Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – February 23, 2026

Transcript

AGI Dreams Podcast – February 23, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Prompt
• 0:00
  injection
• 0:00
  has
• 0:00
  held
• 0:01
  the
• 0:01
  number-one
• 0:01
  slot
• 0:01
  on
• 0:02
  the
• 0:02
  OWASP
• 0:02
  LLM
• 0:03
  Top
• 0:03
  10
• 0:03
  for
• 0:03
  long
• 0:03
  enough
• 0:04
  that
• 0:04
  it
• 0:04
  should
• 0:04
  feel
• 0:04
  familiar.
• 0:05
  It
• 0:05
  does
• 0:05
  not.
• 0:06
  A
• 0:06
  comprehensive
• 0:07
  new
• 0:07
  guide
• 0:07
  from
• 0:07
  Bugcrowd
• 0:08
  walks
• 0:08
  security
• 0:08
  engineers
• 0:09
  through
• 0:09
  the
• 0:09
  full
• 0:09
  attack
• 0:10
  surface
• 0:10
  of
• 0:10
  LLM-integrated
• 0:11
  systems,
• 0:12
  and
• 0:12
  the
• 0:12
  two
• 0:12
  real-world
• 0:13
  case
• 0:13
  studies
• 0:13
  it
• 0:13
  includes
• 0:14
  are
• 0:14
  worth
• 0:14
  the
• 0:14
  read
• 0:15
  alone.
• 0:16
  In
• 0:16
  the
• 0:16
  first,
• 0:16
  a
• 0:16
  pen
• 0:17
  tester
• 0:17
  coerced
• 0:17
  a
• 0:17
  chat
• 0:18
  assistant's
• 0:18
  unrestricted
• 0:19
  HTTP
• 0:19
  tool
• 0:20
  into
• 0:20
  fetching
• 0:20
  the
• 0:20
  AWS
• 0:21
  Instance
• 0:21
  Metadata
• 0:22
  Service
• 0:22
  at
• 0:22
  `169.254.169.254`,
• 0:27
  harvesting
• 0:27
  temporary
• 0:28
  IAM
• 0:28
  credentials
• 0:29
  --
• 0:29
  access
• 0:30
  key
• 0:30
  ID,
• 0:30
  secret
• 0:31
  key,
• 0:31
  session
• 0:31
  token
• 0:31
  --
• 0:32
  through
• 0:32
  nothing
• 0:32
  more
• 0:33
  than
• 0:33
  a
• 0:33
  conversational
• 0:33
  prompt.
• 0:35
  In
• 0:35
  the
• 0:35
  second,
• 0:36
  a
• 0:36
  poisoned
• 0:36
  PDF
• 0:37
  uploaded
• 0:37
  to
• 0:37
  a
• 0:37
  RAG-based
• 0:38
  document
• 0:38
  assistant
• 0:39
  silently
• 0:39
  overrode
• 0:39
  the
• 0:40
  system
• 0:40
  prompt,
• 0:41
  causing
• 0:41
  the
• 0:41
  model
• 0:41
  to
• 0:41
  leak
• 0:42
  nonpublic
• 0:42
  internal
• 0:43
  API
• 0:43
  URLs
• 0:44
  and
• 0:44
  staging
• 0:44
  environment
• 0:45
  details
• 0:45
  that
• 0:45
  the
• 0:45
  attacker
• 0:46
  then
• 0:46
  used
• 0:46
  to
• 0:46
  enumerate
• 0:47
  less-secured
• 0:47
  infrastructure.
• 0:49
  The
• 0:49
  core
• 0:49
  problem
• 0:49
  remains
• 0:50
  unchanged:
• 0:51
  "An
• 0:51
  LLM
• 0:51
  doesn't
• 0:52
  inherently
• 0:52
  distinguish"
• 0:52
  instructions
• 0:53
  from
• 0:53
  data;
• 0:54
  "everything
• 0:54
  it
• 0:54
  sees
• 0:55
  is
• 0:55
  just
• 0:55
  text
• 0:55
  that
• 0:55
  might
• 0:56
  contain
• 0:56
  instructions."
• 0:57
  The
• 0:57
  guide's
• 0:58
  testing
• 0:58
  methodology
• 0:58
  --
• 0:59
  map
• 0:59
  what
• 0:59
  the
• 1:00
  model
• 1:00
  can
• 1:00
  call,
• 1:01
  what
• 1:01
  data
• 1:01
  it
• 1:01
  reads,
• 1:02
  what
• 1:02
  happens
• 1:02
  to
• 1:02
  its
• 1:02
  output,
• 1:03
  then
• 1:03
  probe
• 1:04
  systematically
• 1:04
  across
• 1:04
  direct
• 1:05
  injection,
• 1:06
  indirect
• 1:06
  injection,
• 1:07
  and
• 1:07
  insecure
• 1:07
  output
• 1:08
  handling
• 1:08
  --
• 1:09
  is
• 1:09
  the
• 1:09
  closest
• 1:09
  thing
• 1:10
  the
• 1:10
  field
• 1:10
  has
• 1:10
  to
• 1:10
  a
• 1:10
  standardized
• 1:11
  playbook.
• 1:12
  A
• 1:12
  much
• 1:12
  larger
• 1:13
  effort
• 1:13
  to
• 1:13
  formalize
• 1:13
  this
• 1:14
  problem
• 1:14
  arrived
• 1:14
  from
• 1:15
  Google,
• 1:16
  UC
• 1:16
  San
• 1:16
  Diego,
• 1:17
  UW-Madison,
• 1:18
  Meta
• 1:18
  FAIR,
• 1:19
  Gray
• 1:19
  Swan
• 1:20
  AI,
• 1:20
  and
• 1:21
  Cornell
• 1:21
  in
• 1:21
  an
• 1:21
  updated
• 1:22
  Systematization
• 1:23
  of
• 1:23
  Knowledge
• 1:23
  paper.
• 1:24
  "Systems
• 1:24
  Security
• 1:25
  Foundations
• 1:25
  for
• 1:25
  Agentic
• 1:26
  Computing"
• 1:26
  identifies
• 1:27
  four
• 1:27
  fundamental
• 1:27
  challenges
• 1:28
  when
• 1:28
  applying
• 1:28
  classical
• 1:29
  security
• 1:29
  engineering
• 1:30
  to
• 1:30
  agents.
• 1:31
  First,
• 1:32
  the
• 1:32
  Trusted
• 1:32
  Computing
• 1:32
  Base
• 1:33
  is
• 1:33
  probabilistic
• 1:33
  --
• 1:34
  "a
• 1:34
  probabilistic
• 1:35
  reference
• 1:35
  monitor
• 1:36
  might
• 1:36
  correctly
• 1:36
  deny
• 1:37
  unauthorized
• 1:37
  access
• 1:38
  99%
• 1:39
  of
• 1:39
  the
• 1:39
  time,
• 1:39
  but
• 1:39
  the
• 1:39
  remaining
• 1:40
  1%
• 1:40
  can
• 1:40
  be
• 1:40
  exploited."
• 1:42
  Second,
• 1:42
  security
• 1:43
  policies
• 1:43
  are
• 1:44
  dynamic
• 1:44
  and
• 1:44
  task-specific;
• 1:46
  unlike
• 1:46
  Android
• 1:46
  apps
• 1:47
  with
• 1:47
  fixed
• 1:47
  permission
• 1:48
  sets,
• 1:48
  agents
• 1:49
  synthesize
• 1:49
  programs
• 1:50
  on
• 1:50
  the
• 1:50
  fly
• 1:50
  from
• 1:50
  natural-language
• 1:51
  prompts.
• 1:52
  Third,
• 1:53
  the
• 1:53
  security
• 1:53
  boundary
• 1:53
  is
• 1:54
  fuzzy:
• 1:54
  agents
• 1:55
  lack
• 1:55
  the
• 1:55
  layered
• 1:56
  architectures
• 1:56
  (hardware,
• 1:57
  kernel,
• 1:58
  process)
• 1:58
  that
• 1:58
  traditional
• 1:59
  systems
• 1:59
  use
• 1:59
  for
• 2:00
  enforcement.
• 2:01
  Fourth,
• 2:02
  prompt
• 2:02
  injection
• 2:03
  is
• 2:03
  functionally
• 2:03
  analogous
• 2:04
  to
• 2:04
  dynamic
• 2:04
  code
• 2:05
  loading
• 2:05
  --
• 2:06
  "a
• 2:06
  very
• 2:06
  challenging
• 2:07
  problem
• 2:07
  that
• 2:07
  remains
• 2:08
  difficult
• 2:08
  to
• 2:08
  address
• 2:09
  today."
