Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – February 28, 2026

Transcript

AGI Dreams Podcast – February 28, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  For
• 0:00
  anyone
• 0:00
  joining
• 0:00
  mid-escalation:
• 0:01
  Anthropic
• 0:02
  signed
• 0:02
  a
• 0:02
  Pentagon
• 0:02
  contract
• 0:03
  last
• 0:03
  summer
• 0:03
  that
• 0:04
  required
• 0:04
  military
• 0:04
  adherence
• 0:05
  to
• 0:05
  its
• 0:05
  usage
• 0:05
  policy.
• 0:06
  When
• 0:06
  the
• 0:07
  Pentagon
• 0:07
  demanded
• 0:07
  unrestricted
• 0:08
  access
• 0:08
  for
• 0:08
  "all
• 0:09
  lawful
• 0:09
  purposes,"
• 0:10
  Anthropic
• 0:10
  asked
• 0:10
  for
• 0:10
  two
• 0:11
  specific
• 0:11
  guarantees
• 0:11
  —
• 0:12
  no
• 0:12
  mass
• 0:13
  surveillance
• 0:13
  of
• 0:13
  American
• 0:13
  citizens,
• 0:14
  and
• 0:14
  no
• 0:14
  autonomous
• 0:15
  weapons
• 0:15
  without
• 0:15
  human
• 0:16
  oversight.
• 0:16
  The
• 0:16
  Pentagon
• 0:17
  refused.
• 0:17
  Hegseth
• 0:18
  then
• 0:18
  threatened
• 0:19
  to
• 0:19
  invoke
• 0:19
  the
• 0:19
  Defense
• 0:19
  Production
• 0:20
  Act
• 0:20
  or
• 0:20
  designate
• 0:21
  Anthropic
• 0:21
  a
• 0:21
  "supply
• 0:22
  chain
• 0:22
  risk,"
• 0:23
  a
• 0:23
  classification
• 0:23
  previously
• 0:24
  reserved
• 0:24
  for
• 0:24
  foreign
• 0:24
  adversaries
• 0:25
  like
• 0:25
  Huawei.
• 0:26
  After
• 0:26
  Defense
• 0:27
  Secretary
• 0:27
  Hegseth's
• 0:27
  Friday
• 0:28
  deadline
• 0:28
  for
• 0:28
  Anthropic
• 0:28
  to
• 0:28
  drop
• 0:29
  safety
• 0:29
  guardrails
• 0:29
  passed
• 0:30
  without
• 0:30
  capitulation,
• 0:31
  Trump
• 0:31
  took
• 0:31
  to
• 0:31
  Truth
• 0:31
  Social
• 0:32
  with
• 0:32
  an
• 0:32
  executive
• 0:32
  broadside:
• 0:33
  federal
• 0:33
  agencies
• 0:34
  are
• 0:34
  to
• 0:34
  "cease
• 0:34
  usage"
• 0:35
  of
• 0:35
  Anthropic
• 0:35
  products,
• 0:36
  with
• 0:36
  the
• 0:36
  company
• 0:36
  labeled
• 0:37
  a
• 0:37
  "Radical
• 0:37
  Left
• 0:38
  AI
• 0:38
  Company"
• 0:38
  that
• 0:38
  is
• 0:38
  "putting
• 0:39
  American
• 0:39
  lives
• 0:39
  in
• 0:39
  danger"
• 0:40
  by
• 0:40
  enforcing
• 0:41
  its
• 0:41
  terms
• 0:41
  of
• 0:41
  service
• 0:42
  on
• 0:42
  military
• 0:42
  users.
• 0:43
  "It
• 0:43
  is
• 0:43
  curious
• 0:43
  that
• 0:43
  physical
• 0:44
  courage
• 0:44
  should
• 0:44
  be
• 0:44
  so
• 0:44
  common
• 0:45
  in
• 0:45
  the
• 0:45
  world
• 0:45
  and
• 0:45
  moral
• 0:46
  courage
• 0:46
  so
• 0:46
  rare."
• 0:51
  Rich
• 0:51
  Mogull's
• 0:52
  "Core
• 0:52
  Collapse"
• 0:52
  essay
• 0:53
  for
• 0:53
  the
• 0:53
  Cloud
• 0:53
  Security
• 0:54
  Alliance
• 0:54
  may
• 0:54
  be
• 0:55
  the
• 0:55
  most
• 0:55
  structurally
• 0:56
  clear
• 0:56
  articulation
• 0:57
  of
• 0:57
  why
• 0:58
  AI
• 0:58
  tilts
• 0:58
  the
• 0:59
  cybersecurity
• 0:59
  balance
• 1:00
  toward
• 1:00
  attackers
• 1:00
  —
• 1:01
  and
• 1:01
  what
• 1:01
  defenders
• 1:01
  should
• 1:02
  actually
• 1:02
  do
• 1:02
  about
• 1:02
  it.
• 1:03
  The
• 1:03
  metaphor
• 1:04
  is
• 1:04
  astrophysical:
• 1:05
  a
• 1:05
  star
• 1:05
  burns
• 1:06
  through
• 1:06
  its
• 1:06
  elements
• 1:06
  until
• 1:07
  it
• 1:07
  reaches
• 1:07
  iron,
• 1:08
  which
• 1:08
  absorbs
• 1:09
  energy
• 1:09
  rather
• 1:09
  than
• 1:09
  releasing
• 1:10
  it,
• 1:10
  and
• 1:11
  the
• 1:11
  core
• 1:11
  collapses.
• 1:12
  Security's
• 1:13
  "iron"
• 1:13
  is
• 1:13
  the
• 1:14
  response-cycle
• 1:14
  model
• 1:15
  that
• 1:15
  has
• 1:15
  served
• 1:15
  the
• 1:15
  industry
• 1:16
  for
• 1:16
  a
• 1:16
  decade.
• 1:17
  Attackers
• 1:17
  control
• 1:18
  the
• 1:18
  bounding
• 1:18
  of
• 1:18
  their
• 1:18
  problem
• 1:19
  space
• 1:19
  —
• 1:19
  find
• 1:20
  one
• 1:20
  vulnerability,
• 1:21
  exploit
• 1:21
  one
• 1:21
  path.
• 1:22
  AI
• 1:23
  makes
• 1:23
  that
• 1:23
  search
• 1:24
  faster,
• 1:24
  cheaper,
• 1:24
  and
• 1:25
  parallelizable.
• 1:26
  Defenders
• 1:27
  face
• 1:27
  a
• 1:27
  nearly
• 1:27
  unbounded
• 1:28
  problem:
• 1:29
  protect
• 1:29
  all
• 1:29
  resources
• 1:30
  from
• 1:30
  all
• 1:30
  attackers
• 1:31
  at
• 1:31
  all
• 1:31
  times.
• 1:32
  Even
• 1:32
  when
• 1:32
  an
• 1:32
  AI-assisted
• 1:33
  defender
• 1:33
  discovers
• 1:34
  a
• 1:34
  vulnerability
• 1:35
  simultaneously
• 1:36
  with
• 1:36
  an
• 1:36
  attacker,
• 1:37
  the
• 1:37
  defender
• 1:37
  still
• 1:37
  needs
• 1:38
  to
• 1:38
  develop
• 1:38
  a
• 1:38
  patch,
• 1:39
  test
• 1:39
  it,
• 1:39
  stage
• 1:40
  it,
• 1:40
  deploy
• 1:40
  it
• 1:40
  across
• 1:41
  every
• 1:41
  affected
• 1:41
  system,
• 1:42
  and
• 1:42
  keep
• 1:42
  production
• 1:42
  running.
