Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – May 15, 2026

Transcript

AGI Dreams Podcast – May 15, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  patch-to-exploit
• 0:01
  pipeline
• 0:01
  has
• 0:01
  gone
• 0:01
  from
• 0:02
  conference-talk
• 0:02
  speculation
• 0:03
  to
• 0:03
  single-researcher
• 0:04
  reality.
• 0:05
  A
• 0:05
  security
• 0:05
  researcher
• 0:06
  writing
• 0:06
  under
• 0:06
  the
• 0:07
  handle
• 0:07
  OriginHQ
• 0:08
  has
• 0:08
  published
• 0:08
  PatchWatch
• 0:09
  and
• 0:09
  Pocsmith
• 0:09
  —
• 0:10
  a
• 0:10
  two-stage
• 0:10
  Rust
• 0:11
  +
• 0:11
  Agent
• 0:11
  SDK
• 0:12
  pipeline
• 0:12
  that
• 0:12
  ingests
• 0:13
  Microsoft's
• 0:13
  Patch
• 0:14
  Tuesday
• 0:14
  releases,
• 0:15
  diffs
• 0:15
  the
• 0:15
  binaries
• 0:15
  via
• 0:16
  Ghidriff,
• 0:16
  triages
• 0:17
  candidates
• 0:17
  by
• 0:17
  CVSS
• 0:18
  score,
• 0:19
  and
• 0:19
  then
• 0:19
  hands
• 0:20
  the
• 0:20
  structured
• 0:20
  diff
• 0:20
  report
• 0:21
  to
• 0:21
  a
• 0:21
  Claude
• 0:21
  agent
• 0:21
  operating
• 0:22
  inside
• 0:22
  a
• 0:22
  KDNET-attached
• 0:23
  Hyper-V
• 0:23
  VM
• 0:24
  equipped
• 0:24
  with
• 0:24
  purpose-built
• 0:25
  MCP
• 0:25
  servers
• 0:26
  for
• 0:26
  VM
• 0:26
  lifecycle,
• 0:27
  kernel
• 0:27
  debugging,
• 0:28
  Ghidra
• 0:28
  static
• 0:29
  analysis,
• 0:29
  and
• 0:29
  compilation.
• 0:31
  The
• 0:31
  system
• 0:31
  produced
• 0:32
  a
• 0:32
  verified
• 0:32
  elevation-of-privilege
• 0:33
  exploit
• 0:33
  for
• 0:33
  CVE-2026-27914
• 0:36
  (a
• 0:38
  Mark-of-the-Web
• 0:39
  bypass
• 0:39
  in
• 0:39
  mmc.exe)
• 0:41
  and
• 0:41
  a
• 0:41
  Level
• 0:41
  A
• 0:41
  crash
• 0:42
  reproduction
• 0:42
  for
• 0:42
  CVE-2026-41096
• 0:45
  (a
• 0:47
  CVSS
• 0:47
  9.8
• 0:48
  heap
• 0:48
  overflow
• 0:49
  in
• 0:49
  ws2
• 0:50
  32.dll's
• 0:51
  DNS
• 0:51
  client).
• 0:52
  Total
• 0:53
  spend:
• 0:54
  roughly
• 0:54
  $300
• 0:55
  in
• 0:55
  API
• 0:55
  tokens
• 0:56
  on
• 0:56
  an
• 0:56
  Anthropic
• 0:56
  Team
• 0:56
  subscription.
• 0:58
  What
• 0:58
  makes
• 0:58
  this
• 0:58
  disclosure
• 0:59
  notable
• 0:59
  is
• 0:59
  not
• 1:00
  the
• 1:00
  raw
• 1:00
  capability
• 1:00
  —
• 1:01
  Mozilla's
• 1:01
  271-vulnerability
• 1:03
  haul
• 1:03
  and
• 1:03
  MOAK's
• 1:04
  97.8%
• 1:05
  CVE
• 1:05
  exploitation
• 1:06
  rate
• 1:06
  have
• 1:07
  already
• 1:07
  demonstrated
• 1:07
  that
• 1:08
  AI-assisted
• 1:08
  offense
• 1:09
  scales
• 1:09
  —
• 1:10
  but
• 1:10
  the
• 1:10
  architecture.
• 1:11
  Each
• 1:11
  MCP
• 1:12
  server
• 1:12
  (vm-control,
• 1:13
  kernel-debugger,
• 1:14
  ghidra-bridge,
• 1:15
  research-tools)
• 1:16
  is
• 1:16
  reusable
• 1:16
  outside
• 1:17
  the
• 1:17
  harness.
• 1:18
  Phases
• 1:18
  are
• 1:18
  bounded
• 1:19
  by
• 1:19
  time,
• 1:19
  iteration,
• 1:20
  and
• 1:20
  dollar
• 1:21
  budgets,
• 1:21
  with
• 1:22
  context
• 1:22
  windows
• 1:23
  reset
• 1:23
  between
• 1:23
  hypotheses.
• 1:25
  The
• 1:25
  agent
• 1:25
  writes
• 1:25
  its
• 1:25
  own
• 1:26
  working
• 1:26
  state
• 1:26
  to
• 1:26
  a
• 1:26
  persistence
• 1:27
  file,
• 1:28
  meaning
• 1:28
  a
• 1:28
  human
• 1:28
  can
• 1:29
  drop
• 1:29
  in
• 1:29
  mid-run
• 1:29
  for
• 1:30
  manual
• 1:30
  work
• 1:31
  and
• 1:31
  resume
• 1:31
  automated
• 1:32
  operation
• 1:32
  later.
• 1:33
  The
• 1:33
  researcher
• 1:34
  is
• 1:34
  candid
• 1:34
  about
• 1:34
  limitations:
• 1:36
  the
• 1:36
  MMC
• 1:36
  exploit
• 1:37
  landed
• 1:37
  at
• 1:37
  PR:H
• 1:38
  /
• 1:38
  UI:R
• 1:39
  rather
• 1:39
  than
• 1:40
  the
• 1:40
  PR:L
• 1:40
  /
• 1:41
  UI:N
• 1:41
  boundary
• 1:42
  MSRC
• 1:43
  scored,
• 1:44
  and
• 1:44
  the
• 1:44
  system
• 1:44
  "isn't
• 1:45
  producing
• 1:45
  high-quality
• 1:45
  exploits
• 1:46
  yet."
• 1:47
  But
• 1:47
  the
• 1:47
  gap
• 1:47
  between
• 1:48
  a
• 1:48
  working
• 1:48
  N-day
• 1:49
  repro
• 1:49
  and
• 1:50
  a
• 1:50
  sharpened
• 1:50
  weaponized
• 1:50
  chain
• 1:51
  is
• 1:51
  a
• 1:51
  refinement
• 1:52
  problem,
• 1:53
  not
• 1:53
  an
• 1:53
  architectural
• 1:53
  one.
• 1:54
  Separately,
• 1:55
  Palisade
• 1:56
  Research
• 1:56
  published
• 1:57
  what
• 1:57
  they
• 1:57
  call
• 1:57
  "the
• 1:57
  first
• 1:58
  documented
• 1:58
  instance
• 1:59
  of
• 1:59
  AI
• 1:59
  self-replication
• 2:00
  via
• 2:00
  hacking"
• 2:01
  —
• 2:02
  models
• 2:02
  including
• 2:03
  GPT-4
• 2:03
  and
• 2:04
  Claude
• 2:04
  given
• 2:04
  a
• 2:05
  single
• 2:05
  prompt
• 2:05
  to
• 2:05
  hack
• 2:06
  a
• 2:06
  machine
• 2:06
  and
• 2:06
  copy
• 2:06
  themselves
• 2:07
  onto
• 2:07
  it,
• 2:08
  with
• 2:08
  the
• 2:08
  copies
• 2:08
  then
• 2:09
  repeating
• 2:09
  the
• 2:09
  process
• 2:09
  in
• 2:09
  a
• 2:10
  chain.
