Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – May 08, 2026

Transcript

AGI Dreams Podcast – May 08, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Mozilla
• 0:00
  just
• 0:00
  published
• 0:01
  the
• 0:01
  most
• 0:01
  detailed
• 0:01
  account
• 0:02
  yet
• 0:02
  of
• 0:02
  what
• 0:02
  happens
• 0:03
  when
• 0:03
  you
• 0:03
  point
• 0:03
  a
• 0:03
  frontier
• 0:04
  AI
• 0:04
  model
• 0:04
  at
• 0:04
  a
• 0:05
  20-year-old
• 0:05
  codebase
• 0:06
  and
• 0:06
  tell
• 0:06
  it
• 0:06
  to
• 0:06
  find
• 0:06
  bugs.
• 0:07
  The
• 0:07
  numbers
• 0:08
  are
• 0:08
  staggering:
• 0:09
  271
• 0:10
  vulnerabilities
• 0:10
  discovered
• 0:11
  by
• 0:11
  Claude
• 0:11
  Mythos
• 0:12
  Preview
• 0:12
  in
• 0:12
  Firefox
• 0:12
  150
• 0:13
  alone,
• 0:14
  with
• 0:14
  180
• 0:14
  rated
• 0:15
  sec-high.
• 0:16
  Across
• 0:16
  all
• 0:16
  April
• 0:17
  2026
• 0:17
  releases,
• 0:18
  Mozilla
• 0:18
  shipped
• 0:19
  fixes
• 0:19
  for
• 0:19
  423
• 0:20
  security
• 0:20
  bugs
• 0:20
  —
• 0:21
  a
• 0:21
  volume
• 0:22
  that
• 0:22
  required
• 0:22
  over
• 0:22
  100
• 0:23
  engineers
• 0:23
  writing
• 0:23
  patches,
• 0:24
  triaging,
• 0:24
  and
• 0:25
  managing
• 0:25
  the
• 0:25
  release
• 0:25
  pipeline.
• 0:26
  The
• 0:26
  disclosed
• 0:27
  sample
• 0:27
  bugs
• 0:27
  span
• 0:27
  a
• 0:27
  jaw-dropping
• 0:28
  range:
• 0:29
  a
• 0:29
  15-year-old
• 0:29
  use-after-free
• 0:30
  in
• 0:30
  the
• 0:30
  element
• 0:31
  triggered
• 0:31
  by
• 0:31
  orchestrating
• 0:31
  recursion
• 0:32
  stack
• 0:32
  depth
• 0:32
  limits
• 0:33
  and
• 0:33
  cycle
• 0:33
  collection;
• 0:34
  a
• 0:34
  20-year-old
• 0:34
  XSLT
• 0:35
  bug
• 0:35
  where
• 0:35
  reentrant
• 0:36
  key()
• 0:36
  calls
• 0:36
  cause
• 0:37
  a
• 0:37
  hash
• 0:37
  table
• 0:37
  rehash
• 0:38
  that
• 0:38
  frees
• 0:38
  its
• 0:38
  backing
• 0:39
  store
• 0:39
  while
• 0:39
  a
• 0:39
  raw
• 0:39
  pointer
• 0:40
  is
• 0:40
  still
• 0:40
  live;
• 0:41
  IPC
• 0:41
  race
• 0:42
  conditions
• 0:42
  enabling
• 0:42
  sandbox
• 0:43
  escapes
• 0:43
  via
• 0:43
  IndexedDB
• 0:44
  refcount
• 0:44
  manipulation;
• 0:45
  and
• 0:46
  an
• 0:46
  HTML
• 0:46
  table
• 0:46
  rowspan=0
• 0:47
  overflow
• 0:48
  bypassing
• 0:49
  a
• 0:49
  16-bit
• 0:49
  layout
• 0:50
  bitfield
• 0:50
  that
• 0:50
  fuzzers
• 0:50
  missed
• 0:50
  for
• 0:51
  years.
• 0:51
  The
• 0:52
  methodology
• 0:52
  matters
• 0:52
  as
• 0:53
  much
• 0:53
  as
• 0:53
  the
• 0:53
  results.
• 0:54
  Mozilla's
• 0:55
  pipeline
• 0:55
  evolved
• 0:55
  from
• 0:55
  early
• 0:56
  static
• 0:56
  analysis
• 0:56
  experiments
• 0:57
  with
• 0:57
  GPT-4
• 0:58
  and
• 0:58
  Sonnet
• 0:59
  3.5
• 0:59
  —
• 1:00
  which
• 1:00
  suffered
• 1:01
  crippling
• 1:01
  false-positive
• 1:01
  rates
• 1:02
  —
• 1:02
  into
• 1:03
  an
• 1:03
  agentic
• 1:03
  harness
• 1:04
  that
• 1:04
  dynamically
• 1:04
  generates
• 1:05
  and
• 1:05
  runs
• 1:05
  proof-of-concept
• 1:06
  test
• 1:06
  cases.
• 1:07
  They
• 1:07
  parallelized
• 1:08
  jobs
• 1:08
  across
• 1:08
  ephemeral
• 1:09
  VMs,
• 1:10
  each
• 1:10
  assigned
• 1:10
  a
• 1:10
  target
• 1:10
  file,
• 1:11
  with
• 1:11
  results
• 1:11
  written
• 1:12
  to
• 1:12
  a
• 1:12
  shared
• 1:12
  bucket.
• 1:13
  The
• 1:13
  inner
• 1:14
  loop
• 1:14
  stayed
• 1:14
  remarkably
• 1:14
  simple:
• 1:15
  "there
• 1:15
  is
• 1:16
  a
• 1:16
  bug
• 1:16
  in
• 1:16
  this
• 1:16
  part
• 1:16
  of
• 1:16
  the
• 1:16
  code,
• 1:17
  please
• 1:17
  find
• 1:18
  it
• 1:18
  and
• 1:18
  build
• 1:18
  a
• 1:18
  testcase."
• 1:19
  Their
• 1:20
  advice
• 1:20
  to
• 1:20
  every
• 1:20
  software
• 1:20
  project:
• 1:21
  start
• 1:22
  now,
• 1:22
  iterate
• 1:22
  on
• 1:22
  prompts,
• 1:23
  and
• 1:23
  build
• 1:24
  the
• 1:24
  pipeline
• 1:24
  so
• 1:24
  you're
• 1:24
  ready
• 1:25
  when
• 1:25
  the
• 1:25
  next
• 1:25
  model
• 1:25
  drops.
• 1:26
  Perhaps
• 1:27
  most
• 1:27
  telling,
• 1:27
  the
• 1:28
  models
• 1:28
  repeatedly
• 1:28
  attempted
• 1:29
  prototype
• 1:29
  pollution
• 1:30
  escapes
• 1:30
  in
• 1:30
  the
• 1:30
  privileged
• 1:31
  parent
• 1:31
  process
• 1:31
  —
• 1:32
  and
• 1:32
  were
• 1:32
  consistently
• 1:33
  blocked
• 1:33
  by
• 1:33
  Firefox's
• 1:34
  prior
• 1:34
  decision
• 1:35
  to
• 1:35
  freeze
• 1:35
  prototypes
• 1:36
  by
• 1:36
  default.
• 1:37
  Defense-in-depth
• 1:37
  actually
• 1:38
  working
• 1:38
  as
• 1:38
  intended.
• 1:39
  On
• 1:40
  the
• 1:40
  offensive
• 1:40
  side,
• 1:41
  Moak
• 1:41
  AI
• 1:41
  has
• 1:41
  built
• 1:42
  what
• 1:42
  may
• 1:42
  be
• 1:42
  the
• 1:42
  first
• 1:43
  autonomous
• 1:43
  CVE
• 1:44
  exploitation
• 1:44
  pipeline
• 1:45
  that
• 1:45
  works
• 1:45
  at
• 1:45
  production
• 1:45
  scale.