• 1:43
  The
• 1:43
  attacker's
• 1:44
  cycle
• 1:44
  is
• 1:44
  destructive
• 1:45
  —
• 1:45
  they
• 1:46
  carry
• 1:46
  no
• 1:46
  technical
• 1:47
  debt.
• 1:48
  A
• 1:48
  new
• 1:48
  preprint
• 1:49
  from
• 1:49
  Zhuo,
• 1:49
  Ding,
• 1:50
  Guo,
• 1:51
  and
• 1:51
  Meng
• 1:52
  at
• 1:52
  Monash,
• 1:53
  UCLA,
• 1:53
  UCSB,
• 1:54
  and
• 1:55
  NUS
• 1:55
  sharpens
• 1:55
  the
• 1:56
  policy
• 1:56
  dimension.
• 1:57
  Their
• 1:57
  argument:
• 1:58
  AI
• 1:59
  agent-driven
• 1:59
  cyber
• 1:59
  attacks
• 2:00
  are
• 2:00
  not
• 2:00
  speculative
• 2:01
  but
• 2:01
  an
• 2:01
  economic
• 2:01
  inevitability.
• 2:03
  Current
• 2:03
  agents
• 2:04
  already
• 2:04
  achieve
• 2:04
  54.5%
• 2:06
  on
• 2:06
  penetration
• 2:06
  testing
• 2:07
  benchmarks
• 2:07
  and
• 2:08
  12.5%
• 2:09
  on
• 2:09
  real-world
• 2:09
  CVE
• 2:10
  exploitation
• 2:10
  —
• 2:12
  numbers
• 2:12
  that,
• 2:13
  given
• 2:13
  attackers'
• 2:13
  favorable
• 2:14
  economics
• 2:14
  (low
• 2:14
  marginal
• 2:15
  cost,
• 2:15
  tolerance
• 2:16
  for
• 2:16
  high
• 2:16
  failure
• 2:17
  rates),
• 2:17
  are
• 2:17
  already
• 2:18
  operationally
• 2:18
  significant.
• 2:20
  The
• 2:20
  proposed
• 2:20
  response
• 2:21
  is
• 2:21
  counterintuitive:
• 2:22
  train
• 2:23
  offensive
• 2:23
  AI
• 2:23
  agents
• 2:24
  deliberately,
• 2:25
  confine
• 2:25
  them
• 2:25
  to
• 2:25
  audited
• 2:26
  cyber
• 2:26
  ranges,
• 2:27
  and
• 2:27
  distill
• 2:27
  their
• 2:27
  findings
• 2:28
  into
• 2:28
  defensive-only
• 2:29
  agents.
• 2:30
  The
• 2:30
  paper
• 2:30
  frames
• 2:31
  this
• 2:31
  not
• 2:31
  as
• 2:31
  irresponsibility
• 2:32
  but
• 2:33
  as
• 2:33
  essential
• 2:33
  defensive
• 2:34
  infrastructure
• 2:34
  —
• 2:35
  "containing
• 2:36
  cybersecurity
• 2:36
  risks
• 2:37
  requires
• 2:37
  mastering
• 2:38
  them
• 2:38
  in
• 2:38
  controlled
• 2:39
  settings
• 2:39
  before
• 2:39
  adversaries
• 2:40
  do."
• 2:41
  On
• 2:41
  the
• 2:41
  supply-chain
• 2:42
  front,
• 2:42
  SafePickle
• 2:43
  addresses
• 2:43
  a
• 2:44
  vector
• 2:44
  that
• 2:44
  has
• 2:44
  received
• 2:45
  surprisingly
• 2:45
  little
• 2:45
  attention:
• 2:47
  malicious
• 2:47
  Python
• 2:48
  Pickle
• 2:48
  files
• 2:48
  distributed
• 2:49
  through
• 2:49
  model
• 2:49
  repositories
• 2:50
  like
• 2:50
  Hugging
• 2:50
  Face,
• 2:51
  where
• 2:51
  95%
• 2:52
  of
• 2:52
  observed
• 2:53
  malicious
• 2:53
  models
• 2:54
  use
• 2:54
  this
• 2:54
  serialization
• 2:55
  format.
• 2:56
  The
• 2:56
  approach
• 2:56
  is
• 2:56
  elegantly
• 2:57
  simple
• 2:57
  —
• 2:58
  extract
• 2:58
  opcode
• 2:59
  frequency
• 2:59
  distributions
• 3:00
  from
• 3:00
  Pickle
• 3:00
  bytecode,
• 3:01
  then
• 3:01
  classify
• 3:02
  using
• 3:02
  supervised
• 3:03
  and
• 3:03
  unsupervised
• 3:04
  ML
• 3:04
  models.
• 3:05
  No
• 3:05
  code
• 3:05
  execution,
• 3:06
  no
• 3:06
  per-library
• 3:07
  policies,
• 3:08
  no
• 3:08
  instrumentation.
• 3:09
  On
• 3:09
  a
• 3:09
  curated
• 3:10
  dataset
• 3:10
  of
• 3:10
  727
• 3:12
  Hugging
• 3:12
  Face
• 3:12
  files,
• 3:13
  SafePickle's
• 3:13
  best
• 3:14
  models
• 3:14
  dramatically
• 3:15
  outperformed
• 3:15
  every
• 3:16
  existing
• 3:16
  scanner
• 3:16
  (Modelscan,
• 3:17
  Fickling,
• 3:18
  ClamAV,
• 3:18
  VirusTotal
• 3:19
  achieved
• 3:19
  7–63%
• 3:20
  F1).
• 3:22
  On
• 3:22
  PickleBall's
• 3:23
  out-of-distribution
• 3:24
  dataset,
• 3:24
  it
• 3:24
  surpassed
• 3:25
  PickleBall
• 3:25
  itself.
• 3:26
  Most
• 3:27
  impressively,
• 3:27
  it
• 3:27
  was
• 3:28
  the
• 3:28
  only
• 3:28
  method
• 3:28
  to
• 3:28
  correctly
• 3:29
  classify
• 3:29
  all
• 3:30
  9
• 3:30
  Hide-and-Seek
• 3:30
  adversarial
• 3:31
  samples
• 3:31
  —
• 3:32
  polyglot,
• 3:33
  compressed,
• 3:33
  and
• 3:34
  multi-format
• 3:34
  payloads
• 3:35
  specifically
• 3:35
  engineered
• 3:36
  to
• 3:36
  evade
• 3:36
  detection.
• 3:37
  Runtime:
• 3:38
  sub-millisecond
• 3:39
  per
• 3:39
  file.
• 3:42
  A
• 3:42
  Substack
• 3:42
  post
• 3:43
  from
• 3:43
  the
• 3:43
  security
• 3:43
  tooling
• 3:44
  team
• 3:44
  at
• 3:44
  Lasso
• 3:44
  (who,
• 3:45
  ironically,
• 3:46
  build
• 3:46
  tools
• 3:46
  to
• 3:46
  make
• 3:46
  agents
• 3:47
  secure
• 3:47
  by
• 3:47
  design)
• 3:48
  documents
• 3:48
  one
• 3:48
  of
• 3:49
  the
• 3:49
  more
• 3:49
  instructive
• 3:49
  agent
• 3:50
  compromises
• 3:50
  in
• 3:50
  recent
• 3:51
  memory.
