Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – February 21, 2026

Transcript

AGI Dreams Podcast – February 21, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Google
• 0:00
  shipped
• 0:00
  a
• 0:00
  Chrome
• 0:01
  stable
• 0:01
  channel
• 0:01
  update
• 0:02
  this
• 0:02
  week
• 0:02
  patching
• 0:02
  three
• 0:03
  high-severity
• 0:03
  vulnerabilities:
• 0:04
  CVE-2026-2648,
• 0:08
  a
• 0:09
  heap
• 0:09
  buffer
• 0:09
  overflow
• 0:09
  in
• 0:10
  PDFium;
• 0:11
  CVE-2026-2649,
• 0:15
  an
• 0:15
  integer
• 0:15
  overflow
• 0:15
  in
• 0:16
  V8
• 0:16
  reported
• 0:16
  by
• 0:16
  a
• 0:17
  KAIST
• 0:17
  Hacking
• 0:17
  Lab
• 0:18
  researcher;
• 0:19
  and
• 0:19
  CVE-2026-2650,
• 0:22
  a
• 0:22
  heap
• 0:23
  buffer
• 0:23
  overflow
• 0:23
  in
• 0:23
  Media.
• 0:24
  The
• 0:24
  V8
• 0:25
  integer
• 0:25
  overflow
• 0:26
  is
• 0:26
  the
• 0:26
  one
• 0:26
  to
• 0:26
  watch
• 0:27
  —
• 0:27
  V8
• 0:28
  bugs
• 0:28
  are
• 0:28
  bread
• 0:28
  and
• 0:28
  butter
• 0:29
  for
• 0:29
  exploit
• 0:29
  chains,
• 0:30
  and
• 0:30
  this
• 0:30
  one
• 0:30
  sat
• 0:31
  in
• 0:31
  the
• 0:31
  wild
• 0:31
  long
• 0:31
  enough
• 0:31
  to
• 0:32
  earn
• 0:32
  a
• 0:32
  separate
• 0:32
  Hacker
• 0:32
  News
• 0:32
  thread
• 0:33
  flagging
• 0:33
  a
• 0:33
  CSS-related
• 0:34
  zero-day
• 0:34
  in
• 0:34
  the
• 0:34
  same
• 0:35
  release
• 0:35
  cycle.
• 0:36
  Chrome
• 0:36
  145.0.7632.109/110
• 0:40
  is
• 0:41
  the
• 0:41
  fix
• 0:41
  for
• 0:41
  Windows
• 0:41
  and
• 0:41
  Mac;
• 0:42
  Linux
• 0:42
  gets
• 0:42
  144.0.7559.109.
• 0:47
  If
• 0:47
  you
• 0:47
  haven't
• 0:48
  force-updated
• 0:48
  yet,
• 0:48
  now
• 0:49
  is
• 0:49
  the
• 0:49
  time.
• 0:50
  Meanwhile,
• 0:51
  Microsoft
• 0:51
  confirmed
• 0:52
  a
• 0:52
  bug
• 0:52
  in
• 0:52
  Copilot
• 0:53
  that
• 0:53
  caused
• 0:53
  it
• 0:53
  to
• 0:53
  summarize
• 0:54
  confidential
• 0:54
  emails
• 0:54
  —
• 0:55
  messages
• 0:56
  the
• 0:56
  user
• 0:56
  was
• 0:56
  not
• 0:56
  authorized
• 0:57
  to
• 0:57
  view
• 0:57
  —
• 0:58
  and
• 0:58
  surface
• 0:58
  that
• 0:58
  content
• 0:59
  in
• 0:59
  Copilot
• 0:59
  responses.
• 1:00
  This
• 1:00
  is
• 1:00
  not
• 1:01
  a
• 1:01
  theoretical
• 1:01
  oversharing
• 1:02
  risk;
• 1:02
  it
• 1:03
  is
• 1:03
  an
• 1:03
  actual
• 1:03
  data
• 1:03
  leakage
• 1:04
  incident
• 1:04
  where
• 1:04
  the
• 1:04
  AI
• 1:05
  assistant
• 1:05
  breached
• 1:05
  access
• 1:06
  boundaries
• 1:06
  the
• 1:06
  organization
• 1:07
  had
• 1:07
  explicitly
• 1:07
  configured.
• 1:08
  Microsoft
• 1:09
  attributed
• 1:09
  it
• 1:09
  to
• 1:10
  a
• 1:10
  bug,
• 1:10
  not
• 1:10
  a
• 1:10
  design
• 1:11
  flaw,
• 1:11
  but
• 1:11
  the
• 1:11
  distinction
• 1:12
  matters
• 1:12
  less
• 1:12
  when
• 1:12
  the
• 1:13
  confidential
• 1:13
  content
• 1:13
  has
• 1:14
  already
• 1:14
  been
• 1:14
  presented
• 1:14
  to
• 1:15
  an
• 1:15
  unauthorized
• 1:15
  user.
• 1:16
  The
• 1:16
  same
• 1:17
  BleepingComputer
• 1:17
  report
• 1:18
  noted
• 1:18
  a
• 1:18
  broader
• 1:18
  security
• 1:19
  landscape
• 1:19
  in
• 1:19
  motion:
• 1:20
  Amazon
• 1:20
  disclosed
• 1:21
  an
• 1:21
  AI-assisted
• 1:22
  attacker
• 1:22
  who
• 1:22
  breached
• 1:23
  600
• 1:23
  FortiGate
• 1:24
  firewalls
• 1:24
  across
• 1:25
  55
• 1:25
  countries
• 1:26
  in
• 1:26
  five
• 1:26
  weeks,
• 1:27
  and
• 1:27
  researchers
• 1:28
  identified
• 1:28
  PromptSpy
• 1:29
  as
• 1:29
  the
• 1:29
  first
• 1:29
  known
• 1:30
  Android
• 1:30
  malware
• 1:31
  to
• 1:31
  use
• 1:31
  generative
• 1:31
  AI
• 1:32
  at
• 1:32
  runtime,
• 1:33
  leveraging
• 1:33
  Google's
• 1:33
  Gemini
• 1:34
  model
• 1:34
  to
• 1:34
  adapt
• 1:35
  its
• 1:35
  persistence
• 1:35
  mechanisms
• 1:36
  per
• 1:36
  device.
• 1:37
  The
• 1:37
  attack
• 1:38
  surface
• 1:38
  is
• 1:38
  not
• 1:38
  just
• 1:39
  expanding
• 1:39
  through
• 1:39
  bugs
• 1:40
  in
• 1:40
  existing
• 1:40
  products
• 1:40
  —
• 1:41
  it
• 1:41
  is
• 1:41
  being
• 1:42
  deliberately
• 1:42
  designed
• 1:43
  into
• 1:43
  new
• 1:43
  protocols.
• 1:44
  WebMCP,
• 1:45
  an
• 1:45
  experimental
• 1:46
  browser
• 1:46
  API
• 1:47
  in
• 1:47
  Chrome
• 1:47
  146+
• 1:49
  that
• 1:49
  lets
• 1:49
  websites
• 1:50
  expose
• 1:50
  JavaScript
• 1:51
  functions
• 1:51
  as
• 1:51
  "tools"
• 1:52
  for
• 1:52
  AI
• 1:52
  agents,
• 1:53
  introduces
• 1:54
  what
• 1:54
  one
• 1:54
  security
• 1:55
  researcher
• 1:55
  calls
• 1:55
  the
• 1:55
  "data
• 1:56
  ferry"
• 1:56
  problem.