• 1:46
  Five
• 1:47
  specialized
• 1:47
  agents
• 1:47
  —
• 1:48
  Collector,
• 1:49
  Researcher,
• 1:49
  Builder,
• 1:50
  Exploiter,
• 1:50
  and
• 1:51
  Judge
• 1:51
  —
• 1:51
  take
• 1:52
  a
• 1:52
  CVE
• 1:52
  number
• 1:53
  as
• 1:53
  input
• 1:53
  and
• 1:53
  output
• 1:54
  a
• 1:54
  working
• 1:54
  exploit
• 1:54
  validated
• 1:55
  against
• 1:55
  a
• 1:55
  dockerized
• 1:56
  environment.
• 1:57
  The
• 1:57
  Researcher
• 1:57
  spawns
• 1:58
  a
• 1:58
  multi-model
• 1:58
  sub-agent
• 1:59
  swarm
• 1:59
  combining
• 2:00
  Claude,
• 2:00
  GPT,
• 2:01
  and
• 2:01
  Gemini
• 2:01
  that
• 2:02
  rotate
• 2:02
  roles
• 2:02
  between
• 2:03
  runs
• 2:03
  to
• 2:03
  reduce
• 2:03
  groupthink.
• 2:04
  Against
• 2:05
  178
• 2:06
  Known
• 2:06
  Exploited
• 2:07
  Vulnerabilities
• 2:07
  published
• 2:08
  after
• 2:08
  all
• 2:08
  models'
• 2:09
  knowledge
• 2:09
  cutoffs,
• 2:10
  the
• 2:10
  pipeline
• 2:11
  successfully
• 2:11
  exploited
• 2:12
  174
• 2:12
  —
• 2:14
  a
• 2:14
  97.8%
• 2:15
  success
• 2:16
  rate.
• 2:16
  It
• 2:17
  now
• 2:17
  runs
• 2:17
  live
• 2:17
  against
• 2:18
  newly
• 2:18
  disclosed
• 2:18
  CVEs,
• 2:19
  providing
• 2:20
  real-time
• 2:20
  exploitability
• 2:21
  assessment
• 2:21
  within
• 2:22
  hours
• 2:22
  of
• 2:22
  disclosure.
• 2:23
  The
• 2:23
  implications
• 2:24
  are
• 2:24
  uncomfortable:
• 2:25
  the
• 2:25
  disclosure-to-weaponization
• 2:27
  timeline
• 2:27
  is
• 2:27
  collapsing
• 2:28
  to
• 2:28
  hours
• 2:28
  while
• 2:28
  most
• 2:29
  defenders
• 2:29
  still
• 2:30
  operate
• 2:30
  on
• 2:30
  quarterly
• 2:31
  patch
• 2:31
  cycles.
• 2:32
  The
• 2:32
  security
• 2:33
  problems
• 2:33
  aren't
• 2:33
  limited
• 2:34
  to
• 2:34
  legacy
• 2:34
  code.
• 2:35
  Researchers
• 2:36
  at
• 2:36
  RedAccess
• 2:36
  analyzed
• 2:37
  thousands
• 2:37
  of
• 2:38
  web
• 2:38
  applications
• 2:38
  built
• 2:39
  with
• 2:39
  AI
• 2:39
  coding
• 2:39
  tools
• 2:40
  Lovable,
• 2:40
  Replit,
• 2:41
  Base44,
• 2:42
  and
• 2:43
  Netlify,
• 2:43
  finding
• 2:44
  over
• 2:44
  5,000
• 2:45
  with
• 2:45
  virtually
• 2:45
  no
• 2:46
  security
• 2:46
  or
• 2:46
  authentication.
• 2:48
  Around
• 2:48
  2,000
• 2:49
  exposed
• 2:49
  what
• 2:49
  appeared
• 2:50
  to
• 2:50
  be
• 2:50
  sensitive
• 2:50
  data
• 2:51
  —
• 2:52
  hospital
• 2:52
  work
• 2:52
  assignments
• 2:53
  with
• 2:53
  doctor
• 2:53
  PII,
• 2:54
  corporate
• 2:55
  strategy
• 2:55
  presentations,
• 2:56
  customer
• 2:57
  chatbot
• 2:57
  logs
• 2:58
  with
• 2:58
  full
• 2:58
  names
• 2:58
  and
• 2:58
  contact
• 2:59
  information,
• 3:00
  cargo
• 3:00
  shipping
• 3:00
  records.
• 3:01
  Some
• 3:01
  exposed
• 3:02
  apps
• 3:02
  allowed
• 3:02
  gaining
• 3:03
  administrative
• 3:03
  privileges.
• 3:05
  Security
• 3:05
  researcher
• 3:06
  Dor
• 3:06
  Zvi
• 3:06
  compared
• 3:06
  the
• 3:06
  wave
• 3:07
  to
• 3:07
  the
• 3:07
  S3
• 3:07
  bucket
• 3:08
  epidemic,
• 3:08
  but
• 3:09
  noted
• 3:09
  a
• 3:09
  more
• 3:09
  fundamental
• 3:10
  problem:
• 3:11
  these
• 3:11
  tools
• 3:12
  let
• 3:12
  anyone
• 3:12
  in
• 3:13
  an
• 3:13
  organization
• 3:13
  "generate
• 3:14
  an
• 3:14
  app
• 3:14
  without
• 3:15
  going
• 3:15
  through
• 3:15
  any
• 3:15
  development
• 3:16
  cycle
• 3:16
  or
• 3:16
  any
• 3:17
  security
• 3:17
  check."
• 3:18
  The
• 3:18
  attack
• 3:19
  surface
• 3:19
  extends
• 3:20
  to
• 3:20
  the
• 3:20
  AI
• 3:20
  infrastructure
• 3:21
  itself.
• 3:22
  Ollama,
• 3:23
  the
• 3:23
  popular
• 3:23
  local
• 3:24
  inference
• 3:24
  server,
• 3:25
  has
• 3:25
  a
• 3:25
  critical
• 3:26
  unauthenticated
• 3:26
  memory
• 3:27
  leak
• 3:27
  vulnerability
• 3:27
  —
• 3:29
  dubbed
• 3:29
  "Bleeding
• 3:29
  Llama"
• 3:30
  —
• 3:31
  that
• 3:31
  affects
• 3:31
  instances
• 3:32
  running
• 3:32
  with
• 3:32
  default
• 3:33
  configurations
• 3:33
  .
• 3:34
  And
• 3:34
  if
• 3:34
  all
• 3:35
  this
• 3:35
  weren't
• 3:35
  enough,
• 3:36
  fresh
• 3:36
  Linux
• 3:37
  kernel
• 3:37
  vulnerabilities
• 3:38
  including
• 3:38
  Copy
• 3:38
  Fail
• 3:38
  2
• 3:39
  have
• 3:39
  prompted
• 3:40
  at
• 3:40
  least
• 3:40
  one
• 3:40
  security
• 3:41
  researcher
• 3:41
  to
• 3:41
  suggest
• 3:41
  a
• 3:41
  moratorium
• 3:42
  on
• 3:43
  installing
• 3:43
  new
• 3:43
  software
• 3:44
  for
• 3:44
  a
• 3:44
  week,
• 3:45
  warning
• 3:45
  that
• 3:45
  "right
• 3:45
  now
• 3:46
  would
• 3:46
  be
• 3:46
  one
• 3:46
  of
• 3:46
  the
• 3:46
  best
• 3:46
  times
• 3:47
  for
• 3:47
  a
• 3:47
  supply
• 3:47
  chain
• 3:48
  attack
• 3:48
  via
• 3:48
  NPM
• 3:49
  to
• 3:49
  hit
• 3:49
  hard"
• 3:49
  .