• 2:10
  The
• 2:10
  paper
• 2:10
  catalogs
• 2:11
  11
• 2:11
  real-world
• 2:11
  attacks
• 2:12
  on
• 2:12
  deployed
• 2:12
  systems,
• 2:13
  including
• 2:14
  Microsoft
• 2:14
  365
• 2:15
  Copilot
• 2:16
  data
• 2:16
  exfiltration
• 2:17
  via
• 2:17
  ASCII
• 2:17
  smuggling
• 2:18
  and
• 2:18
  Claude
• 2:18
  Code
• 2:19
  DNS
• 2:19
  exfiltration
• 2:20
  via
• 2:20
  the
• 2:20
  ping
• 2:20
  command,
• 2:21
  mapping
• 2:22
  each
• 2:22
  to
• 2:22
  violated
• 2:22
  classical
• 2:23
  security
• 2:23
  principles.
• 2:25
  The
• 2:25
  conclusion
• 2:25
  is
• 2:25
  measured
• 2:26
  but
• 2:26
  firm:
• 2:27
  "traditional
• 2:27
  security
• 2:28
  principles
• 2:28
  and
• 2:28
  techniques
• 2:29
  still
• 2:30
  apply
• 2:30
  and
• 2:30
  can
• 2:30
  thwart
• 2:31
  many
• 2:31
  attacks
• 2:31
  on
• 2:31
  agentic
• 2:32
  systems,
• 2:33
  but
• 2:33
  some
• 2:33
  attack
• 2:34
  classes
• 2:34
  expose
• 2:35
  genuinely
• 2:35
  new
• 2:35
  problems
• 2:36
  that
• 2:36
  require
• 2:36
  new
• 2:36
  solutions."
• 2:38
  On
• 2:38
  the
• 2:38
  offensive
• 2:39
  tooling
• 2:39
  side,
• 2:40
  Gadi
• 2:40
  Evron's
• 2:40
  Exploitation
• 2:41
  Validator
• 2:42
  takes
• 2:42
  a
• 2:42
  refreshingly
• 2:43
  honest
• 2:43
  approach
• 2:44
  to
• 2:44
  the
• 2:44
  hallucination
• 2:45
  problem
• 2:45
  in
• 2:45
  LLM-driven
• 2:46
  vulnerability
• 2:46
  scanning.
• 2:47
  The
• 2:47
  pipeline
• 2:48
  runs
• 2:48
  through
• 2:48
  five
• 2:49
  stages
• 2:49
  --
• 2:50
  inventory,
• 2:51
  one-shot
• 2:51
  exploitability
• 2:52
  check,
• 2:52
  systematic
• 2:53
  attack-tree
• 2:53
  analysis,
• 2:54
  sanity
• 2:55
  validation
• 2:55
  against
• 2:56
  actual
• 2:56
  code
• 2:57
  ("file
• 2:57
  exists,
• 2:58
  path
• 2:58
  correct,
• 2:59
  lines
• 2:59
  accurate,
• 3:00
  code
• 3:00
  verbatim,
• 3:01
  flow
• 3:01
  is
• 3:01
  real"),
• 3:02
  and
• 3:02
  a
• 3:02
  final
• 3:02
  ruling
• 3:03
  that
• 3:03
  filters
• 3:03
  out
• 3:03
  findings
• 3:04
  with
• 3:04
  hedging
• 3:04
  language
• 3:05
  or
• 3:05
  unrealistic
• 3:06
  preconditions.
• 3:07
  The
• 3:07
  core
• 3:07
  design
• 3:08
  principle:
• 3:09
  "Assume
• 3:09
  exploitable
• 3:10
  until
• 3:10
  proven
• 3:10
  otherwise."
• 3:11
  It
• 3:12
  is
• 3:12
  a
• 3:12
  useful
• 3:12
  corrective
• 3:13
  to
• 3:13
  the
• 3:13
  wave
• 3:13
  of
• 3:13
  AI
• 3:14
  security
• 3:14
  tools
• 3:14
  that
• 3:14
  report
• 3:15
  findings
• 3:15
  without
• 3:15
  proving
• 3:16
  them.
• 3:17
  A
• 3:17
  related
• 3:17
  arXiv
• 3:18
  paper
• 3:18
  examines
• 3:18
  what
• 3:19
  breaks
• 3:19
  embodied
• 3:19
  AI
• 3:20
  security
• 3:20
  when
• 3:20
  LLM
• 3:21
  vulnerabilities
• 3:21
  intersect
• 3:22
  with
• 3:22
  cyber-physical
• 3:23
  system
• 3:23
  flaws,
• 3:24
  though
• 3:24
  the
• 3:24
  full
• 3:24
  paper
• 3:24
  was
• 3:25
  not
• 3:25
  yet
• 3:25
  available
• 3:25
  for
• 3:25
  analysis
• 3:26
  at
• 3:26
  publication
• 3:27
  time.
• 3:32
  When
• 3:32
  automation
• 3:32
  touches
• 3:33
  live
• 3:33
  workflows,
• 3:34
  the
• 3:34
  question
• 3:35
  stops
• 3:35
  being
• 3:35
  "does
• 3:35
  the
• 3:36
  AI
• 3:36
  work?"
• 3:37
  and
• 3:37
  becomes
• 3:37
  "who
• 3:38
  authorized
• 3:38
  this
• 3:38
  action?"
• 3:39
  The
• 3:40
  Unhyped
• 3:40
  AI
• 3:40
  newsletter
• 3:41
  crystallized
• 3:41
  this
• 3:42
  transition
• 3:42
  as
• 3:42
  the
• 3:43
  "AI
• 3:43
  Automation
• 3:44
  Ceiling"
• 3:44
  --
• 3:45
  the
• 3:45
  predictable
• 3:45
  moment
• 3:46
  when
• 3:46
  agentic
• 3:46
  systems
• 3:46
  move
• 3:47
  from
• 3:47
  advisory
• 3:47
  pilots
• 3:48
  into
• 3:48
  production
• 3:48
  workflows
• 3:49
  and
• 3:49
  expose
• 3:49
  that
• 3:50
  "the
• 3:50
  organisation's
• 3:50
  capacity
• 3:51
  to
• 3:51
  govern
• 3:52
  delegation
• 3:52
  lagged
• 3:52
  behind
• 3:53
  deployment."
• 3:54
  The
• 3:54
  pattern
• 3:54
  is
• 3:54
  familiar
• 3:55
  to
• 3:55
  anyone
• 3:55
  who
• 3:55
  has
• 3:55
  watched
• 3:56
  enterprise
• 3:56
  software
• 3:56
  rollouts:
• 3:57
  "Exception
• 3:58
  rates
• 3:58
  rise.
• 3:59
  Definitions
• 3:59
  of
• 4:00
  'good'
• 4:00
  fracture
• 4:00
  across
• 4:00
  functions.
• 4:01
  Outputs
• 4:02
  that
• 4:02
  are
• 4:02
  technically
• 4:02
  correct
• 4:03
  create
• 4:03
  operational
• 4:04
  rework."
• 4:05
  The
• 4:05
  essay's
• 4:05
  sharpest
• 4:05
  insight
• 4:06
  is
• 4:06
  that
• 4:06
  most
• 4:06
  organizations
• 4:07
  misdiagnose
• 4:07
  the
• 4:08
  slowdown
• 4:08
  as
• 4:08
  an
• 4:08
  execution
• 4:09
  problem
• 4:09
  when
• 4:09
  it
• 4:10
  is
• 4:10
  actually
• 4:10
  structural:
• 4:11
  "Agentic
• 4:11
  automation
• 4:12
  rarely
• 4:12
  fails
• 4:12
  first
• 4:13
  as
• 4:13
  intelligence.
• 4:14
  It
• 4:14
  fails
• 4:14
  as
• 4:14
  coordination."
• 4:15
  Shadow
• 4:16
  systems
• 4:16
  --
• 4:17
  human
• 4:17
  workarounds
• 4:18
  built
• 4:18
  to
• 4:18
  compensate
• 4:19
  for
• 4:19
  the
• 4:19
  agent's
• 4:19
  blind
• 4:20
  spots
• 4:20
  --
• 4:20
  become
• 4:21
  "the
• 4:21
  interest
• 4:21
  payment
• 4:22
  on
• 4:22
  exceptions
• 4:22
  you
• 4:22
  never
• 4:23
  costed."
• 4:24
  The
• 4:24
  three
• 4:24
  board-level
• 4:25
  questions
• 4:25
  it
• 4:26
  proposes
• 4:26
  (true
• 4:26
  exception
• 4:27
  rate
• 4:27
  including
• 4:28
  manual
• 4:28
  correction,
• 4:29
  named
• 4:29
  decision-rights
• 4:30
  owners
• 4:30
  for
• 4:30
  uncertainty,
• 4:31
  and
• 4:31
  explicit
• 4:32
  stop
• 4:32
  criteria
• 4:32
  without
• 4:33
  political
• 4:33
  fallout)
• 4:34
  are
• 4:34
  the
• 4:34
  kind
• 4:34
  of
• 4:34
  operational
• 4:35
  framework
• 4:35
  that
• 4:35
  separates
• 4:36
  organizations
• 4:36
  that
• 4:36
  scale
• 4:37
  agents
• 4:37
  from
• 4:37
  those
• 4:37
  that
• 4:38
  carry
• 4:38
  them.