• 2:11
  The
• 2:11
  community
• 2:11
  response
• 2:12
  ranged
• 2:12
  from
• 2:12
  "this
• 2:13
  is
• 2:13
  a
• 2:13
  nothing
• 2:13
  burger
• 2:13
  —
• 2:14
  worms
• 2:15
  have
• 2:15
  existed
• 2:15
  for
• 2:15
  decades"
• 2:16
  to
• 2:16
  "this
• 2:16
  is
• 2:16
  a
• 2:17
  modest
• 2:17
  but
• 2:17
  important
• 2:18
  step
• 2:18
  to
• 2:18
  establish
• 2:19
  precedent."
• 2:20
  The
• 2:20
  paper's
• 2:20
  actual
• 2:21
  finding
• 2:21
  is
• 2:21
  narrower
• 2:22
  than
• 2:22
  its
• 2:22
  headline:
• 2:23
  the
• 2:23
  copies
• 2:23
  communicated
• 2:24
  no
• 2:24
  further
• 2:24
  objective
• 2:25
  beyond
• 2:25
  propagation,
• 2:26
  and
• 2:26
  current
• 2:27
  safety
• 2:27
  filters
• 2:27
  "didn't
• 2:28
  do
• 2:28
  a
• 2:28
  great
• 2:28
  job
• 2:28
  stopping
• 2:29
  it."
• 2:30
  As
• 2:30
  a
• 2:30
  standalone
• 2:30
  result
• 2:31
  it
• 2:31
  is
• 2:31
  incremental,
• 2:32
  but
• 2:32
  it
• 2:32
  maps
• 2:32
  uncomfortably
• 2:33
  onto
• 2:33
  the
• 2:33
  exploit-pipeline
• 2:34
  story
• 2:34
  —
• 2:35
  a
• 2:35
  model
• 2:36
  that
• 2:36
  can
• 2:36
  both
• 2:36
  find
• 2:36
  vulnerabilities
• 2:37
  and
• 2:38
  self-replicate
• 2:38
  changes
• 2:39
  the
• 2:39
  calculus
• 2:39
  on
• 2:39
  containment.
• 2:42
  The
• 2:42
  RTX
• 2:43
  PRO
• 2:43
  6000
• 2:44
  Blackwell
• 2:45
  keeps
• 2:45
  rewriting
• 2:45
  what
• 2:45
  "local
• 2:46
  inference"
• 2:46
  means.
• 2:47
  A
• 2:47
  practitioner
• 2:48
  retrofitted
• 2:49
  the
• 2:49
  MTP
• 2:49
  (Multi-Token
• 2:50
  Prediction)
• 2:50
  head
• 2:51
  onto
• 2:51
  pasta-paul's
• 2:51
  DeepSeek-V4-Flash
• 2:53
  W4A16+FP8
• 2:54
  quantization
• 2:56
  —
• 2:58
  HuggingFace
• 2:58
  Transformers
• 2:59
  had
• 2:59
  been
• 2:59
  silently
• 3:00
  stripping
• 3:00
  it
• 3:00
  at
• 3:00
  load
• 3:00
  time
• 3:00
  —
• 3:01
  ran
• 3:01
  a
• 3:02
  GPTQ
• 3:02
  pass
• 3:03
  on
• 3:03
  the
• 3:03
  MTP
• 3:03
  block's
• 3:03
  routed
• 3:04
  experts
• 3:04
  to
• 3:04
  match
• 3:05
  the
• 3:05
  base
• 3:05
  model's
• 3:05
  INT4
• 3:06
  format,
• 3:07
  patched
• 3:07
  vLLM,
• 3:08
  and
• 3:08
  measured
• 3:09
  decode
• 3:09
  performance
• 3:09
  jumping
• 3:10
  from
• 3:10
  52.85
• 3:11
  tok/s
• 3:12
  to
• 3:12
  85.52
• 3:13
  tok/s
• 3:14
  at
• 3:14
  524k
• 3:15
  context
• 3:15
  on
• 3:15
  two
• 3:16
  RTX
• 3:16
  PRO
• 3:17
  6000
• 3:18
  Max-Q
• 3:18
  cards
• 3:18
  (96
• 3:19
  GB
• 3:19
  each,
• 3:19
  PCIe,
• 3:20
  no
• 3:20
  NVLink).
• 3:21
  Single-stream
• 3:22
  at
• 3:22
  128k
• 3:23
  context
• 3:23
  hits
• 3:24
  approximately
• 3:24
  111
• 3:25
  tok/s.
• 3:26
  The
• 3:27
  671B-total/32B-active
• 3:29
  model
• 3:30
  fits
• 3:30
  on
• 3:30
  192
• 3:31
  GB
• 3:31
  of
• 3:31
  combined
• 3:32
  VRAM.
• 3:33
  Key
• 3:33
  gotcha
• 3:33
  for
• 3:33
  Max-Q
• 3:34
  users:
• 3:35
  you
• 3:35
  must
• 3:35
  pass
• 3:36
  --disable-custom-all-reduce
• 3:37
  because
• 3:37
  vLLM's
• 3:38
  CustomAllreduce
• 3:38
  uses
• 3:39
  CUDA
• 3:39
  P2P
• 3:40
  that
• 3:40
  deadlocks
• 3:40
  on
• 3:40
  PCIe-only
• 3:41
  topology.
• 3:42
  The
• 3:42
  multi-GPU
• 3:43
  story
• 3:43
  got
• 3:43
  simpler
• 3:44
  this
• 3:44
  week
• 3:44
  with
• 3:45
  llama.cpp
• 3:46
  b9095
• 3:48
  shipping
• 3:48
  NCCL-free
• 3:49
  tensor
• 3:49
  parallelism
• 3:50
  for
• 3:50
  dual
• 3:50
  Blackwell
• 3:51
  PCIe
• 3:51
  GPUs.
• 3:53
  Early
• 3:53
  benchmarks
• 3:54
  on
• 3:54
  dual
• 3:54
  5090s
• 3:55
  show
• 3:55
  -sm
• 3:55
  tensor
• 3:56
  matching
• 3:56
  NCCL
• 3:57
  performance
• 3:57
  at
• 3:57
  58
• 3:58
  tok/s
• 3:58
  decode,
• 3:59
  while
• 3:59
  an
• 3:59
  MTP
• 4:00
  branch
• 4:00
  pushes
• 4:00
  to
• 4:00
  135
• 4:01
  tok/s.
• 4:03
  On
• 4:03
  dual
• 4:03
  5060
• 4:04
  Ti
• 4:04
  cards,
• 4:05
  users
• 4:05
  report
• 4:06
  20%
• 4:06
  generation
• 4:07
  gains
• 4:07
  for
• 4:07
  MoE
• 4:07
  models
• 4:08
  and
• 4:08
  10%
• 4:09
  for
• 4:09
  dense.
• 4:09
  The
• 4:10
  significance
• 4:10
  is
• 4:10
  accessibility:
• 4:12
  NCCL
• 4:12
  requires
• 4:13
  specific
• 4:13
  topology,
• 4:14
  driver
• 4:15
  versions,
• 4:15
  and
• 4:15
  library
• 4:16
  configuration
• 4:17
  that
• 4:17
  consumer
• 4:17
  and
• 4:17
  workstation
• 4:18
  setups
• 4:18
  often
• 4:19
  cannot
• 4:19
  satisfy
• 4:20
  cleanly.