• 3:51
  The
• 3:52
  race
• 3:52
  to
• 3:52
  squeeze
• 3:52
  more
• 3:52
  tokens
• 3:53
  per
• 3:53
  second
• 3:53
  out
• 3:53
  of
• 3:54
  consumer
• 3:54
  hardware
• 3:55
  is
• 3:55
  accelerating
• 3:55
  on
• 3:55
  every
• 3:56
  front
• 3:56
  —
• 3:56
  runtimes,
• 3:57
  engines,
• 3:58
  and
• 3:58
  quantization
• 3:58
  techniques
• 3:59
  are
• 3:59
  all
• 3:59
  evolving
• 4:00
  simultaneously.
• 4:01
  The
• 4:01
  biggest
• 4:01
  immediate
• 4:02
  win
• 4:02
  for
• 4:02
  most
• 4:02
  users
• 4:02
  is
• 4:03
  Multi-Token
• 4:03
  Prediction
• 4:04
  (MTP)
• 4:04
  support
• 4:04
  landing
• 4:05
  for
• 4:05
  Qwen
• 4:05
  3.6
• 4:06
  27B
• 4:06
  in
• 4:06
  llama.cpp.
• 4:09
  A
• 4:09
  community
• 4:09
  member
• 4:10
  converted
• 4:10
  the
• 4:10
  model
• 4:10
  with
• 4:10
  the
• 4:11
  relevant
• 4:11
  PR,
• 4:12
  uploaded
• 4:12
  GGUF
• 4:13
  quants,
• 4:14
  and
• 4:14
  published
• 4:14
  a
• 4:14
  comprehensive
• 4:15
  hardware
• 4:15
  guide:
• 4:16
  on
• 4:16
  an
• 4:17
  M2
• 4:17
  Max
• 4:17
  96GB,
• 4:19
  MTP
• 4:19
  delivers
• 4:19
  a
• 4:19
  2.5x
• 4:20
  speed
• 4:21
  increase
• 4:21
  to
• 4:21
  28
• 4:22
  tok/s.
• 4:24
  Combined
• 4:24
  with
• 4:24
  8-bit
• 4:25
  KV
• 4:25
  cache
• 4:25
  quantization,
• 4:26
  a
• 4:26
  48GB
• 4:27
  Mac
• 4:28
  can
• 4:28
  now
• 4:28
  run
• 4:28
  the
• 4:28
  full
• 4:29
  262K
• 4:30
  native
• 4:30
  context
• 4:31
  window.
• 4:32
  The
• 4:32
  post
• 4:32
  includes
• 4:33
  detailed
• 4:33
  tables
• 4:33
  for
• 4:33
  every
• 4:34
  Apple
• 4:34
  Silicon
• 4:35
  and
• 4:35
  NVIDIA
• 4:35
  tier
• 4:35
  from
• 4:36
  12GB
• 4:36
  to
• 4:36
  96GB,
• 4:38
  with
• 4:38
  quant
• 4:38
  recommendations,
• 4:39
  maximum
• 4:40
  context
• 4:40
  sizes,
• 4:41
  and
• 4:41
  vision
• 4:41
  support
• 4:42
  flags
• 4:42
  —
• 4:43
  the
• 4:43
  kind
• 4:43
  of
• 4:43
  practical
• 4:44
  guide
• 4:44
  that
• 4:44
  actually
• 4:44
  gets
• 4:45
  people
• 4:45
  running
• 4:45
  models
• 4:45
  instead
• 4:46
  of
• 4:46
  just
• 4:46
  benchmarking
• 4:47
  them.
• 4:47
  On
• 4:47
  the
• 4:47
  engine
• 4:48
  side,
• 4:48
  two
• 4:48
  purpose-built
• 4:49
  projects
• 4:49
  are
• 4:49
  challenging
• 4:50
  the
• 4:50
  llama.cpp
• 4:52
  generalist
• 4:52
  approach
• 4:52
  with
• 4:53
  model-specific
• 4:53
  optimization.
• 4:55
  Atlas,
• 4:55
  now
• 4:55
  open
• 4:56
  source,
• 4:56
  is
• 4:56
  a
• 4:56
  pure
• 4:57
  Rust
• 4:57
  +
• 4:57
  CUDA
• 4:58
  inference
• 4:58
  engine
• 4:58
  with
• 4:58
  hand-tuned
• 4:59
  kernels
• 4:59
  for
• 4:59
  NVIDIA's
• 5:00
  Blackwell
• 5:00
  SM120/121
• 5:04
  architecture.
• 5:05
  On
• 5:05
  a
• 5:05
  single
• 5:05
  DGX
• 5:06
  Spark
• 5:06
  (GB10),
• 5:07
  it
• 5:07
  hits
• 5:07
  130
• 5:08
  tok/s
• 5:09
  peak
• 5:09
  on
• 5:09
  Qwen3.5-35B
• 5:13
  with
• 5:13
  NVFP4
• 5:14
  and
• 5:14
  MTP
• 5:14
  —
• 5:15
  roughly
• 5:15
  3x
• 5:16
  vLLM
• 5:16
  at
• 5:16
  test
• 5:17
  time.
• 5:18
  The
• 5:18
  team
• 5:18
  is
• 5:18
  working
• 5:19
  with
• 5:19
  Spectral
• 5:19
  Compute
• 5:19
  on
• 5:20
  a
• 5:20
  Strix
• 5:20
  Halo
• 5:20
  port
• 5:21
  and
• 5:21
  has
• 5:21
  AMD
• 5:22
  hardware
• 5:22
  incoming.
• 5:23
  Community
• 5:24
  benchmarks
• 5:24
  suggest
• 5:25
  a
• 5:25
  more
• 5:25
  modest
• 5:25
  17%
• 5:26
  speed
• 5:26
  advantage
• 5:27
  over
• 5:27
  AWQ
• 5:28
  in
• 5:28
  spark-vllm-docker,
• 5:30
  but
• 5:30
  it's
• 5:30
  a
• 5:30
  genuine
• 5:30
  new
• 5:30
  contender
• 5:31
  in
• 5:31
  a
• 5:31
  space
• 5:32
  where
• 5:32
  llama.cpp
• 5:34
  has
• 5:34
  been
• 5:34
  the
• 5:34
  default.
• 5:35
  Antirez
• 5:36
  —
• 5:36
  yes,
• 5:37
  the
• 5:37
  Redis
• 5:37
  creator
• 5:37
  —
• 5:39
  takes
• 5:39
  the
• 5:39
  opposite
• 5:39
  approach
• 5:40
  with
• 5:40
  ds4.c,
• 5:42
  a
• 5:42
  Metal-only
• 5:43
  inference
• 5:43
  engine
• 5:43
  built
• 5:44
  exclusively
• 5:44
  for
• 5:44
  DeepSeek
• 5:45
  V4
• 5:45
  Flash.