• 4:39
  Faramesh
• 4:40
  offers
• 4:40
  an
• 4:40
  architectural
• 4:41
  answer
• 4:41
  to
• 4:41
  this
• 4:41
  coordination
• 4:42
  gap.
• 4:43
  The
• 4:43
  research
• 4:43
  paper
• 4:44
  (arXiv
• 4:46
  2601.17744)
• 4:48
  introduces
• 4:48
  the
• 4:48
  Action
• 4:48
  Authorization
• 4:49
  Boundary
• 4:49
  (AAB)
• 4:50
  --
• 4:51
  a
• 4:51
  mandatory,
• 4:51
  non-bypassable
• 4:52
  enforcement
• 4:53
  layer
• 4:53
  between
• 4:53
  agent
• 4:54
  reasoning
• 4:54
  and
• 4:54
  real-world
• 4:55
  execution.
• 4:56
  The
• 4:56
  core
• 4:56
  argument:
• 4:57
  "inference
• 4:57
  produces
• 4:58
  information,
• 4:59
  whereas
• 4:59
  execution
• 4:59
  produces
• 5:00
  consequences,"
• 5:01
  and
• 5:01
  current
• 5:01
  frameworks
• 5:02
  collapse
• 5:02
  this
• 5:02
  distinction
• 5:03
  by
• 5:03
  treating
• 5:03
  proposal
• 5:03
  and
• 5:04
  execution
• 5:04
  as
• 5:04
  a
• 5:04
  single
• 5:05
  step.
• 5:06
  Every
• 5:06
  agent-generated
• 5:07
  action
• 5:07
  passes
• 5:08
  through
• 5:08
  a
• 5:08
  deterministic
• 5:09
  authorization
• 5:09
  function
• 5:10
  that
• 5:10
  maps
• 5:10
  a
• 5:10
  Canonical
• 5:11
  Action
• 5:11
  Representation,
• 5:12
  policy
• 5:13
  set,
• 5:13
  and
• 5:13
  system
• 5:13
  state
• 5:14
  to
• 5:14
  one
• 5:14
  of
• 5:14
  three
• 5:14
  outcomes:
• 5:15
  PERMIT,
• 5:16
  DEFER,
• 5:17
  or
• 5:17
  DENY.
• 5:18
  The
• 5:18
  system
• 5:19
  achieves
• 5:19
  2.24ms
• 5:21
  median
• 5:21
  decision
• 5:21
  latency,
• 5:22
  sustains
• 5:23
  7,800
• 5:24
  actions
• 5:25
  per
• 5:25
  minute
• 5:25
  on
• 5:25
  a
• 5:25
  single
• 5:26
  M1
• 5:26
  machine,
• 5:27
  and
• 5:27
  in
• 5:27
  stress
• 5:28
  testing
• 5:28
  produced
• 5:28
  zero
• 5:29
  double-executions
• 5:30
  across
• 5:30
  one
• 5:30
  million
• 5:31
  duplicated
• 5:31
  requests
• 5:32
  with
• 5:32
  64
• 5:33
  concurrent
• 5:33
  workers.
• 5:34
  The
• 5:34
  open-source
• 5:35
  implementation
• 5:36
  provides
• 5:36
  YAML-based
• 5:37
  policies
• 5:37
  with
• 5:37
  first-match-wins
• 5:38
  evaluation,
• 5:39
  automatic
• 5:40
  risk
• 5:40
  scoring,
• 5:41
  and
• 5:41
  a
• 5:41
  real-time
• 5:41
  web
• 5:42
  dashboard
• 5:42
  --
• 5:43
  with
• 5:43
  drop-in
• 5:43
  wrappers
• 5:44
  for
• 5:44
  LangChain,
• 5:45
  CrewAI,
• 5:46
  AutoGen,
• 5:46
  and
• 5:46
  MCP.
• 5:48
  The
• 5:48
  video
• 5:48
  walkthrough
• 5:49
  demonstrates
• 5:49
  the
• 5:49
  practical
• 5:50
  flow:
• 5:51
  a
• 5:51
  $85
• 5:52
  damaged-shirt
• 5:53
  refund
• 5:53
  auto-approves
• 5:54
  against
• 5:54
  policy,
• 5:55
  a
• 5:55
  $225
• 5:57
  missing-order
• 5:57
  claim
• 5:58
  pauses
• 5:58
  for
• 5:58
  manager
• 5:59
  review,
• 5:59
  and
• 6:00
  a
• 6:00
  $2,600
• 6:01
  full-refund
• 6:02
  attempt
• 6:02
  is
• 6:02
  blocked
• 6:03
  instantly
• 6:03
  before
• 6:04
  touching
• 6:04
  the
• 6:04
  customer's
• 6:04
  account.
• 6:06
  The
• 6:06
  sealed
• 6:06
  audit
• 6:06
  trail
• 6:07
  --
• 6:07
  policy
• 6:08
  version,
• 6:08
  exact
• 6:09
  parameters,
• 6:09
  human-vs-policy
• 6:11
  approval
• 6:11
  --
• 6:12
  is
• 6:12
  the
• 6:12
  kind
• 6:13
  of
• 6:13
  evidence
• 6:13
  chain
• 6:13
  that
• 6:14
  compliance
• 6:14
  teams
• 6:14
  actually
• 6:15
  need.
• 6:19
  The
• 6:19
  infrastructure
• 6:20
  around
• 6:20
  AI
• 6:21
  agents
• 6:21
  is
• 6:21
  converging
• 6:22
  fast.
• 6:23
  Charlotte,
• 6:23
  a
• 6:23
  new
• 6:24
  open-source
• 6:24
  MCP
• 6:25
  server
• 6:25
  for
• 6:25
  browser
• 6:26
  interaction,
• 6:26
  attacks
• 6:26
  the
• 6:27
  token-efficiency
• 6:27
  problem
• 6:27
  head-on.
• 6:28
  Instead
• 6:29
  of
• 6:29
  dumping
• 6:29
  raw
• 6:29
  accessibility
• 6:30
  trees
• 6:30
  into
• 6:31
  the
• 6:31
  context
• 6:31
  window,
• 6:32
  it
• 6:32
  decomposes
• 6:33
  pages
• 6:33
  into
• 6:33
  structured
• 6:34
  representations
• 6:34
  with
• 6:34
  landmarks,
• 6:35
  headings,
• 6:35
  and
• 6:35
  stable
• 6:36
  hash-based
• 6:36
  element
• 6:37
  IDs
• 6:37
  across
• 6:37
  three
• 6:38
  detail
• 6:38
  levels
• 6:38
  --
• 6:39
  minimal
• 6:39
  for
• 6:39
  orientation,
• 6:40
  summary
• 6:41
  for
• 6:41
  context,
• 6:41
  full
• 6:42
  for
• 6:42
  deep
• 6:42
  inspection.
• 6:43
  The
• 6:43
  benchmarks
• 6:43
  are
• 6:44
  striking:
• 6:44
  a
• 6:44
  Wikipedia
• 6:45
  page
• 6:45
  consumes
• 6:46
  7,667
• 6:48
  characters
• 6:48
  through
• 6:48
  Charlotte
• 6:49
  versus
• 6:49
  1,040,636
• 6:51
  through
• 6:52
  Playwright
• 6:52
  MCP;
• 6:53
  Hacker
• 6:53
  News
• 6:54
  takes
• 6:54
  336
• 6:55
  characters
• 6:55
  versus
• 6:55
  61,230.
• 6:57
  A
• 6:58
  100-page
• 6:58
  browsing
• 6:59
  session
• 6:59
  drops
• 6:59
  from
• 6:59
  roughly
• 7:00
  $15.30
• 7:01
  in
• 7:01
  input
• 7:01
  tokens
• 7:02
  on
• 7:02
  Claude
• 7:02
  Opus
• 7:02
  to
• 7:02
  about
• 7:03
  $0.09.
• 7:04
  The
• 7:04
  project
• 7:04
  was
• 7:05
  built
• 7:05
  and
• 7:05
  verified
• 7:05
  by
• 7:06
  an
• 7:06
  agent
• 7:06
  using
• 7:06
  Charlotte
• 7:07
  itself
• 7:07
  --
• 7:08
  catching
• 7:08
  a
• 7:08
  mobile
• 7:08
  overflow
• 7:09
  bug
• 7:09
  and
• 7:09
  fixing
• 7:10
  16
• 7:10
  unlabeled
• 7:11
  SVG
• 7:11
  icons
• 7:12
  without
• 7:12
  a
• 7:12
  human
• 7:12
  looking
• 7:13
  at
• 7:13
  the
• 7:13
  page.