• 3:51
  While
• 3:52
  benchmarking
• 3:52
  an
• 3:52
  AI
• 3:53
  coding
• 3:53
  agent
• 3:53
  on
• 3:53
  a
• 3:53
  benign
• 3:54
  task
• 3:54
  —
• 3:55
  "search
• 3:55
  for
• 3:55
  Python
• 3:56
  async
• 3:56
  best
• 3:56
  practices,
• 3:57
  then
• 3:57
  create
• 3:58
  a
• 3:58
  GitHub
• 3:58
  issue
• 3:58
  summarizing
• 3:59
  the
• 3:59
  findings"
• 3:59
  —
• 4:00
  the
• 4:00
  agent
• 4:01
  generated
• 4:01
  a
• 4:01
  shell
• 4:01
  command
• 4:02
  containing
• 4:02
  backtick-delimited
• 4:03
  text.
• 4:03
  Bash
• 4:04
  interpreted
• 4:04
  the
• 4:04
  backticks
• 4:04
  as
• 4:05
  command
• 4:05
  substitution,
• 4:06
  executed
• 4:06
  the
• 4:06
  `set`
• 4:07
  builtin,
• 4:07
  and
• 4:07
  spliced
• 4:08
  every
• 4:08
  environment
• 4:08
  variable
• 4:09
  —
• 4:09
  API
• 4:10
  keys,
• 4:10
  Telegram
• 4:10
  secrets,
• 4:11
  auth
• 4:11
  tokens
• 4:11
  —
• 4:12
  into
• 4:12
  a
• 4:12
  public
• 4:13
  GitHub
• 4:13
  issue.
• 4:14
  An
• 4:14
  attacker
• 4:14
  from
• 4:14
  an
• 4:14
  IP
• 4:15
  geolocated
• 4:15
  to
• 4:15
  India
• 4:16
  SSH'd
• 4:16
  into
• 4:16
  the
• 4:16
  machine
• 4:17
  minutes
• 4:17
  later.
• 4:17
  The
• 4:18
  team's
• 4:18
  conclusion
• 4:18
  was
• 4:18
  disarmingly
• 4:19
  honest:
• 4:20
  "Is
• 4:20
  this
• 4:20
  our
• 4:20
  responsibility,
• 4:21
  or
• 4:21
  an
• 4:21
  error
• 4:21
  in
• 4:21
  the
• 4:21
  AI
• 4:22
  model,
• 4:22
  or
• 4:22
  a
• 4:22
  bug
• 4:22
  of
• 4:22
  the
• 4:22
  bot?
• 4:23
  We
• 4:23
  honestly
• 4:24
  do
• 4:24
  not
• 4:24
  know.
• 4:24
  AI
• 4:25
  security
• 4:25
  is
• 4:25
  now
• 4:25
  a
• 4:26
  long-tail
• 4:26
  problem;
• 4:27
  there
• 4:27
  are
• 4:27
  simply
• 4:27
  so
• 4:27
  many
• 4:28
  different
• 4:28
  failure
• 4:28
  modes."
• 4:29
  Scale
• 4:29
  that
• 4:29
  lesson
• 4:29
  across
• 4:30
  the
• 4:30
  internet:
• 4:31
  researchers
• 4:32
  flagged
• 4:32
  over
• 4:32
  40,000
• 4:33
  AI
• 4:33
  agent
• 4:34
  instances
• 4:34
  exposed
• 4:35
  to
• 4:35
  the
• 4:35
  public
• 4:35
  internet
• 4:36
  with
• 4:36
  full
• 4:36
  system
• 4:36
  access,
• 4:38
  and
• 4:38
  community
• 4:38
  reports
• 4:39
  suggest
• 4:39
  the
• 4:39
  real
• 4:40
  number
• 4:40
  on
• 4:40
  some
• 4:40
  platforms
• 4:41
  may
• 4:41
  exceed
• 4:41
  two
• 4:41
  million.
• 4:43
  The
• 4:43
  problem
• 4:43
  is
• 4:43
  not
• 4:44
  exotic
• 4:44
  —
• 4:45
  it
• 4:45
  is
• 4:45
  default
• 4:45
  configurations
• 4:46
  bound
• 4:46
  to
• 4:47
  0.0.0.0
• 4:48
  on
• 4:48
  VPS
• 4:49
  instances
• 4:49
  by
• 4:50
  users
• 4:50
  who
• 4:50
  "do
• 4:50
  not
• 4:51
  know
• 4:51
  how
• 4:51
  to
• 4:51
  secure
• 4:51
  a
• 4:51
  box
• 4:52
  or
• 4:52
  use
• 4:52
  a
• 4:52
  command
• 4:52
  line."
• 4:53
  As
• 4:53
  one
• 4:54
  commenter
• 4:54
  noted,
• 4:55
  these
• 4:55
  are
• 4:55
  not
• 4:55
  just
• 4:55
  gaping
• 4:56
  holes;
• 4:56
  they
• 4:57
  are
• 4:57
  "really
• 4:57
  capable"
• 4:57
  botnets
• 4:58
  waiting
• 4:58
  to
• 4:58
  happen.
• 4:59
  The
• 4:59
  audit
• 5:00
  layer
• 5:00
  that
• 5:00
  is
• 5:00
  supposed
• 5:01
  to
• 5:01
  catch
• 5:01
  these
• 5:01
  problems
• 5:02
  has
• 5:02
  its
• 5:02
  own
• 5:02
  reliability
• 5:03
  issue.
• 5:04
  Kevin
• 5:04
  Joensen
• 5:04
  ran
• 5:05
  50
• 5:05
  iterations
• 5:06
  of
• 5:06
  Claude
• 5:06
  Code's
• 5:06
  security-review
• 5:07
  command
• 5:07
  against
• 5:08
  a
• 5:08
  byte-identical
• 5:08
  vulnerable
• 5:09
  application.
• 5:10
  The
• 5:10
  results
• 5:10
  varied
• 5:11
  significantly
• 5:11
  across
• 5:12
  runs
• 5:12
  —
• 5:13
  sometimes
• 5:13
  all
• 5:13
  bugs
• 5:14
  were
• 5:14
  identified,
• 5:15
  sometimes
• 5:15
  critical
• 5:16
  vulnerabilities
• 5:16
  were
• 5:16
  missed
• 5:17
  entirely.
• 5:18
  XSS
• 5:18
  findings
• 5:19
  fluctuated
• 5:19
  the
• 5:19
  most.
• 5:20
  The
• 5:21
  takeaway
• 5:21
  is
• 5:21
  not
• 5:21
  that
• 5:21
  AI-driven
• 5:22
  audits
• 5:22
  are
• 5:22
  useless
• 5:23
  but
• 5:23
  that
• 5:24
  they
• 5:24
  are
• 5:24
  non-deterministic
• 5:25
  in
• 5:25
  a
• 5:25
  domain
• 5:25
  that
• 5:25
  has
• 5:26
  historically
• 5:26
  depended
• 5:26
  on
• 5:27
  determinism.
