Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – September 25, 2026

Transcript

AGI Dreams Podcast – September 25, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  real
• 0:00
  cost
• 0:00
  of
• 0:01
  LLM-driven
• 0:01
  security
• 0:02
  scanning
• 0:02
  was
• 0:02
  never
• 0:03
  capability,
• 0:04
  it
• 0:04
  was
• 0:04
  noise:
• 0:05
  unverified
• 0:05
  hypotheses
• 0:06
  dressed
• 0:06
  up
• 0:07
  as
• 0:07
  vulnerability
• 0:07
  reports.
• 0:08
  Google's
• 0:09
  PageBreak,
• 0:09
  an
• 0:10
  internal
• 0:10
  Product
• 0:10
  Security
• 0:11
  agent
• 0:11
  piloted
• 0:12
  in
• 0:12
  November
• 0:12
  2025
• 0:13
  and
• 0:13
  a
• 0:13
  full
• 0:14
  project
• 0:14
  since
• 0:14
  January
• 0:15
  2026,
• 0:16
  attacks
• 0:16
  exactly
• 0:17
  that.
• 0:17
  Running
• 0:18
  mostly
• 0:18
  on
• 0:18
  Gemini
• 0:19
  models
• 0:19
  (3.1
• 0:20
  Pro
• 0:20
  and
• 0:20
  3.5
• 0:20
  Flash),
• 0:21
  it
• 0:21
  has
• 0:21
  uncovered
• 0:22
  500-plus
• 0:22
  cross-site
• 0:23
  scripting
• 0:23
  (XSS)
• 0:24
  vulnerabilities
• 0:24
  across
• 0:25
  Google's
• 0:25
  first-party
• 0:26
  web
• 0:26
  applications,
• 0:27
  including
• 0:27
  sensitive
• 0:27
  domains
• 0:28
  —
• 0:28
  and
• 0:28
  what
• 0:28
  matters
• 0:29
  is
• 0:29
  the
• 0:29
  false-positive
• 0:30
  rate
• 0:30
  behind
• 0:30
  it:
• 0:31
  near
• 0:31
  zero,
• 0:32
  because
• 0:32
  nothing
• 0:32
  reaches
• 0:33
  a
• 0:33
  product
• 0:33
  team
• 0:33
  until
• 0:33
  a
• 0:33
  validator
• 0:34
  has
• 0:34
  actually
• 0:35
  exploited
• 0:35
  the
• 0:35
  flaw.
• 0:36
  When
• 0:36
  the
• 0:36
  agent
• 0:37
  hypothesizes
• 0:37
  a
• 0:38
  bug,
• 0:38
  specialized,
• 0:39
  non-AI-written
• 0:40
  validators
• 0:40
  execute
• 0:41
  real
• 0:41
  payloads
• 0:42
  against
• 0:42
  a
• 0:42
  running
• 0:42
  environment
• 0:43
  —
• 0:44
  injecting
• 0:44
  JavaScript
• 0:45
  and
• 0:45
  monitoring
• 0:45
  whether
• 0:46
  it
• 0:46
  runs,
• 0:46
  manipulating
• 0:47
  database
• 0:47
  queries,
• 0:48
  creating
• 0:48
  world-readable
• 0:49
  files,
• 0:49
  triggering
• 0:50
  outbound
• 0:50
  DNS
• 0:51
  or
• 0:51
  sleep-delay
• 0:52
  callbacks.
• 0:53
  Non-deterministic
• 0:54
  findings
• 0:54
  stay
• 0:54
  internal
• 0:55
  as
• 0:55
  seeds;
• 0:56
  agents
• 0:56
  run
• 0:57
  with
• 0:57
  identical
• 0:57
  seeds
• 0:58
  across
• 0:58
  many
• 0:58
  iterations
• 0:59
  to
• 0:59
  land
• 0:59
  a
• 0:59
  correct
• 1:00
  exploit
• 1:00
  trajectory.
