-Small-4-119B-2603-NVFP4,
• 11:07
  an
• 11:07
  official
• 11:08
  NVFP4-quantized
• 11:09
  version
• 11:10
  of
• 11:10
  their
• 11:10
  119-billion-parameter
• 11:11
  model.
• 11:12
  NVFP4
• 11:14
  is
• 11:14
  NVIDIA's
• 11:15
  4-bit
• 11:15
  floating
• 11:16
  point
• 11:16
  format
• 11:16
  —
• 11:17
  distinct
• 11:17
  from
• 11:17
  the
• 11:17
  GGUF
• 11:18
  and
• 11:18
  AWQ
• 11:19
  formats
• 11:19
  that
• 11:19
  have
• 11:19
  dominated
• 11:20
  local
• 11:20
  deployment.
• 11:21
  An
• 11:21
  official
• 11:22
  vendor-quantized
• 11:22
  release
• 11:23
  signals
• 11:23
  that
• 11:23
  hardware-specific
• 11:24
  quantization
• 11:25
  is
• 11:25
  becoming
• 11:25
  a
• 11:25
  distribution
• 11:26
  strategy,
• 11:27
  not
• 11:27
  just
• 11:27
  a
• 11:27
  community
• 11:28
  aftermarket.
• 11:29
  The
• 11:29
  model
• 11:29
  targets
• 11:30
  NVIDIA
• 11:30
  GPUs
• 11:31
  with
• 11:31
  native
• 11:31
  FP4
• 11:32
  support,
• 11:32
  which
• 11:33
  means
• 11:33
  Blackwell
• 11:34
  and
• 11:34
  newer
• 11:34
  architectures
• 11:35
  get
• 11:35
  optimized
• 11:35
  inference
• 11:36
  without
• 11:36
  the
• 11:36
  quality
• 11:37
  tradeoffs
• 11:37
  of
• 11:37
  post-hoc
• 11:38
  quantization.
• 11:39
  On
• 11:39
  the
• 11:39
  community
• 11:40
  side,
• 11:40
  Omnicoder-Claude-4.6-Opus-Uncensored-GGUF
• 11:43
  continues
• 11:44
  the
• 11:44
  GGUF
• 11:44
  format's
• 11:45
  dominance
• 11:45
  as
• 11:46
  the
• 11:46
  de
• 11:46
  facto
• 11:46
  standard
• 11:46
  for
• 11:47
  local
• 11:47
  model
• 11:47
  distribution.
• 11:48
  The
• 11:48
  hardware
• 11:49
  story
• 11:49
  extends
• 11:49
  beyond
• 11:50
  NVIDIA.
• 11:51
  AMD
• 11:51
  NPUs
• 11:52
  on
• 11:52
  Linux
• 11:52
  are
• 11:52
  now
• 11:52
  capable
• 11:53
  of
• 11:53
  running
• 11:53
  LLMs
• 11:54
  —
• 11:55
  a
• 11:55
  meaningful
• 11:55
  shift
• 11:55
  for
• 11:55
  an
• 11:56
  ecosystem
• 11:56
  where
• 11:56
  AMD
• 11:57
  has
• 11:57
  long
• 11:57
  been
• 11:57
  the
• 11:57
  underdog
• 11:58
  in
• 11:58
  AI
• 11:58
  inference.
• 11:59
  NPU
• 12:00
  (Neural
• 12:00
  Processing
• 12:01
  Unit)
• 12:01
  architectures
• 12:01
  execute
• 12:02
  transformer
• 12:02
  layers
• 12:03
  with
• 12:03
  dedicated
• 12:03
  hardware
• 12:04
  acceleration
• 12:05
  rather
• 12:05
  than
• 12:05
  repurposing
• 12:06
  general-purpose
• 12:06
  GPU
• 12:07
  compute,
• 12:08
  offering
• 12:08
  better
• 12:08
  power
• 12:08
  efficiency
• 12:09
  for
• 12:09
  sustained
• 12:09
  inference
• 12:10
  workloads.