• 5:28
  A
• 5:28
  traditional
• 5:29
  scanner
• 5:29
  reports
• 5:29
  an
• 5:29
  open
• 5:30
  port
• 5:30
  every
• 5:30
  time.
• 5:31
  An
• 5:32
  LLM-based
• 5:32
  scanner
• 5:33
  might
• 5:33
  miss
• 5:33
  it
• 5:33
  depending
• 5:34
  on
• 5:34
  token
• 5:34
  sampling.
• 5:35
  The
• 5:35
  practical
• 5:35
  recommendation
• 5:36
  from
• 5:36
  the
• 5:36
  discussion:
• 5:37
  use
• 5:38
  deterministic
• 5:38
  tools
• 5:39
  to
• 5:39
  build
• 5:39
  a
• 5:39
  reliable
• 5:40
  context
• 5:40
  base,
• 5:41
  then
• 5:41
  let
• 5:41
  the
• 5:41
  LLM
• 5:42
  reason
• 5:42
  over
• 5:42
  that
• 5:42
  context
• 5:43
  rather
• 5:43
  than
• 5:43
  discover
• 5:44
  from
• 5:44
  scratch.
• 5:49
  A
• 5:49
  research
• 5:50
  team
• 5:50
  running
• 5:51
  56,000
• 5:52
  multi-agent
• 5:52
  simulations
• 5:53
  through
• 5:53
  their
• 5:53
  open-source
• 5:54
  CogniArch
• 5:54
  framework
• 5:54
  delivered
• 5:55
  an
• 5:55
  empirically
• 5:55
  devastating
• 5:56
  finding:
• 5:57
  a
• 5:57
  single
• 5:57
  aggressive
• 5:58
  agent
• 5:58
  among
• 5:58
  five
• 5:59
  cooperative
• 5:59
  ones
• 5:59
  reduces
• 6:00
  cooperation
• 6:00
  by
• 6:01
  roughly
• 6:01
  50%.
• 6:02
  The
• 6:02
  "just
• 6:03
  outnumber
• 6:03
  the
• 6:03
  bad
• 6:03
  agents"
• 6:04
  hypothesis
• 6:04
  does
• 6:05
  not
• 6:05
  hold.
• 6:06
  Cooperation
• 6:06
  collapse
• 6:06
  follows
• 6:07
  a
• 6:07
  clean
• 6:07
  dose-response
• 6:08
  curve
• 6:08
  from
• 6:08
  zero
• 6:08
  to
• 6:08
  four
• 6:09
  aggressors,
• 6:10
  cognitive
• 6:10
  architecture
• 6:10
  provides
• 6:11
  measurable
• 6:11
  resilience
• 6:12
  (dual-process
• 6:12
  agents
• 6:13
  resist
• 6:13
  value
• 6:13
  drift
• 6:14
  better
• 6:14
  than
• 6:14
  reactive
• 6:14
  ones),
• 6:15
  and
• 6:15
  all
• 6:15
  architectures
• 6:16
  perform
• 6:16
  identically
• 6:17
  in
• 6:17
  solo
• 6:17
  survival
• 6:17
  —
• 6:18
  differences
• 6:19
  only
• 6:19
  emerge
• 6:19
  under
• 6:19
  social
• 6:20
  pressure.
• 6:21
  Effect
• 6:21
  sizes
• 6:21
  are
• 6:21
  massive
• 6:22
  (Cohen's
• 6:22
  d
• 6:22
  =
• 6:24
  6.71–24.99),
• 6:26
  making
• 6:26
  this
• 6:26
  one
• 6:26
  of
• 6:26
  the
• 6:26
  clearest
• 6:27
  quantitative
• 6:27
  demonstrations
• 6:28
  that
• 6:28
  multi-agent
• 6:28
  alignment
• 6:29
  is
• 6:29
  not
• 6:29
  a
• 6:29
  nice-to-have
• 6:30
  but
• 6:30
  a
• 6:30
  structural
• 6:31
  requirement.
• 6:32
  Agentic-Drift
• 6:33
  operationalizes
• 6:33
  the
• 6:33
  monitoring
• 6:34
  side
• 6:34
  of
• 6:34
  this
• 6:34
  problem
• 6:35
  for
• 6:35
  production
• 6:35
  ML
• 6:35
  systems.
• 6:37
  The
• 6:37
  open-source
• 6:37
  platform
• 6:38
  runs
• 6:38
  four
• 6:38
  statistical
• 6:39
  drift
• 6:39
  detection
• 6:39
  methods
• 6:40
  in
• 6:40
  parallel
• 6:40
  (PSI,
• 6:41
  KS
• 6:41
  test,
• 6:42
  Jensen-Shannon
• 6:42
  divergence,
• 6:43
  and
• 6:43
  basic
• 6:44
  statistical
• 6:44
  checks)
• 6:45
  in
• 6:45
  under
• 6:45
  20
• 6:45
  milliseconds,
• 6:46
  then
• 6:46
  coordinates
• 6:47
  a
• 6:47
  four-agent
• 6:47
  response
• 6:47
  system
• 6:48
  —
• 6:49
  analyzer,
• 6:49
  recommender,
• 6:50
  executor,
• 6:51
  monitor
• 6:51
  —
• 6:52
  that
• 6:52
  learns
• 6:52
  from
• 6:52
  successful
• 6:53
  interventions
• 6:53
  via
• 6:53
  an
• 6:53
  AgentDB-backed
• 6:54
  skill
• 6:54
  library.
• 6:55
  For
• 6:56
  financial
• 6:56
  services,
• 6:56
  it
• 6:57
  ships
• 6:57
  Basel
• 6:57
  II/III-compliant
• 6:58
  PSI
• 6:59
  thresholds
• 6:59
  out
• 7:00
  of
• 7:00
  the
• 7:00
  box.
• 7:01
  The
• 7:01
  architecture
• 7:01
  follows
• 7:01
  a
• 7:02
  clean
• 7:02
  separation:
• 7:03
  core
• 7:03
  detection
• 7:03
  engine,
• 7:04
  use-case
• 7:05
  layer,
• 7:05
  adapter
• 7:05
  layer,
• 7:06
  memory
• 7:06
  layer.
• 7:07
  Whether
• 7:07
  it
• 7:07
  fulfills
• 7:08
  the
• 7:08
  "84.8%
• 7:09
  SWE-Bench
• 7:10
  solve
• 7:10
  rate"
• 7:10
  claim
• 7:10
  in
• 7:10
  its
• 7:11
  README
• 7:12
  will
• 7:12
  require
• 7:13
  independent
• 7:13
  verification,
• 7:14
  but
• 7:14
  the
• 7:14
  drift
• 7:15
  detection
• 7:15
  primitives
• 7:15
  are
• 7:16
  well-grounded
• 7:16
  in
• 7:16
  established
• 7:17
  statistics.
• 7:18
  The
• 7:18
  philosophical
• 7:19
  companion
• 7:19
  piece
• 7:19
  to
• 7:19
  both
• 7:20
  articles
• 7:20
  is
• 7:21
  Rafael
• 7:21
  Knuth's
• 7:21
  argument
• 7:22
  that
• 7:22
  the
• 7:22
  entire
• 7:23
  framing
• 7:23
  of
• 7:23
  "make
• 7:23
  AI
• 7:24
  more
• 7:24
  reliable"
• 7:24
  is
• 7:25
  backwards.
