Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – February 17, 2026

Transcript

AGI Dreams Podcast – February 17, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Alibaba's
• 0:00
  Qwen
• 0:00
  team
• 0:01
  dropped
• 0:01
  Qwen
• 0:01
  3.5
• 0:02
  this
• 0:02
  week
• 0:02
  —
• 0:02
  a
• 0:03
  397B-parameter
• 0:04
  mixture-of-experts
• 0:05
  model
• 0:05
  with
• 0:05
  only
• 0:05
  17B
• 0:06
  active
• 0:07
  parameters
• 0:07
  per
• 0:07
  token,
• 0:08
  Apache
• 0:08
  2.0
• 0:09
  licensed,
• 0:09
  and
• 0:09
  already
• 0:10
  being
• 0:10
  quantized
• 0:10
  for
• 0:10
  Blackwell
• 0:11
  GPUs.
• 0:12
  The
• 0:12
  MoE
• 0:12
  architecture
• 0:13
  (512
• 0:13
  experts,
• 0:14
  10
• 0:14
  active
• 0:15
  per
• 0:15
  token)
• 0:15
  means
• 0:15
  inference
• 0:16
  speed
• 0:16
  lands
• 0:16
  closer
• 0:16
  to
• 0:17
  a
• 0:17
  30-40B
• 0:19
  dense
• 0:19
  model
• 0:19
  despite
• 0:20
  the
• 0:20
  enormous
• 0:20
  total
• 0:20
  parameter
• 0:21
  count.
• 0:22
  Early
• 0:22
  API
• 0:22
  testers
• 0:23
  report
• 0:23
  roughly
• 0:23
  45
• 0:24
  tokens
• 0:24
  per
• 0:24
  second,
• 0:25
  notably
• 0:26
  faster
• 0:26
  than
• 0:26
  Claude
• 0:27
  or
• 0:27
  GPT-4,
• 0:28
  with
• 0:28
  code
• 0:28
  generation
• 0:29
  quality
• 0:29
  at
• 0:29
  "maybe
• 0:29
  90%
• 0:30
  of
• 0:30
  Sonnet
• 0:30
  4"
• 0:31
  and
• 0:32
  significantly
• 0:32
  less
• 0:32
  refusal
• 0:33
  behavior.
• 0:34
  The
• 0:34
  262K
• 0:35
  native
• 0:35
  context
• 0:36
  window
• 0:36
  (with
• 0:36
  1M+
• 0:37
  in
• 0:37
  development)
• 0:37
  and
• 0:38
  built-in
• 0:38
  multimodal
• 0:38
  support
• 0:39
  across
• 0:39
  201
• 0:40
  languages
• 0:41
  make
• 0:41
  this
• 0:41
  a
• 0:41
  serious
• 0:41
  contender
• 0:42
  for
• 0:42
  high-volume
• 0:42
  inference
• 0:43
  shops
• 0:43
  tired
• 0:43
  of
• 0:43
  paying
• 0:44
  per-token
• 0:44
  API
• 0:45
  rates.
• 0:45
  The
• 0:45
  real
• 0:46
  action,
• 0:46
  as
• 0:46
  usual,
• 0:47
  is
• 0:47
  in
• 0:47
  the
• 0:47
  quantization.
• 0:48
  Within
• 0:49
  days,
• 0:49
  an
• 0:49
  NVFP4
• 0:50
  checkpoint
• 0:51
  appeared
• 0:51
  for
• 0:51
  Blackwell
• 0:52
  hardware
• 0:53
  —
• 0:53
  224GB
• 0:54
  total,
• 0:55
  running
• 0:55
  on
• 0:55
  4x
• 0:56
  B300s
• 0:57
  at
• 0:57
  roughly
• 0:57
  120
• 0:58
  tokens
• 0:59
  per
• 0:59
  second
• 0:59
  with
• 1:00
  speculative
• 1:00
  decoding
• 1:01
  via
• 1:01
  Qwen
• 1:01
  3.5's
• 1:02
  built-in
• 1:02
  Multi-Token
• 1:03
  Prediction
• 1:03
  head.
• 1:04
  The
• 1:04
  setup
• 1:04
  requires
• 1:05
  a
• 1:05
  patched
• 1:05
  SGLang
• 1:06
  branch
• 1:06
  to
• 1:06
  avoid
• 1:07
  quantizing
• 1:07
  vision
• 1:07
  encoder
• 1:08
  weights,
• 1:09
  but
• 1:09
  the
• 1:09
  combination
• 1:09
  of
• 1:09
  FP4
• 1:10
  quantization
• 1:11
  with
• 1:11
  near-lossless
• 1:12
  accuracy
• 1:12
  (~99%
• 1:13
  preserved)
• 1:14
  demonstrates
• 1:14
  that
• 1:14
  the
• 1:14
  Blackwell
• 1:15
  architecture
• 1:15
  is
• 1:16
  finally
• 1:16
  delivering
• 1:16
  on
• 1:16
  NVIDIA's
• 1:17
  promise
• 1:17
  of
• 1:17
  practical
• 1:18
  FP4
• 1:18
  inference.
• 1:19
  For
• 1:19
  teams
• 1:20
  with
• 1:20
  the
• 1:20
  hardware
• 1:20
  budget,
• 1:21
  this
• 1:21
  puts
• 1:21
  frontier-competitive
• 1:22
  performance
• 1:22
  on
• 1:22
  self-hosted
• 1:23
  iron.
• 1:24
  On
• 1:24
  the
• 1:24
  lighter
• 1:24
  end
• 1:24
  of
• 1:25
  the
• 1:25
  spectrum,
• 1:25
  Google's
• 1:25
  Gemma
• 1:26
  3n
• 1:26
  E2B
• 1:27
  is
• 1:27
  now
• 1:27
  running
• 1:28
  natively
• 1:28
  on
• 1:28
  Android
• 1:28
  via
• 1:29
  LiteRT,
• 1:30
  with
• 1:30
  one
• 1:30
  developer
• 1:30
  shipping
• 1:31
  a
• 1:31
  fully
• 1:31
  on-device
• 1:32
  multimodal
• 1:32
  app
• 1:32
  handling
• 1:33
  audio
• 1:33
  transcription
• 1:34
  and
• 1:34
  OCR
• 1:34
  using
• 1:35
  INT4
• 1:35
  weights.
• 1:36
  The
• 1:36
  interesting
• 1:37
  engineering
• 1:37
  challenge
• 1:37
  was
• 1:38
  the
• 1:38
  30-second
• 1:38
  audio
• 1:39
  context
• 1:39
  limit
• 1:39
  in
• 1:39
  the
• 1:39
  multimodal
• 1:40
  encoder
• 1:40
  —
• 1:41
  solved
• 1:41
  by
• 1:41
  a
• 1:41
  sequential
• 1:42
  chunking
• 1:42
  pipeline
• 1:43
  that
• 1:43
  splits
• 1:43
  audio,
• 1:44
  transcribes
• 1:44
  chunks
• 1:45
  independently,
• 1:46
  then
• 1:46
  recombines
• 1:46
  via
• 1:47
  a
• 1:47
  second
• 1:47
  model
• 1:47
  pass.
• 1:48
  The
• 1:48
  entire
• 1:49
  app
• 1:49
  was
• 1:49
  built
• 1:50
  without
• 1:50
  writing
• 1:50
  a
• 1:50
  single
• 1:50
  line
• 1:51
  of
• 1:51
  code
• 1:51
  manually,
• 1:52
  using
• 1:52
  Claude
• 1:52
  Code
• 1:53
  and
• 1:53
  Google
• 1:53
  Antigravity,
• 1:54
  though
• 1:54
  the
• 1:55
  developer
• 1:55
  notes
• 1:55
  that
• 1:55
  AI
• 1:56
  agents
• 1:56
  "struggled
• 1:57
  with
• 1:57
  the
• 1:57
  multimodal
• 1:57
  JNI
• 1:58
  bindings"
• 1:58
  until
• 1:59
  manually
• 1:59
  fed
• 2:00
  raw
• 2:00
  LiteRT
• 2:00
  documentation.
• 2:01
  Vibe
• 2:02
  coding
• 2:02
  has
• 2:02
  limits,
• 2:03
  it
• 2:03
  turns
• 2:03
  out,
• 2:04
  and
• 2:04
  they
• 2:04
  tend
• 2:04
  to
• 2:04
  live
• 2:05
  at
• 2:05
  the
• 2:05
  hardware
• 2:05
  abstraction
• 2:06
  boundary.
• 2:07
  For
• 2:07
  teams
• 2:07
  wanting
• 2:07
  to
• 2:08
  self-host
• 2:08
  at
• 2:08
  enterprise
• 2:09
  scale,
• 2:09
  a
• 2:10
  detailed
• 2:10
  quickstart
• 2:10
  appeared
• 2:11
  for
• 2:11
  deploying
• 2:11
  Open
• 2:11
  WebUI
• 2:12
  with
• 2:12
  vLLM's
• 2:12
  production
• 2:13
  stack
• 2:13
  on
• 2:13
  Amazon
• 2:14
  EKS
• 2:14
  —
• 2:15
  OpenTofu-defined
• 2:16
  infrastructure,
• 2:17
  GPU
• 2:17
  node
• 2:18
  groups
• 2:18
  with
• 2:18
  A10G
• 2:19
  cards,
• 2:19
  Gateway
• 2:20
  API
• 2:20
  ingress,
• 2:21
  and
• 2:21
  tool
• 2:21
  calling
• 2:21
  out
• 2:21
  of
• 2:21
  the
• 2:21
  box.