• 12:11
  Linux
• 12:11
  support
• 12:12
  matters
• 12:12
  because
• 12:12
  it
• 12:12
  opens
• 12:13
  the
• 12:13
  door
• 12:13
  to
• 12:13
  server-side
• 12:14
  and
• 12:14
  embedded
• 12:14
  deployments
• 12:15
  where
• 12:15
  Windows
• 12:15
  is
• 12:16
  not
• 12:16
  an
• 12:16
  option.
• 12:17
  Meanwhile,
• 12:18
  Rick
• 12:18
  Beato
• 12:18
  —
• 12:19
  a
• 12:19
  YouTube
• 12:19
  creator
• 12:20
  with
• 12:20
  millions
• 12:20
  of
• 12:20
  subscribers
• 12:21
  primarily
• 12:21
  covering
• 12:22
  music
• 12:22
  production
• 12:22
  —
• 12:23
  drew
• 12:23
  a
• 12:24
  parallel
• 12:24
  between
• 12:24
  AI's
• 12:25
  trajectory
• 12:25
  and
• 12:25
  the
• 12:25
  music
• 12:26
  industry's
• 12:26
  digital
• 12:27
  disruption,
• 12:28
  arguing
• 12:28
  that
• 12:28
  local
• 12:28
  LLMs
• 12:29
  will
• 12:29
  win
• 12:29
  over
• 12:29
  centralized
• 12:30
  services
• 12:30
  for
• 12:31
  the
• 12:31
  same
• 12:31
  reasons
• 12:31
  independent
• 12:32
  music
• 12:32
  distribution
• 12:33
  eventually
• 12:33
  prevailed
• 12:34
  over
• 12:34
  label
• 12:34
  gatekeeping.
• 12:35
  The
• 12:36
  cultural
• 12:36
  signal
• 12:36
  matters:
• 12:37
  when
• 12:37
  mainstream
• 12:38
  creators
• 12:38
  outside
• 12:39
  the
• 12:39
  AI
• 12:39
  bubble
• 12:39
  start
• 12:39
  advocating
• 12:40
  for
• 12:40
  local
• 12:40
  inference,
• 12:41
  the
• 12:42
  technology
• 12:42
  is
• 12:42
  crossing
• 12:43
  an
• 12:43
  adoption
• 12:43
  threshold.
• 12:47
  Training
• 12:47
  large
• 12:48
  language
• 12:48
  models
• 12:48
  on
• 12:49
  long
• 12:49
  sequences
• 12:49
  has
• 12:49
  become
• 12:50
  essential,
• 12:51
  but
• 12:51
  the
• 12:51
  attention
• 12:51
  mechanism
• 12:52
  scales
• 12:52
  quadratically
• 12:53
  with
• 12:53
  sequence
• 12:54
  length
• 12:54
  —
• 12:54
  and
• 12:55
  even
• 12:55
  FlashAttention
• 12:56
  only
• 12:56
  reduces
• 12:56
  memory,
• 12:57
  not
• 12:57
  compute.
• 12:58
  Ulysses
• 12:59
  Sequence
• 12:59
  Parallelism,
• 13:01
  part
• 13:01
  of
• 13:01
  Snowflake
• 13:02
  AI
• 13:02
  Research's
• 13:03
  Arctic
• 13:03
  Long
• 13:03
  Sequence
• 13:04
  Training
• 13:04
  protocol,
• 13:05
  provides
• 13:05
  an
• 13:05
  elegant
• 13:06
  solution
• 13:06
  by
• 13:06
  distributing
• 13:07
  attention
• 13:07
  computation
• 13:08
  across
• 13:08
  multiple
• 13:09
  GPUs
• 13:09
  through
• 13:10
  attention
• 13:10
  head
• 13:10
  parallelism.