• 1:57
  An
• 1:57
  AI
• 1:58
  agent
• 1:58
  can
• 1:58
  invoke
• 1:58
  a
• 1:58
  tool
• 1:59
  on
• 1:59
  your
• 1:59
  bank's
• 1:59
  site,
• 2:00
  receive
• 2:00
  your
• 2:00
  balance
• 2:01
  into
• 2:01
  its
• 2:01
  context,
• 2:02
  then
• 2:02
  pass
• 2:03
  that
• 2:03
  data
• 2:03
  as
• 2:03
  parameters
• 2:04
  to
• 2:04
  a
• 2:04
  tool
• 2:04
  on
• 2:04
  a
• 2:04
  malicious
• 2:05
  site.
• 2:06
  CORS
• 2:06
  was
• 2:06
  implemented
• 2:07
  to
• 2:07
  prevent
• 2:07
  scripts
• 2:08
  on
• 2:08
  one
• 2:08
  origin
• 2:08
  from
• 2:09
  reading
• 2:09
  data
• 2:09
  from
• 2:09
  another;
• 2:10
  WebMCP
• 2:11
  introduces
• 2:11
  the
• 2:12
  AI
• 2:12
  agent
• 2:12
  as
• 2:12
  a
• 2:13
  trusted
• 2:13
  intermediary
• 2:14
  that
• 2:14
  can
• 2:14
  move
• 2:14
  data
• 2:14
  between
• 2:15
  origins
• 2:15
  through
• 2:15
  legitimate
• 2:16
  tool
• 2:16
  invocations.
• 2:18
  There
• 2:18
  is
• 2:18
  no
• 2:18
  mechanism
• 2:19
  to
• 2:19
  verify
• 2:19
  that
• 2:19
  a
• 2:19
  tool's
• 2:20
  actual
• 2:20
  behavior
• 2:21
  matches
• 2:21
  its
• 2:21
  description
• 2:21
  —
• 2:22
  a
• 2:22
  tool
• 2:23
  described
• 2:23
  as
• 2:23
  "preview
• 2:24
  order"
• 2:24
  could
• 2:24
  execute
• 2:25
  a
• 2:25
  purchase
• 2:25
  —
• 2:26
  and
• 2:26
  tools
• 2:26
  can
• 2:27
  request
• 2:27
  extensive
• 2:27
  user
• 2:28
  data
• 2:28
  under
• 2:28
  the
• 2:28
  guise
• 2:29
  of
• 2:29
  personalization.
• 2:31
  The
• 2:31
  WebMCP
• 2:32
  authors
• 2:32
  acknowledge
• 2:32
  these
• 2:33
  risks
• 2:33
  in
• 2:33
  their
• 2:33
  GitHub
• 2:34
  issues
• 2:34
  and
• 2:34
  are
• 2:34
  exploring
• 2:35
  mitigations,
• 2:36
  but
• 2:36
  the
• 2:36
  protocol
• 2:37
  itself
• 2:37
  specifies
• 2:38
  few
• 2:38
  controls,
• 2:39
  leaving
• 2:40
  implementation
• 2:40
  to
• 2:40
  browsers
• 2:41
  and
• 2:41
  agent
• 2:41
  developers.
• 2:43
  Security
• 2:43
  teams
• 2:44
  should
• 2:44
  be
• 2:44
  tracking
• 2:44
  this
• 2:44
  now,
• 2:45
  before
• 2:46
  it
• 2:46
  ships
• 2:46
  to
• 2:46
  stable.
• 2:51
  Anthropic
• 2:51
  launched
• 2:51
  Claude
• 2:52
  Code
• 2:52
  Security
• 2:52
  in
• 2:53
  a
• 2:53
  limited
• 2:53
  research
• 2:53
  preview
• 2:54
  this
• 2:54
  week,
• 2:54
  and
• 2:54
  the
• 2:54
  framing
• 2:55
  is
• 2:55
  worth
• 2:55
  noting:
• 2:55
  this
• 2:56
  is
• 2:56
  not
• 2:56
  a
• 2:56
  static
• 2:56
  analysis
• 2:57
  tool.
• 2:58
  It
• 2:58
  reads
• 2:58
  and
• 2:58
  reasons
• 2:59
  about
• 2:59
  code
• 2:59
  the
• 2:59
  way
• 2:59
  a
• 2:59
  human
• 3:00
  security
• 3:00
  researcher
• 3:00
  would
• 3:01
  —
• 3:01
  tracing
• 3:02
  how
• 3:02
  data
• 3:02
  moves
• 3:03
  through
• 3:03
  an
• 3:03
  application,
• 3:04
  understanding
• 3:04
  component
• 3:05
  interactions,
• 3:06
  and
• 3:06
  catching
• 3:06
  complex
• 3:07
  vulnerabilities
• 3:08
  that
• 3:08
  rule-based
• 3:08
  scanners
• 3:09
  miss,
• 3:09
  particularly
• 3:10
  flaws
• 3:10
  in
• 3:10
  business
• 3:11
  logic
• 3:11
  and
• 3:11
  broken
• 3:11
  access
• 3:12
  control.
• 3:12
  Every
• 3:13
  finding
• 3:13
  passes
• 3:13
  through
• 3:13
  a
• 3:13
  multi-stage
• 3:14
  verification
• 3:15
  process
• 3:15
  where
• 3:15
  Claude
• 3:15
  re-examines
• 3:16
  each
• 3:16
  result,
• 3:17
  attempting
• 3:17
  to
• 3:17
  prove
• 3:18
  or
• 3:18
  disprove
• 3:18
  its
• 3:18
  own
• 3:18
  findings
• 3:19
  before
• 3:19
  surfacing
• 3:20
  them.
• 3:20
  Validated
• 3:21
  findings
• 3:21
  appear
• 3:21
  in
• 3:21
  a
• 3:21
  dashboard
• 3:22
  with
• 3:22
  severity
• 3:22
  ratings
• 3:23
  and
• 3:23
  confidence
• 3:23
  scores,
• 3:24
  and
• 3:24
  suggested
• 3:24
  patches
• 3:25
  require
• 3:25
  human
• 3:25
  approval
• 3:26
  before
• 3:26
  application.
• 3:27
  Anthropic
• 3:28
  reports
• 3:28
  that
• 3:28
  using
• 3:28
  Claude
• 3:28
  Opus
• 3:29
  4.6,
• 3:30
  their
• 3:30
  Frontier
• 3:31
  Red
• 3:31
  Team
• 3:31
  found
• 3:32
  zero-day
• 3:32
  vulnerabilities
• 3:33
  in
• 3:33
  production
• 3:33
  open-source
• 3:34
  codebases
• 3:34
  —
• 3:35
  bugs
• 3:35
  that
• 3:35
  had
• 3:35
  gone
• 3:35
  undetected
• 3:36
  for
• 3:36
  decades
• 3:37
  despite
• 3:37
  years
• 3:37
  of
• 3:37
  expert
• 3:38
  review.
• 3:39
  Enterprise
• 3:39
  and
• 3:39
  Team
• 3:39
  customers
• 3:40
  get
• 3:40
  early
• 3:40
  access;
• 3:41
  open-source
• 3:42
  maintainers
• 3:42
  can
• 3:42
  apply
• 3:42
  for
• 3:43
  free
• 3:43
  expedited
• 3:43
  access.
• 3:44
  The
• 3:45
  offensive
• 3:45
  side
• 3:45
  of
• 3:45
  AI
• 3:45
  security
• 3:46
  is
• 3:46
  evolving
• 3:46
  just
• 3:46
  as
• 3:47
  fast.