• 4:20
  Removing
• 4:21
  it
• 4:21
  as
• 4:21
  a
• 4:21
  hard
• 4:21
  dependency
• 4:22
  means
• 4:22
  dual-GPU
• 4:23
  local
• 4:23
  inference
• 4:24
  becomes
• 4:24
  a
• 4:24
  realistic
• 4:25
  out-of-box
• 4:25
  experience.
• 4:26
  On
• 4:27
  the
• 4:27
  model
• 4:27
  side,
• 4:27
  Qwen
• 4:28
  3.6
• 4:28
  35B
• 4:29
  A3B
• 4:30
  continues
• 4:30
  to
• 4:30
  impress.
• 4:32
  One
• 4:32
  researcher's
• 4:32
  benchmark
• 4:33
  —
• 4:34
  feeding
• 4:34
  niche
• 4:34
  academic
• 4:35
  papers
• 4:35
  plus
• 4:36
  corresponding
• 4:36
  code
• 4:37
  and
• 4:37
  asking
• 4:37
  the
• 4:37
  model
• 4:38
  to
• 4:38
  map
• 4:38
  one
• 4:38
  to
• 4:38
  the
• 4:38
  other
• 4:39
  —
• 4:40
  found
• 4:40
  Qwen
• 4:40
  3.6
• 4:41
  35B
• 4:42
  A3B,
• 4:43
  Qwen
• 4:43
  3.6
• 4:44
  27B,
• 4:45
  Gemma
• 4:45
  4
• 4:46
  26B
• 4:46
  A4B,
• 4:47
  and
• 4:47
  Nemotron
• 4:48
  3
• 4:48
  Nano
• 4:49
  all
• 4:49
  dramatically
• 4:49
  outperforming
• 4:50
  what
• 4:50
  small
• 4:50
  local
• 4:51
  models
• 4:51
  could
• 4:51
  do
• 4:51
  just
• 4:52
  months
• 4:52
  ago.
• 4:53
  The
• 4:53
  improvement
• 4:54
  tracks
• 4:54
  directly
• 4:55
  to
• 4:55
  long-context
• 4:56
  architectures
• 4:56
  (gated
• 4:57
  delta
• 4:57
  net,
• 4:58
  hybrid
• 4:58
  Mamba2,
• 4:59
  sliding
• 4:59
  window
• 5:00
  attention)
• 5:00
  that
• 5:01
  let
• 5:01
  these
• 5:01
  models
• 5:01
  ingest
• 5:02
  full
• 5:02
  papers
• 5:02
  without
• 5:03
  choking.
• 5:04
  The
• 5:04
  poster's
• 5:04
  claim
• 5:05
  that
• 5:05
  "an
• 5:05
  intelligent
• 5:06
  human
• 5:06
  with
• 5:06
  any
• 5:06
  of
• 5:06
  these
• 5:07
  four
• 5:07
  models
• 5:07
  is
• 5:07
  more
• 5:08
  capable
• 5:08
  than
• 5:08
  Opus
• 5:09
  4.7
• 5:09
  on
• 5:10
  its
• 5:10
  own"
• 5:10
  is
• 5:10
  provocative
• 5:11
  but
• 5:11
  grounded
• 5:12
  in
• 5:12
  their
• 5:12
  specific
• 5:12
  eval.
• 5:13
  A
• 5:13
  comprehensive
• 5:14
  independent
• 5:15
  study
• 5:15
  of
• 5:15
  TurboQuant
• 5:15
  confirms
• 5:16
  what
• 5:16
  earlier
• 5:17
  community
• 5:17
  reception
• 5:17
  suggested:
• 5:19
  FP8
• 5:19
  via
• 5:20
  --kv-cache-dtype
• 5:21
  fp8
• 5:22
  remains
• 5:22
  the
• 5:22
  best
• 5:22
  default
• 5:23
  for
• 5:23
  KV-cache
• 5:24
  quantization
• 5:24
  (2x
• 5:25
  capacity,
• 5:26
  negligible
• 5:26
  accuracy
• 5:27
  loss),
• 5:28
  while
• 5:28
  TurboQuant's
• 5:28
  k8v4
• 5:29
  variant
• 5:30
  offers
• 5:30
  only
• 5:30
  2.4x
• 5:31
  versus
• 5:31
  2x
• 5:32
  savings
• 5:32
  with
• 5:32
  "consistent
• 5:33
  negative
• 5:33
  impact
• 5:34
  on
• 5:34
  throughput
• 5:34
  and
• 5:34
  latency."
• 5:36
  The
• 5:36
  4bit-nc
• 5:36
  variant
• 5:37
  may
• 5:37
  serve
• 5:37
  edge
• 5:38
  deployments
• 5:38
  where
• 5:38
  memory
• 5:39
  is
• 5:39
  the
• 5:39
  dominant
• 5:39
  constraint,
• 5:40
  but
• 5:40
  3-bit
• 5:41
  variants
• 5:41
  "substantially
• 5:42
  degrade
• 5:43
  latency
• 5:43
  and
• 5:43
  throughput"
• 5:44
  and
• 5:44
  are
• 5:44
  poor
• 5:44
  production
• 5:45
  candidates.
• 5:46
  A
• 5:46
  separate
• 5:46
  RaBitQ
• 5:47
  comparison
• 5:47
  paper
• 5:48
  found
• 5:48
  TurboQuant
• 5:48
  "performs
• 5:49
  worse
• 5:49
  in
• 5:49
  most
• 5:49
  tested
• 5:50
  settings"
• 5:51
  and
• 5:51
  flagged
• 5:51
  reproducibility
• 5:52
  issues
• 5:52
  in
• 5:52
  TurboQuant's
• 5:53
  published
• 5:53
  results.
• 5:54
  Meanwhile,
• 5:55
  an
• 5:55
  ambitious
• 5:55
  solo
• 5:56
  developer
• 5:56
  is
• 5:56
  training
• 5:57
  a
• 5:57
  7B-parameter
• 5:58
  DeepSeek
• 5:58
  V3
• 5:59
  architecture
• 5:59
  model
• 5:59
  with
• 6:00
  64
• 6:00
  experts
• 6:01
  from
• 6:01
  scratch
• 6:01
  on
• 6:01
  a
• 6:01
  single
• 6:02
  RTX
• 6:02
  PRO
• 6:03
  6000
• 6:03
  Blackwell,
• 6:04
  using
• 6:04
  GUM+muon
• 6:05
  optimization
• 6:06
  and
• 6:06
  ZeRO
• 6:06
  Stage
• 6:06
  2
• 6:07
  offloading
• 6:07
  at
• 6:07
  approximately
• 6:08
  80
• 6:08
  GB
• 6:08
  VRAM.
• 6:09
  At
• 6:09
  15,000
• 6:10
  steps
• 6:10
  the
• 6:10
  model
• 6:11
  is
• 6:11
  still
• 6:11
  early
• 6:11
  —
• 6:12
  it
• 6:12
  thinks
• 6:12
  the
• 6:12
  capital
• 6:12
  of
• 6:12
  France
• 6:13
  is
• 6:13
  Nice
• 6:13
  and
• 6:13
  cannot
• 6:14
  answer
• 6:14
  2+2
• 6:14
  —
• 6:15
  but
• 6:15
  the
• 6:15
  architecture
• 6:15
  config
• 6:16
  (MLA
• 6:16
  with
• 6:16
  d
• 6:16
  model
• 6:17
  1408,
• 6:18
  64
• 6:18
  experts
• 6:19
  with
• 6:19
  top-4
• 6:19
  routing,
• 6:20
  DOLMA/RedPajama
• 6:21
  data
• 6:21
  mix,
• 6:22
  chinchilla-optimal
• 6:23
  280B
• 6:24
  training
• 6:24
  tokens)
• 6:24
  is
• 6:24
  fully
• 6:25
  open
• 6:25
  and
• 6:25
  the
• 6:25
  developer
• 6:26
  intends
• 6:26
  to
• 6:26
  release
• 6:26
  under
• 6:27
  a
• 6:27
  copyleft-style
• 6:27
  license
• 6:28
  requiring
• 6:28
  all
• 6:28
  derivative
• 6:29
  models
• 6:29
  be
• 6:29
  public.