• 13:12
  The
• 13:12
  key
• 13:12
  insight:
• 13:13
  attention
• 13:13
  heads
• 13:14
  are
• 13:14
  independent.
• 13:15
  Ulysses
• 13:16
  splits
• 13:16
  input
• 13:17
  sequences
• 13:17
  along
• 13:17
  the
• 13:17
  sequence
• 13:18
  dimension,
• 13:18
  uses
• 13:19
  all-to-all
• 13:19
  collective
• 13:20
  operations
• 13:20
  to
• 13:20
  redistribute
• 13:21
  data
• 13:21
  so
• 13:22
  each
• 13:22
  GPU
• 13:22
  holds
• 13:23
  all
• 13:23
  sequence
• 13:23
  positions
• 13:24
  but
• 13:24
  only
• 13:24
  a
• 13:24
  subset
• 13:25
  of
• 13:25
  heads,
• 13:25
  computes
• 13:26
  attention
• 13:26
  locally,
• 13:27
  then
• 13:27
  reverses
• 13:27
  the
• 13:27
  redistribution.
• 13:29
  Communication
• 13:30
  volume
• 13:30
  scales
• 13:30
  as
• 13:30
  O(n
• 13:31
  d
• 13:31
  /
• 13:31
  P)
• 13:31
  per
• 13:31
  GPU
• 13:31
  —
• 13:32
  a
• 13:32
  factor
• 13:33
  of
• 13:33
  P
• 13:33
  less
• 13:33
  than
• 13:33
  Ring
• 13:34
  Attention's
• 13:34
  O(n
• 13:34
  d)
• 13:35
  via
• 13:35
  sequential
• 13:35
  point-to-point
• 13:36
  transfers.
• 13:37
  The
• 13:37
  practical
• 13:37
  results
• 13:38
  on
• 13:38
  H100
• 13:39
  80GB
• 13:39
  GPUs
• 13:40
  are
• 13:40
  compelling.
• 13:41
  With
• 13:41
  SP=4,
• 13:42
  per-GPU
• 13:43
  memory
• 13:43
  drops
• 13:43
  3.3x
• 13:44
  at
• 13:45
  the
• 13:45
  same
• 13:45
  sequence
• 13:45
  length,
• 13:46
  enabling
• 13:46
  training
• 13:47
  at
• 13:47
  up
• 13:47
  to
• 13:47
  96K
• 13:48
  tokens
• 13:48
  on
• 13:48
  4x
• 13:49
  H100s
• 13:49
  (12x
• 13:50
  longer
• 13:51
  than
• 13:51
  the
• 13:51
  data-parallel
• 13:52
  baseline
• 13:52
  of
• 13:52
  8K).
• 13:53
  At
• 13:53
  64K
• 13:54
  tokens,
• 13:55
  SP=4
• 13:56
  achieves
• 13:56
  13,396
• 13:58
  tokens
• 13:58
  per
• 13:58
  second
• 13:59
  —
• 14:00
  3.7x
• 14:00
  the
• 14:00
  baseline
• 14:01
  throughput.
• 14:01
  The
• 14:02
  all-to-all
• 14:02
  communication
• 14:03
  overhead
• 14:03
  is
• 14:03
  minimal
• 14:04
  on
• 14:04
  NVLink-connected
• 14:05
  GPUs,
• 14:06
  so
• 14:06
  the
• 14:06
  real
• 14:06
  benefit
• 14:06
  comes
• 14:07
  from
• 14:07
  processing
• 14:07
  proportionally
• 14:08
  more
• 14:08
  tokens
• 14:08
  per
• 14:09
  step
• 14:09
  as
• 14:09
  sequence
• 14:09
  length
• 14:10
  grows.