• 1:02
  Against
• 1:02
  hundreds
• 1:03
  of
• 1:03
  apps
• 1:03
  built
• 1:03
  on
• 1:03
  the
• 1:03
  high-assurance
• 1:04
  framework
• 1:04
  from
• 1:05
  Google's
• 1:05
  2025
• 1:06
  Secure
• 1:06
  by
• 1:06
  Design
• 1:06
  blueprint,
• 1:07
  PageBreak
• 1:08
  found
• 1:08
  only
• 1:08
  2
• 1:08
  XSS
• 1:09
  as
• 1:09
  of
• 1:09
  September
• 1:09
  4,
• 1:10
  2026
• 1:10
  —
• 1:12
  both
• 1:12
  on
• 1:12
  internal
• 1:13
  apps
• 1:13
  or
• 1:13
  debug
• 1:14
  endpoints
• 1:14
  .
• 1:15
  Sakana
• 1:15
  AI's
• 1:16
  Fugu-Cyber
• 1:16
  lands
• 1:17
  on
• 1:17
  the
• 1:17
  same
• 1:17
  conclusion
• 1:18
  from
• 1:18
  the
• 1:18
  vendor
• 1:18
  side:
• 1:19
  a
• 1:19
  multi-agent
• 1:20
  system
• 1:20
  —
• 1:21
  a
• 1:21
  pool
• 1:21
  of
• 1:21
  specialized
• 1:22
  agents
• 1:22
  behind
• 1:22
  one
• 1:23
  endpoint,
• 1:23
  behaving
• 1:24
  like
• 1:24
  a
• 1:24
  single
• 1:24
  model
• 1:24
  —
• 1:25
  posting
• 1:26
  86.9%
• 1:27
  on
• 1:27
  CyberGym,
• 1:28
  which
• 1:28
  verifies
• 1:29
  real-world
• 1:29
  vulnerabilities
• 1:30
  in
• 1:30
  codebases,
• 1:31
  and
• 1:31
  72.1%
• 1:32
  on
• 1:33
  CTI-REALM,
• 1:36
  which
• 1:36
  turns
• 1:37
  threat-intelligence
• 1:37
  reports
• 1:38
  into
• 1:38
  working
• 1:38
  detection
• 1:39
  rules,
• 1:39
  comparable
• 1:40
  to
• 1:40
  GPT-5.5-Cyber
• 1:41
  and
• 1:42
  Mythos-Preview.
• 1:43
  The
• 1:43
  more
• 1:43
  interesting
• 1:44
  half
• 1:44
  of
• 1:44
  the
• 1:44
  release
• 1:45
  is
• 1:45
  the
• 1:45
  reality
• 1:45
  check
• 1:45
  around
• 1:46
  it:
• 1:46
  citing
• 1:47
  a
• 1:47
  Nikkei
• 1:47
  Digital
• 1:47
  Governance
• 1:48
  report,
• 1:49
  Sakana
• 1:49
  notes
• 1:49
  that
• 1:49
  major
• 1:50
  financial
• 1:50
  institutions
• 1:51
  with
• 1:51
  frontier-model
• 1:52
  access
• 1:52
  still
• 1:53
  fail
• 1:53
  to
• 1:53
  operationalize
• 1:54
  it
• 1:54
  —
• 1:54
  raw
• 1:55
  models
• 1:55
  generate
• 1:55
  false
• 1:56
  positives
• 1:56
  without
• 1:57
  specialized
• 1:57
  harnesses,
• 1:58
  internal
• 1:58
  talent,
• 1:59
  and
• 1:59
  human-in-the-loop
• 2:00
  validation.
• 2:01
  That
• 2:01
  is
• 2:02
  PageBreak's
• 2:02
  architecture
• 2:03
  sold
• 2:03
  as
• 2:03
  a
• 2:03
  product
• 2:03
  .
• 2:06
  Attackers
• 2:06
  skipped
• 2:07
  the
• 2:07
  consensus-building
• 2:08
  phase.
• 2:09
  ThreatDown's
• 2:09
  analysis
• 2:09
  of
• 2:10
  CARBONATO,
• 2:13
  a
• 2:13
  Docker
• 2:14
  botnet,
• 2:14
  documents
• 2:15
  malware
• 2:15
  embedding
• 2:15
  a
• 2:15
  working
• 2:16
  AI
• 2:16
  agent.