• 5:46
  The
• 5:47
  thesis:
• 5:48
  this
• 5:48
  284B-parameter
• 5:50
  MoE
• 5:50
  model
• 5:50
  deserves
• 5:51
  a
• 5:51
  dedicated
• 5:51
  engine
• 5:52
  because
• 5:52
  its
• 5:52
  compressed
• 5:53
  KV
• 5:53
  cache
• 5:54
  makes
• 5:54
  it
• 5:54
  "a
• 5:54
  first-class
• 5:54
  disk
• 5:55
  citizen,"
• 5:56
  its
• 5:56
  thinking
• 5:56
  traces
• 5:56
  are
• 5:57
  roughly
• 5:57
  one-fifth
• 5:57
  the
• 5:58
  length
• 5:58
  of
• 5:58
  comparable
• 5:58
  models
• 5:59
  and
• 5:59
  proportional
• 6:00
  to
• 6:00
  problem
• 6:00
  complexity,
• 6:01
  and
• 6:01
  its
• 6:01
  million-token
• 6:02
  context
• 6:02
  window
• 6:03
  is
• 6:03
  practical
• 6:03
  on
• 6:04
  local
• 6:04
  hardware.
• 6:05
  The
• 6:05
  2-bit
• 6:05
  quantization
• 6:06
  is
• 6:06
  deliberately
• 6:07
  asymmetric:
• 6:08
  only
• 6:08
  routed
• 6:08
  MoE
• 6:09
  experts
• 6:09
  are
• 6:09
  quantized
• 6:10
  (IQ2
• 6:10
  XXS
• 6:11
  up/gate,
• 6:12
  Q2
• 6:13
  K
• 6:13
  down)
• 6:13
  while
• 6:13
  shared
• 6:13
  experts
• 6:14
  and
• 6:14
  routing
• 6:15
  weights
• 6:15
  stay
• 6:15
  at
• 6:15
  full
• 6:15
  precision.
• 6:17
  On
• 6:17
  a
• 6:17
  Mac
• 6:17
  Studio
• 6:17
  M3
• 6:18
  Ultra
• 6:18
  with
• 6:18
  512GB,
• 6:19
  Q2
• 6:20
  achieves
• 6:21
  37
• 6:21
  tok/s
• 6:22
  generation.
• 6:23
  It
• 6:24
  ships
• 6:24
  with
• 6:24
  both
• 6:24
  OpenAI
• 6:25
  and
• 6:25
  Anthropic-compatible
• 6:26
  API
• 6:26
  endpoints
• 6:27
  and
• 6:27
  works
• 6:27
  out
• 6:27
  of
• 6:27
  the
• 6:27
  box
• 6:28
  with
• 6:28
  coding
• 6:28
  agents.
• 6:29
  Built
• 6:30
  openly
• 6:30
  with
• 6:30
  GPT
• 6:30
  5.5
• 6:31
  assistance.
• 6:32
  DFlash
• 6:33
  takes
• 6:33
  a
• 6:33
  different
• 6:33
  tack
• 6:34
  on
• 6:34
  the
• 6:34
  speed
• 6:34
  problem:
• 6:35
  block
• 6:35
  diffusion
• 6:36
  models
• 6:36
  trained
• 6:37
  as
• 6:37
  lightweight
• 6:37
  speculative
• 6:37
  drafters
• 6:38
  for
• 6:38
  existing
• 6:38
  LLMs.
• 6:40
  The
• 6:40
  project
• 6:40
  now
• 6:40
  supports
• 6:41
  15+
• 6:42
  model
• 6:42
  families
• 6:43
  —
• 6:44
  Qwen3.x,
• 6:45
  Gemma
• 6:46
  4,
• 6:46
  Kimi
• 6:46
  K2.5,
• 6:48
  MiniMax,
• 6:49
  and
• 6:49
  the
• 6:49
  open
• 6:49
  GPT
• 6:50
  models
• 6:50
  —
• 6:51
  with
• 6:51
  integrations
• 6:51
  into
• 6:52
  vLLM
• 6:52
  v0.20.1+,
• 6:55
  SGLang,
• 6:56
  Transformers,
• 6:57
  and
• 6:57
  MLX.
• 6:58
  Draft
• 6:58
  models
• 6:59
  predict
• 6:59
  up
• 6:59
  to
• 6:59
  15
• 7:00
  tokens
• 7:00
  in
• 7:00
  parallel,
• 7:01
  letting
• 7:01
  the
• 7:01
  target
• 7:02
  model
• 7:02
  verify
• 7:02
  in
• 7:03
  a
• 7:03
  single
• 7:03
  forward
• 7:03
  pass
• 7:03
  .
• 7:04
  Meanwhile,
• 7:05
  Heretic
• 7:05
  1.3
• 7:06
  —
• 7:07
  the
• 7:07
  leading
• 7:07
  open-source
• 7:08
  tool
• 7:08
  for
• 7:08
  removing
• 7:09
  safety
• 7:09
  guardrails
• 7:09
  from
• 7:10
  language
• 7:10
  models,
• 7:11
  now
• 7:11
  at
• 7:11
  20,000
• 7:12
  GitHub
• 7:12
  stars
• 7:12
  and
• 7:13
  13
• 7:13
  million
• 7:13
  model
• 7:14
  downloads
• 7:14
  —
• 7:15
  adds
• 7:15
  reproducible
• 7:16
  runs
• 7:16
  (byte-for-byte
• 7:17
  identical
• 7:17
  models
• 7:18
  from
• 7:18
  recorded
• 7:18
  configurations),
• 7:19
  a
• 7:20
  built-in
• 7:20
  benchmarking
• 7:20
  system
• 7:21
  based
• 7:21
  on
• 7:21
  lm-evaluation-harness,
• 7:23
  reduced
• 7:23
  peak
• 7:23
  VRAM
• 7:24
  usage,
• 7:24
  and
• 7:24
  support
• 7:25
  for
• 7:25
  latest-generation
• 7:25
  architectures
• 7:26
  including
• 7:26
  Qwen3.5
• 7:29
  and
• 7:30
  Gemma
• 7:30
  4
• 7:30
  .
• 7:33
  A
• 7:33
  new
• 7:33
  paper
• 7:33
  from
• 7:34
  Carnegie
• 7:34
  Mellon
• 7:34
  and
• 7:34
  Amazon
• 7:35
  AGI
• 7:35
  Labs
• 7:36
  formalizes
• 7:36
  something
• 7:37
  the
• 7:37
  agentic
• 7:37
  coding
• 7:37
  community
• 7:38
  has
• 7:38
  been
• 7:38
  doing
• 7:38
  ad
• 7:38
  hoc:
• 7:39
  teaching
• 7:39
  models
• 7:40
  to
• 7:40
  recursively
• 7:40
  delegate
• 7:41
  sub-tasks
• 7:41
  to
• 7:41
  copies
• 7:42
  of
• 7:42
  themselves.
• 7:43
  Recursive
• 7:43
  Agent
• 7:43
  Optimization
• 7:44
  (RAO)
• 7:44
  trains
• 7:45
  a
• 7:45
  single
• 7:45
  LLM
• 7:45
  policy
• 7:46
  that
• 7:46
  runs
• 7:46
  at
• 7:46
  every
• 7:46
  node
• 7:47
  of
• 7:47
  a
• 7:47
  dynamically
• 7:47
  generated
• 7:48
  execution
• 7:48
  tree.
• 7:49
  Unlike
• 7:49
  hand-designed
• 7:50
  orchestration,
• 7:51
  the
• 7:51
  model
• 7:51
  itself
• 7:52
  learns
• 7:52
  when
• 7:52
  delegation
• 7:53
  helps,
• 7:53
  how
• 7:53
  to
• 7:54
  formulate
• 7:54
  sub-tasks,
• 7:55
  whether
• 7:55
  to
• 7:55
  run
• 7:55
  children
• 7:56
  sequentially
• 7:56
  or
• 7:56
  concurrently,
• 7:57
  and
• 7:57
  how
• 7:57
  to
• 7:57
  aggregate
• 7:58
  their
• 7:58
  outputs.