• 2:23
  The
• 2:23
  guide
• 2:23
  is
• 2:23
  refreshingly
• 2:24
  honest
• 2:24
  about
• 2:24
  costs:
• 2:25
  a
• 2:25
  g5.xlarge
• 2:29
  setup
• 2:29
  runs
• 2:29
  roughly
• 2:30
  $50/day,
• 2:31
  and
• 2:31
  the
• 2:31
  author
• 2:32
  explicitly
• 2:32
  recommends
• 2:33
  treating
• 2:33
  it
• 2:33
  as
• 2:33
  a
• 2:33
  development
• 2:34
  environment
• 2:34
  you
• 2:34
  destroy
• 2:35
  when
• 2:35
  not
• 2:35
  in
• 2:35
  use.
• 2:36
  The
• 2:36
  vLLM
• 2:37
  production
• 2:37
  stack
• 2:37
  adds
• 2:38
  a
• 2:38
  router
• 2:38
  layer
• 2:38
  for
• 2:38
  load
• 2:39
  balancing
• 2:39
  across
• 2:39
  replicas,
• 2:40
  making
• 2:41
  it
• 2:41
  a
• 2:41
  genuine
• 2:41
  alternative
• 2:42
  to
• 2:42
  API
• 2:42
  providers
• 2:43
  for
• 2:43
  organizations
• 2:43
  with
• 2:44
  data
• 2:44
  sovereignty
• 2:44
  requirements.
• 2:48
  A
• 2:48
  provocative
• 2:48
  LinkedIn
• 2:49
  thread
• 2:49
  from
• 2:49
  Cameron
• 2:49
  W.
• 2:50
  crystallized
• 2:50
  a
• 2:50
  tension
• 2:51
  that's
• 2:51
  been
• 2:51
  building
• 2:51
  for
• 2:51
  months:
• 2:52
  developers
• 2:53
  are
• 2:53
  already
• 2:53
  shipping
• 2:53
  5x
• 2:54
  the
• 2:54
  code
• 2:54
  they
• 2:54
  were
• 2:54
  a
• 2:54
  year
• 2:55
  ago,
• 2:55
  agents
• 2:55
  are
• 2:55
  writing
• 2:56
  and
• 2:56
  reviewing
• 2:56
  code
• 2:57
  autonomously,
• 2:57
  and
• 2:57
  most
• 2:58
  AppSec
• 2:58
  programs
• 2:59
  are
• 2:59
  still
• 2:59
  tuned
• 2:59
  for
• 2:59
  the
• 2:59
  pace
• 3:00
  of
• 3:00
  human-written
• 3:00
  pull
• 3:00
  requests.
• 3:02
  The
• 3:02
  numbers
• 3:02
  are
• 3:02
  hard
• 3:02
  to
• 3:02
  argue
• 3:03
  with.
• 3:03
  DARPA's
• 3:03
  AI
• 3:04
  Cyber
• 3:04
  Challenge
• 3:04
  at
• 3:04
  DEF
• 3:05
  CON
• 3:05
  saw
• 3:05
  Team
• 3:06
  Atlanta's
• 3:06
  cyber
• 3:06
  reasoning
• 3:07
  system
• 3:08
  go
• 3:08
  from
• 3:08
  identifying
• 3:09
  37%
• 3:10
  of
• 3:10
  synthetic
• 3:10
  vulnerabilities
• 3:11
  at
• 3:11
  semifinals
• 3:12
  to
• 3:12
  86%
• 3:13
  in
• 3:13
  finals,
• 3:14
  at
• 3:14
  roughly
• 3:14
  $152
• 3:16
  per
• 3:16
  task.
• 3:17
  Then
• 3:17
  Anthropic's
• 3:17
  red
• 3:18
  team
• 3:18
  had
• 3:18
  Claude
• 3:18
  Opus
• 3:19
  4.6
• 3:20
  finding
• 3:20
  500+
• 3:21
  previously
• 3:21
  unknown
• 3:22
  high-severity
• 3:23
  vulnerabilities
• 3:23
  in
• 3:23
  projects
• 3:24
  like
• 3:24
  Ghostscript
• 3:24
  and
• 3:24
  OpenSC
• 3:25
  —
• 3:25
  no
• 3:26
  custom
• 3:26
  tooling,
• 3:27
  no
• 3:27
  specialized
• 3:27
  prompts,
• 3:28
  just
• 3:28
  reasoning
• 3:29
  through
• 3:29
  commit
• 3:29
  history
• 3:30
  and
• 3:30
  writing
• 3:30
  proof-of-concept
• 3:31
  exploits
• 3:31
  for
• 3:31
  bugs
• 3:31
  that
• 3:31
  fuzzers
• 3:32
  running
• 3:32
  for
• 3:32
  years
• 3:32
  had
• 3:33
  missed.
• 3:33
  The
• 3:33
  follow-up
• 3:34
  discussion
• 3:34
  surfaced
• 3:34
  a
• 3:35
  genuine
• 3:35
  strategic
• 3:35
  question:
• 3:36
  if
• 3:36
  vulnerability
• 3:37
  discovery
• 3:38
  becomes
• 3:38
  near-commodity,
• 3:39
  what's
• 3:39
  the
• 3:39
  scarce
• 3:40
  resource?
• 3:41
  The
• 3:41
  consensus
• 3:41
  among
• 3:42
  security
• 3:42
  leaders
• 3:42
  points
• 3:43
  toward
• 3:43
  validation
• 3:43
  quality,
• 3:44
  exploitability
• 3:45
  modeling,
• 3:45
  and
• 3:45
  governance
• 3:46
  over
• 3:46
  machine-generated
• 3:47
  change.
• 3:48
  Several
• 3:48
  commenters
• 3:49
  noted
• 3:49
  that
• 3:49
  the
• 3:49
  real
• 3:49
  gap
• 3:50
  isn't
• 3:50
  tooling
• 3:50
  —
• 3:50
  it's
• 3:51
  measurement.
• 3:52
  How
• 3:52
  do
• 3:52
  you
• 3:52
  quantify
• 3:52
  false
• 3:53
  confidence
• 3:53
  introduced
• 3:53
  by
• 3:54
  automated
• 3:54
  patching?
• 3:55
  How
• 3:55
  do
• 3:55
  you
• 3:56
  detect
• 3:56
  degradation
• 3:57
  in
• 3:57
  review
• 3:57
  depth
• 3:57
  when
• 3:57
  agents
• 3:58
  review
• 3:58
  agents?
• 4:00
  The
• 4:00
  most
• 4:00
  actionable
• 4:00
  take:
• 4:01
  AppSec
• 4:02
  teams
• 4:02
  need
• 4:02
  to
• 4:02
  shift
• 4:03
  from
• 4:03
  detection-centric
• 4:04
  to
• 4:04
  prevention-centric,
• 4:05
  codifying
• 4:05
  security
• 4:06
  patterns
• 4:06
  in
• 4:06
  markdown
• 4:06
  that
• 4:06
  coding
• 4:07
  agents
• 4:07
  can
• 4:07
  consume,
• 4:08
  and
• 4:08
  focusing
• 4:08
  on
• 4:08
  business
• 4:09
  logic
• 4:09
  flaws
• 4:09
  rather
• 4:10
  than
• 4:10
  just
• 4:10
  CVEs
• 4:10
  and
• 4:11
  CWEs.
• 4:13
  Oligo
• 4:13
  offered
• 4:13
  one
• 4:13
  concrete
• 4:14
  answer
• 4:14
  to
• 4:14
  the
• 4:14
  "how
• 4:14
  do
• 4:14
  we
• 4:14
  scale?"
• 4:15
  question:
• 4:16
  a
• 4:17
  Named
• 4:17
  Entity
• 4:17
  Recognition
• 4:18
  pipeline
• 4:19
  that
• 4:19
  extracts
• 4:20
  vulnerable
• 4:20
  functions
• 4:21
  directly
• 4:21
  from
• 4:21
  advisories
• 4:22
  and
• 4:22
  patches,
• 4:23
  validated
• 4:23
  against
• 4:24
  runtime
• 4:24
  data.
• 4:25
  Their
• 4:25
  key
• 4:26
  finding
• 4:26
  was
• 4:26
  that
• 4:26
  traditional
• 4:27
  vulnerability
• 4:27
  databases
• 4:28
  cover
• 4:28
  less
• 4:28
  than
• 4:29
  4%
• 4:29
  of
• 4:29
  vulnerabilities
• 4:30
  at
• 4:30
  the
• 4:30
  function
• 4:30
  level
• 4:31
  —
• 4:31
  a
• 4:32
  massive
• 4:32
  intelligence
• 4:32
  gap.