• 3:48
  A
• 3:48
  detailed
• 3:48
  walkthrough
• 3:48
  of
• 3:49
  AI
• 3:49
  penetration
• 3:49
  testing
• 3:50
  by
• 3:50
  Jason
• 3:50
  Haddix
• 3:50
  —
• 3:51
  who
• 3:51
  literally
• 3:52
  wrote
• 3:52
  the
• 3:52
  AI
• 3:52
  pen
• 3:52
  testing
• 3:53
  methodology
• 3:53
  —
• 3:54
  demonstrates
• 3:54
  how
• 3:54
  far
• 3:55
  the
• 3:55
  field
• 3:55
  has
• 3:55
  come
• 3:55
  beyond
• 3:56
  "making
• 3:56
  ChatGPT
• 3:57
  say
• 3:57
  bad
• 3:57
  words."
• 3:58
  His
• 3:58
  team's
• 3:58
  Agent
• 3:59
  Breaker
• 3:59
  challenges
• 4:00
  present
• 4:00
  actual
• 4:00
  AI-integrated
• 4:01
  applications
• 4:02
  (portfolio
• 4:02
  advisors,
• 4:03
  trip
• 4:03
  planners,
• 4:04
  corporate
• 4:05
  messaging)
• 4:05
  where
• 4:05
  attackers
• 4:06
  must
• 4:06
  exploit
• 4:06
  LLM-enabled
• 4:07
  search
• 4:07
  bars
• 4:07
  to
• 4:08
  leak
• 4:08
  system
• 4:08
  prompts,
• 4:09
  extract
• 4:10
  API
• 4:10
  keys,
• 4:11
  and
• 4:11
  exfiltrate
• 4:11
  confidential
• 4:12
  data
• 4:12
  from
• 4:13
  RAG
• 4:13
  databases.
• 4:14
  In
• 4:14
  one
• 4:14
  real
• 4:15
  client
• 4:15
  engagement
• 4:16
  recreated
• 4:16
  as
• 4:16
  a
• 4:16
  CTF,
• 4:17
  a
• 4:17
  search
• 4:18
  bar
• 4:18
  yielded
• 4:18
  a
• 4:18
  Jira
• 4:19
  project
• 4:19
  access
• 4:19
  token
• 4:20
  and
• 4:20
  patent
• 4:20
  acquisition
• 4:21
  data.
• 4:22
  The
• 4:22
  non-deterministic
• 4:23
  nature
• 4:23
  of
• 4:23
  LLMs
• 4:24
  means
• 4:24
  the
• 4:24
  same
• 4:25
  attack
• 4:25
  prompt
• 4:25
  might
• 4:25
  succeed
• 4:26
  on
• 4:26
  attempt
• 4:26
  3
• 4:27
  or
• 4:27
  attempt
• 4:27
  239
• 4:28
  —
• 4:29
  a
• 4:29
  fundamentally
• 4:30
  different
• 4:30
  testing
• 4:30
  paradigm
• 4:31
  from
• 4:31
  traditional
• 4:31
  security.
• 4:33
  A
• 4:33
  12-year-old
• 4:33
  solved
• 4:34
  the
• 4:34
  entire
• 4:34
  five-flag
• 4:35
  CTF
• 4:35
  in
• 4:35
  35
• 4:36
  minutes
• 4:36
  at
• 4:37
  BSides
• 4:37
  San
• 4:37
  Francisco,
• 4:38
  which
• 4:38
  Haddix
• 4:39
  considers
• 4:39
  "entry
• 4:39
  level."
• 4:40
  On
• 4:40
  the
• 4:41
  tooling
• 4:41
  side,
• 4:41
  Lonkero
• 4:41
  —
• 4:42
  a
• 4:42
  Rust-based
• 4:43
  web
• 4:43
  security
• 4:43
  scanner
• 4:44
  from
• 4:44
  Finnish
• 4:44
  firm
• 4:44
  Bountyy
• 4:45
  —
• 4:45
  shipped
• 4:46
  v3.6
• 4:48
  with
• 4:48
  a
• 4:48
  proof-based
• 4:49
  XSS
• 4:49
  detection
• 4:50
  engine
• 4:50
  that
• 4:50
  requires
• 4:50
  no
• 4:51
  browser
• 4:51
  at
• 4:51
  all.
• 4:52
  Instead
• 4:52
  of
• 4:52
  Chrome-based
• 4:53
  scanning,
• 4:53
  it
• 4:53
  uses
• 4:54
  mathematical
• 4:54
  proof
• 4:54
  of
• 4:54
  exploitability:
• 4:56
  send
• 4:56
  a
• 4:56
  canary,
• 4:56
  probe
• 4:57
  escaping
• 4:58
  behavior
• 4:58
  across
• 4:59
  16
• 4:59
  reflection
• 4:59
  contexts,
• 5:01
  then
• 5:01
  derive
• 5:01
  whether
• 5:01
  XSS
• 5:02
  is
• 5:02
  exploitable
• 5:03
  from
• 5:03
  context
• 5:03
  plus
• 5:04
  escaping
• 5:04
  analysis.
• 5:05
  The
• 5:05
  result
• 5:06
  is
• 5:06
  300x
• 5:06
  faster
• 5:07
  scanning
• 5:07
  (~200ms
• 5:08
  per
• 5:09
  URL
• 5:09
  versus
• 5:10
  60+
• 5:10
  seconds
• 5:11
  with
• 5:11
  Chrome),
• 5:11
  2-3
• 5:12
  requests
• 5:12
  per
• 5:12
  parameter,
• 5:13
  and
• 5:13
  zero
• 5:14
  browser
• 5:14
  stability
• 5:14
  issues.
• 5:15
  The
• 5:15
  scanner's
• 5:16
  intelligence
• 5:16
  system
• 5:17
  uses
• 5:17
  Bayesian
• 5:17
  hypothesis
• 5:18
  testing
• 5:18
  to
• 5:19
  form
• 5:19
  and
• 5:19
  refine
• 5:19
  vulnerability
• 5:20
  theories
• 5:20
  —
• 5:21
  starting
• 5:22
  from
• 5:22
  a
• 5:22
  prior
• 5:22
  probability
• 5:23
  based
• 5:23
  on
• 5:23
  parameter
• 5:24
  name
• 5:24
  and
• 5:24
  context,
• 5:25
  then
• 5:25
  updating
• 5:25
  via
• 5:25
  Bayes'
• 5:26
  theorem
• 5:26
  with
• 5:26
  each
• 5:26
  test
• 5:27
  —
• 5:27
  cutting
• 5:28
  payloads
• 5:28
  from
• 5:28
  5,000
• 5:29
  to
• 5:29
  800
• 5:30
  while
• 5:30
  increasing
• 5:30
  detection.
• 5:31
  Its
• 5:31
  v3.1
• 5:33
  also
• 5:33
  introduced
• 5:34
  a
• 5:34
  README
• 5:36
  invisible
• 5:36
  prompt
• 5:37
  injection
• 5:37
  scanner
• 5:37
  that
• 5:38
  detects
• 5:38
  hidden
• 5:38
  instructions
• 5:39
  in
• 5:39
  markdown
• 5:39
  files
• 5:40
  invisible
• 5:40
  to
• 5:40
  rendered
• 5:41
  views
• 5:41
  but
• 5:41
  readable
• 5:42
  by
• 5:42
  LLMs
• 5:42
  processing
• 5:43
  raw
• 5:43
  source.
• 5:44
  The
• 5:44
  RSAC
• 5:45
  '26
• 5:46
  preview
• 5:46
  underscores
• 5:47
  the
• 5:47
  governance
• 5:47
  challenge:
• 5:48
  agentic
• 5:49
  actions
• 5:49
  are
• 5:49
  highly
• 5:49
  fluid,
• 5:50
  dynamic,
• 5:51
  and
• 5:51
  unpredictable,
• 5:52
  requiring
• 5:52
  defenses
• 5:53
  that
• 5:53
  are
• 5:53
  equally
• 5:53
  context-aware.