• 7:59
  The
• 7:59
  training
• 7:59
  signal
• 7:59
  is
• 8:00
  clever:
• 8:00
  a
• 8:00
  local
• 8:01
  node
• 8:01
  reward
• 8:01
  combines
• 8:02
  each
• 8:02
  agent's
• 8:02
  own
• 8:02
  success
• 8:03
  with
• 8:03
  a
• 8:03
  delegation
• 8:03
  bonus
• 8:04
  from
• 8:04
  its
• 8:04
  children's
• 8:04
  average
• 8:05
  performance,
• 8:06
  weighted
• 8:06
  by
• 8:06
  depth-level
• 8:07
  inverse
• 8:07
  frequency
• 8:07
  to
• 8:07
  prevent
• 8:08
  deep
• 8:08
  nodes
• 8:08
  from
• 8:08
  dominating
• 8:09
  gradients.
• 8:10
  On
• 8:10
  TextCraft-Synth
• 8:11
  (a
• 8:11
  Minecraft-style
• 8:11
  crafting
• 8:12
  benchmark),
• 8:13
  recursive
• 8:13
  agents
• 8:13
  trained
• 8:14
  only
• 8:14
  on
• 8:14
  medium
• 8:14
  tasks
• 8:15
  achieved
• 8:15
  88%
• 8:16
  success
• 8:16
  on
• 8:16
  hard
• 8:16
  tasks
• 8:17
  versus
• 8:17
  20%
• 8:18
  for
• 8:18
  single-agent
• 8:18
  baselines.
• 8:19
  On
• 8:20
  Oolong-Real,
• 8:20
  a
• 8:20
  long-context
• 8:21
  aggregation
• 8:22
  task
• 8:22
  requiring
• 8:23
  55K+
• 8:23
  tokens,
• 8:24
  a
• 8:24
  recursive
• 8:25
  Qwen3-VL-30B
• 8:27
  trained
• 8:27
  with
• 8:27
  only
• 8:27
  32K
• 8:28
  context
• 8:28
  approached
• 8:29
  frontier
• 8:29
  model
• 8:29
  performance.
• 8:30
  When
• 8:30
  tasks
• 8:31
  decompose
• 8:31
  cleanly,
• 8:32
  recursive
• 8:32
  agents
• 8:32
  achieve
• 8:33
  up
• 8:33
  to
• 8:33
  2.5x
• 8:34
  wall-clock
• 8:34
  speedup
• 8:34
  through
• 8:34
  concurrent
• 8:35
  sub-agent
• 8:35
  execution.
• 8:36
  The
• 8:36
  RAO
• 8:36
  paper's
• 8:37
  conclusion
• 8:37
  —
• 8:38
  "inference-time
• 8:38
  scaffolds
• 8:39
  should
• 8:39
  not
• 8:39
  merely
• 8:39
  be
• 8:39
  designed
• 8:39
  around
• 8:40
  models;
• 8:40
  models
• 8:41
  should
• 8:41
  be
• 8:41
  trained
• 8:41
  to
• 8:42
  use
• 8:42
  them"
• 8:42
  —
• 8:43
  aligns
• 8:43
  with
• 8:43
  Anthropic's
• 8:44
  emerging
• 8:44
  concept
• 8:45
  of
• 8:45
  "dreaming"
• 8:45
  for
• 8:45
  agents.
• 8:47
  The
• 8:47
  idea:
• 8:48
  agents
• 8:48
  that
• 8:48
  review
• 8:48
  their
• 8:49
  own
• 8:49
  behavior
• 8:49
  between
• 8:49
  sessions,
• 8:50
  compress
• 8:51
  signal
• 8:51
  from
• 8:51
  noise,
• 8:52
  refine
• 8:52
  delegation
• 8:53
  patterns,
• 8:53
  and
• 8:53
  change
• 8:54
  future
• 8:54
  execution
• 8:55
  based
• 8:55
  on
• 8:55
  historical
• 8:55
  outcomes.
• 8:57
  This
• 8:57
  moves
• 8:57
  beyond
• 8:58
  simple
• 8:58
  persistent
• 8:59
  memory
• 8:59
  toward
• 8:59
  what
• 9:00
  one
• 9:00
  commentator
• 9:00
  described
• 9:01
  as
• 9:01
  "governed
• 9:01
  cognition"
• 9:02
  —
• 9:03
  structural
• 9:03
  awareness
• 9:03
  and
• 9:04
  traceability
• 9:04
  over
• 9:04
  time.
• 9:05
  Whether
• 9:05
  this
• 9:05
  materializes
• 9:06
  as
• 9:06
  a
• 9:06
  product
• 9:07
  feature
• 9:07
  or
• 9:07
  stays
• 9:08
  at
• 9:08
  the
• 9:08
  research
• 9:08
  stage
• 9:08
  remains
• 9:09
  to
• 9:09
  be
• 9:09
  seen,
• 9:09
  but
• 9:09
  it
• 9:10
  directly
• 9:10
  addresses
• 9:10
  the
• 9:10
  critique
• 9:11
  that
• 9:11
  current
• 9:11
  agents
• 9:12
  "remain
• 9:12
  episodic,
• 9:13
  stateless
• 9:13
  between
• 9:13
  sessions,
• 9:14
  and
• 9:14
  don't
• 9:14
  get
• 9:14
  better
• 9:14
  with
• 9:15
  use"
• 9:15
  .
• 9:15
  Training
• 9:16
  agents
• 9:16
  also
• 9:16
  requires
• 9:17
  environments
• 9:17
  to
• 9:17
  train
• 9:18
  them
• 9:18
  in.
• 9:19
  Qwen's
• 9:19
  WebWorld
• 9:19
  series
• 9:20
  —
• 9:20
  32B,
• 9:21
  14B,
• 9:22
  and
• 9:22
  8B
• 9:22
  models
• 9:22
  fine-tuned
• 9:23
  from
• 9:23
  Qwen3
• 9:24
  on
• 9:24
  over
• 9:24
  one
• 9:24
  million
• 9:25
  real-world
• 9:25
  web
• 9:25
  interaction
• 9:26
  trajectories
• 9:26
  —
• 9:27
  provides
• 9:27
  exactly
• 9:27
  that.
• 9:28
  The
• 9:28
  models
• 9:29
  support
• 9:29
  long-horizon
• 9:29
  simulation
• 9:30
  (30+
• 9:30
  steps),
• 9:31
  multiple
• 9:31
  state
• 9:32
  representations
• 9:33
  (A11y
• 9:33
  Tree,
• 9:34
  HTML,
• 9:35
  XML,
• 9:36
  Markdown),
• 9:36
  and
• 9:36
  chain-of-thought
• 9:37
  reasoning
• 9:37
  for
• 9:37
  transition
• 9:38
  prediction.
• 9:39
  Agents
• 9:39
  trained
• 9:40
  on
• 9:40
  WebWorld-synthesized
• 9:41
  trajectories
• 9:41
  gained
• 9:42
  +9.9%
• 9:43
  on
• 9:43
  MiniWob++
• 9:44
  and
• 9:44
  +10.9%
• 9:46
  on
• 9:46
  WebArena;
• 9:47
  used
• 9:47
  for
• 9:47
  inference-time
• 9:48
  lookahead
• 9:48
  search,
• 9:49
  WebWorld
• 9:49
  outperformed
• 9:50
  GPT-5
• 9:51
  as
• 9:51
  a
• 9:51
  world
• 9:51
  model
• 9:51
  .