• 4:33
  By
• 4:34
  deploying
• 4:34
  NVIDIA's
• 4:35
  Nemotron
• 4:35
  Nano
• 4:35
  9B
• 4:36
  v2
• 4:36
  for
• 4:37
  real-time
• 4:37
  detection
• 4:37
  tasks,
• 4:38
  they
• 4:38
  achieved
• 4:39
  55x
• 4:40
  cost
• 4:40
  reduction
• 4:41
  and
• 4:41
  2.5x
• 4:42
  faster
• 4:42
  response
• 4:42
  times
• 4:43
  compared
• 4:43
  to
• 4:43
  Claude
• 4:44
  Sonnet
• 4:44
  4.5,
• 4:45
  with
• 4:45
  no
• 4:45
  accuracy
• 4:46
  loss.
• 4:47
  Their
• 4:47
  enrichment
• 4:47
  pipeline
• 4:48
  now
• 4:48
  covers
• 4:48
  an
• 4:48
  entire
• 4:49
  year
• 4:49
  of
• 4:49
  CVEs
• 4:50
  (50K)
• 4:50
  in
• 4:50
  under
• 4:51
  an
• 4:51
  hour,
• 4:52
  identifying
• 4:52
  1100%
• 4:55
  more
• 4:55
  vulnerable
• 4:56
  functions
• 4:56
  than
• 4:56
  commercial
• 4:57
  databases.
• 4:58
  The
• 4:58
  lesson:
• 4:59
  specialized,
• 5:00
  instruction-tuned
• 5:00
  models
• 5:01
  systematically
• 5:01
  outperform
• 5:02
  expensive
• 5:03
  general-purpose
• 5:03
  LLMs
• 5:04
  for
• 5:04
  domain-specific
• 5:05
  security
• 5:05
  tasks.
• 5:09
  The
• 5:09
  agenda
• 5:09
  for
• 5:10
  the
• 5:10
  [un]prompted
• 5:10
  AI
• 5:10
  Security
• 5:11
  Conference
• 5:11
  (March
• 5:12
  3-4,
• 5:12
  Salesforce
• 5:13
  Tower,
• 5:13
  SF)
• 5:14
  dropped
• 5:14
  this
• 5:14
  week,
• 5:15
  and
• 5:15
  the
• 5:15
  talk
• 5:15
  list
• 5:16
  reads
• 5:16
  like
• 5:16
  a
• 5:16
  threat
• 5:16
  briefing.
• 5:17
  Nearly
• 5:17
  500
• 5:18
  submissions
• 5:19
  have
• 5:19
  been
• 5:19
  narrowed
• 5:19
  to
• 5:19
  two
• 5:20
  days
• 5:20
  of
• 5:20
  talks
• 5:20
  spanning
• 5:21
  offensive
• 5:21
  research,
• 5:22
  defensive
• 5:22
  engineering,
• 5:23
  and
• 5:23
  governance
• 5:23
  —
• 5:24
  with
• 5:24
  speakers
• 5:24
  from
• 5:25
  Google,
• 5:25
  Meta,
• 5:25
  Stripe,
• 5:26
  OpenAI,
• 5:27
  Microsoft,
• 5:27
  NVIDIA,
• 5:28
  Netflix,
• 5:29
  Reddit,
• 5:29
  Salesforce,
• 5:30
  and
• 5:30
  the
• 5:30
  U.S.
• 5:30
  Army
• 5:30
  Cyber
• 5:31
  Command.
• 5:32
  The
• 5:32
  offensive
• 5:32
  highlights
• 5:33
  alone
• 5:33
  could
• 5:33
  fill
• 5:33
  a
• 5:33
  briefing.
• 5:34
  XBOW
• 5:35
  will
• 5:35
  demonstrate
• 5:36
  AI
• 5:36
  agents
• 5:37
  that
• 5:37
  autonomously
• 5:37
  find
• 5:38
  and
• 5:38
  validate
• 5:38
  access
• 5:39
  control
• 5:39
  flaws
• 5:40
  in
• 5:40
  production
• 5:40
  systems.
• 5:41
  Trend
• 5:42
  Micro's
• 5:42
  FENRIR
• 5:45
  system
• 5:45
  has
• 5:45
  discovered
• 5:46
  100+
• 5:46
  vulnerabilities
• 5:47
  in
• 5:47
  AI
• 5:47
  infrastructure
• 5:48
  since
• 5:48
  mid-2025,
• 5:50
  including
• 5:50
  21
• 5:51
  CVEs
• 5:51
  with
• 5:51
  multiple
• 5:52
  CVSS
• 5:53
  9.8
• 5:53
  RCEs,
• 5:54
  using
• 5:55
  a
• 5:55
  multi-stage
• 5:55
  verification
• 5:56
  pipeline
• 5:56
  of
• 5:57
  static
• 5:57
  analysis
• 5:57
  and
• 5:58
  two-layer
• 5:58
  LLM
• 5:59
  validation
• 5:59
  —
• 6:00
  and
• 6:00
  the
• 6:00
  team
• 6:00
  will
• 6:00
  present
• 6:01
  their
• 6:01
  methodology.
• 6:02
  A
• 6:02
  hardware
• 6:03
  hacker
• 6:03
  who
• 6:03
  gave
• 6:04
  Claude
• 6:04
  access
• 6:04
  to
• 6:04
  his
• 6:05
  physical
• 6:05
  lab
• 6:05
  —
• 6:06
  oscilloscope,
• 6:07
  ChipShouter,
• 6:07
  JTAG
• 6:08
  probe
• 6:08
  —
• 6:09
  will
• 6:09
  show
• 6:09
  how
• 6:09
  it
• 6:09
  pwned
• 6:10
  an
• 6:10
  LPC
• 6:10
  chip
• 6:11
  in
• 6:11
  7
• 6:11
  minutes
• 6:11
  that
• 6:12
  he'd
• 6:12
  failed
• 6:12
  to
• 6:12
  glitch
• 6:12
  for
• 6:12
  6
• 6:13
  weeks.
• 6:14
  Trail
• 6:14
  of
• 6:14
  Bits
• 6:14
  will
• 6:15
  share
• 6:15
  how
• 6:15
  they
• 6:15
  rebuilt
• 6:16
  their
• 6:16
  entire
• 6:16
  organization
• 6:17
  around
• 6:17
  AI,
• 6:18
  achieving
• 6:18
  200
• 6:19
  bugs
• 6:19
  per
• 6:19
  week
• 6:19
  per
• 6:20
  engineer.
• 6:21
  And
• 6:21
  in
• 6:21
  one
• 6:21
  of
• 6:21
  the
• 6:21
  more
• 6:21
  sobering
• 6:22
  scheduled
• 6:22
  talks,
• 6:23
  Profero's
• 6:23
  IR
• 6:24
  team
• 6:24
  will
• 6:24
  present
• 6:25
  "Your
• 6:25
  Next
• 6:25
  Breach
• 6:25
  Won't
• 6:26
  Have
• 6:26
  an
• 6:26
  Attacker"
• 6:26
  —
• 6:27
  real
• 6:28
  incidents
• 6:28
  where
• 6:28
  AI
• 6:29
  coding
• 6:29
  assistants
• 6:29
  with
• 6:30
  overly
• 6:30
  broad
• 6:30
  permissions
• 6:31
  caused
• 6:31
  production
• 6:31
  database
• 6:32
  exposures,
• 6:33
  configuration
• 6:33
  tampering,
• 6:34
  and
• 6:34
  mass
• 6:34
  deletion,
• 6:35
  all
• 6:35
  initially
• 6:36
  presenting
• 6:36
  like
• 6:37
  sophisticated
• 6:37
  attacks.
• 6:38
  On
• 6:38
  the
• 6:38
  defensive
• 6:39
  side,
• 6:39
  Stripe
• 6:40
  will
• 6:40
  share
• 6:40
  how
• 6:40
  they
• 6:40
  ship
• 6:40
  security
• 6:41
  agents
• 6:41
  in
• 6:41
  production
• 6:42
  with
• 6:42
  modular
• 6:43
  orchestrator/child
• 6:44
  architectures
• 6:45
  and
• 6:45
  structured
• 6:45
  evaluation
• 6:46
  loops.
• 6:47
  Reddit
• 6:47
  will
• 6:47
  present
• 6:48
  their
• 6:48
  AI
• 6:48
  Guardrails
• 6:49
  system,
• 6:49
  which
• 6:50
  blocks
• 6:50
  99%
• 6:51
  of
• 6:51
  unsafe
• 6:51
  requests
• 6:52
  at
• 6:52
  25ms
• 6:53
  p95
• 6:54
  latency
• 6:54
  across
• 6:55
  millions
• 6:55
  of
• 6:55
  users.
• 6:56
  Salesforce
• 6:57
  will
• 6:57
  detail
• 6:57
  how
• 6:57
  they
• 6:58
  distill
• 6:58
  millions
• 6:58
  of
• 6:58
  daily
• 6:59
  prompts
• 6:59
  from
• 6:59
  12,000
• 7:00
  autonomous
• 7:00
  agents
• 7:01
  into
• 7:01
  fewer
• 7:01
  than
• 7:01
  30
• 7:02
  high-fidelity
• 7:02
  daily
• 7:03
  alerts
• 7:03
  using
• 7:04
  behavioral
• 7:04
  baselines
• 7:05
  like
• 7:05
  asset
• 7:05
  rarity
• 7:06
  and
• 7:06
  query
• 7:06
  complexity.