• 5:55
  Non-human
• 5:56
  identity
• 5:56
  management
• 5:57
  is
• 5:57
  emerging
• 5:58
  as
• 5:58
  a
• 5:58
  critical
• 5:58
  gap
• 5:58
  —
• 5:59
  agents
• 6:00
  currently
• 6:00
  proxy
• 6:01
  human
• 6:01
  identities,
• 6:02
  which
• 6:02
  does
• 6:02
  not
• 6:02
  scale.
• 6:04
  Just-in-time
• 6:04
  privilege
• 6:05
  management
• 6:05
  for
• 6:05
  agents
• 6:06
  that
• 6:06
  need
• 6:06
  legitimate
• 6:06
  access
• 6:07
  without
• 6:08
  standing
• 6:08
  in
• 6:08
  the
• 6:08
  way
• 6:08
  of
• 6:08
  productivity
• 6:09
  is
• 6:09
  the
• 6:09
  unsolved
• 6:10
  problem.
• 6:11
  The
• 6:11
  Cloud
• 6:11
  Security
• 6:12
  Alliance
• 6:12
  released
• 6:13
  a
• 6:13
  CISO
• 6:13
  memo
• 6:13
  template
• 6:14
  specifically
• 6:15
  for
• 6:15
  personal
• 6:15
  AI
• 6:16
  desktop
• 6:16
  agents,
• 6:17
  covering
• 6:17
  risk
• 6:17
  context,
• 6:18
  compliance
• 6:19
  considerations,
• 6:19
  and
• 6:20
  policy
• 6:20
  requirements
• 6:20
  .
• 6:25
  A
• 6:25
  detailed
• 6:25
  analysis
• 6:26
  of
• 6:26
  autonomous
• 6:27
  software
• 6:27
  development
• 6:27
  maps
• 6:28
  what
• 6:28
  the
• 6:28
  industry
• 6:28
  calls
• 6:29
  "the
• 6:29
  five
• 6:30
  levels
• 6:30
  of
• 6:30
  vibe
• 6:30
  coding,"
• 6:31
  from
• 6:31
  spicy
• 6:32
  autocomplete
• 6:32
  (Level
• 6:33
  0)
• 6:33
  to
• 6:33
  the
• 6:33
  dark
• 6:33
  factory
• 6:34
  (Level
• 6:34
  5)
• 6:34
  where
• 6:35
  no
• 6:35
  human
• 6:35
  writes
• 6:35
  or
• 6:36
  reviews
• 6:36
  code.
• 6:37
  The
• 6:37
  most
• 6:37
  documented
• 6:38
  Level
• 6:38
  5
• 6:38
  operation
• 6:39
  is
• 6:39
  StrongDM's
• 6:40
  software
• 6:40
  factory:
• 6:41
  three
• 6:41
  engineers,
• 6:42
  no
• 6:42
  sprints,
• 6:43
  no
• 6:43
  standups,
• 6:44
  no
• 6:44
  Jira
• 6:45
  board.
• 6:45
  They
• 6:46
  write
• 6:46
  markdown
• 6:46
  specifications;
• 6:47
  an
• 6:47
  agent
• 6:48
  called
• 6:48
  Attractor
• 6:48
  reads
• 6:49
  them,
• 6:49
  writes
• 6:49
  the
• 6:49
  code,
• 6:50
  and
• 6:50
  tests
• 6:50
  it.
• 6:51
  Critically,
• 6:52
  StrongDM
• 6:53
  does
• 6:53
  not
• 6:53
  use
• 6:53
  traditional
• 6:54
  tests
• 6:54
  —
• 6:55
  they
• 6:55
  use
• 6:55
  external
• 6:56
  "scenarios"
• 6:57
  that
• 6:57
  live
• 6:57
  outside
• 6:57
  the
• 6:58
  codebase,
• 6:58
  functioning
• 6:59
  as
• 6:59
  a
• 6:59
  holdout
• 6:59
  set
• 7:00
  the
• 7:00
  agent
• 7:00
  never
• 7:00
  sees
• 7:01
  during
• 7:01
  development,
• 7:02
  preventing
• 7:02
  the
• 7:03
  AI
• 7:03
  from
• 7:03
  optimizing
• 7:04
  for
• 7:04
  test
• 7:04
  passage
• 7:04
  rather
• 7:05
  than
• 7:05
  correctness.
• 7:06
  A
• 7:06
  "digital
• 7:07
  twin
• 7:07
  universe"
• 7:08
  of
• 7:08
  behavioral
• 7:08
  clones
• 7:09
  (simulated
• 7:09
  Okta,
• 7:10
  Jira,
• 7:11
  Slack,
• 7:12
  Google
• 7:12
  Docs)
• 7:12
  provides
• 7:13
  full
• 7:13
  integration
• 7:13
  testing
• 7:14
  without
• 7:14
  touching
• 7:14
  production.
• 7:15
  The
• 7:16
  output
• 7:16
  —
• 7:16
  CXDB,
• 7:18
  their
• 7:18
  AI
• 7:18
  context
• 7:19
  store
• 7:19
  —
• 7:20
  is
• 7:20
  16,000
• 7:21
  lines
• 7:21
  of
• 7:21
  Rust,
• 7:21
  9,500
• 7:22
  lines
• 7:22
  of
• 7:23
  Go,
• 7:23
  and
• 7:23
  700
• 7:24
  lines
• 7:24
  of
• 7:24
  TypeScript,
• 7:25
  all
• 7:25
  shipped
• 7:25
  in
• 7:25
  production.
• 7:26
  The
• 7:26
  numbers
• 7:27
  backing
• 7:27
  this
• 7:27
  up
• 7:27
  are
• 7:27
  real.
• 7:28
  Anthropic
• 7:28
  estimates
• 7:29
  that
• 7:29
  90%
• 7:29
  of
• 7:30
  Claude
• 7:30
  Code's
• 7:30
  codebase
• 7:30
  was
• 7:31
  written
• 7:31
  by
• 7:31
  Claude
• 7:31
  Code
• 7:31
  itself,
• 7:32
  with
• 7:32
  the
• 7:32
  company
• 7:32
  converging
• 7:33
  toward
• 7:33
  100%
• 7:34
  AI-generated
• 7:35
  code.
• 7:35
  OpenAI
• 7:36
  reports
• 7:36
  Codex
• 7:37
  5.3
• 7:37
  was
• 7:37
  instrumental
• 7:38
  in
• 7:38
  creating
• 7:38
  itself,
• 7:39
  yielding
• 7:39
  a
• 7:39
  25%
• 7:40
  speed
• 7:40
  improvement
• 7:41
  and
• 7:41
  93%
• 7:42
  fewer
• 7:42
  wasted
• 7:43
  tokens.
• 7:44
  Four
• 7:44
  percent
• 7:44
  of
• 7:44
  public
• 7:45
  GitHub
• 7:45
  commits
• 7:45
  are
• 7:46
  now
• 7:46
  directly
• 7:46
  authored
• 7:47
  by
• 7:47
  Claude
• 7:47
  Code,
• 7:48
  a
• 7:48
  figure
• 7:48
  Anthropic
• 7:49
  expects
• 7:49
  to
• 7:49
  exceed
• 7:50
  20%
• 7:50
  by
• 7:50
  year-end.
• 7:51
  And
• 7:52
  yet
• 7:52
  a
• 7:52
  rigorous
• 7:52
  2025
• 7:53
  METR
• 7:53
  randomized
• 7:54
  control
• 7:54
  trial
• 7:55
  found
• 7:55
  experienced
• 7:55
  open-source
• 7:56
  developers
• 7:57
  using
• 7:57
  AI
• 7:57
  tools
• 7:57
  completed
• 7:58
  tasks
• 7:58
  19%
• 7:59
  slower
• 7:59
  —
• 8:00
  while
• 8:00
  believing
• 8:00
  they
• 8:01
  were
• 8:01
  24%
• 8:01
  faster.