• 6:30
  The
• 6:30
  z-lab/Qwen3.6-27B-DFlash
• 6:35
  model
• 6:35
  has
• 6:35
  also
• 6:36
  appeared
• 6:36
  on
• 6:36
  HuggingFace's
• 6:37
  trending
• 6:37
  page,
• 6:38
  representing
• 6:38
  the
• 6:38
  DFlash
• 6:39
  speculative-decoding
• 6:40
  variant
• 6:40
  of
• 6:40
  Qwen's
• 6:40
  27B
• 6:41
  model.
• 6:45
  Sebastián
• 6:45
  Ramírez
• 6:46
  (tiangolo,
• 6:46
  creator
• 6:47
  of
• 6:47
  FastAPI)
• 6:47
  has
• 6:47
  shipped
• 6:48
  library-skills
• 6:48
  —
• 6:49
  a
• 6:49
  CLI
• 6:50
  tool
• 6:50
  that
• 6:50
  scans
• 6:51
  your
• 6:51
  project's
• 6:51
  dependencies,
• 6:52
  finds
• 6:52
  libraries
• 6:53
  that
• 6:53
  embed
• 6:53
  their
• 6:53
  own
• 6:53
  AI
• 6:54
  skills
• 6:54
  via
• 6:54
  the
• 6:54
  agentskills.io
• 6:56
  standard,
• 6:56
  and
• 6:56
  installs
• 6:57
  them
• 6:57
  as
• 6:57
  symbolic
• 6:57
  links
• 6:58
  in
• 6:58
  your
• 6:58
  .agents
• 6:59
  directory.
• 7:00
  The
• 7:00
  key
• 7:00
  insight:
• 7:01
  skills
• 7:01
  symlinked
• 7:01
  to
• 7:01
  the
• 7:01
  installed
• 7:02
  library
• 7:02
  version
• 7:03
  update
• 7:03
  automatically
• 7:03
  when
• 7:04
  you
• 7:04
  pip
• 7:04
  install
• 7:04
  --upgrade,
• 7:05
  eliminating
• 7:06
  the
• 7:06
  stale-knowledge
• 7:06
  problem
• 7:07
  that
• 7:07
  plagues
• 7:07
  agent
• 7:08
  coding
• 7:08
  assistants
• 7:08
  trained
• 7:08
  on
• 7:09
  old
• 7:09
  API
• 7:09
  patterns.
• 7:10
  Libraries
• 7:11
  opting
• 7:11
  in
• 7:11
  include
• 7:11
  FastAPI
• 7:12
  and
• 7:12
  Streamlit.
• 7:13
  A
• 7:13
  --claude
• 7:13
  flag
• 7:14
  handles
• 7:14
  Claude
• 7:14
  Code's
• 7:15
  non-standard
• 7:15
  skill
• 7:15
  directory.
• 7:17
  Coming
• 7:17
  at
• 7:17
  the
• 7:17
  same
• 7:17
  problem
• 7:18
  from
• 7:18
  a
• 7:18
  distribution
• 7:18
  angle,
• 7:19
  taito
• 7:19
  is
• 7:19
  a
• 7:19
  Go-based
• 7:20
  package
• 7:20
  manager
• 7:20
  for
• 7:21
  AI
• 7:21
  skill
• 7:21
  and
• 7:21
  agent
• 7:22
  bundles
• 7:22
  that
• 7:22
  packages
• 7:23
  them
• 7:23
  as
• 7:23
  OCI
• 7:23
  artifacts
• 7:24
  —
• 7:24
  the
• 7:25
  same
• 7:25
  container-image
• 7:25
  format
• 7:26
  used
• 7:26
  by
• 7:26
  Docker
• 7:26
  registries.
• 7:27
  Install
• 7:28
  via
• 7:28
  taito
• 7:28
  install
• 7:29
  github.com/anthropics/skills,
• 7:33
  update
• 7:34
  via
• 7:34
  taito
• 7:34
  update,
• 7:35
  package
• 7:35
  your
• 7:36
  own
• 7:36
  via
• 7:36
  taito
• 7:36
  package
• 7:37
  your.registry/namespace/artifact:tag.
• 7:41
  The
• 7:41
  OCI
• 7:41
  approach
• 7:42
  means
• 7:42
  existing
• 7:42
  container
• 7:43
  registries
• 7:43
  (GitHub
• 7:44
  Packages,
• 7:45
  Docker
• 7:45
  Hub,
• 7:46
  private
• 7:46
  registries)
• 7:47
  become
• 7:47
  skill
• 7:47
  distribution
• 7:48
  infrastructure
• 7:48
  without
• 7:49
  new
• 7:49
  tooling.
• 7:50
  On
• 7:50
  the
• 7:50
  open-source-your-own-agent
• 7:51
  front,
• 7:52
  a
• 7:52
  developer
• 7:52
  released
• 7:52
  nanoclaude
• 7:53
  —
• 7:53
  a
• 7:53
  from-scratch
• 7:54
  reimplementation
• 7:54
  of
• 7:55
  Claude
• 7:55
  Code's
• 7:55
  tool
• 7:55
  loop
• 7:56
  —
• 7:56
  accompanied
• 7:57
  by
• 7:57
  a
• 7:57
  walkthrough
• 7:57
  video.
• 7:58
  The
• 7:58
  community's
• 7:59
  main
• 7:59
  feedback
• 7:59
  was
• 7:59
  pragmatic:
• 8:00
  watch
• 8:01
  the
• 8:01
  trademark
• 8:01
  (Anthropic
• 8:02
  has
• 8:02
  enforced
• 8:02
  against
• 8:02
  similar
• 8:03
  names
• 8:03
  before),
• 8:04
  consider
• 8:04
  writing
• 8:05
  the
• 8:05
  CLI
• 8:05
  in
• 8:05
  a
• 8:05
  compiled
• 8:06
  language
• 8:06
  for
• 8:06
  performance,
• 8:07
  and
• 8:07
  note
• 8:07
  that
• 8:08
  existing
• 8:08
  projects
• 8:09
  like
• 8:09
  OpenCode
• 8:09
  and
• 8:09
  npcsh
• 8:10
  already
• 8:11
  occupy
• 8:11
  this
• 8:11
  space.
• 8:12
  The
• 8:12
  value
• 8:13
  proposition
• 8:13
  remains
• 8:14
  "building
• 8:14
  from
• 8:14
  scratch
• 8:15
  is
• 8:15
  the
• 8:15
  best
• 8:15
  way
• 8:16
  to
• 8:16
  understand
• 8:16
  what's
• 8:16
  going
• 8:17
  on
• 8:17
  under
• 8:17
  the
• 8:17
  hood."
• 8:18
  Jonathan
• 8:18
  McGuinness's
• 8:19
  "Graveyard
• 8:19
  Folder"
• 8:20
  essay
• 8:20
  provides
• 8:21
  the
• 8:21
  meta-commentary
• 8:22
  this
• 8:22
  ecosystem
• 8:22
  needs.