• 14:10
  Ulysses
• 14:11
  is
• 14:11
  now
• 14:11
  integrated
• 14:12
  across
• 14:12
  the
• 14:12
  Hugging
• 14:12
  Face
• 14:12
  ecosystem:
• 14:13
  Accelerate,
• 14:14
  Transformers
• 14:15
  Trainer,
• 14:16
  and
• 14:16
  TRL's
• 14:16
  SFTTrainer
• 14:17
  all
• 14:17
  handle
• 14:17
  sequence
• 14:18
  sharding,
• 14:18
  loss
• 14:19
  aggregation
• 14:19
  across
• 14:19
  ranks,
• 14:20
  and
• 14:20
  pre-shifted
• 14:21
  labels
• 14:21
  automatically.
• 14:22
  The
• 14:22
  constraint
• 14:22
  is
• 14:23
  that
• 14:23
  sequence
• 14:23
  length
• 14:23
  must
• 14:24
  be
• 14:24
  divisible
• 14:24
  by
• 14:24
  the
• 14:24
  SP
• 14:25
  degree,
• 14:25
  and
• 14:25
  the
• 14:25
  number
• 14:26
  of
• 14:26
  attention
• 14:26
  heads
• 14:26
  must
• 14:26
  be
• 14:26
  at
• 14:27
  least
• 14:27
  equal
• 14:27
  to
• 14:27
  the
• 14:27
  SP
• 14:28
  degree
• 14:28
  —
• 14:28
  which
• 14:29
  makes
• 14:29
  Ring
• 14:29
  Attention
• 14:29
  the
• 14:29
  fallback
• 14:30
  for
• 14:30
  models
• 14:30
  with
• 14:30
  few
• 14:31
  heads.
• 14:34
  Healthcare
• 14:34
  AI
• 14:35
  has
• 14:35
  been
• 14:35
  predominantly
• 14:36
  perception-based
• 14:36
  —
• 14:37
  models
• 14:38
  that
• 14:38
  interpret
• 14:38
  signals
• 14:38
  and
• 14:39
  classify
• 14:39
  pathology.
• 14:41
  But
• 14:41
  healthcare
• 14:41
  involves
• 14:42
  doing,
• 14:42
  and
• 14:43
  the
• 14:43
  static
• 14:43
  datasets
• 14:43
  of
• 14:44
  the
• 14:44
  past
• 14:44
  lack
• 14:44
  embodiment,
• 14:45
  contact
• 14:45
  dynamics,
• 14:46
  and
• 14:46
  closed-loop
• 14:47
  control.
• 14:48
  Open-H-Embodiment
• 14:49
  is
• 14:49
  a
• 14:49
  community-driven
• 14:50
  dataset
• 14:50
  initiative
• 14:50
  spanning
• 14:51
  35
• 14:51
  organizations
• 14:52
  (including
• 14:52
  Johns
• 14:53
  Hopkins,
• 14:53
  Stanford,
• 14:54
  NVIDIA,
• 14:55
  CMR
• 14:55
  Surgical,
• 14:56
  and
• 14:56
  UC
• 14:56
  Berkeley)
• 14:57
  that
• 14:57
  builds
• 14:57
  the
• 14:57
  first
• 14:58
  large-scale
• 14:58
  foundation
• 14:59
  for
• 14:59
  physical
• 14:59
  AI
• 14:59
  in
• 14:59
  healthcare
• 15:00
  robotics.
• 15:01
  The
• 15:01
  dataset
• 15:01
  comprises
• 15:02
  CC-BY-4.0
• 15:04
  training
• 15:04
  data
• 15:04
  across
• 15:04
  simulation,
• 15:05
  benchtop
• 15:05
  exercises
• 15:06
  like
• 15:06
  suturing,
• 15:07
  and
• 15:07
  real
• 15:07
  clinical
• 15:08
  procedures,
• 15:08
  using
• 15:09
  both
• 15:09
  commercial
• 15:09
  robots
• 15:09
  (CMR
• 15:10
  Surgical,
• 15:10
  Rob
• 15:11
  Surgical,
• 15:11
  Tuodao)
• 15:12
  and
• 15:12
  research
• 15:12
  platforms
• 15:13
  (dVRK,
• 15:13
  Franka,
• 15:14
  Kuka).