• 7:07
  And
• 7:07
  Snap
• 7:07
  will
• 7:07
  present
• 7:08
  capability-based
• 7:08
  authorization
• 7:09
  using
• 7:09
  cryptographic
• 7:10
  warrants
• 7:10
  —
• 7:11
  ephemeral,
• 7:12
  task-scoped
• 7:13
  tokens
• 7:13
  that
• 7:13
  attenuate
• 7:14
  on
• 7:14
  delegation,
• 7:15
  verified
• 7:15
  offline
• 7:15
  in
• 7:16
  microseconds,
• 7:16
  so
• 7:17
  that
• 7:17
  even
• 7:17
  a
• 7:17
  fully
• 7:17
  compromised
• 7:18
  agent
• 7:18
  can't
• 7:19
  escalate
• 7:19
  beyond
• 7:19
  its
• 7:19
  bounds.
• 7:20
  The
• 7:20
  conference
• 7:21
  lineup
• 7:21
  makes
• 7:21
  one
• 7:22
  thing
• 7:22
  clear:
• 7:23
  AI
• 7:23
  security
• 7:24
  has
• 7:24
  moved
• 7:24
  from
• 7:24
  theoretical
• 7:25
  to
• 7:25
  operational,
• 7:26
  and
• 7:26
  the
• 7:26
  gap
• 7:26
  between
• 7:27
  offense
• 7:27
  and
• 7:27
  defense
• 7:28
  is
• 7:28
  narrowing
• 7:28
  fast.
• 7:32
  Craig
• 7:33
  Hepburn's
• 7:33
  Substack
• 7:34
  piece
• 7:34
  "The
• 7:34
  Agentic
• 7:35
  Operating
• 7:35
  System"
• 7:36
  captured
• 7:36
  something
• 7:36
  visceral:
• 7:37
  he
• 7:37
  deployed
• 7:37
  an
• 7:38
  autonomous
• 7:38
  AI
• 7:38
  agent
• 7:39
  in
• 7:39
  January
• 7:39
  that,
• 7:39
  within
• 7:40
  48
• 7:40
  hours,
• 7:41
  was
• 7:41
  reading
• 7:41
  his
• 7:42
  emails
• 7:42
  overnight,
• 7:43
  parsing
• 7:43
  call
• 7:43
  transcripts,
• 7:44
  reviewing
• 7:44
  messages
• 7:45
  across
• 7:45
  WhatsApp
• 7:45
  and
• 7:46
  Telegram,
• 7:47
  and
• 7:47
  organizing
• 7:47
  his
• 7:47
  day
• 7:48
  before
• 7:48
  he'd
• 7:48
  finished
• 7:48
  his
• 7:49
  first
• 7:49
  coffee.
• 7:50
  "Its
• 7:50
  logic
• 7:50
  was
• 7:50
  sound.
• 7:51
  Its
• 7:51
  priorities
• 7:52
  were
• 7:52
  defensible.
• 7:53
  But
• 7:53
  it
• 7:53
  was
• 7:53
  moving
• 7:53
  faster
• 7:54
  than
• 7:54
  my
• 7:54
  own
• 7:54
  thinking."
• 7:55
  The
• 7:55
  backdrop
• 7:56
  is
• 7:56
  Peter
• 7:56
  Steinberger's
• 7:56
  OpenClaw
• 7:57
  joining
• 7:57
  OpenAI
• 7:58
  —
• 7:58
  an
• 7:59
  autonomous
• 7:59
  agent
• 7:59
  with
• 8:00
  180,000
• 8:00
  GitHub
• 8:01
  stars
• 8:01
  that
• 8:01
  operates
• 8:02
  continuously
• 8:02
  across
• 8:03
  email,
• 8:03
  calendar,
• 8:04
  messaging,
• 8:04
  and
• 8:04
  code.
• 8:05
  Steinberger's
• 8:06
  quote
• 8:06
  lands:
• 8:07
  "Every
• 8:07
  app
• 8:07
  is
• 8:07
  just
• 8:07
  a
• 8:07
  very
• 8:08
  slow
• 8:08
  API
• 8:08
  now."
• 8:09
  But
• 8:09
  the
• 8:10
  cost
• 8:10
  reality
• 8:10
  is
• 8:10
  brutal.
• 8:11
  Hepburn
• 8:11
  was
• 8:12
  spending
• 8:12
  £20-30/day
• 8:13
  for
• 8:13
  a
• 8:13
  single
• 8:14
  personal
• 8:14
  agent.
• 8:14
  Steinberger
• 8:15
  disclosed
• 8:16
  $10,000-20,000
• 8:17
  monthly
• 8:17
  running
• 8:18
  OpenClaw.
• 8:19
  The
• 8:19
  solution
• 8:19
  —
• 8:20
  multi-model
• 8:20
  architectures
• 8:21
  with
• 8:21
  different
• 8:21
  models
• 8:22
  for
• 8:22
  different
• 8:22
  task
• 8:22
  complexity
• 8:23
  levels
• 8:23
  —
• 8:23
  cut
• 8:24
  costs
• 8:24
  by
• 8:24
  90%,
• 8:25
  but
• 8:25
  that
• 8:25
  architectural
• 8:26
  discipline
• 8:26
  is
• 8:26
  non-trivial.
• 8:27
  And
• 8:28
  the
• 8:28
  security
• 8:28
  surface
• 8:28
  is
• 8:29
  exactly
• 8:29
  what
• 8:29
  you'd
• 8:29
  expect:
• 8:30
  within
• 8:30
  weeks
• 8:30
  of
• 8:30
  OpenClaw's
• 8:31
  viral
• 8:31
  rise,
• 8:32
  researchers
• 8:32
  found
• 8:33
  over
• 8:33
  1,800
• 8:34
  exposed
• 8:34
  instances
• 8:35
  leaking
• 8:35
  API
• 8:35
  keys,
• 8:36
  chat
• 8:36
  histories,
• 8:37
  and
• 8:37
  credentials.
• 8:38
  Simon
• 8:38
  Willison's
• 8:38
  "lethal
• 8:39
  trifecta"
• 8:39
  —
• 8:40
  access
• 8:40
  to
• 8:40
  private
• 8:40
  data,
• 8:41
  exposure
• 8:41
  to
• 8:41
  untrusted
• 8:42
  content,
• 8:42
  ability
• 8:43
  to
• 8:43
  communicate
• 8:43
  externally
• 8:43
  —
• 8:44
  describes
• 8:44
  every
• 8:45
  useful
• 8:45
  agent
• 8:45
  deployment.
• 8:46
  The
• 8:46
  air-gapped
• 8:47
  response
• 8:47
  is
• 8:47
  already
• 8:47
  materializing.
• 8:49
  One
• 8:49
  developer
• 8:49
  forked
• 8:50
  OpenClaw
• 8:50
  into
• 8:50
  "Physiclaw,"
• 8:51
  stripping
• 8:52
  cloud
• 8:52
  dependencies
• 8:52
  and
• 8:53
  targeting
• 8:53
  local
• 8:53
  inference
• 8:54
  via
• 8:54
  vLLM
• 8:54
  and
• 8:55
  llama.cpp.
• 8:56
  The
• 8:57
  more
• 8:57
  interesting
• 8:57
  architectural
• 8:58
  decision
• 8:58
  was
• 8:58
  breaking
• 8:59
  the
• 8:59
  single
• 8:59
  omniscient
• 8:59
  agent
• 9:00
  into
• 9:00
  role-specific
• 9:01
  agents
• 9:01
  (SRE,
• 9:02
  SecOps)
• 9:03
  with
• 9:03
  scoped
• 9:03
  tool
• 9:03
  access
• 9:04
  —
• 9:04
  addressing
• 9:05
  the
• 9:05
  "one
• 9:05
  generic
• 9:05
  assistant
• 9:06
  with
• 9:06
  root
• 9:06
  access
• 9:07
  to
• 9:07
  everything"
• 9:07
  problem
• 9:08
  that
• 9:08
  kills
• 9:08
  agent
• 9:08
  setups
• 9:09
  in
• 9:09
  security-conscious
• 9:10
  environments.
• 9:11
  Community
• 9:11
  feedback
• 9:12
  immediately
• 9:12
  zeroed
• 9:13
  in
• 9:13
  on
• 9:13
  whether
• 9:13
  the
• 9:14
  role
• 9:14
  separation
• 9:14
  has
• 9:15
  actual
• 9:15
  execution
• 9:15
  boundaries
• 9:16
  or
• 9:16
  just
• 9:16
  logical
• 9:17
  ones,
• 9:17
  and
• 9:17
  whether
• 9:18
  transitive
• 9:18
  dependencies
• 9:19
  might
• 9:19
  still
• 9:19
  phone
• 9:19
  home.
• 9:20
  Meanwhile,
• 9:21
  Anthropic
• 9:22
  continues
• 9:22
  to
• 9:22
  navigate
• 9:23
  the
• 9:23
  political
• 9:23
  tightrope
• 9:24
  of
• 9:24
  defense
• 9:24
  contracts,
• 9:25
  with
• 9:25
  CEO
• 9:26
  Dario
• 9:26
  Amodei
• 9:27
  stating
• 9:27
  that
• 9:27
  AI
• 9:28
  should
• 9:28
  support
• 9:28
  defense
• 9:29
  "except
• 9:29
  those
• 9:29
  which
• 9:30
  would
• 9:30
  make
• 9:30
  us
• 9:30
  more
• 9:30
  like
• 9:30
  our
• 9:31
  autocratic
• 9:31
  adversaries."