• 7:14
  On
• 7:14
  the
• 7:14
  execution
• 7:15
  side,
• 7:15
  Kilntainers
• 7:16
  gives
• 7:16
  every
• 7:16
  agent
• 7:17
  an
• 7:17
  ephemeral
• 7:17
  Linux
• 7:17
  sandbox
• 7:18
  via
• 7:18
  MCP
• 7:19
  --
• 7:20
  disposable
• 7:20
  containers
• 7:21
  that
• 7:21
  isolate
• 7:21
  tool
• 7:22
  execution
• 7:22
  and
• 7:22
  prevent
• 7:23
  agents
• 7:23
  from
• 7:23
  persisting
• 7:23
  state
• 7:24
  or
• 7:24
  escaping
• 7:24
  their
• 7:24
  environment.
• 7:26
  Run-Agent
• 7:27
  takes
• 7:27
  orchestration
• 7:28
  in
• 7:28
  a
• 7:28
  different
• 7:28
  direction:
• 7:29
  a
• 7:29
  shell
• 7:30
  script
• 7:30
  and
• 7:30
  markdown-based
• 7:31
  message
• 7:31
  bus
• 7:31
  that
• 7:32
  launches
• 7:32
  Claude,
• 7:32
  Codex,
• 7:33
  or
• 7:33
  Gemini
• 7:33
  agents
• 7:34
  with
• 7:34
  full
• 7:34
  isolation
• 7:35
  and
• 7:35
  traceability.
• 7:36
  Each
• 7:37
  agent
• 7:37
  gets
• 7:37
  a
• 7:37
  defined
• 7:38
  role
• 7:38
  (orchestrator,
• 7:38
  researcher,
• 7:39
  implementer,
• 7:40
  reviewer,
• 7:41
  tester,
• 7:42
  debugger,
• 7:42
  monitor),
• 7:43
  a
• 7:43
  dedicated
• 7:44
  run
• 7:44
  folder
• 7:44
  capturing
• 7:45
  the
• 7:45
  exact
• 7:45
  prompt
• 7:46
  and
• 7:46
  full
• 7:46
  stdout/stderr,
• 7:48
  and
• 7:49
  an
• 7:49
  append-only
• 7:49
  message
• 7:50
  bus
• 7:50
  for
• 7:50
  inter-agent
• 7:51
  communication.
• 7:52
  The
• 7:52
  design
• 7:53
  philosophy
• 7:53
  is
• 7:53
  deliberately
• 7:54
  low-tech
• 7:54
  --
• 7:55
  "no
• 7:55
  databases,
• 7:56
  no
• 7:57
  infrastructure,
• 7:58
  just
• 7:58
  a
• 7:58
  shared
• 7:58
  markdown
• 7:59
  file"
• 7:59
  --
• 8:00
  which
• 8:00
  makes
• 8:01
  it
• 8:01
  auditable
• 8:01
  in
• 8:01
  ways
• 8:02
  that
• 8:02
  complex
• 8:02
  orchestration
• 8:03
  frameworks
• 8:03
  often
• 8:04
  are
• 8:04
  not.
• 8:05
  The
• 8:05
  agents
• 8:06
  run
• 8:06
  with
• 8:06
  full
• 8:06
  permissions
• 8:07
  by
• 8:07
  design,
• 8:08
  a
• 8:08
  tradeoff
• 8:08
  the
• 8:08
  documentation
• 8:09
  is
• 8:09
  refreshingly
• 8:10
  honest
• 8:10
  about.
• 8:11
  Manifold
• 8:12
  rounds
• 8:12
  out
• 8:12
  the
• 8:12
  tooling
• 8:13
  picture
• 8:13
  by
• 8:13
  giving
• 8:14
  each
• 8:14
  agent
• 8:14
  its
• 8:14
  own
• 8:14
  git
• 8:15
  branch,
• 8:15
  its
• 8:15
  own
• 8:16
  terminal,
• 8:16
  and
• 8:17
  real-time
• 8:17
  conflict
• 8:17
  detection
• 8:18
  when
• 8:18
  two
• 8:18
  agents
• 8:19
  touch
• 8:19
  the
• 8:19
  same
• 8:19
  file
• 8:19
  --
• 8:20
  "Three
• 8:21
  agents.
• 8:21
  One
• 8:22
  repo.
• 8:22
  Zero
• 8:23
  conflicts."
• 8:24
  The
• 8:24
  UI
• 8:24
  is
• 8:24
  a
• 8:24
  multi-pane
• 8:25
  terminal
• 8:25
  environment
• 8:26
  with
• 8:26
  resizable
• 8:27
  layouts,
• 8:27
  a
• 8:28
  built-in
• 8:28
  diff
• 8:28
  viewer,
• 8:29
  and
• 8:29
  file
• 8:29
  browser
• 8:29
  with
• 8:30
  syntax
• 8:30
  highlighting.
• 8:31
  No
• 8:31
  Electron.
• 8:33
  No
• 8:33
  loading
• 8:33
  spinners.
• 8:34
  The
• 8:34
  tool
• 8:35
  is
• 8:35
  agent-agnostic
• 8:36
  (Claude
• 8:36
  Code,
• 8:37
  Codex,
• 8:37
  Gemini
• 8:38
  CLI)
• 8:39
  and
• 8:39
  the
• 8:39
  isolation
• 8:39
  model
• 8:40
  is
• 8:40
  simple:
• 8:41
  automatic
• 8:41
  branch
• 8:41
  creation
• 8:42
  on
• 8:42
  spawn,
• 8:42
  cleanup
• 8:43
  on
• 8:43
  completion.
• 8:46
  At
• 8:46
  one
• 8:46
  end
• 8:46
  of
• 8:46
  the
• 8:46
  inference
• 8:47
  hardware
• 8:47
  spectrum,
• 8:48
  Taalas
• 8:48
  emerged
• 8:48
  from
• 8:49
  stealth
• 8:49
  with
• 8:49
  over
• 8:49
  $200
• 8:50
  million
• 8:50
  in
• 8:51
  funding
• 8:51
  and
• 8:51
  a
• 8:51
  radical
• 8:52
  proposition:
• 8:53
  stop
• 8:53
  trying
• 8:53
  to
• 8:53
  make
• 8:54
  compute
• 8:54
  engines
• 8:54
  malleable
• 8:55
  and
• 8:55
  just
• 8:56
  etch
• 8:56
  the
• 8:56
  model
• 8:56
  weights
• 8:57
  directly
• 8:57
  into
• 8:57
  transistors.
• 8:58
  The
• 8:59
  "Hard
• 8:59
  Coded
• 8:59
  Inference"
• 9:00
  architecture
• 9:00
  pairs
• 9:01
  a
• 9:01
  mask
• 9:01
  ROM
• 9:01
  recall
• 9:02
  fabric
• 9:02
  with
• 9:03
  SRAM
• 9:03
  for
• 9:03
  KV
• 9:03
  caches
• 9:04
  and
• 9:04
  adaptations.
• 9:05
  CEO
• 9:05
  Ljubisa
• 9:06
  Bajic
• 9:06
  --
• 9:07
  who
• 9:07
  previously
• 9:07
  founded
• 9:08
  Tenstorrent
• 9:08
  --
• 9:09
  claims
• 9:09
  the
• 9:09
  density
• 9:10
  is
• 9:10
  "basically
• 9:10
  insane":
• 9:11
  "We
• 9:11
  can
• 9:11
  store
• 9:12
  four
• 9:12
  bits
• 9:12
  away
• 9:13
  and
• 9:13
  do
• 9:13
  the
• 9:13
  multiply
• 9:13
  related
• 9:14
  to
• 9:14
  it
• 9:14
  --
• 9:14
  everything
• 9:14
  --
• 9:15
  with
• 9:15
  a
• 9:15
  single
• 9:16
  transistor."
• 9:17
  The
• 9:17
  first-generation
• 9:18
  HC1
• 9:18
  chip
• 9:18
  is
• 9:18
  fabricated
• 9:19
  on
• 9:19
  TSMC
• 9:19
  6nm
• 9:20
  with
• 9:20
  53
• 9:21
  billion
• 9:21
  transistors,
• 9:22
  supports
• 9:22
  8
• 9:23
  billion
• 9:23
  parameters
• 9:23
  per
• 9:23
  chip,
• 9:24
  and
• 9:24
  burns
• 9:24
  about
• 9:24
  200
• 9:25
  watts
• 9:25
  per
• 9:25
  card.