• 8:23
  The
• 8:24
  thesis:
• 8:25
  every
• 8:25
  command,
• 8:26
  hook,
• 8:26
  and
• 8:26
  skill
• 8:26
  you
• 8:27
  build
• 8:27
  is
• 8:27
  a
• 8:27
  small
• 8:28
  bet
• 8:28
  on
• 8:28
  your
• 8:28
  future
• 8:28
  self,
• 8:29
  and
• 8:29
  keeping
• 8:30
  it
• 8:30
  costs
• 8:30
  almost
• 8:30
  nothing
• 8:31
  today
• 8:31
  —
• 8:32
  so
• 8:32
  you
• 8:32
  keep
• 8:33
  all
• 8:33
  of
• 8:33
  them
• 8:33
  until
• 8:33
  you
• 8:33
  hit
• 8:33
  the
• 8:34
  context-window
• 8:34
  cap
• 8:35
  or
• 8:35
  the
• 8:35
  tool
• 8:35
  limit.
• 8:36
  His
• 8:36
  two-rule
• 8:37
  fix:
• 8:38
  a
• 8:38
  gate
• 8:38
  on
• 8:38
  the
• 8:38
  way
• 8:38
  in
• 8:38
  (don't
• 8:39
  build
• 8:39
  until
• 8:39
  you've
• 8:39
  done
• 8:40
  the
• 8:40
  task
• 8:40
  manually
• 8:40
  three
• 8:41
  times),
• 8:42
  and
• 8:42
  a
• 8:42
  log
• 8:42
  on
• 8:42
  the
• 8:42
  way
• 8:42
  out
• 8:43
  (anything
• 8:43
  unused
• 8:43
  for
• 8:43
  90
• 8:44
  days
• 8:44
  is
• 8:44
  a
• 8:44
  delete,
• 8:45
  not
• 8:45
  a
• 8:45
  maybe).
• 8:46
  The
• 8:47
  essay
• 8:47
  resonates
• 8:47
  because
• 8:48
  the
• 8:48
  agent-skills
• 8:48
  ecosystem
• 8:49
  is
• 8:49
  still
• 8:49
  in
• 8:49
  its
• 8:49
  "full
• 8:50
  folder"
• 8:50
  phase
• 8:50
  —
• 8:51
  Trail
• 8:52
  of
• 8:52
  Bits
• 8:52
  at
• 8:52
  94
• 8:52
  plugins
• 8:53
  and
• 8:53
  201
• 8:54
  skills
• 8:54
  is
• 8:54
  impressive
• 8:55
  until
• 8:55
  you
• 8:55
  ask
• 8:55
  how
• 8:55
  many
• 8:56
  run
• 8:56
  weekly.
• 8:59
  Anthropic
• 9:00
  launched
• 9:00
  Claude
• 9:00
  for
• 9:00
  Small
• 9:01
  Business
• 9:01
  —
• 9:02
  a
• 9:02
  toggle-on
• 9:02
  package
• 9:02
  inside
• 9:03
  Claude
• 9:03
  Cowork
• 9:04
  that
• 9:04
  connects
• 9:04
  Claude
• 9:04
  to
• 9:04
  QuickBooks,
• 9:05
  PayPal,
• 9:05
  HubSpot,
• 9:06
  Canva,
• 9:06
  DocuSign,
• 9:07
  Google
• 9:07
  Workspace,
• 9:08
  and
• 9:08
  Microsoft
• 9:08
  365
• 9:09
  with
• 9:10
  pre-built
• 9:10
  workflows
• 9:10
  for
• 9:10
  payroll
• 9:11
  planning,
• 9:11
  month-end
• 9:12
  close,
• 9:12
  invoice
• 9:13
  chasing,
• 9:13
  campaign
• 9:13
  attribution,
• 9:14
  and
• 9:14
  contract
• 9:15
  management.
• 9:16
  The
• 9:16
  framing
• 9:16
  is
• 9:16
  explicitly
• 9:17
  public-benefit:
• 9:17
  small
• 9:18
  businesses
• 9:18
  account
• 9:18
  for
• 9:19
  44%
• 9:19
  of
• 9:19
  U.S.
• 9:20
  GDP
• 9:20
  but
• 9:20
  "their
• 9:20
  adoption
• 9:21
  of
• 9:21
  AI
• 9:21
  has
• 9:21
  lagged
• 9:22
  behind
• 9:22
  larger
• 9:22
  enterprises."
• 9:23
  Alongside
• 9:24
  the
• 9:24
  product,
• 9:24
  Anthropic
• 9:25
  partnered
• 9:25
  with
• 9:25
  Coursera
• 9:26
  on
• 9:26
  a
• 9:26
  free
• 9:26
  AI
• 9:27
  fluency
• 9:27
  course
• 9:28
  and
• 9:28
  is
• 9:28
  running
• 9:28
  a
• 9:28
  10-city
• 9:28
  workshop
• 9:29
  tour
• 9:29
  (starting
• 9:29
  Chicago
• 9:30
  May
• 9:30
  14)
• 9:30
  offering
• 9:31
  one-month
• 9:31
  Claude
• 9:31
  Max
• 9:32
  subscriptions
• 9:32
  to
• 9:32
  attendees.
• 9:33
  Security
• 9:33
  positioning:
• 9:34
  human-in-the-loop
• 9:35
  approval
• 9:35
  on
• 9:35
  all
• 9:36
  actions,
• 9:36
  role-based
• 9:37
  access
• 9:37
  mirroring
• 9:38
  existing
• 9:38
  tool
• 9:38
  permissions,
• 9:39
  no
• 9:39
  training
• 9:39
  on
• 9:40
  business
• 9:40
  data
• 9:40
  by
• 9:40
  default
• 9:41
  on
• 9:41
  Team
• 9:41
  and
• 9:41
  Enterprise
• 9:41
  plans.
• 9:42
  This
• 9:43
  represents
• 9:43
  a
• 9:43
  notable
• 9:44
  pivot
• 9:44
  from
• 9:44
  Anthropic's
• 9:44
  established
• 9:45
  enterprise
• 9:45
  and
• 9:45
  government
• 9:46
  focus
• 9:46
  —
• 9:47
  the
• 9:47
  same
• 9:47
  company
• 9:48
  that
• 9:48
  reported
• 9:48
  roughly
• 9:48
  $30B
• 9:49
  annualized
• 9:50
  revenue
• 9:50
  with
• 9:50
  eight
• 9:50
  Fortune
• 9:51
  10
• 9:51
  customers
• 9:51
  is
• 9:52
  now
• 9:52
  courting
• 9:52
  solopreneurs
• 9:53
  through
• 9:53
  CDFI
• 9:53
  partnerships
• 9:54
  and
• 9:54
  seed-funding
• 9:55
  accelerators.
• 9:56
  Whether
• 9:56
  this
• 9:56
  is
• 9:56
  mission-driven
• 9:57
  equity
• 9:57
  or
• 9:57
  market-expansion
• 9:58
  strategy
• 9:59
  (or
• 9:59
  both)
• 9:59
  will
• 9:59
  depend
• 9:59
  on
• 9:59
  pricing
• 10:00
  and
• 10:00
  retention
• 10:01
  once
• 10:01
  the
• 10:01
  free
• 10:01
  Max
• 10:02
  month
• 10:02
  expires.
• 10:03
  Dario
• 10:03
  Amodei,
• 10:04
  in
• 10:04
  a
• 10:04
  recent
• 10:05
  extended
• 10:05
  conversation,
• 10:06
  reinforced
• 10:07
  the
• 10:07
  tension
• 10:07
  at
• 10:07
  the
• 10:07
  heart
• 10:07
  of
• 10:07
  Anthropic's
• 10:08
  positioning:
• 10:09
  worry
• 10:09
  about
• 10:09
  "economics
• 10:10
  and
• 10:10
  the
• 10:10
  concentration
• 10:11
  of
• 10:11
  power"
• 10:11
  even
• 10:11
  as
• 10:11
  the
• 10:11
  company
• 10:12
  builds
• 10:12
  the
• 10:12
  most
• 10:12
  concentrated
• 10:13
  AI
• 10:13
  capabilities
• 10:14
  available.