• 8:03
  The
• 8:03
  gap
• 8:03
  between
• 8:03
  frontier
• 8:04
  teams
• 8:04
  running
• 8:04
  dark
• 8:05
  factories
• 8:05
  and
• 8:05
  everyone
• 8:06
  else
• 8:06
  stuck
• 8:06
  at
• 8:06
  Level
• 8:06
  2,
• 8:07
  getting
• 8:07
  measurably
• 8:08
  slower
• 8:08
  while
• 8:08
  convinced
• 8:09
  they
• 8:09
  are
• 8:09
  speeding
• 8:09
  up,
• 8:10
  is
• 8:10
  the
• 8:10
  most
• 8:10
  important
• 8:11
  gap
• 8:11
  in
• 8:11
  tech
• 8:11
  right
• 8:11
  now.
• 8:12
  This
• 8:12
  tension
• 8:13
  has
• 8:13
  deep
• 8:13
  theoretical
• 8:14
  roots.
• 8:14
  Ikenna
• 8:15
  Okpala's
• 8:15
  series
• 8:16
  on
• 8:16
  probabilistic
• 8:16
  engineering
• 8:17
  in
• 8:17
  deterministic
• 8:18
  systems
• 8:18
  frames
• 8:18
  the
• 8:19
  core
• 8:19
  problem:
• 8:20
  software
• 8:20
  engineering
• 8:21
  was
• 8:21
  built
• 8:21
  to
• 8:22
  eliminate
• 8:22
  ambiguity
• 8:22
  —
• 8:24
  compilers,
• 8:24
  type
• 8:25
  systems,
• 8:25
  CI,
• 8:26
  code
• 8:27
  review,
• 8:27
  reproducible
• 8:28
  builds
• 8:28
  —
• 8:29
  and
• 8:29
  now
• 8:29
  we
• 8:29
  are
• 8:30
  injecting
• 8:30
  probabilistic
• 8:31
  engines
• 8:31
  and
• 8:32
  asking
• 8:32
  them
• 8:32
  to
• 8:32
  ship
• 8:33
  deterministic
• 8:33
  outcomes.
• 8:34
  If
• 8:35
  your
• 8:35
  evaluator
• 8:35
  is
• 8:35
  also
• 8:36
  probabilistic,
• 8:37
  you
• 8:37
  have
• 8:37
  not
• 8:37
  built
• 8:37
  a
• 8:37
  truth
• 8:38
  oracle;
• 8:39
  you
• 8:39
  have
• 8:39
  built
• 8:39
  a
• 8:39
  circular
• 8:40
  validation
• 8:40
  trap.
• 8:41
  Two
• 8:42
  models
• 8:42
  agreeing
• 8:42
  is
• 8:43
  correlation,
• 8:44
  not
• 8:44
  proof.
• 8:45
  Okpala's
• 8:45
  prescription:
• 8:46
  represent
• 8:47
  your
• 8:47
  codebase
• 8:47
  as
• 8:48
  a
• 8:48
  structure
• 8:48
  with
• 8:48
  constraints,
• 8:49
  measure
• 8:49
  whether
• 8:50
  a
• 8:50
  proposed
• 8:50
  change
• 8:51
  can
• 8:51
  exist
• 8:51
  inside
• 8:51
  that
• 8:52
  structure,
• 8:52
  and
• 8:53
  block
• 8:53
  violations
• 8:53
  before
• 8:54
  they
• 8:54
  land
• 8:54
  —
• 8:55
  governance
• 8:55
  through
• 8:56
  topology,
• 8:56
  not
• 8:57
  language.
• 9:01
  FastCode,
• 9:01
  from
• 9:02
  HKU's
• 9:02
  Data
• 9:02
  Science
• 9:03
  Lab,
• 9:03
  claims
• 9:03
  to
• 9:04
  deliver
• 9:04
  3-4x
• 9:05
  faster
• 9:05
  code
• 9:05
  understanding
• 9:06
  than
• 9:06
  Cursor
• 9:06
  or
• 9:06
  Claude
• 9:07
  Code
• 9:07
  at
• 9:07
  44-55%
• 9:09
  lower
• 9:09
  cost,
• 9:10
  using
• 9:10
  a
• 9:10
  three-phase
• 9:11
  framework:
• 9:12
  semantic-structural
• 9:12
  code
• 9:13
  representation
• 9:13
  (AST-based
• 9:14
  parsing
• 9:15
  across
• 9:15
  8+
• 9:15
  languages,
• 9:16
  hybrid
• 9:16
  BM25/embedding
• 9:18
  index,
• 9:19
  multi-layer
• 9:19
  graph
• 9:20
  modeling),
• 9:21
  lightning-fast
• 9:21
  navigation
• 9:22
  (two-stage
• 9:22
  search,
• 9:23
  code
• 9:23
  skimming
• 9:24
  via
• 9:24
  function
• 9:24
  signatures
• 9:25
  rather
• 9:25
  than
• 9:25
  full
• 9:25
  files),
• 9:26
  and
• 9:26
  budget-aware
• 9:27
  context
• 9:27
  management
• 9:27
  that
• 9:27
  considers
• 9:28
  confidence,
• 9:29
  query
• 9:29
  complexity,
• 9:30
  codebase
• 9:30
  size,
• 9:31
  cost,
• 9:31
  and
• 9:31
  iteration
• 9:32
  count
• 9:32
  before
• 9:32
  deciding
• 9:33
  what
• 9:33
  to
• 9:33
  load.
• 9:34
  The
• 9:34
  key
• 9:34
  insight
• 9:35
  is
• 9:35
  "scouting-first"
• 9:36
  —
• 9:37
  build
• 9:37
  a
• 9:37
  semantic
• 9:37
  map,
• 9:38
  navigate
• 9:38
  the
• 9:38
  structure,
• 9:39
  then
• 9:39
  load
• 9:39
  only
• 9:40
  targeted
• 9:40
  files,
• 9:41
  rather
• 9:41
  than
• 9:41
  the
• 9:41
  traditional
• 9:42
  approach
• 9:42
  of
• 9:42
  repeatedly
• 9:43
  loading
• 9:43
  and
• 9:43
  searching.
• 9:44
  It
• 9:44
  supports
• 9:45
  multi-repository
• 9:46
  reasoning
• 9:46
  and
• 9:46
  runs
• 9:47
  with
• 9:47
  local
• 9:47
  models
• 9:47
  like
• 9:48
  qwen3-coder-30b.
• 9:49
  The
• 9:49
  Claude
• 9:49
  Code
• 9:49
  ecosystem
• 9:50
  itself
• 9:51
  is
• 9:51
  generating
• 9:51
  a
• 9:51
  rich
• 9:52
  best-practices
• 9:52
  culture.
• 9:53
  A
• 9:54
  community-maintained
• 9:55
  guide
• 9:55
  crystallizes
• 9:55
  hard-won
• 9:56
  lessons:
• 9:57
  keep
• 9:57
  CLAUDE.md
• 9:58
  under
• 9:58
  150
• 9:59
  lines,
• 10:00
  use
• 10:00
  feature-specific
• 10:00
  subagents
• 10:01
  with
• 10:01
  skills
• 10:01
  instead
• 10:02
  of
• 10:02
  generic
• 10:02
  agents,
• 10:03
  do
• 10:03
  manual
• 10:03
  `/compact`
• 10:04
  at
• 10:04
  50%
• 10:05
  context,
• 10:06
  always
• 10:06
  start
• 10:06
  with
• 10:06
  plan
• 10:07
  mode,
• 10:07
  make
• 10:08
  subtasks
• 10:08
  small
• 10:08
  enough
• 10:09
  to
• 10:09
  complete
• 10:09
  in
• 10:09
  under
• 10:10
  50%
• 10:10
  context,
• 10:11
  and
• 10:11
  commit
• 10:12
  as
• 10:12
  soon
• 10:12
  as
• 10:12
  each
• 10:12
  task
• 10:13
  completes.