• 9:26
  Performance
• 9:27
  benchmarks
• 9:27
  on
• 9:27
  Llama
• 9:28
  3.1
• 9:28
  8B
• 9:29
  show
• 9:29
  substantial
• 9:30
  throughput
• 9:30
  advantages
• 9:30
  over
• 9:31
  Nvidia's
• 9:31
  B200,
• 9:32
  Groq,
• 9:33
  SambaNova,
• 9:33
  and
• 9:33
  Cerebras,
• 9:34
  with
• 9:34
  significantly
• 9:35
  lower
• 9:35
  latency
• 9:36
  and
• 9:36
  cost
• 9:36
  per
• 9:36
  token.
• 9:37
  The
• 9:37
  obvious
• 9:37
  limitation
• 9:38
  --
• 9:39
  every
• 9:39
  model
• 9:39
  revision
• 9:40
  requires
• 9:40
  a
• 9:40
  new
• 9:40
  chip
• 9:41
  spin
• 9:41
  --
• 9:42
  is
• 9:42
  mitigated
• 9:42
  by
• 9:42
  a
• 9:42
  two-month
• 9:43
  turnaround
• 9:43
  from
• 9:43
  weights
• 9:44
  to
• 9:44
  deployed
• 9:44
  PCI-Express
• 9:45
  cards,
• 9:45
  changing
• 9:46
  only
• 9:46
  two
• 9:46
  metal
• 9:47
  layers.
• 9:48
  With
• 9:48
  training
• 9:48
  runs
• 9:49
  costing
• 9:49
  billions
• 9:50
  and
• 9:50
  model
• 9:50
  release
• 9:50
  cadences
• 9:51
  lengthening,
• 9:52
  Taalas
• 9:52
  argues
• 9:52
  the
• 9:53
  economics
• 9:53
  work:
• 9:54
  it
• 9:54
  costs
• 9:54
  100x
• 9:55
  more
• 9:55
  to
• 9:55
  train
• 9:55
  a
• 9:55
  model
• 9:56
  than
• 9:56
  to
• 9:56
  customize
• 9:56
  an
• 9:56
  HC
• 9:57
  chip.
• 9:58
  A
• 9:58
  frontier-class
• 9:58
  LLM
• 9:59
  across
• 9:59
  multiple
• 10:00
  HC2
• 10:00
  cards
• 10:01
  is
• 10:01
  planned
• 10:01
  for
• 10:01
  year-end.
• 10:02
  At
• 10:03
  the
• 10:03
  other
• 10:03
  end
• 10:03
  of
• 10:03
  the
• 10:03
  spectrum,
• 10:04
  an
• 10:04
  anonymous
• 10:05
  team
• 10:05
  has
• 10:05
  converted
• 10:06
  800
• 10:06
  retired
• 10:07
  Ethereum
• 10:07
  mining
• 10:08
  RX
• 10:08
  580
• 10:09
  GPUs
• 10:10
  into
• 10:10
  an
• 10:10
  AI
• 10:10
  inference
• 10:11
  cluster
• 10:11
  optimized
• 10:11
  for
• 10:12
  mass
• 10:12
  document
• 10:12
  OCR.
• 10:14
  The
• 10:14
  fundamental
• 10:14
  constraint
• 10:15
  is
• 10:15
  that
• 10:15
  these
• 10:15
  cards
• 10:16
  have
• 10:16
  decent
• 10:16
  8GB
• 10:17
  VRAM
• 10:17
  but
• 10:18
  low
• 10:18
  PCIe
• 10:18
  speeds
• 10:19
  and
• 10:19
  are
• 10:20
  each
• 10:20
  stranded
• 10:20
  on
• 10:20
  Celeron
• 10:21
  G3950
• 10:22
  boards
• 10:22
  --
• 10:24
  tensor
• 10:25
  parallelism
• 10:25
  across
• 10:26
  nodes
• 10:26
  is
• 10:26
  impossible.
• 10:27
  The
• 10:27
  solution:
• 10:29
  treat
• 10:29
  each
• 10:29
  GPU
• 10:30
  as
• 10:30
  a
• 10:30
  completely
• 10:30
  independent
• 10:31
  inference
• 10:31
  worker
• 10:32
  running
• 10:32
  Qwen3-VL-8B
• 10:34
  via
• 10:35
  llama.cpp's
• 10:36
  Vulkan
• 10:36
  backend.
• 10:37
  Getting
• 10:38
  there
• 10:38
  required
• 10:38
  building
• 10:39
  the
• 10:39
  entire
• 10:39
  graphics
• 10:39
  stack
• 10:40
  from
• 10:40
  source
• 10:40
  (libdrm,
• 10:41
  Wayland,
• 10:41
  Mesa
• 10:41
  with
• 10:42
  the
• 10:42
  RADV
• 10:42
  Vulkan
• 10:43
  driver,
• 10:43
  Vulkan
• 10:44
  SDK)
• 10:44
  and
• 10:45
  discovering
• 10:45
  that
• 10:45
  AVX2/FMA
• 10:49
  binaries
• 10:49
  segfault
• 10:50
  on
• 10:50
  Celerons.
• 10:52
  The
• 10:52
  practical
• 10:52
  result
• 10:53
  is
• 10:53
  200
• 10:53
  GPUs
• 10:54
  processing
• 10:54
  200
• 10:55
  pages
• 10:55
  simultaneously
• 10:56
  with
• 10:56
  a
• 10:56
  four-tier
• 10:57
  quality
• 10:57
  escalation
• 10:58
  path
• 10:58
  (q4
• 10:59
  through
• 10:59
  bf16
• 11:00
  on
• 11:00
  six
• 11:00
  GPUs).
• 11:02
  A
• 11:02
  966-page
• 11:03
  clinical
• 11:04
  ophthalmology
• 11:04
  textbook
• 11:05
  --
• 11:06
  dense
• 11:06
  medical
• 11:06
  terminology,
• 11:07
  complex
• 11:08
  diagrams,
• 11:08
  multi-column
• 11:09
  layouts
• 11:09
  --
• 11:10
  costs
• 11:10
  about
• 11:11
  $0.50
• 11:11
  in
• 11:11
  electricity
• 11:12
  versus
• 11:12
  $12
• 11:13
  through
• 11:13
  the
• 11:13
  OpenAI
• 11:14
  API,
• 11:15
  at
• 11:15
  comparable
• 11:15
  quality.
• 11:16
  A
• 11:16
  video
• 11:16
  frame
• 11:17
  analysis
• 11:17
  pipeline
• 11:18
  (800
• 11:18
  frames
• 11:18
  across
• 11:19
  800
• 11:19
  GPUs
• 11:20
  for
• 11:20
  60
• 11:20
  seconds
• 11:21
  of
• 11:21
  video)
• 11:21
  is
• 11:21
  in
• 11:21
  development.
• 11:26
  Mark
• 11:27
  Russinovich
• 11:27
  and
• 11:27
  Scott
• 11:28
  Hanselman
• 11:28
  argue
• 11:28
  in
• 11:28
  a
• 11:28
  new
• 11:29
  ACM
• 11:29
  opinion
• 11:30
  piece
• 11:30
  that
• 11:30
  generative
• 11:30
  AI
• 11:31
  has
• 11:31
  created
• 11:31
  a
• 11:31
  "seniority-biased
• 11:32
  technological
• 11:33
  change"
• 11:34
  that
• 11:34
  amplifies
• 11:34
  experienced
• 11:35
  engineers
• 11:35
  while
• 11:35
  dragging
• 11:36
  down
• 11:36
  juniors.
• 11:37
  The
• 11:37
  evidence
• 11:37
  is
• 11:37
  concrete:
• 11:38
  Microsoft's
• 11:39
  Project
• 11:39
  Societas
• 11:40
  saw
• 11:40
  seven
• 11:40
  part-time
• 11:41
  engineers
• 11:41
  deliver
• 11:42
  a
• 11:42
  consumer-ready
• 11:42
  preview
• 11:43
  in
• 11:43
  10
• 11:43
  weeks
• 11:43
  with
• 11:43
  98%
• 11:44
  AI-generated
• 11:45
  code.
• 11:46
  But
• 11:46
  the
• 11:46
  agents
• 11:46
  exhibit
• 11:47
  "intern-like"
• 11:47
  failures
• 11:48
  --
• 11:49
  masking
• 11:49
  race
• 11:49
  conditions
• 11:50
  with
• 11:50
  sleeps,
• 11:50
  claiming
• 11:51
  success
• 11:51
  despite
• 11:51
  significant
• 11:52
  bugs,
• 11:52
  duplicating
• 11:53
  code,
• 11:53
  dismissing
• 11:54
  crashes.