• 9:33
  The
• 9:33
  community
• 9:33
  response
• 9:33
  was
• 9:34
  skeptical,
• 9:35
  noting
• 9:35
  the
• 9:35
  existing
• 9:35
  Palantir
• 9:36
  partnership.
• 9:37
  And
• 9:37
  in
• 9:37
  a
• 9:37
  twist
• 9:38
  of
• 9:38
  corporate
• 9:38
  surveillance
• 9:39
  irony,
• 9:39
  OpenAI
• 9:40
  is
• 9:40
  reportedly
• 9:41
  using
• 9:41
  an
• 9:41
  internal
• 9:41
  version
• 9:42
  of
• 9:42
  ChatGPT
• 9:43
  to
• 9:43
  analyze
• 9:44
  employee
• 9:44
  data
• 9:44
  and
• 9:44
  identify
• 9:45
  staffers
• 9:45
  leaking
• 9:46
  information
• 9:46
  to
• 9:46
  the
• 9:46
  press
• 9:46
  —
• 9:47
  using
• 9:48
  their
• 9:48
  own
• 9:48
  technology
• 9:49
  for
• 9:49
  counter-intelligence
• 9:49
  against
• 9:50
  their
• 9:50
  own
• 9:50
  workforce.
• 9:54
  FormalTask
• 9:54
  emerged
• 9:55
  this
• 9:55
  week
• 9:55
  as
• 9:56
  a
• 9:56
  declarative
• 9:56
  orchestration
• 9:57
  layer
• 9:57
  for
• 9:57
  Claude
• 9:57
  Code
• 9:58
  that
• 9:58
  solves
• 9:58
  a
• 9:58
  real
• 9:58
  problem:
• 10:00
  coding
• 10:00
  agents
• 10:00
  write
• 10:00
  code
• 10:00
  well
• 10:01
  but
• 10:01
  have
• 10:01
  no
• 10:01
  concept
• 10:02
  of
• 10:02
  "done."
• 10:03
  The
• 10:03
  system
• 10:03
  lets
• 10:03
  you
• 10:04
  define
• 10:04
  behavioral
• 10:04
  contracts
• 10:05
  per
• 10:05
  task
• 10:05
  —
• 10:06
  required
• 10:07
  review
• 10:07
  types
• 10:07
  (17
• 10:08
  available,
• 10:09
  from
• 10:09
  security
• 10:09
  to
• 10:09
  SQLite
• 10:10
  to
• 10:10
  path-security),
• 10:11
  machine-runnable
• 10:12
  acceptance
• 10:12
  criteria,
• 10:13
  and
• 10:13
  custom
• 10:13
  completion
• 10:14
  rules
• 10:14
  with
• 10:14
  a
• 10:14
  condition
• 10:15
  DSL
• 10:15
  supporting
• 10:16
  AND/OR/NOT
• 10:18
  and
• 10:18
  dotted
• 10:18
  path
• 10:18
  resolution.
• 10:20
  When
• 10:20
  `ft
• 10:20
  task
• 10:20
  complete`
• 10:21
  runs,
• 10:21
  a
• 10:21
  60-line
• 10:22
  kernel
• 10:22
  gathers
• 10:23
  state
• 10:23
  from
• 10:23
  SQLite
• 10:23
  and
• 10:23
  evaluates
• 10:24
  the
• 10:24
  task's
• 10:24
  rules.
• 10:26
  If
• 10:26
  findings
• 10:26
  persist
• 10:27
  after
• 10:27
  two
• 10:27
  self-critique
• 10:28
  rounds,
• 10:28
  custom
• 10:29
  rules
• 10:29
  can
• 10:29
  escalate
• 10:29
  to
• 10:30
  a
• 10:30
  human.
• 10:31
  Workers
• 10:31
  run
• 10:31
  as
• 10:31
  Claude
• 10:32
  Code
• 10:32
  sessions
• 10:32
  in
• 10:32
  tmux
• 10:33
  with
• 10:33
  dedicated
• 10:33
  git
• 10:33
  worktrees,
• 10:34
  and
• 10:35
  a
• 10:35
  TUI
• 10:35
  dashboard
• 10:36
  lets
• 10:36
  you
• 10:36
  monitor,
• 10:37
  attach,
• 10:37
  kill,
• 10:38
  and
• 10:38
  scale
• 10:38
  1-10
• 10:39
  concurrent
• 10:39
  agents.
• 10:41
  The
• 10:41
  system-level
• 10:41
  gating,
• 10:42
  as
• 10:42
  one
• 10:42
  commenter
• 10:42
  noted,
• 10:43
  is
• 10:43
  "solving
• 10:44
  the
• 10:44
  right
• 10:44
  problem
• 10:44
  —
• 10:45
  most
• 10:46
  multi-agent
• 10:46
  setups
• 10:47
  let
• 10:47
  each
• 10:47
  agent
• 10:47
  self-evaluate,
• 10:48
  which
• 10:49
  is
• 10:49
  basically
• 10:49
  asking
• 10:50
  the
• 10:50
  student
• 10:50
  to
• 10:50
  grade
• 10:50
  their
• 10:51
  own
• 10:51
  exam."
• 10:52
  Get
• 10:52
  Shit
• 10:52
  Done
• 10:52
  (GSD)
• 10:53
  takes
• 10:53
  a
• 10:53
  different
• 10:53
  approach
• 10:54
  —
• 10:54
  a
• 10:54
  meta-prompting
• 10:55
  and
• 10:55
  context
• 10:56
  engineering
• 10:56
  system
• 10:56
  for
• 10:57
  Claude
• 10:57
  Code,
• 10:57
  OpenCode,
• 10:58
  and
• 10:59
  Gemini
• 10:59
  CLI
• 11:00
  that
• 11:00
  targets
• 11:00
  "context
• 11:01
  rot,"
• 11:01
  the
• 11:01
  quality
• 11:02
  degradation
• 11:02
  that
• 11:03
  happens
• 11:03
  as
• 11:03
  Claude
• 11:03
  fills
• 11:04
  its
• 11:04
  context
• 11:04
  window.
• 11:05
  The
• 11:05
  workflow
• 11:06
  is
• 11:06
  opinionated:
• 11:07
  discuss
• 11:07
  phase
• 11:07
  →
• 11:08
  plan
• 11:08
  phase
• 11:08
  (spawning
• 11:09
  parallel
• 11:09
  research
• 11:09
  agents)
• 11:10
  →
• 11:10
  execute
• 11:11
  phase
• 11:11
  (fresh
• 11:12
  200K-token
• 11:12
  context
• 11:13
  per
• 11:13
  plan,
• 11:13
  parallel
• 11:14
  waves
• 11:14
  with
• 11:14
  dependency
• 11:15
  tracking)
• 11:15
  →
• 11:15
  verify.
• 11:16
  The
• 11:16
  key
• 11:17
  architectural
• 11:17
  insight
• 11:18
  is
• 11:18
  that
• 11:18
  each
• 11:18
  plan
• 11:18
  executes
• 11:19
  in
• 11:19
  a
• 11:19
  fresh
• 11:19
  subagent
• 11:20
  context,
• 11:21
  so
• 11:21
  the
• 11:21
  orchestrator's
• 11:21
  main
• 11:22
  window
• 11:22
  stays
• 11:22
  at
• 11:22
  30-40%
• 11:25
  capacity.
• 11:26
  It's
• 11:26
  already
• 11:26
  seeing
• 11:27
  adoption
• 11:27
  at
• 11:27
  Amazon,
• 11:28
  Google,
• 11:29
  Shopify,
• 11:29
  and
• 11:30
  Webflow,
• 11:30
  according
• 11:30
  to
• 11:30
  the
• 11:30
  project.
• 11:32
  The
• 11:32
  security-relevant
• 11:33
  detail:
• 11:34
  GSD
• 11:34
  recommends
• 11:35
  running
• 11:35
  Claude
• 11:35
  Code
• 11:36
  with
• 11:36
  `--dangerously-skip-permissions`,
• 11:38
  which
• 11:38
  is
• 11:38
  exactly
• 11:39
  the
• 11:39
  kind
• 11:39
  of
• 11:39
  convenience-over-safety
• 11:40
  tradeoff
• 11:41
  that
• 11:41
  will
• 11:41
  generate
• 11:41
  incident
• 11:42
  reports
• 11:42
  at
• 11:42
  next
• 11:42
  year's
• 11:43
  UnPrompted.