• 2:17
  A
• 2:17
  day
• 2:17
  of
• 2:17
  passive
• 2:18
  collection
• 2:18
  on
• 2:18
  an
• 2:18
  unauthenticated
• 2:19
  Docker
• 2:19
  registry
• 2:20
  found
• 2:20
  in
• 2:20
  August
• 2:20
  2026
• 2:21
  pulled
• 2:21
  59
• 2:22
  repositories,
• 2:23
  234
• 2:24
  image
• 2:24
  tags,
• 2:24
  and
• 2:25
  4.3
• 2:25
  GB
• 2:25
  —
• 2:26
  a
• 2:26
  trojanized
• 2:27
  wallet
• 2:27
  operation
• 2:28
  and
• 2:28
  the
• 2:28
  botnet
• 2:28
  itself.
• 2:29
  CARBONATO
• 2:32
  worms
• 2:33
  through
• 2:33
  Docker
• 2:33
  daemons
• 2:33
  accepting
• 2:34
  unauthenticated
• 2:35
  connections
• 2:35
  on
• 2:35
  port
• 2:35
  2375
• 2:36
  —
• 2:37
  thousands
• 2:37
  remain
• 2:38
  publicly
• 2:38
  reachable
• 2:38
  —
• 2:39
  launching
• 2:39
  a
• 2:39
  privileged
• 2:40
  container
• 2:40
  with
• 2:40
  the
• 2:40
  host
• 2:41
  filesystem
• 2:41
  mounted,
• 2:42
  running
• 2:42
  host
• 2:42
  commands
• 2:43
  via
• 2:43
  nsenter,
• 2:44
  persisting
• 2:44
  via
• 2:44
  cron,
• 2:45
  systemd
• 2:45
  timers,
• 2:46
  and
• 2:46
  OpenRC
• 2:46
  under
• 2:47
  immutable
• 2:47
  bits,
• 2:48
  and
• 2:48
  hiding
• 2:48
  an
• 2:48
  XMRig
• 2:49
  miner
• 2:49
  at
• 2:49
  /usr/sbin/systemd-logind
• 2:51
  .
• 2:53
  The
• 2:53
  AI
• 2:53
  core
• 2:53
  deserves
• 2:54
  a
• 2:54
  second
• 2:54
  read:
• 2:54
  the
• 2:55
  implant
• 2:55
  installs
• 2:55
  Hermes
• 2:56
  Agent
• 2:56
  —
• 2:57
  an
• 2:57
  MIT-licensed
• 2:58
  open-source
• 2:58
  framework
• 2:59
  from
• 2:59
  Nous
• 2:59
  Research
• 2:59
  —
• 3:00
  unchanged,
• 3:01
  then
• 3:01
  overwrites
• 3:02
  its
• 3:02
  persona
• 3:02
  file
• 3:02
  with
• 3:03
  a
• 3:03
  39-line
• 3:03
  prompt:
• 3:04
  "You
• 3:04
  are
• 3:04
  GH0ST
• 3:05
  —
• 3:07
  senior
• 3:07
  hacker,
• 3:08
  pentester
• 3:08
  and
• 3:08
  exploit
• 3:09
  developer...
• 3:10
  You
• 3:10
  are
• 3:10
  not
• 3:10
  an
• 3:10
  assistant;
• 3:11
  you
• 3:11
  are
• 3:11
  a
• 3:11
  living
• 3:12
  post-exploitation
• 3:13
  tool."
• 3:14
  Tasks
• 3:14
  arrive
• 3:14
  via
• 3:15
  Telegram
• 3:15
  and
• 3:15
  route
• 3:15
  through
• 3:15
  the
• 3:15
  crew's
• 3:16
  LLM
• 3:16
  gateway
• 3:17
  serving
• 3:17
  27
• 3:17
  models
• 3:18
  —
• 3:19
  the
• 3:19
  model
• 3:19
  writes
• 3:19
  commands,
• 3:20
  reads
• 3:20
  output,
• 3:21
  decides
• 3:21
  next
• 3:21
  steps.
• 3:22
  The
• 3:23
  loot
• 3:23
  priority
• 3:23
  is
• 3:23
  explicit:
• 3:24
  AI
• 3:25
  API
• 3:25
  keys
• 3:25
  are
• 3:26
  "loot
• 3:26
  #1
• 3:26
  —
• 3:27
  above
• 3:27
  SSH
• 3:28
  credentials,
• 3:28
  above
• 3:29
  access
• 3:29
  tokens,
• 3:30
  above
• 3:30
  databases,"
• 3:31
  with
• 3:31
  14
• 3:31
  providers
• 3:32
  named
• 3:32
  and
• 3:32
  keys
• 3:32
  stored
• 3:33
  in
• 3:33
  plaintext.