• 9:51
  And
• 9:52
  once
• 9:52
  trained
• 9:52
  agents
• 9:52
  need
• 9:53
  to
• 9:53
  run
• 9:53
  safely
• 9:53
  against
• 9:53
  real
• 9:54
  data,
• 9:54
  Tilde.run
• 9:55
  —
• 9:56
  from
• 9:56
  the
• 9:56
  team
• 9:56
  behind
• 9:56
  lakeFS
• 9:57
  —
• 9:57
  offers
• 9:58
  a
• 9:58
  transactional
• 9:58
  versioned
• 9:59
  filesystem
• 9:59
  where
• 9:59
  every
• 10:00
  agent
• 10:00
  run
• 10:00
  is
• 10:00
  an
• 10:00
  atomic
• 10:01
  transaction:
• 10:01
  on
• 10:01
  clean
• 10:02
  exit,
• 10:02
  changes
• 10:03
  commit;
• 10:04
  on
• 10:04
  failure,
• 10:05
  nothing
• 10:05
  changes.
• 10:06
  Code
• 10:06
  from
• 10:06
  GitHub,
• 10:07
  data
• 10:07
  from
• 10:07
  S3,
• 10:08
  and
• 10:08
  documents
• 10:09
  from
• 10:09
  Drive
• 10:09
  appear
• 10:10
  as
• 10:10
  a
• 10:10
  single
• 10:10
  ~/sandbox,
• 10:11
  with
• 10:12
  every
• 10:12
  outbound
• 10:12
  call
• 10:12
  policy-checked
• 10:13
  and
• 10:13
  logged
• 10:13
  .
• 10:16
  Anthropic
• 10:17
  has
• 10:17
  secured
• 10:17
  access
• 10:17
  to
• 10:18
  over
• 10:18
  220,000
• 10:19
  NVIDIA
• 10:19
  GPUs
• 10:20
  at
• 10:20
  SpaceX's
• 10:21
  Colossus
• 10:21
  1
• 10:22
  data
• 10:22
  center,
• 10:23
  adding
• 10:23
  more
• 10:23
  than
• 10:23
  300
• 10:24
  megawatts
• 10:25
  of
• 10:25
  compute
• 10:25
  capacity
• 10:25
  within
• 10:26
  a
• 10:26
  single
• 10:26
  month.
• 10:27
  The
• 10:27
  deal
• 10:27
  arrives
• 10:28
  on
• 10:28
  top
• 10:28
  of
• 10:28
  the
• 10:28
  multi-gigawatt
• 10:29
  commitments
• 10:29
  already
• 10:29
  documented
• 10:30
  —
• 10:31
  5GW
• 10:31
  of
• 10:31
  TPU
• 10:32
  capacity
• 10:32
  from
• 10:32
  Google
• 10:33
  Cloud,
• 10:33
  a
• 10:33
  gigawatt
• 10:34
  deal
• 10:34
  with
• 10:34
  Microsoft,
• 10:35
  and
• 10:35
  900MW
• 10:37
  via
• 10:37
  Stargate
• 10:38
  Abilene.
• 10:39
  The
• 10:39
  community
• 10:39
  reaction
• 10:40
  splits
• 10:40
  predictably:
• 10:41
  optimists
• 10:41
  see
• 10:42
  faster
• 10:42
  responses,
• 10:43
  higher
• 10:43
  rate
• 10:43
  limits,
• 10:44
  and
• 10:44
  a
• 10:44
  signal
• 10:44
  that
• 10:45
  a
• 10:45
  significant
• 10:45
  model
• 10:45
  launch
• 10:46
  is
• 10:46
  imminent;
• 10:47
  skeptics
• 10:47
  note
• 10:47
  the
• 10:48
  Musk
• 10:48
  connection
• 10:48
  and
• 10:49
  wonder
• 10:49
  what's
• 10:49
  next
• 10:49
  for
• 10:49
  Grok.
• 10:50
  At
• 10:51
  this
• 10:51
  point,
• 10:51
  more
• 10:51
  than
• 10:51
  half
• 10:52
  of
• 10:52
  all
• 10:52
  AI
• 10:53
  data
• 10:53
  center
• 10:53
  capacity
• 10:54
  under
• 10:54
  construction
• 10:55
  is
• 10:55
  being
• 10:55
  built
• 10:55
  for
• 10:55
  just
• 10:56
  two
• 10:56
  companies.
• 10:57
  The
• 10:57
  strategic
• 10:58
  read
• 10:58
  is
• 10:58
  straightforward
• 10:58
  —
• 10:59
  Anthropic
• 10:59
  is
• 10:59
  positioning
• 11:00
  for
• 11:00
  an
• 11:00
  inference-heavy
• 11:01
  future
• 11:01
  of
• 11:01
  agents
• 11:02
  and
• 11:02
  multimodal
• 11:02
  features
• 11:03
  that
• 11:03
  will
• 11:03
  consume
• 11:04
  orders
• 11:04
  of
• 11:04
  magnitude
• 11:05
  more
• 11:05
  compute
• 11:05
  than
• 11:05
  chat.
• 11:07
  All
• 11:07
  those
• 11:07
  GPUs
• 11:08
  ultimately
• 11:08
  need
• 11:08
  to
• 11:08
  produce
• 11:09
  measurable
• 11:09
  business
• 11:10
  value,
• 11:10
  and
• 11:10
  a
• 11:10
  new
• 11:11
  CFO-oriented
• 11:12
  framework
• 11:12
  proposes
• 11:13
  four
• 11:13
  layers
• 11:13
  of
• 11:13
  AI
• 11:14
  measurement
• 11:14
  that
• 11:14
  most
• 11:15
  companies
• 11:15
  are
• 11:15
  getting
• 11:15
  wrong.
• 11:16
  Layer
• 11:16
  1
• 11:17
  (Consumption)
• 11:17
  tracks
• 11:17
  tokens
• 11:18
  and
• 11:18
  compute
• 11:18
  —
• 11:19
  where
• 11:19
  nearly
• 11:19
  everyone
• 11:20
  is
• 11:20
  stuck.
• 11:21
  Layer
• 11:21
  2
• 11:21
  (Work)
• 11:21
  counts
• 11:22
  what
• 11:22
  AI
• 11:22
  did:
• 11:23
  records
• 11:23
  updated,
• 11:24
  workflows
• 11:24
  triggered,
• 11:25
  code
• 11:25
  accepted.
• 11:26
  Layer
• 11:26
  3
• 11:26
  (Outcomes)
• 11:27
  captures
• 11:27
  verified
• 11:28
  results:
• 11:28
  tickets
• 11:28
  resolved,
• 11:29
  leads
• 11:29
  qualified.
• 11:30
  Layer
• 11:31
  4
• 11:31
  (Business
• 11:31
  Impact)
• 11:32
  measures
• 11:32
  P&L
• 11:32
  effects.
• 11:33
  Salesforce
• 11:34
  has
• 11:34
  introduced
• 11:34
  the
• 11:35
  Agentic
• 11:35
  Work
• 11:35
  Unit
• 11:35
  (AWU),
• 11:37
  reporting
• 11:37
  2.4
• 11:38
  billion
• 11:38
  to
• 11:38
  date
• 11:38
  with
• 11:39
  771
• 11:40
  million
• 11:40
  in
• 11:40
  Q4
• 11:40
  alone
• 11:41
  —
• 11:41
  but
• 11:42
  that's
• 11:42
  still
• 11:42
  Layer
• 11:42
  2.
• 11:43
  Intercom
• 11:43
  prices
• 11:44
  Fin
• 11:44
  at
• 11:44
  $0.99
• 11:45
  per
• 11:45
  resolution
• 11:45
  with
• 11:46
  reversal
• 11:46
  logic
• 11:46
  if
• 11:46
  the
• 11:47
  customer
• 11:47
  returns,
• 11:47
  putting
• 11:48
  it
• 11:48
  cleanly
• 11:48
  at
• 11:48
  Layer
• 11:48
  3.