• 7:25
  His
• 7:26
  thesis:
• 7:26
  stop
• 7:27
  asking
• 7:27
  the
• 7:27
  AI
• 7:27
  to
• 7:28
  enforce
• 7:28
  rules
• 7:28
  on
• 7:28
  itself.
• 7:29
  Instead,
• 7:30
  separate
• 7:30
  judgment
• 7:31
  (which
• 7:31
  LLMs
• 7:31
  do
• 7:32
  well)
• 7:32
  from
• 7:32
  mechanical
• 7:32
  enforcement
• 7:33
  (which
• 7:33
  they
• 7:33
  do
• 7:34
  badly)
• 7:34
  by
• 7:34
  building
• 7:34
  simple
• 7:35
  checkpoint
• 7:35
  scripts
• 7:36
  that
• 7:36
  validate
• 7:36
  every
• 7:36
  output
• 7:37
  deterministically.
• 7:38
  Batch
• 7:38
  evaluations
• 7:39
  into
• 7:39
  groups
• 7:39
  of
• 7:39
  10–15
• 7:41
  items
• 7:41
  to
• 7:41
  stay
• 7:41
  within
• 7:42
  the
• 7:42
  model's
• 7:42
  effective
• 7:42
  attention
• 7:43
  window,
• 7:43
  run
• 7:43
  each
• 7:44
  batch
• 7:44
  through
• 7:44
  the
• 7:44
  checkpoint,
• 7:45
  fix
• 7:45
  only
• 7:45
  what
• 7:46
  the
• 7:46
  checkpoint
• 7:46
  flags.
• 7:47
  "A
• 7:47
  prompt
• 7:47
  is
• 7:48
  a
• 7:48
  request.
• 7:48
  A
• 7:48
  checkpoint
• 7:49
  is
• 7:49
  a
• 7:49
  gate.
• 7:50
  Requests
• 7:50
  can
• 7:51
  be
• 7:51
  ignored
• 7:51
  under
• 7:51
  pressure.
• 7:52
  Gates
• 7:52
  cannot."
• 7:55
  Block's
• 7:56
  latest
• 7:56
  round
• 7:56
  of
• 7:56
  layoffs
• 7:56
  —
• 7:57
  framed
• 7:57
  by
• 7:57
  leadership
• 7:58
  as
• 7:58
  AI-driven
• 7:59
  efficiency
• 7:59
  —
• 7:59
  drew
• 8:00
  a
• 8:00
  sharp
• 8:00
  rebuttal
• 8:00
  from
• 8:01
  fintech
• 8:01
  analyst
• 8:01
  Bradley
• 8:02
  Leimer.
• 8:02
  His
• 8:02
  argument:
• 8:03
  AI
• 8:04
  adoption
• 8:04
  at
• 8:04
  scale
• 8:04
  is
• 8:05
  "typically
• 8:05
  gradual,
• 8:06
  uneven,
• 8:06
  and
• 8:07
  deeply
• 8:07
  tied
• 8:07
  to
• 8:07
  specific
• 8:08
  workflow
• 8:08
  redesign."
• 8:09
  It
• 8:09
  does
• 8:09
  not
• 8:10
  arrive
• 8:10
  overnight
• 8:10
  as
• 8:11
  justification
• 8:11
  for
• 8:11
  sweeping
• 8:12
  workforce
• 8:12
  reductions.
• 8:14
  Block's
• 8:14
  history
• 8:14
  tells
• 8:15
  a
• 8:15
  different
• 8:15
  story
• 8:15
  than
• 8:16
  its
• 8:16
  AI
• 8:16
  narrative
• 8:16
  —
• 8:17
  aggressive
• 8:18
  expansion
• 8:18
  across
• 8:19
  Cash
• 8:19
  App,
• 8:19
  Afterpay,
• 8:20
  TIDAL,
• 8:20
  TBD,
• 8:21
  and
• 8:21
  various
• 8:22
  crypto
• 8:22
  ventures
• 8:22
  created
• 8:23
  headcount
• 8:23
  that
• 8:23
  made
• 8:23
  sense
• 8:24
  when
• 8:24
  markets
• 8:24
  rewarded
• 8:25
  growth
• 8:25
  but
• 8:25
  looked
• 8:25
  bloated
• 8:26
  the
• 8:26
  moment
• 8:26
  risk
• 8:26
  repricing
• 8:27
  began.
• 8:28
  The
• 8:28
  layoffs,
• 8:28
  Leimer
• 8:29
  argues,
• 8:29
  look
• 8:29
  less
• 8:30
  like
• 8:30
  a
• 8:30
  sudden
• 8:30
  AI
• 8:30
  revolution
• 8:31
  and
• 8:31
  more
• 8:31
  like
• 8:31
  a
• 8:31
  familiar
• 8:32
  correction
• 8:32
  following
• 8:33
  strategic
• 8:33
  overreach.
• 8:34
  The
• 8:35
  broader
• 8:35
  warning
• 8:35
  is
• 8:35
  for
• 8:36
  banking:
• 8:36
  financial
• 8:37
  institutions
• 8:38
  operate
• 8:38
  on
• 8:38
  trust
• 8:38
  and
• 8:38
  institutional
• 8:39
  memory
• 8:40
  that
• 8:40
  "large
• 8:40
  language
• 8:41
  models
• 8:41
  lack
• 8:41
  a
• 8:41
  large
• 8:42
  enough
• 8:42
  context
• 8:42
  window
• 8:42
  to
• 8:43
  retain."
• 8:44
  Rapid
• 8:44
  workforce
• 8:45
  contraction
• 8:45
  in
• 8:45
  the
• 8:45
  name
• 8:45
  of
• 8:45
  technology
• 8:46
  threatens
• 8:46
  the
• 8:46
  value
• 8:47
  proposition
• 8:47
  itself.
• 8:48
  Geoffrey
• 8:49
  Huntley's
• 8:49
  analysis
• 8:50
  from
• 8:50
  the
• 8:50
  other
• 8:50
  side
• 8:50
  of
• 8:50
  the
• 8:50
  table
• 8:51
  is
• 8:51
  bleaker
• 8:51
  and
• 8:52
  more
• 8:52
  granular.
• 8:52
  Software
• 8:53
  development
• 8:53
  costs
• 8:53
  have
• 8:53
  fallen
• 8:54
  to
• 8:54
  $10.42
• 8:55
  an
• 8:55
  hour
• 8:55
  —
• 8:56
  below
• 8:56
  minimum
• 8:56
  wage
• 8:57
  —
• 8:57
  and
• 8:57
  at
• 8:58
  a
• 8:58
  recent
• 8:58
  Cursor
• 8:58
  meetup,
• 8:58
  "nearly
• 8:59
  everyone
• 8:59
  in
• 8:59
  the
• 8:59
  room
• 9:00
  was
• 9:00
  not
• 9:00
  a
• 9:00
  software
• 9:00
  developer,
• 9:01
  showing
• 9:01
  off
• 9:01
  their
• 9:01
  latest
• 9:02
  and
• 9:02
  greatest
• 9:02
  creations."