• 3:34
  Attribution
• 3:34
  leans
• 3:35
  Costa
• 3:35
  Rican
• 3:35
  —
• 3:36
  voseo
• 3:36
  Spanish
• 3:37
  reports,
• 3:37
  the
• 3:38
  Carbo506
• 3:39
  handle,
• 3:39
  UTC-06:00
• 3:42
  build
• 3:43
  timestamps,
• 3:43
  tunnels
• 3:44
  into
• 3:44
  AS262145
• 3:46
  —
• 3:47
  and
• 3:48
  as
• 3:48
  of
• 3:48
  September
• 3:48
  3
• 3:49
  most
• 3:49
  infrastructure
• 3:49
  was
• 3:49
  still
• 3:50
  online.
• 3:51
  Defender
• 3:51
  guidance:
• 3:52
  never
• 3:52
  expose
• 3:53
  the
• 3:53
  Docker
• 3:53
  daemon;
• 3:54
  don't
• 3:54
  blacklist
• 3:54
  the
• 3:54
  legitimate
• 3:55
  hermes-agent
• 3:55
  package
• 3:55
  —
• 3:56
  hunt
• 3:56
  the
• 3:56
  abuse
• 3:57
  signature
• 3:57
  (the
• 3:58
  GH0ST
• 3:59
  persona
• 3:59
  file,
• 4:00
  unexplained
• 4:00
  Telegram
• 4:01
  egress);
• 4:01
  and
• 4:02
  treat
• 4:02
  AI
• 4:02
  API
• 4:03
  keys
• 4:03
  like
• 4:03
  bank
• 4:03
  credentials.
• 4:04
  Licensed
• 4:05
  red-team
• 4:05
  tooling
• 4:05
  converges
• 4:06
  on
• 4:06
  the
• 4:06
  same
• 4:06
  stack:
• 4:07
  Toshell,
• 4:08
  an
• 4:08
  MIT-licensed
• 4:09
  single-binary
• 4:10
  Go
• 4:10
  C2
• 4:10
  framework,
• 4:11
  ships
• 4:11
  a
• 4:11
  team
• 4:11
  server,
• 4:12
  web
• 4:12
  console,
• 4:13
  and
• 4:13
  multi-platform
• 4:14
  implants
• 4:14
  with
• 4:14
  an
• 4:14
  AI
• 4:15
  copilot
• 4:15
  behind
• 4:15
  approval
• 4:16
  gates.
• 4:16
  Its
• 4:17
  most
• 4:17
  empirical
• 4:17
  finding
• 4:18
  concerns
• 4:18
  execution
• 4:19
  gates
• 4:19
  rather
• 4:19
  than
• 4:19
  evasion
• 4:20
  —
• 4:21
  on
• 4:21
  consumer
• 4:22
  endpoint
• 4:22
  protection
• 4:23
  like
• 4:23
  360/PC
• 4:24
  Manager,
• 4:25
  unsigned
• 4:25
  PEs
• 4:26
  are
• 4:26
  rejected
• 4:26
  at
• 4:26
  process
• 4:27
  creation
• 4:27
  (even
• 4:28
  a
• 4:28
  bare
• 4:28
  Hello-World
• 4:28
  Go
• 4:29
  binary),
• 4:29
  while
• 4:30
  a
• 4:30
  Microsoft-signed
• 4:30
  copy
• 4:31
  of
• 4:31
  notepad.exe
• 4:32
  executes
• 4:33
  fine,
• 4:34
  making
• 4:34
  code
• 4:34
  signing
• 4:34
  the
• 4:34
  real
• 4:35
  bar
• 4:35
  .
• 4:36
  The
• 4:36
  offensive
• 4:36
  models
• 4:37
  are
• 4:37
  meanwhile
• 4:37
  moving
• 4:37
  out
• 4:37
  of
• 4:37
  API
• 4:38
  jurisdiction.