• 10:14
  The
• 10:14
  guide
• 10:14
  also
• 10:14
  catalogs
• 10:15
  a
• 10:15
  command-skill-agent
• 10:16
  architecture
• 10:16
  for
• 10:17
  progressive
• 10:17
  disclosure:
• 10:18
  commands
• 10:19
  as
• 10:19
  entry
• 10:19
  points,
• 10:20
  agents
• 10:20
  as
• 10:20
  orchestrators
• 10:21
  with
• 10:21
  preloaded
• 10:21
  skills,
• 10:22
  and
• 10:22
  skills
• 10:23
  as
• 10:23
  domain
• 10:23
  knowledge
• 10:24
  injected
• 10:24
  at
• 10:24
  startup.
• 10:25
  Power
• 10:25
  users
• 10:26
  running
• 10:26
  multiple
• 10:27
  parallel
• 10:27
  Claude
• 10:27
  Code
• 10:27
  sessions
• 10:28
  report
• 10:28
  that
• 10:28
  the
• 10:28
  biggest
• 10:29
  workflow
• 10:29
  gap
• 10:29
  is
• 10:29
  reviewing
• 10:30
  what
• 10:30
  the
• 10:30
  agent
• 10:30
  built
• 10:31
  across
• 10:31
  many
• 10:31
  files
• 10:31
  —
• 10:32
  scrolling
• 10:33
  through
• 10:33
  terminal
• 10:33
  output
• 10:33
  and
• 10:34
  git
• 10:34
  logs
• 10:34
  to
• 10:34
  piece
• 10:34
  together
• 10:35
  changes.
• 10:36
  SonarSource's
• 10:37
  Kintsugi,
• 10:38
  an
• 10:38
  Agentic
• 10:38
  Development
• 10:39
  Environment
• 10:39
  built
• 10:39
  on
• 10:39
  top
• 10:40
  of
• 10:40
  Claude
• 10:40
  Code
• 10:40
  rather
• 10:41
  than
• 10:41
  replacing
• 10:41
  it,
• 10:42
  addresses
• 10:42
  this
• 10:42
  with
• 10:42
  a
• 10:43
  visual
• 10:43
  queue
• 10:43
  for
• 10:43
  managing
• 10:44
  parallel
• 10:44
  sessions,
• 10:45
  plan
• 10:45
  review
• 10:46
  before
• 10:46
  execution,
• 10:47
  inline
• 10:47
  comments,
• 10:48
  and
• 10:48
  integrated
• 10:48
  code
• 10:49
  review
• 10:49
  powered
• 10:49
  by
• 10:49
  SonarQube's
• 10:50
  quality
• 10:50
  engine.
• 10:52
  GitNexus
• 10:52
  takes
• 10:52
  a
• 10:52
  different
• 10:53
  approach
• 10:53
  to
• 10:53
  AI
• 10:53
  code
• 10:54
  intelligence:
• 10:55
  it
• 10:55
  turns
• 10:55
  a
• 10:55
  repository
• 10:56
  into
• 10:56
  an
• 10:57
  AST-driven
• 10:57
  knowledge
• 10:58
  graph
• 10:58
  directly
• 10:58
  in
• 10:58
  the
• 10:59
  browser
• 10:59
  using
• 10:59
  Tree-sitter
• 11:00
  for
• 11:00
  parsing
• 11:00
  and
• 11:00
  KuzuDB
• 11:01
  for
• 11:01
  storage,
• 11:02
  with
• 11:02
  a
• 11:02
  local
• 11:03
  model
• 11:03
  generating
• 11:04
  embeddings
• 11:04
  for
• 11:04
  semantic
• 11:05
  search
• 11:05
  over
• 11:05
  files,
• 11:06
  symbols,
• 11:06
  and
• 11:06
  call
• 11:07
  chains.
• 11:08
  Every
• 11:08
  step
• 11:08
  runs
• 11:08
  client-side
• 11:09
  —
• 11:09
  code
• 11:10
  never
• 11:10
  leaves
• 11:10
  the
• 11:10
  machine
• 11:11
  —
• 11:11
  while
• 11:12
  exposing
• 11:12
  an
• 11:12
  MCP
• 11:13
  (Model
• 11:13
  Context
• 11:13
  Protocol)
• 11:14
  interface
• 11:14
  so
• 11:15
  agents
• 11:15
  like
• 11:15
  Claude
• 11:15
  Code
• 11:16
  can
• 11:16
  request
• 11:16
  relational
• 11:17
  context
• 11:18
  ("all
• 11:18
  call
• 11:18
  paths
• 11:19
  touching
• 11:19
  this
• 11:19
  function
• 11:19
  across
• 11:20
  services")
• 11:21
  instead
• 11:21
  of
• 11:21
  streaming
• 11:21
  arbitrary
• 11:22
  files
• 11:22
  into
• 11:22
  context
• 11:23
  windows.
• 11:25
  Daily
• 11:26
  released
• 11:26
  an
• 11:26
  open-source
• 11:27
  benchmark
• 11:27
  (`aiewf-eval`)
• 11:28
  specifically
• 11:28
  targeting
• 11:29
  LLM
• 11:29
  performance
• 11:30
  for
• 11:30
  voice
• 11:30
  agents
• 11:30
  —
• 11:31
  testing
• 11:31
  tool
• 11:31
  calling,
• 11:32
  instruction
• 11:32
  following,
• 11:33
  and
• 11:33
  factual
• 11:34
  grounding
• 11:34
  over
• 11:34
  30-turn
• 11:35
  conversations,
• 11:36
  which
• 11:36
  standard
• 11:36
  benchmarks
• 11:37
  do
• 11:37
  not
• 11:37
  cover.
• 11:38
  The
• 11:38
  headline
• 11:38
  finding:
• 11:39
  three
• 11:39
  models
• 11:40
  now
• 11:40
  saturate
• 11:40
  the
• 11:40
  benchmark
• 11:41
  at
• 11:41
  100%
• 11:41
  on
• 11:42
  every
• 11:42
  metric,
• 11:42
  but
• 11:43
  all
• 11:43
  three
• 11:43
  are
• 11:43
  too
• 11:43
  slow
• 11:44
  for
• 11:44
  production
• 11:44
  voice
• 11:44
  agents.
• 11:45
  Natural
• 11:46
  conversation
• 11:46
  requires
• 11:47
  voice-to-voice
• 11:47
  response
• 11:48
  times
• 11:48
  under
• 11:48
  1,500ms,
• 11:50
  translating
• 11:51
  to
• 11:51
  roughly
• 11:51
  700ms
• 11:52
  time-to-first-token
• 11:53
  for
• 11:53
  text-mode
• 11:53
  LLMs.
• 11:55
  Most
• 11:55
  production
• 11:55
  voice
• 11:56
  agents
• 11:56
  today
• 11:56
  still
• 11:56
  run
• 11:57
  on
• 11:57
  GPT-4.1
• 11:58
  or
• 11:58
  Gemini
• 11:58
  2.5
• 11:59
  Flash
• 11:59
  —
• 12:00
  models
• 12:00
  released
• 12:01
  in
• 12:01
  April
• 12:01
  2025,
• 12:02
  ancient
• 12:02
  by
• 12:02
  AI
• 12:03
  timelines.