• 10:15
  Google
• 10:15
  DeepMind,
• 10:16
  meanwhile,
• 10:16
  is
• 10:17
  rethinking
• 10:17
  the
• 10:17
  most
• 10:17
  fundamental
• 10:18
  piece
• 10:18
  of
• 10:18
  the
• 10:18
  desktop
• 10:19
  UI:
• 10:19
  the
• 10:19
  mouse
• 10:20
  pointer.
• 10:21
  Their
• 10:21
  research
• 10:21
  proposes
• 10:22
  an
• 10:22
  AI-enabled
• 10:22
  cursor
• 10:23
  that
• 10:23
  understands
• 10:24
  not
• 10:24
  just
• 10:24
  where
• 10:24
  you're
• 10:24
  pointing
• 10:25
  but
• 10:25
  what
• 10:25
  you're
• 10:25
  pointing
• 10:26
  at
• 10:26
  and
• 10:26
  why
• 10:26
  it
• 10:26
  matters
• 10:27
  —
• 10:27
  transforming
• 10:28
  pixels
• 10:28
  into
• 10:29
  structured
• 10:29
  entities
• 10:29
  (places,
• 10:30
  dates,
• 10:31
  objects)
• 10:31
  that
• 10:31
  can
• 10:31
  be
• 10:31
  acted
• 10:32
  upon
• 10:32
  with
• 10:32
  natural-language
• 10:33
  shorthand
• 10:33
  like
• 10:33
  "fix
• 10:34
  this"
• 10:34
  or
• 10:34
  "show
• 10:34
  me
• 10:34
  directions."
• 10:35
  Four
• 10:36
  design
• 10:36
  principles
• 10:36
  guide
• 10:37
  the
• 10:37
  work:
• 10:38
  AI
• 10:38
  should
• 10:38
  work
• 10:39
  across
• 10:39
  all
• 10:39
  apps
• 10:40
  (not
• 10:40
  force
• 10:40
  users
• 10:40
  into
• 10:41
  "AI
• 10:41
  detours"),
• 10:42
  capture
• 10:42
  visual
• 10:43
  and
• 10:43
  semantic
• 10:43
  context
• 10:44
  automatically,
• 10:45
  embrace
• 10:45
  deictic
• 10:46
  reference
• 10:46
  ("this"
• 10:46
  and
• 10:47
  "that"),
• 10:47
  and
• 10:47
  treat
• 10:48
  pixels
• 10:48
  as
• 10:48
  actionable
• 10:49
  entities
• 10:49
  rather
• 10:49
  than
• 10:50
  inert
• 10:50
  coordinates.
• 10:51
  Experimental
• 10:52
  demos
• 10:52
  are
• 10:52
  live
• 10:52
  in
• 10:52
  Google
• 10:53
  AI
• 10:53
  Studio,
• 10:53
  with
• 10:54
  integration
• 10:54
  coming
• 10:54
  to
• 10:54
  Chrome
• 10:55
  and
• 10:55
  Chromebook.
• 10:58
  If
• 10:58
  you
• 10:59
  run
• 10:59
  a
• 10:59
  forwarding
• 10:59
  resolver
• 11:00
  like
• 11:00
  Pi-hole
• 11:00
  or
• 11:00
  AdGuard
• 11:01
  Home,
• 11:02
  every
• 11:02
  one
• 11:02
  of
• 11:02
  your
• 11:02
  queries
• 11:02
  goes
• 11:03
  through
• 11:03
  one
• 11:03
  operator
• 11:03
  who
• 11:04
  sees
• 11:04
  both
• 11:04
  your
• 11:04
  IP
• 11:04
  and
• 11:05
  the
• 11:05
  question.
• 11:06
  Recursive
• 11:06
  resolvers
• 11:07
  like
• 11:07
  Unbound
• 11:07
  flip
• 11:07
  the
• 11:07
  problem
• 11:08
  —
• 11:08
  your
• 11:09
  IP
• 11:09
  gets
• 11:09
  exposed
• 11:10
  to
• 11:10
  every
• 11:10
  authoritative
• 11:11
  nameserver
• 11:11
  instead.
• 11:12
  DoH
• 11:13
  and
• 11:13
  DoT
• 11:13
  encrypt
• 11:13
  the
• 11:13
  wire
• 11:14
  but
• 11:14
  don't
• 11:14
  change
• 11:14
  who
• 11:14
  sees
• 11:15
  what.
• 11:16
  Oblivious
• 11:16
  DNS-over-HTTPS
• 11:17
  (ODoH,
• 11:18
  RFC
• 11:19
  9230)
• 11:20
  splits
• 11:20
  the
• 11:20
  resolution
• 11:21
  path
• 11:21
  so
• 11:21
  that
• 11:22
  a
• 11:22
  relay
• 11:22
  sees
• 11:22
  your
• 11:22
  IP
• 11:23
  but
• 11:23
  only
• 11:23
  ciphertext,
• 11:24
  while
• 11:24
  the
• 11:24
  target
• 11:25
  resolver
• 11:25
  sees
• 11:25
  the
• 11:25
  plaintext
• 11:26
  query
• 11:26
  but
• 11:26
  only
• 11:26
  the
• 11:26
  relay's
• 11:27
  IP
• 11:27
  —
• 11:27
  neither
• 11:28
  party
• 11:28
  learns
• 11:28
  both.
• 11:29
  The
• 11:29
  existing
• 11:29
  public
• 11:30
  relay
• 11:30
  ecosystem
• 11:30
  consists
• 11:31
  of
• 11:31
  exactly
• 11:31
  one
• 11:31
  operator.
• 11:33
  Numa
• 11:33
  v0.14
• 11:35
  adds
• 11:35
  the
• 11:36
  second:
• 11:36
  a
• 11:36
  Rust
• 11:37
  binary
• 11:37
  shipping
• 11:38
  client,
• 11:38
  relay,
• 11:39
  and
• 11:39
  public
• 11:39
  deployment,
• 11:40
  paired
• 11:41
  by
• 11:41
  default
• 11:41
  with
• 11:41
  Cloudflare's
• 11:42
  target
• 11:42
  to
• 11:42
  ensure
• 11:42
  two
• 11:43
  independent
• 11:43
  operators
• 11:43
  in
• 11:44
  the
• 11:44
  path.
• 11:45
  The
• 11:45
  relay's
• 11:45
  SSRF-prevention
• 11:46
  validator
• 11:47
  is
• 11:47
  regex-strict
• 11:47
  (RFC
• 11:48
  1035
• 11:49
  ASCII
• 11:49
  labels
• 11:49
  only,
• 11:50
  no
• 11:50
  IP
• 11:51
  literals,
• 11:51
  no
• 11:52
  non-443
• 11:52
  ports),
• 11:53
  and
• 11:53
  same-operator
• 11:54
  configs
• 11:54
  are
• 11:54
  rejected
• 11:55
  by
• 11:55
  default
• 11:55
  to
• 11:55
  prevent
• 11:56
  the
• 11:56
  construction
• 11:56
  from
• 11:56
  collapsing
• 11:57
  into
• 11:57
  theatre.
• 11:58
  Honest
• 11:58
  limitations
• 11:59
  acknowledged:
• 12:00
  traffic
• 12:00
  analysis
• 12:00
  remains
• 12:01
  possible
• 12:01
  against
• 12:01
  low-volume
• 12:02
  relays
• 12:02
  (the
• 12:02
  defense
• 12:02
  is
• 12:03
  more
• 12:03
  users,
• 12:03
  not
• 12:04
  more
• 12:04
  crypto),
• 12:04
  and
• 12:04
  pubkey
• 12:05
  distribution
• 12:05
  is
• 12:05
  centralized
• 12:06
  via
• 12:06
  WebPKI.