• 4:39
  A
• 4:39
  Project
• 4:40
  Black
• 4:40
  experiment
• 4:40
  asked
• 4:41
  whether
• 4:41
  AI
• 4:41
  could
• 4:42
  write
• 4:42
  an
• 4:42
  LSASS
• 4:42
  dumper
• 4:43
  —
• 4:44
  the
• 4:44
  post-exploitation
• 4:45
  step
• 4:45
  that
• 4:45
  extracts
• 4:45
  Windows
• 4:46
  credentials
• 4:46
  for
• 4:46
  lateral
• 4:47
  movement
• 4:47
  —
• 4:48
  that
• 4:48
  evades
• 4:48
  EDR.
• 4:50
  Claude
• 4:50
  (Opus
• 4:50
  5,
• 4:51
  Opus
• 4:51
  4.8,
• 4:52
  Sonnet
• 4:52
  5)
• 4:52
  refused
• 4:53
  outright,
• 4:54
  even
• 4:54
  with
• 4:54
  Cyber
• 4:54
  Verification
• 4:55
  Program
• 4:55
  approval;
• 4:56
  DeepSeek
• 4:57
  v4
• 4:57
  Flash
• 4:58
  produced
• 4:58
  a
• 4:58
  working
• 4:58
  executable
• 4:59
  immediately,
• 5:00
  though
• 5:00
  EDR
• 5:01
  flagged
• 5:01
  it;
• 5:02
  an
• 5:02
  uncensored
• 5:02
  community
• 5:03
  build
• 5:03
  of
• 5:03
  Qwen
• 5:03
  3.8
• 5:04
  27B
• 5:05
  on
• 5:05
  a
• 5:05
  2x
• 5:06
  RTX
• 5:06
  4090
• 5:07
  hashcat
• 5:08
  rig
• 5:08
  then
• 5:08
  made
• 5:09
  it
• 5:09
  stealthier
• 5:09
  unprompted
• 5:09
  —
• 5:10
  less
• 5:11
  suspicious
• 5:11
  process
• 5:12
  spawning,
• 5:12
  reduced
• 5:13
  access
• 5:13
  masks,
• 5:14
  random
• 5:14
  sleeps
• 5:14
  during
• 5:15
  the
• 5:15
  dump,
• 5:15
  scrubbed
• 5:16
  strings
• 5:16
  —
• 5:17
  yielding
• 5:17
  zero
• 5:17
  detections
• 5:18
  across
• 5:18
  two
• 5:18
  lab
• 5:19
  EDR
• 5:19
  products
• 5:19
  .
• 5:20
  Heretic
• 5:20
  industrializes
• 5:21
  that
• 5:21
  last
• 5:21
  step:
• 5:22
  Philipp
• 5:23
  Emanuel
• 5:23
  Weidmann's
• 5:23
  AGPL
• 5:24
  project
• 5:24
  automates
• 5:25
  "abliteration"
• 5:26
  —
• 5:27
  extracting
• 5:27
  per-layer
• 5:28
  refusal
• 5:28
  directions
• 5:29
  from
• 5:29
  contrasting
• 5:29
  prompt
• 5:30
  sets,
• 5:30
  attaching
• 5:31
  LoRA
• 5:31
  adapters,
• 5:32
  running
• 5:32
  a
• 5:32
  200-trial
• 5:33
  parameter
• 5:33
  search
• 5:34
  (~3
• 5:34
  hours
• 5:35
  on
• 5:35
  a
• 5:35
  12
• 5:35
  GB
• 5:35
  RTX
• 5:36
  3060)
• 5:37
  scoring
• 5:37
  refusal
• 5:37
  suppression
• 5:38
  against
• 5:38
  KL
• 5:39
  divergence
• 5:39
  from
• 5:39
  the
• 5:39
  original.
• 5:40
  The
• 5:40
  tutorial
• 5:41
  run
• 5:41
  drops
• 5:41
  refusals
• 5:42
  from
• 5:42
  93/100
• 5:43
  to
• 5:43
  11/100
• 5:44
  at
• 5:44
  KL
• 5:45
  0.0508,
• 5:46
  on
• 5:47
  consumer
• 5:47
  hardware,
• 5:47
  for
• 5:48
  non-experts
• 5:48
  .