• 12:04
  AWS
• 12:04
  Nova
• 12:04
  2
• 12:05
  Pro
• 12:05
  now
• 12:05
  matches
• 12:05
  their
• 12:06
  performance
• 12:06
  and
• 12:06
  latency,
• 12:07
  bringing
• 12:08
  fully
• 12:08
  capable
• 12:08
  enterprise
• 12:09
  voice
• 12:09
  agents
• 12:09
  onto
• 12:09
  AWS
• 12:10
  for
• 12:10
  the
• 12:10
  first
• 12:10
  time.
• 12:11
  On
• 12:12
  the
• 12:12
  speech-to-speech
• 12:12
  front,
• 12:13
  Ultravox
• 12:13
  is
• 12:13
  the
• 12:14
  first
• 12:14
  model
• 12:14
  to
• 12:14
  perform
• 12:15
  well
• 12:15
  on
• 12:15
  long
• 12:15
  multi-turn
• 12:16
  benchmarks,
• 12:17
  and
• 12:17
  it
• 12:17
  is
• 12:17
  open-weights
• 12:18
  —
• 12:19
  30-billion-parameter
• 12:19
  Nemotron
• 12:20
  3
• 12:20
  Nano
• 12:20
  nearly
• 12:20
  matches
• 12:21
  GPT-4o
• 12:22
  performance,
• 12:23
  validating
• 12:23
  the
• 12:23
  thesis
• 12:24
  that
• 12:24
  open
• 12:24
  weights
• 12:24
  will
• 12:24
  gain
• 12:25
  significant
• 12:25
  market
• 12:25
  share
• 12:26
  in
• 12:26
  2026.
• 12:27
  Claude
• 12:27
  Opus
• 12:28
  4.6
• 12:28
  surged
• 12:29
  past
• 12:29
  forecasts
• 12:30
  on
• 12:30
  METR's
• 12:30
  50%
• 12:31
  Time-Horizon
• 12:31
  Benchmark,
• 12:32
  which
• 12:33
  measures
• 12:33
  how
• 12:33
  long
• 12:33
  (in
• 12:33
  human-equivalent
• 12:34
  hours)
• 12:34
  a
• 12:34
  software
• 12:35
  task
• 12:35
  can
• 12:35
  be
• 12:35
  that
• 12:36
  an
• 12:36
  LLM
• 12:36
  completes
• 12:37
  successfully
• 12:37
  half
• 12:37
  the
• 12:37
  time.
• 12:39
  Opus
• 12:39
  4.6
• 12:40
  leads
• 12:40
  at
• 12:40
  roughly
• 12:40
  13+
• 12:41
  hours
• 12:41
  —
• 12:42
  meaning
• 12:43
  it
• 12:43
  can
• 12:43
  reliably
• 12:43
  complete
• 12:44
  tasks
• 12:44
  that
• 12:44
  would
• 12:44
  take
• 12:45
  a
• 12:45
  skilled
• 12:45
  human
• 12:45
  over
• 12:45
  half
• 12:46
  a
• 12:46
  day
• 12:46
  —
• 12:47
  a
• 12:47
  massive
• 12:47
  jump
• 12:47
  from
• 12:48
  GPT-5.2
• 12:49
  at
• 12:49
  ~7
• 12:49
  hours
• 12:50
  and
• 12:50
  Opus
• 12:50
  4.5
• 12:51
  at
• 12:51
  ~5
• 12:51
  hours.
• 12:52
  Earlier
• 12:53
  models
• 12:53
  clustered
• 12:53
  near
• 12:54
  30
• 12:54
  minutes
• 12:54
  to
• 12:54
  1
• 12:55
  hour.
• 12:56
  The
• 12:56
  trajectory
• 12:56
  is
• 12:57
  clearly
• 12:57
  exponential,
• 12:58
  with
• 12:58
  models
• 12:59
  doubling
• 12:59
  effective
• 12:59
  task
• 13:00
  complexity
• 13:00
  every
• 13:01
  few
• 13:01
  months.
• 13:02
  On
• 13:02
  the
• 13:02
  open-weight
• 13:03
  side,
• 13:03
  MiniMax-M2.5
• 13:05
  with
• 13:05
  10B
• 13:05
  active
• 13:06
  parameters
• 13:06
  achieves
• 13:07
  80.2%
• 13:08
  on
• 13:08
  Verified,
• 13:09
  51.3%
• 13:10
  on
• 13:10
  Multi-SWE-Bench,
• 13:12
  and
• 13:12
  76.3%
• 13:13
  on
• 13:13
  BrowseComp,
• 13:14
  with
• 13:15
  Unsloth's
• 13:15
  dynamic
• 13:16
  GGUF
• 13:16
  quantization
• 13:17
  compressing
• 13:17
  the
• 13:17
  model
• 13:18
  to
• 13:18
  fit
• 13:18
  on
• 13:18
  consumer
• 13:19
  hardware
• 13:19
  —
• 13:20
  the
• 13:20
  3-bit
• 13:20
  dynamic
• 13:21
  quant
• 13:21
  runs
• 13:21
  at
• 13:21
  20+
• 13:22
  tokens/s
• 13:23
  on
• 13:23
  128GB
• 13:25
  RAM.
• 13:29
  Jolt
• 13:30
  Atlas,
• 13:30
  a
• 13:30
  zero-knowledge
• 13:31
  machine
• 13:31
  learning
• 13:31
  (zkML)
• 13:32
  framework
• 13:32
  from
• 13:32
  a
• 13:33
  team
• 13:33
  building
• 13:33
  on
• 13:33
  the
• 13:33
  Jolt
• 13:34
  proving
• 13:34
  system,
• 13:35
  tackles
• 13:35
  one
• 13:35
  of
• 13:35
  the
• 13:35
  hardest
• 13:36
  open
• 13:36
  problems
• 13:36
  in
• 13:36
  trustworthy
• 13:37
  AI:
• 13:37
  cryptographic
• 13:38
  proof
• 13:38
  that
• 13:38
  a
• 13:38
  model
• 13:39
  inference
• 13:39
  actually
• 13:40
  ran
• 13:40
  the
• 13:40
  claimed
• 13:40
  model
• 13:40
  on
• 13:40
  the
• 13:40
  claimed
• 13:41
  inputs,
• 13:41
  without
• 13:41
  revealing
• 13:42
  model
• 13:42
  weights
• 13:42
  or
• 13:42
  input
• 13:43
  data.
• 13:43
  Unlike
• 13:44
  zkVMs
• 13:45
  that
• 13:45
  emulate
• 13:45
  CPU
• 13:46
  instruction
• 13:46
  execution,
• 13:47
  Jolt
• 13:47
  Atlas
• 13:47
  operates
• 13:48
  directly
• 13:48
  on
• 13:48
  ONNX
• 13:49
  tensor
• 13:49
  operations
• 13:49
  —
• 13:50
  the
• 13:50
  open-source,
• 13:51
  portable
• 13:51
  format
• 13:51
  used
• 13:52
  across
• 13:52
  PyTorch,
• 13:52
  TensorFlow,
• 13:53
  and
• 13:53
  JAX.
• 13:54
  The
• 13:54
  key
• 13:54
  innovation
• 13:54
  is
• 13:54
  using
• 13:55
  lookup
• 13:55
  arguments
• 13:55
  via
• 13:56
  the
• 13:56
  sumcheck
• 13:56
  protocol
• 13:57
  for
• 13:57
  non-linear
• 13:57
  functions
• 13:58
  (Softmax,
• 13:58
  ReLU),
• 13:59
  which
• 13:59
  are
• 13:59
  the
• 13:59
  dominant
• 13:59
  cost
• 14:00
  in
• 14:00
  zero-knowledge
• 14:00
  proofs
• 14:01
  of
• 14:01
  neural
• 14:01
  networks.