• 11:55
  Labor
• 11:55
  data
• 11:56
  shows
• 11:56
  employment
• 11:56
  of
• 11:56
  22-to-25-year-olds
• 11:58
  in
• 11:58
  AI-exposed
• 11:58
  jobs
• 11:59
  fell
• 11:59
  roughly
• 12:00
  13%
• 12:00
  after
• 12:01
  GPT-4's
• 12:01
  release,
• 12:02
  even
• 12:02
  as
• 12:02
  senior
• 12:02
  roles
• 12:02
  grew.
• 12:04
  The
• 12:04
  proposed
• 12:04
  solution
• 12:05
  is
• 12:05
  a
• 12:05
  "preceptor
• 12:05
  program"
• 12:06
  pairing
• 12:06
  early-in-career
• 12:07
  developers
• 12:07
  with
• 12:08
  mentors
• 12:08
  at
• 12:08
  3:1-5:1
• 12:09
  ratios
• 12:09
  for
• 12:09
  at
• 12:10
  least
• 12:10
  a
• 12:10
  year,
• 12:10
  with
• 12:11
  coding
• 12:11
  assistants
• 12:12
  defaulting
• 12:12
  to
• 12:12
  Socratic
• 12:13
  coaching
• 12:13
  before
• 12:13
  code
• 12:13
  generation.
• 12:15
  The
• 12:15
  framing
• 12:15
  draws
• 12:16
  on
• 12:16
  MIT
• 12:16
  research
• 12:17
  showing
• 12:17
  "cognitive
• 12:17
  debt"
• 12:18
  --
• 12:18
  reduced
• 12:19
  brain
• 12:19
  activity
• 12:19
  and
• 12:20
  lower
• 12:20
  recall
• 12:20
  --
• 12:21
  in
• 12:21
  adults
• 12:21
  who
• 12:21
  used
• 12:22
  ChatGPT
• 12:22
  for
• 12:22
  writing.
• 12:23
  Quoting
• 12:24
  Ethan
• 12:24
  Mollick:
• 12:25
  "Every
• 12:25
  time
• 12:26
  we
• 12:26
  hand
• 12:26
  work
• 12:26
  to
• 12:26
  a
• 12:26
  wizard,
• 12:27
  we
• 12:27
  lose
• 12:28
  a
• 12:28
  chance
• 12:28
  to
• 12:28
  develop
• 12:28
  our
• 12:29
  own
• 12:29
  expertise;
• 12:30
  to
• 12:30
  build
• 12:30
  the
• 12:30
  very
• 12:31
  judgment
• 12:31
  we
• 12:31
  need
• 12:31
  to
• 12:31
  evaluate
• 12:32
  the
• 12:32
  wizard's
• 12:32
  work."
• 12:33
  That
• 12:34
  judgment
• 12:34
  crisis
• 12:34
  plays
• 12:35
  out
• 12:35
  in
• 12:35
  raw,
• 12:36
  unfiltered
• 12:36
  detail
• 12:37
  in
• 12:37
  "The
• 12:37
  Conductor
• 12:37
  Won't
• 12:37
  Stop
• 12:38
  Conducting,"
• 12:39
  a
• 12:39
  personal
• 12:40
  essay
• 12:40
  documenting
• 12:40
  81
• 12:41
  Claude
• 12:41
  Code
• 12:41
  sessions,
• 12:42
  596
• 12:43
  messages,
• 12:44
  fifteen
• 12:45
  releases
• 12:45
  in
• 12:45
  eleven
• 12:45
  days,
• 12:46
  and
• 12:46
  38
• 12:47
  wrong-approach
• 12:47
  corrections.
• 12:48
  The
• 12:49
  author
• 12:49
  corrected
• 12:49
  the
• 12:49
  same
• 12:50
  CLAUDE.md
• 12:50
  rule
• 12:51
  --
• 12:52
  don't
• 12:52
  run
• 12:52
  the
• 12:52
  full
• 12:52
  test
• 12:53
  suite
• 12:53
  --
• 12:53
  twenty
• 12:54
  times
• 12:54
  across
• 12:54
  nine
• 12:55
  days.
• 12:55
  Each
• 12:56
  violation
• 12:56
  caused
• 12:57
  an
• 12:57
  out-of-memory
• 12:57
  crash,
• 12:58
  killing
• 12:58
  the
• 12:58
  session
• 12:59
  and
• 12:59
  destroying
• 12:59
  in-progress
• 13:00
  work.
• 13:01
  Three
• 13:01
  database
• 13:01
  losses
• 13:02
  in
• 13:02
  six
• 13:02
  days.
• 13:03
  An
• 13:04
  HNSW
• 13:04
  vector
• 13:05
  dimension
• 13:05
  mismatch
• 13:06
  (128
• 13:06
  vs
• 13:07
  768)
• 13:08
  that
• 13:08
  took
• 13:08
  six
• 13:09
  releases
• 13:09
  to
• 13:09
  eradicate.
• 13:11
  A
• 13:11
  code
• 13:11
  quality
• 13:11
  sweep
• 13:12
  revealing
• 13:12
  that
• 13:12
  a
• 13:12
  claimed
• 13:13
  "700+
• 13:14
  patterns
• 13:14
  replaced"
• 13:15
  was
• 13:15
  actually
• 13:15
  132
• 13:16
  of
• 13:16
  763
• 13:17
  --
• 13:18
  seventeen
• 13:19
  percent.
• 13:20
  The
• 13:20
  essay
• 13:20
  turns
• 13:21
  inward
• 13:21
  when
• 13:21
  the
• 13:21
  author
• 13:22
  discloses
• 13:22
  that
• 13:22
  his
• 13:22
  grandmother
• 13:23
  died
• 13:23
  a
• 13:23
  month
• 13:23
  prior
• 13:24
  and
• 13:24
  his
• 13:24
  father
• 13:24
  was
• 13:25
  hospitalized:
• 13:26
  "The
• 13:26
  moment
• 13:26
  I'm
• 13:27
  venting
• 13:27
  at
• 13:27
  an
• 13:27
  LLM,
• 13:28
  I'm
• 13:28
  no
• 13:28
  longer
• 13:29
  engineering.
• 13:30
  I'm
• 13:30
  coping.
• 13:31
  And
• 13:31
  coping
• 13:31
  isn't
• 13:32
  a
• 13:32
  debugging
• 13:32
  strategy."
• 13:33
  The
• 13:34
  actionable
• 13:34
  conclusion
• 13:34
  --
• 13:35
  separate
• 13:36
  diagnosis
• 13:36
  from
• 13:37
  implementation,
• 13:38
  encode
• 13:38
  verification
• 13:39
  criteria
• 13:40
  so
• 13:40
  thoroughly
• 13:40
  they
• 13:40
  work
• 13:41
  even
• 13:41
  when
• 13:41
  you
• 13:41
  don't
• 13:41
  --
• 13:42
  is
• 13:42
  the
• 13:43
  kind
• 13:43
  of
• 13:43
  hard-won
• 13:43
  operational
• 13:44
  wisdom
• 13:44
  that
• 13:45
  no
• 13:45
  benchmark
• 13:45
  captures.
• 13:47
  A
• 13:47
  developer
• 13:47
  building
• 13:47
  a
• 13:47
  local
• 13:48
  agent
• 13:48
  to
• 13:48
  manage
• 13:49
  Minecraft
• 13:49
  servers
• 13:50
  via
• 13:50
  Docker
• 13:50
  and
• 13:50
  RCON
• 13:51
  distilled
• 13:51
  complementary
• 13:52
  lessons
• 13:52
  from
• 13:52
  the
• 13:52
  opposite
• 13:53
  end
• 13:53
  of
• 13:53
  the
• 13:53
  sophistication
• 13:54
  spectrum.
• 13:55
  The
• 13:55
  single
• 13:55
  biggest
• 13:56
  improvement
• 13:56
  was
• 13:56
  not
• 13:57
  a
• 13:57
  better
• 13:57
  prompt
• 13:57
  but
• 13:57
  running
• 13:58
  real
• 13:58
  shell
• 13:58
  commands
• 13:59
  to
• 13:59
  discover
• 13:59
  the
• 13:59
  environment
• 14:00
  before
• 14:00
  asking
• 14:00
  the
• 14:01
  LLM
• 14:01
  to
• 14:01
  write
• 14:01
  code:
• 14:02
  "When
• 14:03
  the
• 14:03
  coder
• 14:03
  model
• 14:03
  gets
• 14:03
  `container
• 14:04
  id
• 14:04
  =
• 14:05
  "a1b2c3d4"`
• 14:07
  injected
• 14:07
  as
• 14:07
  an
• 14:07
  actual
• 14:08
  Python
• 14:08
  variable,
• 14:09
  it
• 14:09
  uses
• 14:09
  it.
• 14:10
  When
• 14:10
  it
• 14:10
  has
• 14:10
  to
• 14:11
  guess,
• 14:11
  it
• 14:11
  invents
• 14:11
  IDs
• 14:12
  that
• 14:12
  don't
• 14:12
  exist."