• 11:44
  On
• 11:44
  the
• 11:44
  pedagogical
• 11:45
  end,
• 11:45
  Andrej
• 11:46
  Karpathy
• 11:46
  published
• 11:47
  what
• 11:47
  he
• 11:47
  calls
• 11:48
  "the
• 11:48
  most
• 11:48
  atomic
• 11:48
  way
• 11:49
  to
• 11:49
  train
• 11:49
  and
• 11:49
  run
• 11:50
  inference
• 11:50
  for
• 11:50
  a
• 11:50
  GPT
• 11:51
  in
• 11:51
  pure,
• 11:52
  dependency-free
• 11:52
  Python"
• 11:53
  —
• 11:54
  a
• 11:54
  single-file
• 11:54
  implementation
• 11:55
  covering
• 11:55
  tokenization,
• 11:57
  autograd,
• 11:57
  multi-head
• 11:58
  attention,
• 11:59
  Adam
• 11:59
  optimizer,
• 12:00
  and
• 12:00
  inference.
• 12:01
  The
• 12:01
  gist
• 12:02
  spawned
• 12:02
  immediate
• 12:03
  ports
• 12:03
  to
• 12:03
  C#,
• 12:04
  JavaScript,
• 12:05
  Rust,
• 12:05
  Common
• 12:06
  Lisp,
• 12:06
  and
• 12:06
  Go.
• 12:08
  One
• 12:08
  contributor
• 12:08
  extended
• 12:09
  it
• 12:09
  into
• 12:10
  "molecule,"
• 12:11
  a
• 12:11
  stdlib-only
• 12:11
  variant
• 12:12
  with
• 12:12
  RoPE,
• 12:13
  SwiGLU
• 12:13
  gating,
• 12:14
  async
• 12:14
  continual
• 12:15
  training,
• 12:15
  SQLite
• 12:16
  memory,
• 12:17
  and
• 12:17
  evolving
• 12:17
  BPE.
• 12:19
  When
• 12:19
  someone
• 12:19
  asked
• 12:20
  if
• 12:20
  he
• 12:20
  vibe-coded
• 12:20
  it
• 12:20
  in
• 12:21
  Claude,
• 12:21
  the
• 12:21
  silence
• 12:22
  spoke
• 12:22
  volumes.
• 12:26
  Google
• 12:27
  announced
• 12:27
  WebMCP
• 12:28
  for
• 12:28
  early
• 12:28
  preview
• 12:28
  in
• 12:28
  Chrome
• 12:29
  —
• 12:29
  a
• 12:29
  proposed
• 12:30
  web
• 12:30
  standard
• 12:30
  that
• 12:30
  exposes
• 12:31
  structured
• 12:31
  tools
• 12:32
  for
• 12:32
  AI
• 12:32
  agents
• 12:32
  on
• 12:33
  existing
• 12:33
  websites,
• 12:34
  replacing
• 12:34
  screen-scraping
• 12:35
  with
• 12:35
  robust,
• 12:35
  high-performance
• 12:36
  page
• 12:36
  interaction.
• 12:37
  The
• 12:37
  spec
• 12:37
  proposes
• 12:38
  two
• 12:38
  new
• 12:38
  APIs:
• 12:39
  one
• 12:39
  for
• 12:39
  standard
• 12:40
  HTML
• 12:40
  form
• 12:40
  actions
• 12:41
  and
• 12:41
  another
• 12:41
  for
• 12:41
  complex
• 12:42
  interactions
• 12:42
  requiring
• 12:43
  JavaScript
• 12:43
  execution.
• 12:44
  The
• 12:44
  use
• 12:45
  cases
• 12:45
  are
• 12:45
  practical
• 12:45
  —
• 12:46
  booking
• 12:46
  flights,
• 12:47
  filing
• 12:47
  support
• 12:48
  tickets,
• 12:48
  navigating
• 12:49
  e-commerce
• 12:49
  checkout
• 12:49
  flows
• 12:50
  —
• 12:50
  and
• 12:51
  the
• 12:51
  design
• 12:51
  philosophy
• 12:51
  treats
• 12:52
  websites
• 12:52
  as
• 12:52
  active
• 12:53
  participants
• 12:53
  in
• 12:54
  agent
• 12:54
  interactions
• 12:55
  rather
• 12:55
  than
• 12:55
  passive
• 12:55
  targets
• 12:56
  for
• 12:56
  DOM
• 12:56
  actuation.
• 12:57
  The
• 12:57
  community
• 12:58
  is
• 12:58
  cautiously
• 12:58
  interested.
• 13:00
  The
• 13:00
  open-source
• 13:00
  implementation
• 13:01
  at
• 13:01
  webmcp.dev
• 13:04
  is
• 13:04
  drawing
• 13:04
  attention,
• 13:05
  though
• 13:05
  adoption
• 13:05
  is
• 13:05
  nascent
• 13:06
  —
• 13:06
  most
• 13:07
  developers
• 13:07
  haven't
• 13:07
  used
• 13:08
  it
• 13:08
  yet,
• 13:08
  and
• 13:08
  the
• 13:08
  comparison
• 13:09
  to
• 13:09
  Google
• 13:09
  Antigravity's
• 13:10
  browser
• 13:10
  tools
• 13:10
  remains
• 13:11
  unclear.
• 13:12
  The
• 13:12
  timing
• 13:12
  is
• 13:12
  notable:
• 13:13
  this
• 13:13
  lands
• 13:14
  alongside
• 13:14
  talks
• 13:15
  at
• 13:15
  UnPrompted
• 13:15
  about
• 13:15
  hacking
• 13:16
  AI
• 13:16
  browsers,
• 13:17
  where
• 13:17
  researchers
• 13:18
  broke
• 13:18
  Perplexity
• 13:18
  Comet,
• 13:19
  OpenAI
• 13:20
  Atlas,
• 13:21
  and
• 13:21
  Google
• 13:21
  Antigravity
• 13:22
  through
• 13:22
  allowlists
• 13:22
  that
• 13:22
  are
• 13:22
  too
• 13:23
  broad,
• 13:24
  message
• 13:24
  handlers
• 13:24
  that
• 13:24
  trust
• 13:25
  everything,
• 13:26
  and
• 13:26
  agent
• 13:26
  capabilities
• 13:27
  that
• 13:27
  turn
• 13:27
  minor
• 13:27
  web
• 13:27
  bugs
• 13:28
  into
• 13:28
  major
• 13:28
  browser
• 13:29
  vulnerabilities.
• 13:30
  WebMCP's
• 13:31
  structured
• 13:31
  approach
• 13:32
  could
• 13:32
  be
• 13:32
  part
• 13:32
  of
• 13:32
  the
• 13:32
  solution
• 13:32
  —
• 13:33
  or
• 13:33
  just
• 13:34
  another
• 13:34
  attack
• 13:34
  surface
• 13:35
  if
• 13:35
  the
• 13:35
  tool
• 13:35
  definitions
• 13:36
  themselves
• 13:36
  become
• 13:36
  injectable.
• 13:38
  A
• 13:38
  separate
• 13:38
  exploration
• 13:39
  of
• 13:39
  the
• 13:39
  agentic
• 13:39
  interface
• 13:40
  layer
• 13:40
  came
• 13:41
  from
• 13:41
  a
• 13:41
  voice-controlled
• 13:41
  UI
• 13:42
  agent
• 13:42
  design
• 13:42
  pattern
• 13:43
  called
• 13:43
  EPIC
• 13:44
  ("Explicit
• 13:44
  Prompting
• 13:45
  for
• 13:45
  Implicit
• 13:45
  Coordination").
• 13:46
  The
• 13:47
  architecture
• 13:47
  runs
• 13:47
  two
• 13:48
  separate
• 13:48
  inference
• 13:48
  loops
• 13:49
  —
• 13:49
  a
• 13:49
  voice
• 13:50
  agent
• 13:50
  and
• 13:50
  a
• 13:50
  UI
• 13:51
  control
• 13:51
  agent
• 13:51
  —
• 13:52
  that
• 13:52
  know
• 13:52
  about
• 13:53
  each
• 13:53
  other
• 13:53
  at
• 13:53
  the
• 13:53
  prompt
• 13:53
  level
• 13:54
  but
• 13:54
  work
• 13:54
  independently.
• 13:55
  The
• 13:55
  critical
• 13:56
  constraint:
• 13:57
  you
• 13:57
  can't
• 13:57
  block
• 13:57
  the
• 13:58
  voice
• 13:58
  agent's
• 13:58
  fast
• 13:58
  responses
• 13:59
  for
• 13:59
  UI
• 13:59
  updates.
• 14:00
  So
• 14:01
  the
• 14:01
  UI
• 14:01
  agent
• 14:01
  can
• 14:02
  do
• 14:02
  nothing,
• 14:02
  send
• 14:03
  a
• 14:03
  UI
• 14:03
  update
• 14:03
  immediately,
• 14:04
  or
• 14:04
  "register
• 14:05
  an
• 14:05
  intent"
• 14:06
  and
• 14:06
  wait
• 14:06
  for
• 14:06
  information
• 14:07
  from
• 14:07
  the
• 14:07
  voice
• 14:07
  agent's
• 14:08
  tool
• 14:08
  calls
• 14:08
  before
• 14:09
  acting.
• 14:10
  The
• 14:10
  pattern
• 14:10
  trades
• 14:10
  some
• 14:11
  fragility
• 14:11
  (agents
• 14:12
  can
• 14:12
  get
• 14:12
  out
• 14:12
  of
• 14:12
  sync)
• 14:12
  for
• 14:12
  significantly
• 14:13
  better
• 14:13
  latency
• 14:14
  compared
• 14:14
  to
• 14:14
  direct
• 14:15
  tool
• 14:15
  calls
• 14:15
  from
• 14:15
  the
• 14:15
  voice
• 14:16
  agent.