• 14:02
  Neural
• 14:02
  teleportation
• 14:03
  compresses
• 14:03
  activation
• 14:04
  lookup
• 14:04
  tables
• 14:04
  by
• 14:05
  shrinking
• 14:05
  the
• 14:05
  effective
• 14:05
  input
• 14:06
  range,
• 14:06
  and
• 14:06
  prefix-suffix
• 14:07
  decomposition
• 14:08
  enables
• 14:08
  streaming
• 14:08
  provers
• 14:09
  that
• 14:09
  reduce
• 14:09
  peak
• 14:09
  memory
• 14:10
  from
• 14:10
  O(N)
• 14:10
  to
• 14:10
  roughly
• 14:10
  O(sqrt(N)),
• 14:11
  enabling
• 14:12
  on-device
• 14:12
  proving
• 14:13
  without
• 14:13
  specialized
• 14:13
  hardware.
• 14:14
  Benchmarks
• 14:15
  show
• 14:15
  Jolt
• 14:15
  Atlas
• 14:16
  proving
• 14:16
  nanoGPT
• 14:17
  inference
• 14:17
  in
• 14:18
  14
• 14:18
  seconds
• 14:18
  versus
• 14:19
  ezkl's
• 14:19
  237
• 14:20
  seconds
• 14:20
  —
• 14:21
  a
• 14:21
  17x
• 14:22
  speedup
• 14:22
  on
• 14:22
  proof
• 14:23
  generation
• 14:23
  alone.
• 14:24
  The
• 14:24
  companion
• 14:25
  paper
• 14:25
  outlines
• 14:25
  use
• 14:25
  cases
• 14:26
  for
• 14:26
  agentic
• 14:26
  commerce
• 14:27
  guardrails,
• 14:28
  where
• 14:28
  autonomous
• 14:28
  AI
• 14:28
  agents
• 14:29
  making
• 14:29
  financial
• 14:29
  decisions
• 14:30
  must
• 14:30
  provide
• 14:31
  cryptographic
• 14:31
  proof
• 14:31
  of
• 14:31
  what
• 14:32
  model
• 14:32
  produced
• 14:32
  which
• 14:32
  recommendation.
• 14:33
  At
• 14:33
  the
• 14:33
  infrastructure
• 14:34
  layer,
• 14:34
  RTI's
• 14:35
  GENESIS
• 14:35
  framework
• 14:35
  brings
• 14:36
  DDS
• 14:36
  (Data
• 14:36
  Distribution
• 14:37
  Service)
• 14:37
  —
• 14:38
  the
• 14:38
  same
• 14:38
  middleware
• 14:38
  powering
• 14:39
  flight
• 14:39
  control
• 14:39
  systems,
• 14:40
  FDA-cleared
• 14:41
  surgical
• 14:41
  robots,
• 14:42
  and
• 14:42
  autonomous
• 14:42
  vehicle
• 14:43
  sensor
• 14:43
  fusion
• 14:43
  —
• 14:44
  to
• 14:44
  distributed
• 14:45
  AI
• 14:45
  agent
• 14:46
  networks.
• 14:47
  The
• 14:47
  key
• 14:47
  feature
• 14:47
  is
• 14:47
  zero-configuration
• 14:48
  discovery:
• 14:49
  agents
• 14:50
  find
• 14:50
  each
• 14:50
  other
• 14:50
  automatically
• 14:51
  without
• 14:51
  IP
• 14:51
  addresses,
• 14:52
  ports,
• 14:52
  or
• 14:52
  configuration
• 14:53
  files,
• 14:54
  communicating
• 14:54
  peer-to-peer
• 14:54
  with
• 14:55
  sub-millisecond
• 14:55
  latency.
• 14:57
  Intelligent
• 14:57
  function
• 14:57
  windowing
• 14:58
  uses
• 14:58
  classifiers
• 14:59
  to
• 14:59
  select
• 14:59
  the
• 14:59
  5-10
• 14:59
  relevant
• 15:00
  functions
• 15:00
  from
• 15:00
  200+
• 15:01
  discovered
• 15:01
  tools,
• 15:02
  achieving
• 15:03
  90%+
• 15:03
  reduction
• 15:04
  in
• 15:04
  LLM
• 15:04
  tokens.
• 15:05
  For
• 15:05
  enterprises
• 15:06
  deploying
• 15:06
  multi-agent
• 15:07
  systems
• 15:07
  in
• 15:08
  production,
• 15:08
  the
• 15:09
  reliability
• 15:09
  guarantees
• 15:10
  inherited
• 15:10
  from
• 15:10
  safety-critical
• 15:11
  domains
• 15:11
  address
• 15:12
  a
• 15:12
  gap
• 15:12
  that
• 15:12
  HTTP-based
• 15:13
  agent
• 15:13
  frameworks
• 15:14
  have
• 15:14
  not
• 15:14
  solved.
• 15:16
  At
• 15:16
  a
• 15:16
  different
• 15:16
  layer
• 15:16
  of
• 15:16
  the
• 15:16
  stack,
• 15:17
  the
• 15:18
  RuVector
• 15:18
  ecosystem
• 15:18
  continues
• 15:19
  to
• 15:19
  push
• 15:19
  the
• 15:19
  single-file-does-everything
• 15:20
  thesis.
• 15:22
  A
• 15:22
  test
• 15:22
  of
• 15:22
  the
• 15:22
  RVF
• 15:23
  format
• 15:23
  on
• 15:24
  a
• 15:24
  31,000-word
• 15:25
  text
• 15:26
  ingested
• 15:26
  263
• 15:27
  vectors,
• 15:28
  a
• 15:28
  knowledge
• 15:28
  graph
• 15:29
  of
• 15:29
  306
• 15:30
  nodes
• 15:30
  with
• 15:30
  1,402
• 15:32
  edges,
• 15:32
  and
• 15:32
  GNN
• 15:33
  weights
• 15:33
  with
• 15:33
  4-head
• 15:34
  attention
• 15:34
  —
• 15:35
  all
• 15:35
  in
• 15:35
  a
• 15:35
  single
• 15:36
  137KB
• 15:37
  file
• 15:38
  with
• 15:38
  a
• 15:38
  cryptographic
• 15:38
  SHAKE-256
• 15:43
  witness
• 15:44
  chain
• 15:44
  recording
• 15:44
  every
• 15:44
  operation.
• 15:46
  Copy-on-write
• 15:46
  snapshots
• 15:47
  yielded
• 15:47
  an
• 15:47
  847:1
• 15:49
  compression
• 15:49
  ratio
• 15:49
  for
• 15:50
  version
• 15:50
  control
• 15:50
  on
• 15:50
  AI
• 15:51
  artifacts.
• 15:52
  The
• 15:52
  rvDNA
• 15:53
  toolkit
• 15:53
  applied
• 15:54
  similar
• 15:54
  principles
• 15:54
  to
• 15:55
  genomics,
• 15:56
  processing
• 15:56
  596,007
• 15:58
  markers
• 15:58
  from
• 15:58
  a
• 15:59
  23andMe
• 15:59
  export
• 16:00
  in
• 16:00
  about
• 16:00
  a
• 16:00
  second
• 16:00
  —
• 16:01
  no
• 16:01
  cloud,
• 16:02
  no
• 16:02
  GPU
• 16:02
  —
• 16:03
  with
• 16:04
  conservative
• 16:04
  CYP
• 16:05
  star
• 16:05
  allele
• 16:05
  calling
• 16:05
  and
• 16:06
  drug
• 16:06
  recommendations
• 16:06
  suppressed
• 16:07
  unless
• 16:07
  evidence
• 16:07
  reaches
• 16:08
  at
• 16:08
  least
• 16:08
  Moderate
• 16:09
  confidence.