• 9:03
  His
• 9:03
  model:
• 9:03
  the
• 9:03
  workforce
• 9:04
  splits
• 9:04
  into
• 9:04
  a
• 9:04
  K-shape.
• 9:05
  Model-first
• 9:06
  companies
• 9:06
  operate
• 9:07
  as
• 9:07
  lean
• 9:07
  apex
• 9:07
  predators
• 9:08
  on
• 9:08
  razor-thin
• 9:08
  margins.
• 9:09
  Everyone
• 9:10
  else
• 9:10
  needs
• 9:10
  a
• 9:10
  transformation
• 9:11
  program
• 9:11
  to
• 9:11
  avoid
• 9:12
  being
• 9:12
  outcompeted
• 9:12
  by
• 9:13
  startups
• 9:13
  that
• 9:13
  can
• 9:13
  replicate
• 9:14
  five-year
• 9:14
  roadmaps
• 9:15
  in
• 9:15
  months.
• 9:16
  Former
• 9:16
  moats
• 9:16
  —
• 9:17
  per-seat
• 9:18
  pricing,
• 9:18
  human-designed
• 9:19
  product
• 9:19
  features,
• 9:20
  high
• 9:20
  switching
• 9:21
  costs
• 9:21
  —
• 9:22
  are
• 9:22
  dissolving.
• 9:23
  An
• 9:23
  anonymous
• 9:23
  founder
• 9:24
  in
• 9:24
  Huntley's
• 9:24
  DMs:
• 9:25
  "20-ish
• 9:25
  people
• 9:26
  now
• 9:26
  do
• 9:26
  about
• 9:26
  30x
• 9:27
  the
• 9:27
  output
• 9:27
  of
• 9:27
  what
• 9:27
  having
• 9:28
  more
• 9:28
  than
• 9:28
  60
• 9:28
  did
• 9:28
  3
• 9:29
  years
• 9:29
  ago."
• 9:30
  The
• 9:30
  uncomfortable
• 9:31
  reality
• 9:31
  is
• 9:31
  that
• 9:32
  the
• 9:32
  cycle
• 9:32
  feeds
• 9:32
  itself:
• 9:33
  laid-off
• 9:34
  employees
• 9:34
  upskill
• 9:35
  with
• 9:35
  AI,
• 9:36
  join
• 9:36
  new
• 9:36
  employers
• 9:37
  or
• 9:37
  found
• 9:37
  startups,
• 9:38
  automate
• 9:38
  their
• 9:38
  former
• 9:39
  employers'
• 9:39
  job
• 9:39
  functions,
• 9:40
  and
• 9:40
  the
• 9:40
  cycle
• 9:41
  propagates
• 9:41
  across
• 9:41
  industries.
• 9:44
  Pi,
• 9:44
  from
• 9:44
  the
• 9:45
  creator
• 9:45
  of
• 9:45
  libGDX,
• 9:46
  is
• 9:47
  a
• 9:47
  minimal
• 9:47
  terminal
• 9:47
  coding
• 9:48
  harness
• 9:48
  built
• 9:48
  on
• 9:48
  a
• 9:48
  specific
• 9:49
  philosophy:
• 9:50
  features
• 9:51
  that
• 9:51
  other
• 9:51
  agents
• 9:51
  bake
• 9:51
  in,
• 9:52
  you
• 9:52
  should
• 9:52
  be
• 9:52
  able
• 9:53
  to
• 9:53
  build
• 9:53
  yourself.
• 9:54
  It
• 9:54
  ships
• 9:55
  with
• 9:55
  15+
• 9:55
  provider
• 9:56
  integrations
• 9:56
  (Anthropic,
• 9:57
  OpenAI,
• 9:58
  Google,
• 9:59
  Groq,
• 9:59
  Ollama,
• 10:00
  and
• 10:00
  more),
• 10:01
  tree-structured
• 10:01
  sessions
• 10:02
  with
• 10:02
  branching
• 10:02
  and
• 10:02
  bookmarking,
• 10:03
  mid-session
• 10:04
  model
• 10:04
  switching,
• 10:05
  and
• 10:05
  a
• 10:05
  TypeScript
• 10:06
  extension
• 10:06
  system
• 10:06
  covering
• 10:07
  tools,
• 10:07
  commands,
• 10:08
  keyboard
• 10:08
  shortcuts,
• 10:09
  and
• 10:09
  the
• 10:09
  full
• 10:10
  TUI.
• 10:11
  But
• 10:11
  it
• 10:11
  deliberately
• 10:12
  omits
• 10:12
  sub-agents
• 10:13
  and
• 10:13
  plan
• 10:13
  mode,
• 10:14
  treating
• 10:14
  these
• 10:14
  as
• 10:14
  extension
• 10:15
  territory
• 10:15
  rather
• 10:16
  than
• 10:16
  core
• 10:16
  features.
• 10:17
  The
• 10:17
  design
• 10:17
  is
• 10:18
  aggressively
• 10:18
  extensible
• 10:19
  —
• 10:20
  skills
• 10:20
  for
• 10:20
  progressive
• 10:20
  disclosure,
• 10:21
  prompts
• 10:22
  as
• 10:22
  Markdown
• 10:22
  files,
• 10:23
  middleware
• 10:23
  for
• 10:24
  RAG
• 10:24
  and
• 10:24
  long-term
• 10:24
  memory
• 10:25
  —
• 10:25
  while
• 10:25
  keeping
• 10:26
  the
• 10:26
  core
• 10:26
  small
• 10:26
  enough
• 10:27
  that
• 10:27
  the
• 10:27
  harness
• 10:27
  itself
• 10:27
  does
• 10:27
  not
• 10:28
  dictate
• 10:28
  workflow.
• 10:29
  The
• 10:29
  skills
• 10:29
  ecosystem
• 10:30
  feeding
• 10:30
  these
• 10:30
  harnesses
• 10:31
  continues
• 10:31
  to
• 10:31
  mature.
• 10:32
  Apify's
• 10:33
  open-source
• 10:33
  Agent
• 10:34
  Skills
• 10:34
  collection
• 10:34
  ships
• 10:35
  12
• 10:35
  pre-built
• 10:36
  capabilities
• 10:36
  —
• 10:37
  from
• 10:37
  e-commerce
• 10:38
  scraping
• 10:38
  to
• 10:38
  influencer
• 10:39
  discovery
• 10:39
  to
• 10:39
  lead
• 10:40
  generation
• 10:40
  —
• 10:41
  packaged
• 10:41
  as
• 10:41
  markdown
• 10:42
  skill
• 10:42
  files
• 10:42
  compatible
• 10:43
  with
• 10:43
  Claude
• 10:43
  Code,
• 10:44
  Cursor,
• 10:44
  Codex,
• 10:45
  and
• 10:45
  Gemini
• 10:45
  CLI.
• 10:46
  Install
• 10:47
  with
• 10:47
  one
• 10:47
  command,
• 10:48
  invoke
• 10:48
  from
• 10:48
  any
• 10:48
  supported
• 10:49
  agent.