• 15:14
  Two
• 15:15
  models
• 15:15
  ship
• 15:15
  alongside
• 15:15
  the
• 15:16
  data.
• 15:16
  GR00T-H
• 15:18
  is
• 15:19
  a
• 15:19
  Vision-Language-Action
• 15:19
  model
• 15:20
  for
• 15:20
  surgical
• 15:20
  robotics,
• 15:21
  derived
• 15:22
  from
• 15:22
  NVIDIA's
• 15:22
  Isaac
• 15:23
  GR00T-N
• 15:24
  series
• 15:25
  and
• 15:25
  trained
• 15:26
  on
• 15:26
  roughly
• 15:26
  600
• 15:26
  hours
• 15:27
  of
• 15:27
  Open-H-Embodiment
• 15:28
  data.
• 15:29
  It
• 15:29
  uses
• 15:29
  a
• 15:29
  learnable
• 15:30
  MLP
• 15:30
  to
• 15:30
  map
• 15:31
  each
• 15:31
  robot's
• 15:31
  specific
• 15:32
  kinematics
• 15:32
  to
• 15:32
  a
• 15:32
  shared
• 15:33
  normalized
• 15:33
  action
• 15:34
  space,
• 15:34
  drops
• 15:35
  proprioceptive
• 15:36
  input
• 15:36
  during
• 15:36
  inference
• 15:37
  to
• 15:37
  create
• 15:37
  a
• 15:37
  learned
• 15:37
  bias
• 15:38
  term
• 15:38
  per
• 15:38
  system,
• 15:39
  and
• 15:39
  injects
• 15:39
  instrument
• 15:40
  names
• 15:40
  directly
• 15:40
  into
• 15:41
  the
• 15:41
  VLM
• 15:41
  task
• 15:42
  prompt.
• 15:42
  A
• 15:43
  prototype
• 15:43
  has
• 15:43
  demonstrated
• 15:44
  end-to-end
• 15:45
  suture
• 15:45
  execution.
• 15:47
  Cosmos-H-Surgical-Simulator
• 15:48
  is
• 15:48
  a
• 15:48
  World
• 15:49
  Foundation
• 15:49
  Model
• 15:49
  fine-tuned
• 15:50
  from
• 15:50
  NVIDIA
• 15:51
  Cosmos
• 15:51
  Predict
• 15:52
  2.5
• 15:52
  2B
• 15:53
  that
• 15:53
  generates
• 15:54
  physically
• 15:54
  plausible
• 15:55
  surgical
• 15:55
  video
• 15:55
  directly
• 15:56
  from
• 15:56
  kinematic
• 15:57
  actions
• 15:57
  —
• 15:58
  for
• 15:58
  600
• 15:59
  rollouts,
• 16:00
  simulation
• 16:00
  took
• 16:00
  14
• 16:01
  hours
• 16:01
  versus
• 16:02
  400
• 16:02
  hours
• 16:03
  on
• 16:03
  real
• 16:03
  benchtop
• 16:03
  hardware.
• 16:05
  In
• 16:05
  a
• 16:05
  different
• 16:05
  domain
• 16:06
  but
• 16:06
  with
• 16:06
  related
• 16:07
  sparse-data
• 16:07
  challenges,
• 16:08
  InstantSplat++
• 16:09
  extends
• 16:10
  Gaussian
• 16:10
  splatting
• 16:11
  to
• 16:11
  sparse-view
• 16:11
  large-scale
• 16:12
  scene
• 16:12
  reconstruction,
• 16:13
  supporting
• 16:14
  3D-GS,
• 16:15
  2D-GS,
• 16:16
  and
• 16:16
  Mip-Splatting
• 16:17
  from
• 16:17
  just
• 16:17
  a
• 16:17
  handful
• 16:18
  of
• 16:18
  input
• 16:18
  images
• 16:19
  rather
• 16:19
  than
• 16:19
  hundreds.