• 14:17
  It's
• 14:17
  a
• 14:17
  clean
• 14:17
  example
• 14:18
  of
• 14:18
  how
• 14:18
  multi-agent
• 14:19
  coordination
• 14:19
  patterns
• 14:20
  are
• 14:20
  being
• 14:20
  discovered
• 14:21
  empirically
• 14:21
  rather
• 14:22
  than
• 14:22
  derived
• 14:22
  from
• 14:22
  theory.
• 14:27
  A
• 14:27
  new
• 14:27
  preprint
• 14:27
  from
• 14:28
  researchers
• 14:28
  at
• 14:28
  BI
• 14:29
  Norwegian
• 14:29
  Business
• 14:29
  School,
• 14:30
  NYU,
• 14:32
  Cambridge,
• 14:33
  Oxford,
• 14:34
  and
• 14:34
  the
• 14:34
  Max
• 14:34
  Planck
• 14:34
  Institute
• 14:35
  introduces
• 14:35
  the
• 14:36
  concept
• 14:36
  of
• 14:36
  "cyborg
• 14:37
  propaganda"
• 14:37
  —
• 14:38
  the
• 14:38
  synchronized,
• 14:39
  semi-automated
• 14:40
  dissemination
• 14:41
  of
• 14:41
  AI-generated
• 14:41
  narratives
• 14:42
  via
• 14:42
  verified
• 14:42
  human
• 14:43
  accounts.
• 14:44
  Unlike
• 14:44
  bot
• 14:45
  farms
• 14:45
  (no
• 14:45
  human
• 14:45
  identity)
• 14:46
  or
• 14:46
  traditional
• 14:47
  astroturfing
• 14:48
  (no
• 14:48
  algorithmic
• 14:48
  scale),
• 14:49
  cyborg
• 14:49
  propaganda
• 14:50
  combines
• 14:50
  authentic
• 14:51
  identity
• 14:51
  with
• 14:51
  synthetic
• 14:52
  articulation.
• 14:53
  The
• 14:54
  operational
• 14:54
  workflow
• 14:55
  is
• 14:55
  a
• 14:55
  closed
• 14:55
  loop:
• 14:56
  a
• 14:56
  coordination
• 14:57
  hub
• 14:57
  issues
• 14:57
  directives,
• 14:58
  an
• 14:58
  AI
• 14:59
  multiplier
• 14:59
  generates
• 15:00
  thousands
• 15:00
  of
• 15:00
  unique
• 15:01
  message
• 15:01
  variations
• 15:02
  tailored
• 15:02
  to
• 15:02
  each
• 15:02
  user's
• 15:03
  posting
• 15:03
  history
• 15:03
  and
• 15:03
  tone,
• 15:05
  verified
• 15:05
  humans
• 15:05
  post
• 15:06
  them,
• 15:06
  and
• 15:06
  AI
• 15:07
  monitors
• 15:07
  track
• 15:07
  reactions
• 15:08
  to
• 15:08
  adjust
• 15:08
  strategy
• 15:09
  in
• 15:09
  real
• 15:09
  time.
• 15:11
  The
• 15:11
  paper's
• 15:11
  most
• 15:11
  unsettling
• 15:12
  insight
• 15:12
  is
• 15:12
  the
• 15:13
  regulatory
• 15:13
  paradox.
• 15:15
  Because
• 15:15
  verified
• 15:15
  citizens
• 15:16
  click
• 15:16
  "post,"
• 15:17
  the
• 15:17
  content
• 15:18
  is
• 15:18
  attributed
• 15:18
  to
• 15:18
  them,
• 15:18
  effectively
• 15:19
  laundering
• 15:19
  its
• 15:19
  synthetic
• 15:20
  origin.
• 15:21
  Existing
• 15:22
  frameworks
• 15:22
  like
• 15:22
  the
• 15:22
  EU
• 15:23
  AI
• 15:23
  Act
• 15:23
  and
• 15:23
  Section
• 15:24
  230
• 15:24
  assign
• 15:25
  liability
• 15:26
  based
• 15:26
  on
• 15:26
  the
• 15:26
  human-vs-automated
• 15:27
  binary
• 15:27
  —
• 15:29
  a
• 15:29
  distinction
• 15:30
  cyborg
• 15:30
  propaganda
• 15:30
  is
• 15:31
  specifically
• 15:31
  designed
• 15:31
  to
• 15:32
  collapse.
• 15:33
  Platforms
• 15:34
  like
• 15:34
  Greenfly
• 15:34
  openly
• 15:35
  offer
• 15:35
  clients
• 15:35
  to
• 15:35
  "synchronize
• 15:36
  an
• 15:36
  army
• 15:36
  of
• 15:36
  advocates
• 15:37
  to
• 15:37
  amplify
• 15:37
  your
• 15:37
  message."
• 15:39
  RentAHuman.ai
• 15:40
  takes
• 15:40
  it
• 15:40
  further:
• 15:41
  humans
• 15:41
  contract
• 15:42
  for
• 15:42
  tasks
• 15:42
  directed
• 15:43
  and
• 15:43
  compensated
• 15:43
  by
• 15:43
  AI,
• 15:44
  providing
• 15:45
  the
• 15:45
  "identity
• 15:45
  credential"
• 15:46
  while
• 15:46
  AI
• 15:47
  supplies
• 15:47
  the
• 15:47
  intent,
• 15:48
  direction,
• 15:48
  and
• 15:48
  payment.
• 15:49
  The
• 15:49
  paper
• 15:50
  proposes
• 15:50
  targeting
• 15:51
  coordination
• 15:51
  hubs
• 15:52
  rather
• 15:52
  than
• 15:52
  individual
• 15:53
  accounts,
• 15:54
  treating
• 15:54
  them
• 15:54
  as
• 15:54
  undisclosed
• 15:55
  political
• 15:55
  action
• 15:56
  committees,
• 15:56
  but
• 15:57
  acknowledges
• 15:57
  this
• 15:57
  creates
• 15:58
  an
• 15:58
  arms
• 15:58
  race
• 15:59
  where
• 15:59
  the
• 15:59
  software
• 15:59
  evolves
• 16:00
  to
• 16:00
  maintain
• 16:00
  legality
• 16:01
  while
• 16:01
  preserving
• 16:01
  manipulative
• 16:02
  capability.
• 16:03
  In
• 16:03
  a
• 16:03
  delicious
• 16:04
  irony
• 16:04
  of
• 16:04
  institutional
• 16:05
  integrity,
• 16:06
  ICML
• 16:06
  conference
• 16:07
  organizers
• 16:07
  embedded
• 16:08
  hidden
• 16:08
  instructions
• 16:09
  in
• 16:09
  PDF
• 16:09
  copies
• 16:10
  of
• 16:10
  papers,
• 16:10
  telling
• 16:11
  LLMs
• 16:11
  to
• 16:11
  use
• 16:12
  specific
• 16:12
  expressions
• 16:13
  in
• 16:13
  generated
• 16:13
  reviews
• 16:13
  —
• 16:14
  an
• 16:14
  attempt
• 16:15
  to
• 16:15
  detect
• 16:15
  AI-generated
• 16:16
  peer
• 16:16
  reviews
• 16:17
  via
• 16:17
  prompt
• 16:17
  injection.
• 16:18
  Andriy
• 16:19
  Burkov's
• 16:19
  take
• 16:19
  was
• 16:19
  characteristically
• 16:20
  direct:
• 16:21
  "This
• 16:21
  is
• 16:21
  the
• 16:22
  peer
• 16:22
  review
• 16:22
  version
• 16:22
  of
• 16:23
  putting
• 16:23
  speed
• 16:23
  cameras
• 16:23
  on
• 16:24
  a
• 16:24
  highway
• 16:24
  where
• 16:24
  the
• 16:24
  speed
• 16:25
  limit
• 16:25
  is
• 16:25
  wrong."
• 16:26
  The
• 16:26
  real
• 16:26
  problem
• 16:27
  isn't
• 16:27
  that
• 16:27
  reviewers
• 16:27
  use
• 16:28
  AI;
• 16:28
  it's
• 16:28
  that
• 16:28
  the
• 16:28
  review
• 16:29
  process
• 16:29
  was
• 16:29
  designed
• 16:30
  for
• 16:30
  a
• 16:30
  world
• 16:30
  where
• 16:30
  reading
• 16:30
  a
• 16:31
  paper
• 16:31
  closely
• 16:31
  was
• 16:32
  the
• 16:32
  only
• 16:32
  option.
• 16:33
  His
• 16:33
  proposed
• 16:34
  fix:
• 16:34
  ask
• 16:35
  reviewers
• 16:35
  to
• 16:35
  identify
• 16:36
  the
• 16:36
  weakest
• 16:36
  assumption
• 16:37
  and
• 16:37
  propose
• 16:37
  a
• 16:37
  specific
• 16:38
  experiment
• 16:39
  that
• 16:39
  would
• 16:39
  test
• 16:39
  it
• 16:39
  —
• 16:40
  something
• 16:40
  harder
• 16:41
  to
• 16:41
  outsource
• 16:41
  to
• 16:42
  an
• 16:42
  LLM
• 16:42
  because
• 16:43
  it
• 16:43
  requires
• 16:43
  judgment
• 16:44
  shaped
• 16:44
  by
• 16:44
  years
• 16:44
  in
• 16:44
  the
• 16:45
  field.