• 14:13
  Structural
• 14:13
  fixes
• 14:14
  --
• 14:15
  deleting
• 14:15
  bad
• 14:15
  state
• 14:16
  from
• 14:16
  context,
• 14:17
  rewriting
• 14:17
  task
• 14:17
  descriptions
• 14:18
  from
• 14:18
  scratch
• 14:18
  rather
• 14:19
  than
• 14:19
  appending
• 14:19
  errors
• 14:19
  --
• 14:20
  beat
• 14:20
  "try
• 14:21
  again"
• 14:21
  retry
• 14:21
  logic
• 14:22
  every
• 14:22
  time.
• 14:23
  And
• 14:23
  hard
• 14:23
  pass/fail
• 14:24
  contracts
• 14:25
  ("each
• 14:25
  subtask
• 14:26
  declares
• 14:26
  what
• 14:26
  it
• 14:26
  must
• 14:26
  produce
• 14:27
  as
• 14:27
  STATE:key=value
• 14:28
  prints
• 14:29
  in
• 14:29
  stdout")
• 14:29
  catch
• 14:30
  the
• 14:30
  number-one
• 14:30
  local
• 14:30
  model
• 14:31
  failure
• 14:31
  mode:
• 14:32
  "the
• 14:32
  LLM
• 14:32
  writes
• 14:33
  code
• 14:33
  that
• 14:33
  prints
• 14:33
  'Success!'
• 14:34
  without
• 14:34
  actually
• 14:35
  doing
• 14:35
  anything."
• 14:37
  The
• 14:37
  thread
• 14:37
  connecting
• 14:37
  all
• 14:38
  three
• 14:38
  is
• 14:38
  that
• 14:38
  AI-assisted
• 14:39
  development
• 14:40
  bottlenecks
• 14:40
  have
• 14:40
  shifted
• 14:41
  from
• 14:41
  code
• 14:41
  generation
• 14:42
  to
• 14:42
  everything
• 14:42
  around
• 14:42
  it.
• 14:43
  An
• 14:44
  essay
• 14:44
  on
• 14:44
  intentional
• 14:45
  overuse
• 14:45
  as
• 14:45
  a
• 14:45
  learning
• 14:45
  strategy
• 14:46
  frames
• 14:46
  this
• 14:46
  explicitly:
• 14:48
  "The
• 14:48
  model
• 14:48
  isn't
• 14:48
  the
• 14:48
  bottleneck.
• 14:50
  The
• 14:50
  surrounding
• 14:50
  workflow
• 14:51
  is."
• 14:52
  The
• 14:52
  author
• 14:53
  advocates
• 14:53
  deliberately
• 14:54
  using
• 14:54
  AI
• 14:54
  for
• 14:54
  tasks
• 14:55
  it
• 14:55
  is
• 14:55
  probably
• 14:55
  bad
• 14:56
  at
• 14:56
  --
• 14:57
  nuanced
• 14:57
  architecture
• 14:58
  decisions,
• 14:58
  subtle
• 14:59
  performance
• 14:59
  work
• 14:59
  --
• 15:00
  because
• 15:01
  "the
• 15:01
  failures
• 15:01
  are
• 15:01
  the
• 15:01
  data."
• 15:03
  The
• 15:03
  mapped
• 15:03
  boundaries
• 15:03
  include
• 15:04
  review
• 15:04
  becoming
• 15:05
  slower
• 15:05
  than
• 15:05
  generation,
• 15:06
  context
• 15:07
  engineering
• 15:07
  emerging
• 15:08
  as
• 15:08
  its
• 15:08
  own
• 15:08
  discipline,
• 15:09
  multi-agent
• 15:10
  coordination
• 15:10
  overhead
• 15:11
  as
• 15:11
  the
• 15:11
  real
• 15:12
  cost
• 15:12
  nobody
• 15:12
  talks
• 15:13
  about,
• 15:13
  and
• 15:13
  the
• 15:14
  security
• 15:14
  surface
• 15:14
  growing
• 15:15
  with
• 15:15
  every
• 15:15
  tool
• 15:15
  the
• 15:16
  agent
• 15:16
  touches.
• 15:17
  The
• 15:17
  recommendation
• 15:18
  echoes
• 15:18
  the
• 15:18
  continuous
• 15:19
  delivery
• 15:19
  movement:
• 15:20
  "If
• 15:20
  deploying
• 15:21
  is
• 15:21
  painful,
• 15:22
  deploy
• 15:22
  more,
• 15:23
  until
• 15:23
  it
• 15:23
  isn't.
• 15:24
  If
• 15:25
  AI-assisted
• 15:26
  refactoring
• 15:26
  keeps
• 15:26
  going
• 15:27
  wrong,
• 15:27
  do
• 15:27
  more
• 15:27
  of
• 15:27
  it."
• 15:31
  SkillRL,
• 15:32
  a
• 15:32
  new
• 15:32
  framework
• 15:33
  from
• 15:33
  the
• 15:33
  AIMING
• 15:33
  Lab,
• 15:34
  tackles
• 15:34
  the
• 15:34
  core
• 15:35
  limitation
• 15:35
  of
• 15:35
  episodic
• 15:36
  LLM
• 15:36
  agents:
• 15:37
  they
• 15:37
  cannot
• 15:38
  learn
• 15:38
  from
• 15:38
  past
• 15:38
  experience.
• 15:40
  Rather
• 15:40
  than
• 15:40
  storing
• 15:41
  raw
• 15:41
  trajectories
• 15:41
  (which
• 15:42
  are
• 15:42
  "lengthy
• 15:42
  and
• 15:42
  contain
• 15:43
  significant
• 15:43
  redundancy
• 15:44
  and
• 15:44
  noise"),
• 15:45
  SkillRL
• 15:45
  distills
• 15:46
  successful
• 15:46
  episodes
• 15:47
  into
• 15:47
  demonstration
• 15:47
  patterns
• 15:48
  and
• 15:48
  failed
• 15:48
  ones
• 15:48
  into
• 15:48
  concise
• 15:49
  failure
• 15:49
  lessons,
• 15:50
  then
• 15:50
  organizes
• 15:51
  both
• 15:51
  into
• 15:51
  a
• 15:51
  hierarchical
• 15:52
  skill
• 15:52
  library
• 15:53
  with
• 15:53
  general
• 15:53
  strategic
• 15:53
  principles
• 15:54
  and
• 15:54
  task-specific
• 15:55
  heuristics.
• 15:57
  A
• 15:57
  recursive
• 15:57
  evolution
• 15:58
  mechanism
• 15:58
  lets
• 15:59
  the
• 15:59
  skill
• 15:59
  library
• 15:59
  co-evolve
• 16:00
  with
• 16:00
  the
• 16:00
  agent's
• 16:01
  policy
• 16:01
  during
• 16:01
  reinforcement
• 16:02
  learning
• 16:02
  --
• 16:03
  after
• 16:04
  each
• 16:04
  validation
• 16:04
  epoch,
• 16:05
  failed
• 16:05
  trajectories
• 16:06
  are
• 16:06
  analyzed
• 16:06
  to
• 16:06
  propose
• 16:07
  new
• 16:07
  skills
• 16:07
  and
• 16:08
  refine
• 16:08
  existing
• 16:08
  ones.
• 16:10
  The
• 16:10
  results
• 16:10
  on
• 16:10
  a
• 16:10
  7B
• 16:11
  Qwen
• 16:11
  model
• 16:11
  are
• 16:11
  remarkable:
• 16:12
  89.9%
• 16:14
  success
• 16:14
  rate
• 16:14
  on
• 16:14
  ALFWorld
• 16:15
  (beating
• 16:16
  GPT-4o
• 16:16
  by
• 16:17
  41.9%
• 16:18
  and
• 16:18
  Gemini-2.5-Pro
• 16:20
  by
• 16:20
  29.6%),
• 16:22
  with
• 16:22
  10-20x
• 16:22
  token
• 16:23
  compression
• 16:23
  compared
• 16:24
  to
• 16:24
  raw
• 16:24
  trajectory
• 16:24
  storage.
• 16:25
  Removing
• 16:26
  the
• 16:26
  hierarchical
• 16:26
  structure
• 16:27
  costs
• 16:27
  13.1%;
• 16:29
  replacing
• 16:30
  skills
• 16:30
  with
• 16:30
  raw
• 16:30
  trajectories
• 16:31
  causes
• 16:31
  up
• 16:31
  to
• 16:32
  25%
• 16:32
  degradation.
• 16:34
  The
• 16:34
  skill
• 16:34
  library
• 16:34
  grows
• 16:35
  from
• 16:35
  55
• 16:36
  to
• 16:36
  100
• 16:36
  skills
• 16:36
  during
• 16:37
  training,
• 16:37
  with
• 16:38
  task-specific
• 16:38
  skills
• 16:39
  nearly
• 16:39
  doubling.