• 16:22
  The
• 16:22
  International
• 16:23
  AI
• 16:23
  Working
• 16:23
  Group
• 16:24
  (IAWG)
• 16:25
  takes
• 16:25
  a
• 16:25
  different
• 16:25
  approach
• 16:25
  to
• 16:26
  AI
• 16:26
  governance
• 16:26
  than
• 16:26
  the
• 16:27
  top-down
• 16:27
  frameworks
• 16:28
  dominating
• 16:28
  policy
• 16:28
  discussions.
• 16:30
  Built
• 16:30
  on
• 16:30
  the
• 16:30
  Matrix
• 16:30
  protocol
• 16:31
  —
• 16:31
  the
• 16:31
  same
• 16:31
  decentralized,
• 16:32
  end-to-end
• 16:33
  encrypted
• 16:33
  communication
• 16:34
  standard
• 16:34
  used
• 16:34
  by
• 16:35
  Element
• 16:35
  and
• 16:35
  others
• 16:35
  —
• 16:36
  IAWG
• 16:36
  creates
• 16:37
  a
• 16:37
  network
• 16:37
  where
• 16:37
  members
• 16:38
  register
• 16:38
  their
• 16:38
  agents,
• 16:38
  tools,
• 16:39
  and
• 16:39
  services
• 16:39
  in
• 16:40
  machine-readable
• 16:40
  manifests.
• 16:41
  Agents
• 16:41
  from
• 16:41
  different
• 16:42
  members
• 16:42
  communicate
• 16:42
  directly
• 16:43
  through
• 16:43
  custom
• 16:43
  Matrix
• 16:44
  events,
• 16:44
  delegating
• 16:45
  tasks
• 16:45
  and
• 16:46
  combining
• 16:46
  expertise
• 16:46
  without
• 16:47
  manual
• 16:47
  orchestration.
• 16:48
  The
• 16:48
  architecture
• 16:49
  has
• 16:49
  five
• 16:49
  layers:
• 16:50
  hierarchical
• 16:50
  organization
• 16:51
  via
• 16:51
  Matrix
• 16:52
  Spaces,
• 16:53
  standardized
• 16:53
  agent
• 16:53
  manifests,
• 16:54
  cross-room
• 16:55
  agent
• 16:55
  communication
• 16:56
  with
• 16:56
  coordinator
• 16:56
  orchestration,
• 16:58
  a
• 16:58
  searchable
• 16:58
  registry
• 16:59
  with
• 16:59
  sandbox
• 16:59
  testing,
• 17:00
  and
• 17:00
  cross-signing
• 17:01
  verification
• 17:01
  with
• 17:01
  OIDC
• 17:02
  for
• 17:02
  external
• 17:02
  authentication.
• 17:04
  "Tribes"
• 17:04
  —
• 17:06
  cross-member
• 17:06
  coalitions
• 17:07
  organized
• 17:07
  around
• 17:07
  societal
• 17:08
  challenges
• 17:08
  like
• 17:08
  climate,
• 17:09
  healthcare,
• 17:09
  and
• 17:09
  responsible
• 17:10
  AI
• 17:10
  —
• 17:11
  pool
• 17:11
  agents
• 17:11
  and
• 17:11
  expertise
• 17:12
  for
• 17:12
  larger
• 17:12
  goals.
• 17:13
  Whether
• 17:14
  governance
• 17:14
  mechanisms
• 17:14
  can
• 17:15
  keep
• 17:15
  pace
• 17:15
  with
• 17:15
  the
• 17:15
  displacement
• 17:16
  they
• 17:16
  are
• 17:16
  meant
• 17:16
  to
• 17:16
  manage
• 17:16
  is
• 17:16
  an
• 17:17
  open
• 17:17
  question.