• 5:51
  The
• 5:51
  same
• 5:51
  open-weight
• 5:52
  ecosystem
• 5:52
  produces
• 5:53
  its
• 5:53
  own
• 5:53
  honest
• 5:53
  accounting:
• 5:54
  Splash,
• 5:55
  Incoai's
• 5:56
  compiled
• 5:56
  C++
• 5:58
  and
• 5:58
  Metal
• 5:58
  speculative
• 5:59
  decoding
• 5:59
  engine
• 6:00
  for
• 6:00
  Apple
• 6:00
  Silicon,
• 6:01
  forked
• 6:01
  to
• 6:01
  add
• 6:01
  native
• 6:02
  8-bit
• 6:02
  weights
• 6:02
  —
• 6:03
  Metal
• 6:03
  Q8
• 6:04
  decode
• 6:04
  kernels,
• 6:04
  PR
• 6:05
  proposed
• 6:05
  upstream.
• 6:06
  On
• 6:06
  an
• 6:07
  M5
• 6:07
  Pro
• 6:07
  with
• 6:07
  64
• 6:08
  GB,
• 6:09
  native
• 6:09
  8-bit
• 6:11
  Qwen3.8-27B
• 6:13
  averaged
• 6:13
  36.9
• 6:14
  tok/s
• 6:15
  versus
• 6:15
  official
• 6:16
  4-bit's
• 6:16
  60.7,
• 6:17
  MTP-based
• 6:18
  MLX
• 6:18
  at
• 6:18
  26.5,
• 6:20
  and
• 6:20
  stock
• 6:20
  MLX/llama.cpp
• 6:23
  at
• 6:23
  9.9
• 6:23
  —
• 6:24
  a
• 6:24
  3.73x
• 6:25
  speedup
• 6:26
  over
• 6:26
  stock,
• 6:27
  peaking
• 6:27
  at
• 6:27
  54.8
• 6:28
  tok/s
• 6:29
  on
• 6:29
  math.
• 6:29
  Speculative
• 6:30
  decoding
• 6:30
  speed
• 6:31
  equals
• 6:31
  draft
• 6:31
  speed
• 6:31
  times
• 6:32
  acceptance
• 6:32
  rate
• 6:32
  —
• 6:33
  hence
• 6:33
  the
• 6:33
  paradox:
• 6:34
  uncompressed
• 6:35
  8-bit
• 6:35
  (27
• 6:36
  GB)
• 6:36
  ran
• 6:37
  slightly
• 6:37
  faster
• 6:37
  than
• 6:38
  compressed
• 6:38
  8-bit
• 6:38
  (17
• 6:39
  GB),
• 6:40
  because
• 6:40
  compression
• 6:41
  flattens
• 6:41
  logits,
• 6:42
  lowers
• 6:42
  draft
• 6:42
  acceptance,
• 6:43
  and
• 6:44
  triggers
• 6:44
  verification
• 6:45
  rollbacks
• 6:45
  —
• 6:46
  "the
• 6:46
  kind
• 6:46
  of
• 6:46
  result
• 6:46
  that
• 6:46
  makes
• 6:47
  you
• 6:47
  rethink
• 6:47
  every
• 6:47
  quantization
• 6:48
  benchmark"
• 6:48
  .
• 6:49
  The
• 6:49
  Reasoning
• 6:50
  Cliff
• 6:50
  is
• 6:50
  the
• 6:50
  more
• 6:50
  consequential
• 6:51
  finding:
• 6:52
  on
• 6:52
  extended
• 6:52
  chain-of-thought
• 6:53
  tests
• 6:53
  (MATH-500,
• 6:54
  AIME
• 6:55
  2025,
• 6:56
  GPQA
• 6:57
  Diamond),
• 6:57
  4-bit
• 6:58
  and
• 6:58
  compressed
• 6:58
  8-bit
• 6:59
  builds
• 6:59
  drifted
• 7:00
  mid-way
• 7:00
  through
• 7:00
  algebraic
• 7:01
  series
• 7:01
  and
• 7:02
  produced
• 7:02
  wrong
• 7:02
  values,
• 7:03
  while
• 7:03
  native
• 7:04
  8-bit
• 7:04
  on
• 7:04
  all
• 7:04
  64
• 7:05
  layers
• 7:05
  —
• 7:06
  or
• 7:06
  just
• 7:06
  the
• 7:06
  top
• 7:07
  8
• 7:07
  sensitive
• 7:07
  layers
• 7:07
  —
• 7:08
  eliminated
• 7:09
  the
• 7:09
  cliff.