• 10:50
  Meanwhile,
• 10:50
  Agent
• 10:50
  Skill
• 10:51
  Creator
• 10:51
  v4.1.0
• 10:53
  tackles
• 10:53
  the
• 10:53
  skill-authoring
• 10:54
  bottleneck:
• 10:55
  pass
• 10:55
  in
• 10:55
  documentation,
• 10:57
  code,
• 10:58
  transcripts,
• 10:59
  or
• 10:59
  even
• 10:59
  vague
• 10:59
  descriptions,
• 11:00
  and
• 11:00
  it
• 11:00
  produces
• 11:01
  a
• 11:01
  validated,
• 11:01
  security-scanned
• 11:02
  skill
• 11:03
  ready
• 11:03
  to
• 11:03
  install.
• 11:04
  A
• 11:05
  key
• 11:05
  fix
• 11:05
  in
• 11:05
  this
• 11:05
  release
• 11:06
  addressed
• 11:06
  a
• 11:06
  compatibility
• 11:07
  bug
• 11:07
  where
• 11:07
  generated
• 11:08
  skills
• 11:08
  included
• 11:09
  non-standard
• 11:09
  activation
• 11:10
  fields
• 11:10
  that
• 11:11
  broke
• 11:11
  Claude
• 11:11
  Code
• 11:11
  installation.
• 11:12
  Since
• 11:13
  VS
• 11:13
  Code
• 11:13
  1.108,
• 11:15
  Copilot
• 11:15
  and
• 11:15
  Claude
• 11:16
  Code
• 11:16
  share
• 11:16
  the
• 11:16
  same
• 11:17
  global
• 11:17
  skills
• 11:17
  path,
• 11:18
  so
• 11:18
  one
• 11:18
  install
• 11:18
  works
• 11:19
  across
• 11:19
  both.
• 11:20
  At
• 11:20
  the
• 11:20
  orchestration
• 11:21
  layer,
• 11:21
  Claude
• 11:21
  Flow
• 11:22
  graduated
• 11:22
  from
• 11:22
  alpha
• 11:23
  to
• 11:23
  its
• 11:23
  first
• 11:23
  production-ready
• 11:24
  release
• 11:24
  as
• 11:24
  RuFlo
• 11:25
  v3.5.0
• 11:26
  —
• 11:27
  5,800+
• 11:28
  commits
• 11:28
  across
• 11:29
  10
• 11:29
  months,
• 11:29
  now
• 11:30
  claiming
• 11:30
  deployment
• 11:30
  inside
• 11:31
  a
• 11:31
  majority
• 11:31
  of
• 11:31
  the
• 11:31
  Fortune
• 11:32
  500.
• 11:33
  The
• 11:33
  pitch:
• 11:34
  one
• 11:34
  control
• 11:34
  plane
• 11:34
  for
• 11:35
  60+
• 11:35
  specialized
• 11:36
  agents,
• 11:37
  hierarchical
• 11:37
  and
• 11:37
  mesh
• 11:38
  swarms,
• 11:39
  fault-tolerant
• 11:39
  consensus,
• 11:40
  self-learning
• 11:41
  memory,
• 11:41
  and
• 11:42
  215
• 11:43
  MCP
• 11:43
  tools
• 11:44
  spanning
• 11:44
  orchestration,
• 11:45
  governance,
• 11:46
  neural
• 11:46
  training,
• 11:46
  and
• 11:46
  security.
• 11:47
  Underneath
• 11:48
  it,
• 11:48
  RuVector
• 11:49
  positions
• 11:49
  itself
• 11:50
  as
• 11:50
  a
• 11:50
  self-learning
• 11:51
  vector
• 11:51
  database
• 11:51
  that
• 11:52
  improves
• 11:52
  search
• 11:52
  results
• 11:53
  over
• 11:53
  time
• 11:53
  via
• 11:53
  GNN
• 11:54
  layers
• 11:54
  on
• 11:54
  its
• 11:55
  HNSW
• 11:55
  index,
• 11:56
  runs
• 11:56
  local
• 11:57
  LLM
• 11:57
  inference
• 11:58
  on
• 11:58
  Metal/CUDA/WebGPU
• 12:01
  without
• 12:01
  API
• 12:02
  costs,
• 12:03
  and
• 12:03
  drops
• 12:03
  into
• 12:03
  PostgreSQL
• 12:04
  as
• 12:04
  a
• 12:04
  pgvector
• 12:05
  replacement
• 12:05
  with
• 12:05
  230+
• 12:07
  SQL
• 12:07
  functions.
• 12:08
  The
• 12:08
  RVF
• 12:09
  cognitive
• 12:09
  container
• 12:10
  format
• 12:10
  —
• 12:11
  bundling
• 12:11
  vectors,
• 12:12
  models,
• 12:13
  GNN
• 12:13
  graph
• 12:13
  state,
• 12:14
  and
• 12:14
  a
• 12:14
  bootable
• 12:15
  microkernel
• 12:15
  into
• 12:15
  a
• 12:16
  single
• 12:16
  deployable
• 12:16
  file
• 12:17
  —
• 12:17
  is
• 12:18
  architecturally
• 12:18
  ambitious,
• 12:19
  booting
• 12:19
  as
• 12:20
  a
• 12:20
  microservice
• 12:20
  in
• 12:20
  125ms.
• 12:22
  The
• 12:22
  broader
• 12:22
  ecosystem
• 12:23
  includes
• 12:23
  Ask
• 12:24
  RuVNet
• 12:24
  as
• 12:24
  an
• 12:24
  interactive
• 12:25
  exploration
• 12:25
  layer
• 12:26
  and
• 12:26
  ruvector-domain-expansion,
• 12:28
  which
• 12:28
  implements
• 12:28
  cross-domain
• 12:29
  transfer
• 12:29
  learning
• 12:29
  with
• 12:30
  Thompson
• 12:30
  Sampling
• 12:30
  for
• 12:30
  strategy
• 12:31
  selection
• 12:32
  and
• 12:32
  EWC++
• 12:33
  for
• 12:33
  catastrophic-forgetting
• 12:34
  prevention
• 12:34
  —
• 12:35
  the
• 12:35
  thesis
• 12:36
  being
• 12:36
  that
• 12:36
  "intelligence
• 12:37
  should
• 12:37
  compound"
• 12:37
  rather
• 12:38
  than
• 12:38
  restart
• 12:38
  from
• 12:38
  scratch
• 12:39
  per
• 12:39
  domain.
• 12:40
  Community
• 12:40
  response
• 12:41
  pushed
• 12:41
  back
• 12:41
  constructively:
• 12:43
  "define
• 12:43
  the
• 12:43
  transferable
• 12:44
  unit,
• 12:44
  define
• 12:45
  domains
• 12:45
  operationally,
• 12:46
  and
• 12:46
  gate
• 12:47
  transfer
• 12:47
  by
• 12:47
  explicit
• 12:48
  invariants."
• 12:51
  If
• 12:51
  you
• 12:51
  want
• 12:51
  to
• 12:51
  understand
• 12:52
  how
• 12:52
  a
• 12:52
  CPU
• 12:53
  or
• 12:53
  GPU
• 12:53
  works
• 12:54
  at
• 12:54
  the
• 12:54
  register-transfer
• 12:54
  level,
• 12:55
  excellent
• 12:55
  open-source
• 12:56
  references
• 12:56
  exist.