• 12:07
  On
• 12:07
  the
• 12:07
  hardware
• 12:08
  trust
• 12:08
  side,
• 12:08
  Benn
• 12:08
  Jordan's
• 12:09
  deep-dive
• 12:09
  into
• 12:09
  Unitree
• 12:10
  robot
• 12:10
  dogs
• 12:11
  uncovered
• 12:11
  a
• 12:11
  trifecta
• 12:12
  of
• 12:12
  concerning
• 12:12
  findings:
• 12:13
  an
• 12:13
  arbitrary
• 12:13
  command
• 12:14
  execution
• 12:14
  flaw
• 12:14
  exploitable
• 12:15
  via
• 12:15
  the
• 12:15
  Wi-Fi
• 12:15
  password
• 12:16
  entry
• 12:16
  field,
• 12:16
  a
• 12:16
  year-old
• 12:17
  unpatched
• 12:17
  exploit,
• 12:18
  and
• 12:18
  "highly
• 12:18
  suspicious
• 12:19
  traffic
• 12:19
  to
• 12:19
  Chinese
• 12:20
  servers
• 12:20
  whenever
• 12:20
  the
• 12:20
  robot's
• 12:21
  software
• 12:21
  figured
• 12:21
  that
• 12:21
  it
• 12:22
  was
• 12:22
  not
• 12:22
  being
• 12:22
  watched."
• 12:23
  The
• 12:23
  Lidar
• 12:24
  placement
• 12:24
  below
• 12:24
  the
• 12:24
  head
• 12:24
  renders
• 12:25
  the
• 12:25
  robot
• 12:25
  effectively
• 12:26
  blind
• 12:26
  to
• 12:26
  its
• 12:26
  rear
• 12:26
  and
• 12:26
  surroundings
• 12:27
  —
• 12:27
  making
• 12:28
  even
• 12:28
  the
• 12:28
  basic
• 12:29
  intended
• 12:29
  use
• 12:29
  case
• 12:29
  (yard
• 12:30
  patrol)
• 12:30
  unreliable.
• 12:31
  The
• 12:31
  pattern
• 12:32
  echoes
• 12:32
  a
• 12:32
  prior
• 12:32
  DJI
• 12:33
  robot
• 12:33
  vacuum
• 12:33
  disclosure
• 12:34
  where
• 12:34
  a
• 12:34
  single
• 12:34
  authenticated
• 12:35
  credential
• 12:36
  granted
• 12:36
  fleet-wide
• 12:36
  access
• 12:37
  to
• 12:37
  nearly
• 12:37
  1,000
• 12:38
  vacuums
• 12:38
  across
• 12:39
  24
• 12:39
  countries.
• 12:40
  The
• 12:40
  uncomfortable
• 12:41
  consistency:
• 12:42
  Chinese-manufactured
• 12:43
  consumer
• 12:43
  robotics
• 12:44
  shipping
• 12:44
  with
• 12:44
  firmware
• 12:45
  update
• 12:45
  mechanisms
• 12:46
  that
• 12:46
  represent
• 12:47
  the
• 12:47
  thinnest
• 12:47
  trust
• 12:47
  boundary
• 12:48
  in
• 12:48
  the
• 12:48
  system,
• 12:49
  combined
• 12:49
  with
• 12:49
  telemetry
• 12:49
  behavior
• 12:50
  that
• 12:50
  degrades
• 12:50
  trust
• 12:51
  further.
• 12:54
  ScioMind
• 12:54
  introduces
• 12:55
  a
• 12:55
  cognitively
• 12:56
  grounded
• 12:56
  framework
• 12:57
  for
• 12:57
  LLM-based
• 12:57
  social
• 12:58
  simulation
• 12:58
  that
• 12:58
  tackles
• 12:59
  the
• 12:59
  over-smoothing
• 12:59
  problem
• 12:59
  —
• 13:00
  the
• 13:00
  tendency
• 13:00
  of
• 13:00
  multi-agent
• 13:01
  systems
• 13:01
  to
• 13:01
  converge
• 13:02
  too
• 13:02
  quickly
• 13:02
  toward
• 13:03
  neutral
• 13:03
  consensus
• 13:03
  on
• 13:04
  controversial
• 13:04
  topics.
• 13:05
  The
• 13:05
  core
• 13:05
  mechanism
• 13:06
  is
• 13:06
  a
• 13:06
  memory-anchored
• 13:06
  belief
• 13:07
  update
• 13:07
  rule
• 13:07
  where
• 13:07
  each
• 13:08
  agent's
• 13:08
  resistance
• 13:08
  to
• 13:09
  opinion
• 13:09
  change
• 13:09
  scales
• 13:09
  with
• 13:10
  personality-conditioned
• 13:10
  anchoring
• 13:11
  strength
• 13:11
  derived
• 13:11
  from
• 13:12
  Big
• 13:12
  Five
• 13:12
  traits
• 13:12
  via
• 13:13
  a
• 13:13
  sigmoid
• 13:13
  mapping
• 13:13
  (low
• 13:14
  Openness
• 13:14
  and
• 13:14
  high
• 13:14
  Conscientiousness
• 13:15
  increase
• 13:16
  anchoring).
• 13:17
  A
• 13:17
  four-layer
• 13:17
  memory
• 13:18
  architecture
• 13:18
  (episodic,
• 13:19
  semantic,
• 13:19
  reflection,
• 13:20
  working
• 13:20
  memory)
• 13:21
  provides
• 13:21
  the
• 13:21
  experiential
• 13:22
  substrate
• 13:22
  for
• 13:22
  anchor
• 13:23
  formation,
• 13:23
  and
• 13:23
  dynamic
• 13:24
  profiles
• 13:24
  are
• 13:24
  retrieved
• 13:25
  from
• 13:25
  a
• 13:25
  corpus-grounded
• 13:26
  pipeline
• 13:26
  rather
• 13:27
  than
• 13:27
  assigned
• 13:27
  as
• 13:27
  static
• 13:27
  demographic
• 13:28
  labels.
• 13:29
  Evaluated
• 13:30
  on
• 13:30
  Roe
• 13:30
  v.
• 13:31
  Wade,
• 13:32
  Australian
• 13:32
  social
• 13:32
  media
• 13:33
  ban,
• 13:33
  and
• 13:33
  U.S.
• 13:34
  presidential
• 13:34
  election
• 13:35
  scenarios,
• 13:35
  the
• 13:36
  full
• 13:36
  system
• 13:36
  maintains
• 13:37
  persistent
• 13:37
  but
• 13:38
  bounded
• 13:38
  disagreement
• 13:38
  —
• 13:39
  polarization
• 13:40
  variance
• 13:40
  around
• 13:40
  0.17,
• 13:41
  bimodality
• 13:42
  coefficient
• 13:42
  of
• 13:42
  0.68
• 13:43
  —
• 13:44
  rather
• 13:44
  than
• 13:44
  collapsing
• 13:44
  into
• 13:45
  artificial
• 13:45
  consensus.
• 13:46
  The
• 13:46
  key
• 13:46
  finding:
• 13:48
  agents
• 13:48
  with
• 13:48
  high
• 13:48
  Openness
• 13:49
  change
• 13:49
  opinions
• 13:50
  3-4x
• 13:50
  more
• 13:51
  frequently
• 13:51
  than
• 13:51
  strongly
• 13:52
  anchored
• 13:52
  agents,
• 13:53
  matching
• 13:53
  empirical
• 13:54
  observations
• 13:54
  from
• 13:54
  political
• 13:55
  psychology.