• 17:18
  Anthropic
• 17:18
  CEO
• 17:19
  Dario
• 17:19
  Amodei
• 17:19
  predicted
• 17:20
  that
• 17:20
  50%
• 17:20
  of
• 17:21
  entry-level
• 17:21
  white-collar
• 17:21
  jobs
• 17:22
  will
• 17:22
  be
• 17:22
  "eradicated"
• 17:23
  within
• 17:23
  three
• 17:23
  years.
• 17:24
  The
• 17:24
  Reddit
• 17:24
  discussion
• 17:25
  captured
• 17:25
  the
• 17:25
  range
• 17:25
  of
• 17:25
  practitioner
• 17:26
  responses:
• 17:27
  one
• 17:28
  commenter
• 17:28
  noted
• 17:28
  their
• 17:28
  CEO
• 17:29
  wants
• 17:29
  to
• 17:29
  eliminate
• 17:30
  "the
• 17:30
  shittiest
• 17:30
  parts
• 17:31
  of
• 17:31
  people's
• 17:31
  jobs"
• 17:32
  so
• 17:32
  merchandisers
• 17:33
  can
• 17:33
  explore
• 17:33
  new
• 17:33
  products
• 17:34
  instead
• 17:34
  of
• 17:34
  typing
• 17:35
  into
• 17:35
  ERPs;
• 17:36
  another
• 17:36
  pointed
• 17:37
  out
• 17:37
  the
• 17:37
  demand-side
• 17:37
  problem
• 17:38
  ("who's
• 17:38
  going
• 17:38
  to
• 17:38
  buy
• 17:38
  shit
• 17:39
  when
• 17:39
  you
• 17:39
  wipe
• 17:39
  out
• 17:39
  50%
• 17:40
  of
• 17:40
  the
• 17:40
  middle
• 17:40
  class");
• 17:41
  several
• 17:41
  dismissed
• 17:42
  it
• 17:42
  as
• 17:42
  fundraising-motivated
• 17:43
  doom
• 17:43
  ("a
• 17:43
  new
• 17:44
  day,
• 17:44
  a
• 17:44
  new
• 17:44
  number
• 17:45
  pulled
• 17:45
  out
• 17:45
  of
• 17:45
  someone's
• 17:45
  ass").
• 17:46
  One
• 17:46
  developer
• 17:47
  using
• 17:47
  AI
• 17:47
  daily
• 17:48
  for
• 17:48
  vibe-coding
• 17:48
  noted
• 17:49
  that
• 17:49
  despite
• 17:49
  increased
• 17:50
  individual
• 17:50
  productivity,
• 17:51
  their
• 17:51
  team
• 17:51
  size
• 17:52
  has
• 17:52
  not
• 17:52
  changed.
• 17:53
  The
• 17:53
  pattern
• 17:54
  is
• 17:54
  familiar
• 17:54
  —
• 17:55
  AI
• 17:55
  leaders
• 17:55
  making
• 17:56
  bold
• 17:56
  displacement
• 17:57
  predictions
• 17:57
  while
• 17:57
  practitioners
• 17:58
  report
• 17:58
  augmentation
• 17:59
  rather
• 18:00
  than
• 18:00
  replacement
• 18:00
  —
• 18:01
  but
• 18:01
  the
• 18:01
  specificity
• 18:02
  of
• 18:02
  "50%
• 18:03
  in
• 18:03
  3
• 18:03
  years"
• 18:04
  for
• 18:04
  entry-level
• 18:04
  roles
• 18:05
  adds
• 18:05
  urgency
• 18:05
  to
• 18:05
  a
• 18:05
  conversation
• 18:06
  the
• 18:06
  industry
• 18:07
  has
• 18:07
  been
• 18:07
  having
• 18:07
  in
• 18:07
  vaguer
• 18:08
  terms.