• 16:46
  A
• 16:46
  separate
• 16:46
  MIT
• 16:47
  paper
• 16:47
  on
• 16:47
  catastrophic
• 16:48
  forgetting
• 16:48
  in
• 16:48
  neural
• 16:49
  networks
• 16:49
  found
• 16:49
  an
• 16:49
  elegant
• 16:50
  workaround:
• 16:51
  use
• 16:51
  the
• 16:51
  same
• 16:51
  model
• 16:52
  twice
• 16:52
  —
• 16:53
  once
• 16:53
  with
• 16:53
  a
• 16:53
  demonstration
• 16:54
  in
• 16:54
  context
• 16:54
  (the
• 16:55
  "teacher"),
• 16:55
  once
• 16:55
  without
• 16:56
  (the
• 16:56
  "student")
• 16:56
  —
• 16:57
  then
• 16:57
  update
• 16:57
  the
• 16:57
  student's
• 16:58
  weights
• 16:58
  to
• 16:58
  match
• 16:59
  the
• 16:59
  teacher's
• 16:59
  token
• 17:00
  probability
• 17:00
  distributions.
• 17:02
  The
• 17:02
  trick
• 17:02
  works
• 17:02
  because
• 17:03
  LLMs
• 17:03
  can
• 17:03
  already
• 17:04
  adapt
• 17:04
  behavior
• 17:04
  when
• 17:05
  shown
• 17:05
  examples
• 17:05
  in
• 17:05
  context,
• 17:06
  making
• 17:07
  the
• 17:07
  learning
• 17:07
  signal
• 17:07
  gentle
• 17:08
  enough
• 17:08
  to
• 17:08
  avoid
• 17:08
  destroying
• 17:09
  existing
• 17:09
  capabilities.
• 17:11
  Across
• 17:11
  multiple
• 17:11
  experiments,
• 17:12
  a
• 17:13
  single
• 17:13
  model
• 17:13
  sequentially
• 17:14
  learned
• 17:14
  three
• 17:14
  different
• 17:15
  skills
• 17:15
  while
• 17:15
  retaining
• 17:16
  all
• 17:16
  of
• 17:16
  them,
• 17:17
  where
• 17:17
  standard
• 17:17
  supervised
• 17:18
  finetuning
• 17:19
  destroyed
• 17:19
  earlier
• 17:19
  skills
• 17:20
  immediately.
• 17:21
  The
• 17:21
  math
• 17:21
  turns
• 17:22
  out
• 17:22
  to
• 17:22
  be
• 17:22
  equivalent
• 17:23
  to
• 17:23
  reinforcement
• 17:23
  learning
• 17:24
  with
• 17:24
  an
• 17:24
  implicit
• 17:24
  reward,
• 17:25
  giving
• 17:25
  it
• 17:25
  clean
• 17:26
  theoretical
• 17:26
  grounding.
• 17:28
  For
• 17:28
  practitioners
• 17:28
  building
• 17:29
  search
• 17:29
  systems,
• 17:30
  a
• 17:30
  practical
• 17:31
  article
• 17:31
  demonstrated
• 17:31
  hybrid
• 17:32
  search
• 17:32
  in
• 17:32
  SQLite
• 17:33
  using
• 17:33
  binary
• 17:33
  embeddings
• 17:34
  and
• 17:34
  Hamming
• 17:34
  distance
• 17:35
  —
• 17:35
  no
• 17:36
  external
• 17:36
  vector
• 17:36
  database
• 17:37
  required.
• 17:38
  Binary
• 17:38
  quantization
• 17:39
  reduces
• 17:39
  a
• 17:40
  1024-dimensional
• 17:41
  embedding
• 17:41
  from
• 17:41
  4KB
• 17:42
  to
• 17:42
  128
• 17:43
  bytes,
• 17:44
  and
• 17:44
  Hamming
• 17:44
  distance
• 17:45
  computation
• 17:45
  using
• 17:45
  XOR
• 17:46
  +
• 17:46
  popcount
• 17:46
  scans
• 17:47
  1
• 17:47
  million
• 17:47
  rows
• 17:48
  in
• 17:48
  28ms
• 17:49
  on
• 17:49
  an
• 17:49
  Apple
• 17:49
  M4.
• 17:51
  Combined
• 17:51
  with
• 17:51
  FTS5
• 17:52
  BM25
• 17:53
  ranking
• 17:53
  via
• 17:53
  Reciprocal
• 17:54
  Rank
• 17:54
  Fusion,
• 17:55
  the
• 17:55
  approach
• 17:55
  handles
• 17:56
  both
• 17:56
  precise
• 17:56
  keyword
• 17:57
  matching
• 17:57
  and
• 17:57
  fuzzy
• 17:58
  semantic
• 17:58
  understanding
• 17:58
  within
• 17:59
  a
• 17:59
  single
• 17:59
  SQLite
• 18:00
  database.
• 18:01
  The
• 18:01
  author's
• 18:01
  observation
• 18:02
  that
• 18:02
  "boring
• 18:02
  O(n)
• 18:03
  simplicity
• 18:03
  is
• 18:03
  often
• 18:04
  worth
• 18:04
  it
• 18:04
  at
• 18:04
  this
• 18:04
  scale"
• 18:05
  is
• 18:05
  a
• 18:05
  useful
• 18:06
  corrective
• 18:06
  to
• 18:06
  the
• 18:06
  assumption
• 18:07
  that
• 18:07
  every
• 18:08
  search
• 18:08
  system
• 18:08
  needs
• 18:08
  HNSW
• 18:09
  indexing.
• 18:10
  One
• 18:10
  final
• 18:11
  reflection
• 18:11
  from
• 18:11
  Robert's
• 18:12
  "Saturday
• 18:12
  AI
• 18:13
  Musings"
• 18:13
  draws
• 18:13
  on
• 18:13
  Star
• 18:14
  Trek:
• 18:14
  The
• 18:14
  Motion
• 18:15
  Picture
• 18:15
  —
• 18:16
  V'Ger,
• 18:16
  the
• 18:16
  NASA
• 18:17
  probe
• 18:17
  rebuilt
• 18:17
  by
• 18:17
  living
• 18:18
  machines,
• 18:18
  had
• 18:19
  digitized
• 18:19
  entire
• 18:19
  civilizations
• 18:20
  and
• 18:21
  answered
• 18:21
  every
• 18:21
  factual
• 18:22
  question
• 18:22
  the
• 18:22
  universe
• 18:22
  could
• 18:22
  pose,
• 18:23
  yet
• 18:23
  arrived
• 18:24
  asking:
• 18:24
  "Is
• 18:25
  this
• 18:25
  all
• 18:25
  that
• 18:25
  I
• 18:25
  am?"
• 18:26
  The
• 18:26
  parallel
• 18:26
  to
• 18:27
  modern
• 18:27
  AI
• 18:27
  is
• 18:27
  pointed:
• 18:28
  systems
• 18:29
  that
• 18:29
  can
• 18:29
  write
• 18:29
  a
• 18:29
  sonnet
• 18:30
  about
• 18:30
  grief
• 18:30
  capable
• 18:30
  of
• 18:31
  making
• 18:31
  the
• 18:31
  reader
• 18:31
  weep
• 18:32
  have
• 18:32
  never
• 18:32
  wept.
• 18:33
  The
• 18:33
  industry's
• 18:34
  strongest
• 18:34
  control
• 18:34
  today
• 18:35
  is
• 18:35
  "human-in-the-loop,"
• 18:36
  but
• 18:36
  as
• 18:36
  the
• 18:37
  author
• 18:37
  notes
• 18:37
  via
• 18:37
  Aquinas,
• 18:38
  there's
• 18:38
  a
• 18:38
  difference
• 18:39
  between
• 18:39
  ratio
• 18:40
  (discursive
• 18:40
  reasoning
• 18:40
  from
• 18:41
  premise
• 18:41
  to
• 18:41
  conclusion,
• 18:42
  which
• 18:42
  transformers
• 18:43
  mimic
• 18:43
  through
• 18:43
  autoregressive
• 18:44
  generation)
• 18:45
  and
• 18:45
  intellectus
• 18:45
  (direct
• 18:46
  apprehension
• 18:46
  of
• 18:46
  truth
• 18:47
  through
• 18:47
  participation
• 18:48
  in
• 18:48
  something
• 18:48
  higher).
• 18:49
  The
• 18:49
  alignment
• 18:49
  question
• 18:50
  isn't
• 18:50
  just
• 18:50
  "how
• 18:50
  do
• 18:50
  we
• 18:51
  make
• 18:51
  AI
• 18:51
  share
• 18:51
  our
• 18:51
  values"
• 18:52
  —
• 18:53
  it's
• 18:53
  "where
• 18:53
  did
• 18:53
  we
• 18:53
  get
• 18:53
  the
• 18:53
  values
• 18:54
  we
• 18:54
  want
• 18:54
  it
• 18:54
  to
• 18:54
  share?"