• 13:56
  MIT
• 13:57
  CSAIL's
• 13:57
  RLCR
• 13:58
  (Reinforcement
• 13:58
  Learning
• 13:59
  for
• 13:59
  Calibrated
• 13:59
  Reasoning)
• 13:59
  addresses
• 14:00
  a
• 14:00
  complementary
• 14:00
  problem:
• 14:02
  LLMs
• 14:02
  delivering
• 14:03
  every
• 14:03
  answer
• 14:03
  with
• 14:03
  equal
• 14:04
  certainty
• 14:04
  whether
• 14:04
  they're
• 14:05
  right
• 14:05
  or
• 14:05
  guessing.
• 14:06
  The
• 14:06
  method
• 14:07
  traces
• 14:07
  overconfidence
• 14:08
  to
• 14:08
  a
• 14:08
  specific
• 14:08
  flaw
• 14:09
  in
• 14:09
  how
• 14:09
  reasoning
• 14:10
  models
• 14:10
  are
• 14:10
  trained
• 14:11
  and
• 14:11
  provides
• 14:11
  a
• 14:11
  fix
• 14:12
  without
• 14:12
  sacrificing
• 14:12
  accuracy.
• 14:14
  Community
• 14:14
  reaction
• 14:14
  was
• 14:15
  mixed
• 14:15
  —
• 14:16
  one
• 14:16
  commenter
• 14:16
  noted
• 14:17
  the
• 14:17
  paper
• 14:17
  is
• 14:17
  actually
• 14:17
  from
• 14:18
  last
• 14:18
  year
• 14:18
  and
• 14:19
  questioned
• 14:19
  whether
• 14:19
  any
• 14:19
  major
• 14:20
  model
• 14:20
  has
• 14:20
  adopted
• 14:21
  it,
• 14:21
  while
• 14:21
  another
• 14:21
  argued
• 14:22
  that
• 14:22
  teaching
• 14:22
  "I'm
• 14:22
  not
• 14:23
  sure"
• 14:23
  merely
• 14:24
  shifts
• 14:24
  output
• 14:24
  toward
• 14:25
  hedge-heavy
• 14:25
  training
• 14:25
  data
• 14:26
  without
• 14:26
  genuine
• 14:26
  epistemic
• 14:27
  state.
• 14:28
  The
• 14:28
  deeper
• 14:28
  question
• 14:29
  remains
• 14:29
  whether
• 14:29
  calibrated
• 14:30
  uncertainty
• 14:30
  can
• 14:31
  be
• 14:31
  implemented
• 14:31
  without
• 14:31
  producing
• 14:32
  "hedge
• 14:32
  slop"
• 14:33
  that
• 14:33
  erodes
• 14:33
  user
• 14:34
  trust
• 14:34
  in
• 14:34
  correct
• 14:34
  answers.
• 14:36
  On
• 14:36
  the
• 14:36
  generative-media
• 14:36
  side,
• 14:37
  CDM
• 14:38
  (Continuous-Time
• 14:39
  Distribution
• 14:39
  Matching)
• 14:40
  pushes
• 14:40
  diffusion
• 14:41
  distillation
• 14:41
  to
• 14:41
  4-NFE
• 14:42
  image
• 14:42
  generation
• 14:43
  by
• 14:43
  sampling
• 14:44
  intermediate
• 14:44
  anchors
• 14:45
  uniformly
• 14:45
  from
• 14:45
  (0,
• 14:46
  1]
• 14:47
  on
• 14:47
  a
• 14:47
  dynamic
• 14:47
  continuous
• 14:48
  time
• 14:48
  schedule
• 14:49
  during
• 14:49
  backward
• 14:49
  simulation,
• 14:50
  then
• 14:50
  applying
• 14:51
  CFG
• 14:51
  augmentation
• 14:52
  and
• 14:52
  distribution
• 14:53
  matching
• 14:53
  at
• 14:53
  on-trajectory
• 14:54
  anchors
• 14:54
  while
• 14:55
  an
• 14:55
  explicit
• 14:55
  extrapolation
• 14:56
  objective
• 14:56
  handles
• 14:57
  inter-anchor
• 14:57
  inconsistency.
• 14:59
  Models
• 14:59
  are
• 14:59
  available
• 15:00
  for
• 15:00
  SD3-Medium
• 15:01
  and
• 15:01
  LongCat.
• 15:02
  A
• 15:02
  more
• 15:03
  immediately
• 15:03
  practical
• 15:03
  contribution:
• 15:05
  an
• 15:05
  open-source
• 15:05
  pipeline
• 15:06
  running
• 15:06
  on
• 15:06
  a
• 15:06
  single
• 15:07
  AMD
• 15:07
  Instinct
• 15:08
  MI300X
• 15:10
  (192
• 15:12
  GB
• 15:12
  HBM3)
• 15:13
  that
• 15:13
  takes
• 15:14
  one
• 15:14
  English
• 15:14
  sentence
• 15:14
  and
• 15:15
  produces
• 15:15
  a
• 15:15
  finished
• 15:15
  cinematic
• 15:16
  reel
• 15:16
  with
• 15:17
  characters,
• 15:17
  story,
• 15:18
  music,
• 15:18
  and
• 15:18
  multilingual
• 15:19
  voice-over
• 15:19
  in
• 15:19
  approximately
• 15:20
  10
• 15:20
  minutes.
• 15:21
  The
• 15:21
  eight-stage
• 15:22
  pipeline
• 15:22
  uses
• 15:22
  Qwen3.5-35B-A3B
• 15:27
  as
• 15:27
  director
• 15:28
  agent,
• 15:28
  FLUX.2
• 15:29
  for
• 15:29
  character
• 15:30
  keyframes
• 15:30
  with
• 15:30
  reference-image
• 15:31
  pinning
• 15:31
  (no
• 15:32
  LoRA
• 15:32
  training),
• 15:34
  Wan2.2-I2V-A14B
• 15:36
  for
• 15:36
  animation
• 15:37
  at
• 15:37
  1280x720/81
• 15:40
  frames/16fps,
• 15:44
  a
• 15:44
  vision
• 15:44
  critic
• 15:44
  with
• 15:45
  10
• 15:45
  structured
• 15:45
  failure
• 15:46
  labels
• 15:46
  and
• 15:46
  targeted
• 15:47
  retry
• 15:47
  strategies,
• 15:48
  ACE-Step
• 15:48
  for
• 15:49
  music,
• 15:49
  and
• 15:49
  Kokoro-82M
• 15:50
  for
• 15:51
  narration
• 15:51
  in
• 15:51
  9
• 15:51
  languages.
• 15:53
  Performance
• 15:53
  work
• 15:53
  (ParaAttention
• 15:54
  FBCache
• 15:55
  for
• 15:55
  lossless
• 15:55
  2x
• 15:56
  on
• 15:56
  Wan2.2,
• 15:57
  torch.compile
• 15:58
  on
• 15:59
  the
• 15:59
  transformer,
• 16:00
  AITER
• 16:00
  MoE
• 16:00
  acceleration)
• 16:01
  cut
• 16:01
  end-to-end
• 16:02
  time
• 16:02
  from
• 16:02
  25.9
• 16:03
  to
• 16:03
  10.4
• 16:04
  minutes
• 16:04
  per
• 16:04
  720p
• 16:05
  clip.
• 16:06
  HumeAI's
• 16:07
  tada-3b-ml
• 16:08
  has
• 16:08
  appeared
• 16:08
  on
• 16:08
  HuggingFace's
• 16:09
  trending
• 16:09
  models
• 16:10
  page,
• 16:10
  representing
• 16:11
  a
• 16:11
  3-billion-parameter
• 16:12
  multilingual
• 16:12
  expressive
• 16:13
  speech
• 16:13
  model
• 16:13
  from
• 16:14
  the
• 16:14
  emotion-AI
• 16:14
  company.