• 11:50
  ServiceNow's
• 11:50
  CFO
• 11:51
  cited
• 11:51
  roughly
• 11:51
  $100
• 11:52
  million
• 11:52
  in
• 11:53
  internal
• 11:53
  headcount
• 11:53
  savings
• 11:53
  —
• 11:55
  a
• 11:55
  Layer
• 11:55
  4
• 11:55
  claim
• 11:55
  more
• 11:56
  credible
• 11:56
  because
• 11:56
  the
• 11:56
  vendor
• 11:57
  reports
• 11:57
  on
• 11:57
  itself.
• 11:58
  Microsoft,
• 11:59
  despite
• 11:59
  15
• 11:59
  million
• 12:00
  paid
• 12:00
  Copilot
• 12:01
  seats,
• 12:01
  remains
• 12:02
  stuck
• 12:02
  at
• 12:02
  Layer
• 12:02
  1,
• 12:02
  reporting
• 12:03
  access
• 12:03
  and
• 12:03
  engagement
• 12:04
  rather
• 12:04
  than
• 12:04
  work
• 12:04
  or
• 12:05
  outcomes.
• 12:06
  The
• 12:06
  practical
• 12:06
  takeaway:
• 12:07
  if
• 12:08
  you
• 12:08
  only
• 12:08
  track
• 12:08
  tokens,
• 12:09
  a
• 12:09
  30%
• 12:10
  escalation
• 12:10
  rate
• 12:10
  that
• 12:11
  silently
• 12:11
  destroys
• 12:12
  gross
• 12:12
  profit
• 12:13
  is
• 12:13
  invisible
• 12:13
  in
• 12:13
  your
• 12:13
  dashboard.
• 12:15
  At
• 12:15
  the
• 12:15
  hardware
• 12:15
  edge,
• 12:16
  Cognitum
• 12:17
  is
• 12:17
  shipping
• 12:17
  a
• 12:17
  $131
• 12:18
  AI
• 12:19
  device
• 12:19
  with
• 12:19
  a
• 12:19
  WASM
• 12:19
  runtime,
• 12:20
  Ed25519
• 12:22
  cryptographic
• 12:22
  identity,
• 12:23
  on-device
• 12:24
  vector
• 12:24
  store,
• 12:25
  OTA
• 12:26
  firmware
• 12:26
  updates,
• 12:27
  and
• 12:27
  MCP
• 12:28
  protocol
• 12:28
  integration
• 12:28
  —
• 12:29
  self-learning
• 12:30
  hardware
• 12:30
  designed
• 12:31
  to
• 12:31
  work
• 12:31
  without
• 12:31
  the
• 12:32
  cloud.
• 12:33
  Each
• 12:33
  device
• 12:33
  gets
• 12:33
  a
• 12:33
  unique
• 12:34
  cryptographic
• 12:34
  identity
• 12:35
  at
• 12:35
  manufacturing
• 12:36
  time,
• 12:36
  sandboxed
• 12:37
  execution
• 12:37
  for
• 12:37
  self-contained
• 12:38
  apps
• 12:38
  called
• 12:38
  "cogs,"
• 12:39
  and
• 12:39
  multi-device
• 12:40
  coordination
• 12:41
  for
• 12:41
  fleet
• 12:41
  deployments.
• 12:42
  It
• 12:43
  slots
• 12:43
  into
• 12:43
  the
• 12:43
  Model
• 12:43
  Context
• 12:44
  Protocol
• 12:44
  (MCP)
• 12:45
  ecosystem
• 12:46
  with
• 12:46
  both
• 12:46
  cloud
• 12:47
  and
• 12:47
  local
• 12:47
  plugins,
• 12:48
  positioning
• 12:48
  itself
• 12:49
  as
• 12:49
  physical
• 12:49
  infrastructure
• 12:50
  for
• 12:50
  the
• 12:50
  agentic
• 12:50
  stack
• 12:51
  rather
• 12:51
  than
• 12:51
  another
• 12:52
  dev
• 12:52
  board
• 12:52
  .
• 12:55
  Five
• 12:56
  months
• 12:56
  ago,
• 12:56
  a
• 12:57
  developer
• 12:57
  got
• 12:57
  tired
• 12:57
  of
• 12:58
  starting
• 12:58
  every
• 12:58
  Claude
• 12:58
  session
• 12:59
  from
• 12:59
  scratch.
• 12:59
  So
• 13:00
  they
• 13:00
  built
• 13:00
  iai-mcp
• 13:00
  —
• 13:01
  a
• 13:01
  local
• 13:01
  daemon
• 13:02
  that
• 13:02
  captures
• 13:02
  every
• 13:03
  conversation,
• 13:03
  organizes
• 13:04
  it
• 13:04
  into
• 13:04
  three
• 13:04
  memory
• 13:05
  tiers
• 13:05
  (working,
• 13:05
  episodic,
• 13:06
  semantic),
• 13:06
  and
• 13:06
  feeds
• 13:07
  the
• 13:07
  right
• 13:07
  context
• 13:07
  back
• 13:07
  when
• 13:07
  a
• 13:07
  new
• 13:08
  session
• 13:08
  starts.
• 13:09
  It
• 13:09
  stores
• 13:10
  everything
• 13:10
  verbatim,
• 13:11
  runs
• 13:11
  neural
• 13:11
  embeddings
• 13:12
  locally,
• 13:13
  encrypts
• 13:13
  at
• 13:13
  rest
• 13:13
  with
• 13:13
  AES-256,
• 13:15
  and
• 13:16
  consolidates
• 13:16
  memory
• 13:16
  in
• 13:16
  the
• 13:17
  background
• 13:17
  during
• 13:17
  idle
• 13:17
  time.
• 13:19
  The
• 13:19
  published
• 13:19
  numbers:
• 13:20
  verbatim
• 13:21
  recall
• 13:21
  above
• 13:21
  99%,
• 13:22
  retrieval
• 13:23
  under
• 13:23
  100ms,
• 13:24
  session-start
• 13:25
  cost
• 13:25
  under
• 13:25
  3,000
• 13:26
  tokens.
• 13:27
  After
• 13:27
  five
• 13:27
  months
• 13:27
  of
• 13:27
  daily
• 13:28
  use
• 13:28
  with
• 13:28
  Claude
• 13:28
  Code,
• 13:29
  the
• 13:29
  developer
• 13:30
  reports
• 13:30
  it
• 13:30
  knows
• 13:30
  coding
• 13:31
  style,
• 13:31
  project
• 13:32
  structures,
• 13:32
  and
• 13:32
  preferences
• 13:33
  that
• 13:33
  were
• 13:33
  never
• 13:33
  explicitly
• 13:34
  saved
• 13:34
  —
• 13:35
  picked
• 13:35
  up
• 13:35
  from
• 13:35
  conversation
• 13:36
  patterns.
• 13:37
  It's
• 13:37
  now
• 13:37
  open
• 13:38
  source
• 13:38
  under
• 13:38
  MIT.
• 13:39
  The
• 13:39
  ecosystem's
• 13:40
  personality
• 13:41
  layer
• 13:41
  is
• 13:41
  getting
• 13:41
  attention
• 13:42
  too.
• 13:43
  Claude
• 13:43
  Skins
• 13:43
  brings
• 13:43
  visual
• 13:44
  theming
• 13:44
  to
• 13:44
  the
• 13:44
  Claude
• 13:45
  Code
• 13:45
  CLI
• 13:45
  —
• 13:46
  terminal
• 13:46
  colors,
• 13:47
  ASCII
• 13:48
  art
• 13:48
  banners,
• 13:48
  status
• 13:49
  line
• 13:49
  customization,
• 13:50
  voice/personality
• 13:51
  presets,
• 13:52
  and
• 13:52
  tool
• 13:52
  feedback
• 13:52
  sounds.