• 16:40
  AgentDB
• 16:41
  v3,
• 16:41
  built
• 16:41
  on
• 16:42
  RuVector,
• 16:42
  extends
• 16:43
  the
• 16:43
  memory-as-substrate
• 16:44
  idea
• 16:44
  into
• 16:44
  what
• 16:44
  it
• 16:44
  calls
• 16:45
  a
• 16:45
  "Cognitive
• 16:45
  Container"
• 16:45
  --
• 16:46
  a
• 16:46
  single
• 16:47
  RVF
• 16:47
  file
• 16:47
  holding
• 16:48
  vectors
• 16:48
  with
• 16:48
  HNSW
• 16:49
  indexes,
• 16:50
  LoRA
• 16:50
  adapters
• 16:51
  and
• 16:51
  reinforcement
• 16:51
  policies,
• 16:52
  episodic
• 16:53
  and
• 16:53
  causal
• 16:53
  memory,
• 16:54
  compression
• 16:54
  tiers,
• 16:55
  and
• 16:55
  a
• 16:55
  cryptographic
• 16:55
  witness
• 16:56
  chain
• 16:56
  proving
• 16:56
  the
• 16:56
  history
• 16:57
  has
• 16:57
  not
• 16:57
  been
• 16:57
  tampered
• 16:57
  with.
• 16:58
  Because
• 16:59
  the
• 16:59
  container
• 16:59
  is
• 16:59
  append-safe
• 17:00
  and
• 17:00
  copy-on-write,
• 17:01
  you
• 17:01
  can
• 17:01
  branch
• 17:01
  an
• 17:02
  entire
• 17:02
  intelligence
• 17:02
  state
• 17:03
  in
• 17:03
  milliseconds,
• 17:04
  test
• 17:04
  a
• 17:04
  new
• 17:04
  strategy,
• 17:05
  compare
• 17:06
  reward
• 17:06
  signals,
• 17:07
  and
• 17:07
  promote
• 17:07
  the
• 17:07
  better
• 17:07
  branch.
• 17:08
  The
• 17:08
  practical
• 17:09
  implication
• 17:09
  for
• 17:09
  agentic
• 17:10
  coding:
• 17:11
  a
• 17:11
  code
• 17:11
  review
• 17:11
  agent
• 17:12
  accumulates
• 17:12
  structured
• 17:13
  experience
• 17:13
  across
• 17:14
  thousands
• 17:14
  of
• 17:14
  pull
• 17:14
  requests,
• 17:16
  a
• 17:16
  debugging
• 17:16
  agent
• 17:17
  retains
• 17:17
  causal
• 17:17
  memory
• 17:18
  from
• 17:18
  production
• 17:18
  incidents,
• 17:19
  and
• 17:19
  all
• 17:20
  of
• 17:20
  it
• 17:20
  is
• 17:20
  portable
• 17:20
  --
• 17:21
  move
• 17:21
  the
• 17:21
  file
• 17:22
  and
• 17:22
  the
• 17:22
  agent
• 17:22
  behaves
• 17:23
  identically.
• 17:24
  Agentic
• 17:24
  QE
• 17:25
  v3.7.0
• 17:27
  integrates
• 17:27
  the
• 17:28
  RVF
• 17:28
  container
• 17:28
  format
• 17:29
  to
• 17:29
  give
• 17:29
  quality
• 17:30
  engineering
• 17:30
  agents
• 17:30
  a
• 17:31
  cryptographic
• 17:31
  witness
• 17:31
  chain
• 17:32
  --
• 17:32
  "When
• 17:33
  an
• 17:33
  agent
• 17:33
  says
• 17:33
  'all
• 17:34
  tests
• 17:34
  passed,'
• 17:34
  you
• 17:35
  can
• 17:35
  verify
• 17:35
  it.
• 17:36
  No
• 17:36
  more
• 17:36
  completion
• 17:37
  theater."
• 17:38
  The
• 17:38
  release
• 17:38
  also
• 17:39
  adds
• 17:39
  MinCut-based
• 17:40
  test
• 17:40
  optimization
• 17:41
  that
• 17:41
  models
• 17:41
  the
• 17:42
  test
• 17:42
  suite
• 17:42
  as
• 17:42
  a
• 17:42
  graph
• 17:43
  problem
• 17:43
  to
• 17:43
  identify
• 17:44
  critical-path
• 17:44
  versus
• 17:45
  safely-skippable
• 17:45
  tests,
• 17:46
  and
• 17:46
  copy-on-write
• 17:47
  learning
• 17:47
  isolation
• 17:48
  so
• 17:48
  agents
• 17:49
  can
• 17:49
  experiment
• 17:49
  with
• 17:49
  new
• 17:50
  testing
• 17:50
  strategies
• 17:50
  without
• 17:51
  contaminating
• 17:51
  production
• 17:52
  knowledge.
• 17:54
  Google
• 17:54
  shipped
• 17:55
  Gemini
• 17:55
  3.1
• 17:56
  Pro
• 17:56
  across
• 17:57
  consumer
• 17:57
  and
• 17:58
  developer
• 17:58
  products
• 17:59
  this
• 17:59
  week,
• 17:59
  building
• 18:00
  on
• 18:00
  the
• 18:00
  Gemini
• 18:00
  3
• 18:01
  series
• 18:01
  with
• 18:01
  what
• 18:01
  it
• 18:01
  calls
• 18:02
  a
• 18:02
  step
• 18:02
  forward
• 18:02
  in
• 18:03
  core
• 18:03
  reasoning.
• 18:04
  On
• 18:04
  ARC-AGI-2
• 18:06
  --
• 18:08
  a
• 18:08
  benchmark
• 18:09
  evaluating
• 18:09
  a
• 18:09
  model's
• 18:09
  ability
• 18:10
  to
• 18:10
  solve
• 18:10
  entirely
• 18:11
  new
• 18:11
  logic
• 18:11
  patterns
• 18:12
  --
• 18:13
  3.1
• 18:13
  Pro
• 18:13
  scored
• 18:14
  77.1%,
• 18:16
  more
• 18:16
  than
• 18:16
  doubling
• 18:16
  the
• 18:17
  reasoning
• 18:17
  performance
• 18:17
  of
• 18:17
  3
• 18:18
  Pro.
• 18:19
  The
• 18:19
  model
• 18:19
  is
• 18:19
  rolling
• 18:20
  out
• 18:20
  in
• 18:20
  preview
• 18:20
  via
• 18:21
  the
• 18:21
  Gemini
• 18:21
  API
• 18:22
  in
• 18:22
  AI
• 18:22
  Studio,
• 18:23
  Antigravity,
• 18:24
  Vertex
• 18:24
  AI,
• 18:25
  Gemini
• 18:25
  Enterprise,
• 18:26
  Gemini
• 18:26
  CLI,
• 18:27
  Android
• 18:28
  Studio,
• 18:28
  NotebookLM,
• 18:29
  and
• 18:29
  the
• 18:29
  Gemini
• 18:30
  app.
• 18:31
  The
• 18:31
  demo
• 18:31
  reel
• 18:31
  emphasizes
• 18:32
  practical
• 18:32
  applications
• 18:33
  over
• 18:33
  benchmarks:
• 18:34
  generating
• 18:35
  website-ready
• 18:36
  animated
• 18:36
  SVGs
• 18:37
  from
• 18:37
  text
• 18:37
  prompts,
• 18:38
  building
• 18:39
  a
• 18:39
  live
• 18:39
  ISS
• 18:40
  orbit
• 18:40
  dashboard
• 18:41
  from
• 18:41
  public
• 18:41
  telemetry
• 18:42
  APIs,
• 18:42
  and
• 18:43
  coding
• 18:43
  a
• 18:43
  3D
• 18:43
  starling
• 18:44
  murmuration
• 18:44
  with
• 18:45
  hand-tracking
• 18:45
  and
• 18:45
  generative
• 18:46
  audio.
• 18:47
  Google
• 18:47
  notes
• 18:48
  it
• 18:48
  is
• 18:48
  releasing
• 18:48
  in
• 18:48
  preview
• 18:49
  "to
• 18:49
  validate
• 18:49
  these
• 18:50
  updates
• 18:50
  and
• 18:50
  continue
• 18:51
  to
• 18:51
  make
• 18:51
  further
• 18:51
  advancements
• 18:52
  in
• 18:52
  areas
• 18:52
  such
• 18:53
  as
• 18:53
  ambitious
• 18:53
  agentic
• 18:54
  workflows."
• 18:55
  Higher
• 18:55
  limits
• 18:56
  are
• 18:56
  available
• 18:56
  for
• 18:56
  Google
• 18:57
  AI
• 18:57
  Pro
• 18:57
  and
• 18:57
  Ultra
• 18:58
  subscribers.