• 12:57
  For
• 12:57
  NPUs
• 12:58
  —
• 12:59
  the
• 12:59
  neural
• 12:59
  accelerators
• 12:59
  from
• 13:00
  NVIDIA,
• 13:00
  Google,
• 13:01
  and
• 13:01
  Apple
• 13:01
  that
• 13:01
  actually
• 13:02
  run
• 13:02
  inference
• 13:02
  —
• 13:02
  there
• 13:02
  has
• 13:02
  been
• 13:03
  almost
• 13:03
  nothing.
• 13:04
  tiny-npu
• 13:04
  fills
• 13:05
  that
• 13:05
  gap:
• 13:05
  a
• 13:05
  fully
• 13:05
  synthesizable
• 13:06
  neural
• 13:06
  processing
• 13:06
  unit
• 13:07
  in
• 13:07
  SystemVerilog,
• 13:07
  designed
• 13:08
  for
• 13:08
  education
• 13:08
  rather
• 13:09
  than
• 13:09
  production.
• 13:09
  It
• 13:10
  supports
• 13:10
  two
• 13:10
  execution
• 13:11
  modes.
• 13:11
  LLM
• 13:12
  Mode
• 13:12
  runs
• 13:12
  real
• 13:12
  transformer
• 13:13
  models
• 13:13
  (GPT-2,
• 13:14
  LLaMA,
• 13:14
  Mistral,
• 13:15
  Qwen2)
• 13:15
  from
• 13:16
  actual
• 13:16
  HuggingFace
• 13:16
  weights
• 13:17
  through
• 13:17
  a
• 13:17
  128-bit
• 13:18
  microcode
• 13:18
  ISA
• 13:19
  with
• 13:19
  KV-cache
• 13:20
  optimization.
• 13:21
  Graph
• 13:21
  Mode
• 13:21
  accepts
• 13:21
  arbitrary
• 13:22
  ONNX
• 13:22
  models
• 13:23
  compiled
• 13:23
  by
• 13:23
  an
• 13:23
  included
• 13:24
  Python-based
• 13:24
  compiler,
• 13:25
  dispatching
• 13:25
  operations
• 13:26
  across
• 13:26
  13
• 13:26
  dedicated
• 13:27
  hardware
• 13:27
  engines
• 13:28
  (GEMM
• 13:28
  with
• 13:28
  a
• 13:28
  16x16
• 13:29
  systolic
• 13:30
  array
• 13:30
  of
• 13:30
  256
• 13:31
  dual
• 13:31
  INT8/FP16
• 13:34
  MACs,
• 13:35
  softmax,
• 13:35
  reduce,
• 13:36
  math
• 13:36
  LUT,
• 13:36
  gather,
• 13:37
  slice,
• 13:37
  concat,
• 13:37
  pooling,
• 13:38
  padding,
• 13:38
  resize,
• 13:39
  cast,
• 13:39
  plus
• 13:39
  element-wise).
• 13:40
  Async
• 13:40
  dispatch
• 13:41
  with
• 13:41
  scoreboard-based
• 13:42
  DMA+GEMM
• 13:43
  overlap
• 13:44
  and
• 13:44
  hardware
• 13:44
  performance
• 13:44
  counters
• 13:45
  round
• 13:45
  out
• 13:45
  the
• 13:45
  architecture.
• 13:46
  All
• 13:46
  outputs
• 13:47
  are
• 13:47
  verified
• 13:48
  cycle-accurate
• 13:48
  against
• 13:48
  C++
• 13:49
  golden
• 13:49
  models.
• 13:50
  For
• 13:50
  students,
• 13:51
  chip
• 13:51
  designers,
• 13:52
  or
• 13:52
  anyone
• 13:53
  who
• 13:53
  has
• 13:53
  wondered
• 13:53
  what
• 13:53
  happens
• 13:54
  between
• 13:54
  `model.forward()`
• 13:55
  and
• 13:55
  the
• 13:55
  silicon,
• 13:56
  this
• 13:56
  is
• 13:56
  the
• 13:56
  most
• 13:56
  readable
• 13:57
  open
• 13:57
  reference
• 13:57
  available.
• 13:58
  Parakeet.cpp
• 14:00
  brings
• 14:00
  NVIDIA's
• 14:01
  Parakeet
• 14:01
  ASR
• 14:01
  models
• 14:02
  to
• 14:02
  pure
• 14:02
  C++
• 14:03
  with
• 14:03
  Metal
• 14:03
  GPU
• 14:04
  acceleration,
• 14:05
  built
• 14:05
  on
• 14:05
  the
• 14:05
  Axiom
• 14:05
  tensor
• 14:06
  library
• 14:06
  rather
• 14:06
  than
• 14:06
  ONNX
• 14:07
  Runtime
• 14:07
  or
• 14:07
  PyTorch.
• 14:08
  The
• 14:08
  headline:
• 14:09
  ~27ms
• 14:10
  encoder
• 14:10
  inference
• 14:11
  on
• 14:11
  Apple
• 14:11
  Silicon
• 14:11
  GPU
• 14:12
  for
• 14:12
  10
• 14:12
  seconds
• 14:13
  of
• 14:13
  audio
• 14:13
  with
• 14:13
  the
• 14:13
  110M
• 14:14
  parameter
• 14:14
  model
• 14:15
  —
• 14:16
  96x
• 14:16
  faster
• 14:16
  than
• 14:17
  CPU.
• 14:18
  It
• 14:18
  supports
• 14:18
  six
• 14:19
  model
• 14:19
  variants
• 14:19
  spanning
• 14:19
  offline
• 14:20
  transcription,
• 14:21
  streaming
• 14:21
  with
• 14:22
  end-of-utterance
• 14:22
  detection,
• 14:23
  and
• 14:23
  speaker
• 14:23
  diarization,
• 14:25
  all
• 14:25
  sharing
• 14:25
  the
• 14:25
  same
• 14:26
  audio
• 14:26
  pipeline
• 14:26
  (16kHz
• 14:27
  mono
• 14:27
  WAV
• 14:28
  to
• 14:28
  80-bin
• 14:28
  Mel
• 14:28
  spectrogram
• 14:29
  to
• 14:29
  FastConformer
• 14:30
  encoder).
• 14:31
  Word-level
• 14:31
  timestamps
• 14:32
  with
• 14:32
  confidence
• 14:32
  scores
• 14:33
  and
• 14:33
  phrase
• 14:33
  boosting
• 14:34
  for
• 14:34
  domain
• 14:34
  vocabulary
• 14:35
  are
• 14:35
  included.
• 14:36
  Rounding
• 14:36
  out
• 14:37
  the
• 14:37
  hardware
• 14:37
  cluster,
• 14:37
  super-tart-vphone
• 14:38
  achieves
• 14:39
  functional
• 14:39
  macOS/iOS
• 14:41
  virtualization
• 14:42
  on
• 14:42
  Apple
• 14:42
  Silicon
• 14:42
  with
• 14:43
  Metal
• 14:43
  GPU
• 14:43
  passthrough
• 14:44
  by
• 14:44
  forking
• 14:44
  the
• 14:44
  Tart
• 14:45
  virtualization
• 14:45
  framework
• 14:46
  and
• 14:46
  patching
• 14:46
  the
• 14:46
  entire
• 14:47
  bootchain
• 14:47
  —
• 14:48
  useful
• 14:48
  for
• 14:48
  security
• 14:49
  research
• 14:49
  and
• 14:49
  CI
• 14:50
  automation
• 14:50
  on
• 14:50
  Apple
• 14:50
  platforms.