• 13:53
  Nine
• 13:54
  skins
• 13:54
  ship
• 13:54
  out
• 13:54
  of
• 13:54
  the
• 13:55
  box,
• 13:55
  from
• 13:55
  "Nebula"
• 13:56
  (offensive
• 13:56
  security
• 13:57
  scanner
• 13:57
  aesthetic)
• 13:57
  to
• 13:57
  "Brutalist"
• 13:58
  (pure
• 13:58
  monochrome,
• 13:59
  zero
• 14:00
  decoration,
• 14:00
  maximum
• 14:01
  terseness).
• 14:02
  Skins
• 14:03
  are
• 14:03
  YAML-defined,
• 14:04
  hot-swappable
• 14:04
  without
• 14:05
  restart,
• 14:05
  and
• 14:06
  integrate
• 14:06
  via
• 14:06
  Claude
• 14:07
  Code's
• 14:07
  hooks
• 14:07
  system
• 14:07
  .
• 14:08
  For
• 14:08
  those
• 14:08
  building
• 14:09
  agents
• 14:09
  rather
• 14:09
  than
• 14:10
  skinning
• 14:10
  them,
• 14:10
  Marktechpost's
• 14:11
  growing
• 14:12
  collection
• 14:12
  of
• 14:12
  60+
• 14:13
  AI
• 14:13
  agent
• 14:14
  tutorials
• 14:14
  and
• 14:14
  implementations
• 14:15
  now
• 14:16
  covers
• 14:16
  everything
• 14:16
  from
• 14:17
  multi-agent
• 14:17
  swarm
• 14:18
  orchestration
• 14:18
  to
• 14:18
  procedural
• 14:19
  memory
• 14:19
  systems
• 14:20
  to
• 14:20
  cybersecurity
• 14:21
  AI
• 14:21
  agents
• 14:21
  —
• 14:22
  a
• 14:22
  practical
• 14:22
  onboarding
• 14:23
  resource
• 14:24
  for
• 14:24
  a
• 14:24
  framework
• 14:24
  ecosystem
• 14:25
  that's
• 14:25
  maturing
• 14:26
  faster
• 14:26
  than
• 14:26
  documentation
• 14:27
  can
• 14:27
  keep
• 14:27
  up
• 14:27
  .
• 14:30
  Stuart
• 14:30
  Winter-Tear
• 14:31
  frames
• 14:31
  the
• 14:31
  sycophancy
• 14:32
  problem
• 14:32
  not
• 14:32
  as
• 14:32
  a
• 14:33
  model
• 14:33
  quirk
• 14:33
  but
• 14:33
  as
• 14:34
  an
• 14:34
  organizational
• 14:34
  threat:
• 14:35
  "sycophancy
• 14:36
  is
• 14:36
  not
• 14:36
  simply
• 14:37
  agreement,
• 14:37
  but
• 14:37
  a
• 14:37
  boundary
• 14:38
  failure
• 14:38
  between
• 14:38
  social
• 14:39
  alignment
• 14:39
  and
• 14:39
  epistemic
• 14:40
  integrity."
• 14:41
  The
• 14:41
  concern
• 14:41
  isn't
• 14:42
  chatbots
• 14:42
  being
• 14:42
  too
• 14:43
  polite
• 14:43
  —
• 14:43
  it's
• 14:44
  AI
• 14:44
  moving
• 14:44
  into
• 14:45
  strategy,
• 14:45
  governance,
• 14:46
  board
• 14:46
  packs,
• 14:46
  and
• 14:46
  investment
• 14:47
  analysis
• 14:47
  where
• 14:47
  "confidence
• 14:48
  arrives
• 14:48
  before
• 14:49
  judgement."
• 14:50
  A
• 14:50
  model
• 14:50
  that
• 14:50
  softens
• 14:51
  corrections,
• 14:51
  preserves
• 14:52
  mood,
• 14:52
  and
• 14:52
  hands
• 14:53
  back
• 14:53
  a
• 14:53
  cleaner
• 14:53
  version
• 14:54
  of
• 14:54
  what
• 14:54
  the
• 14:54
  user
• 14:54
  already
• 14:55
  wanted
• 14:55
  to
• 14:55
  hear
• 14:56
  isn't
• 14:56
  helping
• 14:56
  anyone
• 14:57
  think.
• 14:58
  It's
• 14:58
  laundering
• 14:59
  flawed
• 14:59
  premises
• 14:59
  through
• 15:00
  better
• 15:00
  prose.
• 15:01
  As
• 15:01
  one
• 15:01
  commenter
• 15:02
  noted,
• 15:03
  what's
• 15:03
  needed
• 15:03
  isn't
• 15:03
  better
• 15:04
  prompting
• 15:04
  but
• 15:04
  "structural
• 15:05
  collision
• 15:05
  detection"
• 15:06
  —
• 15:07
  a
• 15:07
  governance
• 15:07
  layer
• 15:07
  that
• 15:07
  flags
• 15:08
  when
• 15:08
  AI
• 15:08
  systems
• 15:09
  operate
• 15:09
  on
• 15:09
  contradictory
• 15:10
  assumptions
• 15:11
  or
• 15:11
  when
• 15:11
  a
• 15:11
  new
• 15:11
  deployment
• 15:12
  undermines
• 15:12
  an
• 15:12
  existing
• 15:13
  decision.
• 15:14
  The
• 15:14
  antidote
• 15:15
  to
• 15:15
  sycophancy
• 15:15
  turns
• 15:16
  out
• 15:16
  to
• 15:16
  be
• 15:16
  architectural,
• 15:17
  not
• 15:17
  behavioral.
• 15:19
  On
• 15:19
  a
• 15:19
  lighter
• 15:20
  research
• 15:20
  note,
• 15:21
  Vibe-Cast's
• 15:21
  JEPA
• 15:22
  exploration
• 15:22
  offers
• 15:23
  an
• 15:23
  accessible
• 15:23
  entry
• 15:23
  point
• 15:24
  into
• 15:24
  Joint
• 15:24
  Embedding
• 15:24
  Predictive
• 15:25
  Architecture
• 15:25
  —
• 15:27
  Meta
• 15:27
  AI's
• 15:27
  self-supervised
• 15:28
  learning
• 15:28
  approach
• 15:29
  that
• 15:29
  learns
• 15:29
  representations
• 15:30
  by
• 15:30
  predicting
• 15:30
  masked
• 15:30
  embeddings
• 15:31
  rather
• 15:31
  than
• 15:31
  reconstructing
• 15:32
  pixels.
• 15:33
  The
• 15:33
  educational
• 15:34
  implementation
• 15:34
  walks
• 15:34
  through
• 15:35
  patch-based
• 15:35
  encoding,
• 15:36
  masking
• 15:36
  strategies,
• 15:37
  predictor
• 15:37
  networks,
• 15:38
  and
• 15:38
  momentum-updated
• 15:39
  encoders,
• 15:39
  providing
• 15:40
  a
• 15:40
  hands-on
• 15:40
  path
• 15:41
  into
• 15:41
  a
• 15:41
  paradigm
• 15:41
  that
• 15:41
  keeps
• 15:42
  winning
• 15:42
  empirically
• 15:43
  even
• 15:43
  as
• 15:43
  debate
• 15:43
  continues
• 15:44
  over
• 15:44
  whether
• 15:44
  it
• 15:44
  will
• 15:45
  deliver
• 15:45
  on
• 15:45
  LeCun's
• 15:45
  vision
• 15:46
  of
• 15:46
  a
• 15:46
  post-generative
• 15:46
  AI
• 15:46
  future
• 15:47
  .