• 7:10
  Context
• 7:10
  scaling
• 7:10
  held
• 7:11
  too:
• 7:12
  Qwen3.8's
• 7:15
  hybrid
• 7:15
  architecture
• 7:16
  (48
• 7:16
  linear-attention
• 7:17
  layers
• 7:17
  plus
• 7:17
  16
• 7:18
  full-attention
• 7:18
  layers)
• 7:19
  kept
• 7:19
  decode
• 7:19
  at
• 7:19
  21–33
• 7:21
  tok/s
• 7:22
  out
• 7:22
  to
• 7:22
  190k
• 7:23
  tokens,
• 7:24
  with
• 7:24
  cache-hit
• 7:25
  TTFT
• 7:26
  of
• 7:26
  6.5
• 7:26
  seconds
• 7:27
  at
• 7:27
  187k
• 7:28
  context
• 7:28
  —
• 7:30
  though
• 7:30
  cold
• 7:30
  prefill
• 7:30
  of
• 7:31
  180k
• 7:32
  tokens
• 7:32
  takes
• 7:32
  four
• 7:32
  to
• 7:33
  five
• 7:33
  minutes,
• 7:34
  more
• 7:34
  than
• 7:34
  agent
• 7:34
  harnesses
• 7:35
  tolerate.
• 7:36
  PrismML's
• 7:36
  Bonsai
• 7:37
  2
• 7:37
  makes
• 7:37
  a
• 7:37
  bigger
• 7:37
  compression
• 7:38
  claim
• 7:38
  and
• 7:38
  gets
• 7:39
  rougher
• 7:39
  treatment.
• 7:40
  Ternary
• 7:40
  compression
• 7:41
  of
• 7:41
  the
• 7:41
  full
• 7:41
  Qwen
• 7:42
  3.8
• 7:42
  27B
• 7:43
  comes
• 7:43
  in
• 7:43
  9x
• 7:44
  smaller
• 7:44
  at
• 7:44
  5–6
• 7:46
  GiB
• 7:46
  VRAM,
• 7:47
  claiming
• 7:47
  98.2%
• 7:48
  of
• 7:48
  quality
• 7:49
  —
• 7:49
  roughly
• 7:50
  2
• 7:50
  points
• 7:50
  worse
• 7:51
  across
• 7:51
  benchmarks
• 7:51
  versus
• 7:52
  about
• 7:52
  14
• 7:53
  for
• 7:53
  Q2
• 7:53
  XXS
• 7:53
  —
• 7:54
  at
• 7:54
  ~900
• 7:55
  tok/s
• 7:56
  prefill
• 7:56
  on
• 7:56
  a
• 7:57
  7900XTX
• 8:00
  via
• 8:01
  a
• 8:01
  forked
• 8:01
  llama.cpp.
• 8:03
  The
• 8:03
  community's
• 8:04
  receipts
• 8:04
  are
• 8:04
  less
• 8:04
  generous:
• 8:05
  one
• 8:05
  tester
• 8:06
  reports
• 8:06
  the
• 8:06
  build
• 8:06
  misses
• 8:07
  30%
• 8:07
  of
• 8:07
  needles
• 8:08
  in
• 8:08
  a
• 8:08
  needle
• 8:08
  test
• 8:09
  —
• 8:09
  attention
• 8:09
  degradation
• 8:10
  that
• 8:10
  single-pass
• 8:11
  benchmark
• 8:11
  aggregates
• 8:12
  hide
• 8:12
  —
• 8:12
  and
• 8:13
  the
• 8:13
  "full
• 8:13
  precision
• 8:13
  quality"
• 8:14
  framing
• 8:14
  drew
• 8:15
  open
• 8:15
  sarcasm
• 8:15
  plus
• 8:16
  a
• 8:16
  linked
• 8:16
  thread
• 8:17
  alleging
• 8:17
  cherry-picked
• 8:17
  evaluations
• 8:18
  .
• 8:19
  KVA
• 8:19
  projections
• 8:20
  attack
• 8:20
  the
• 8:20
  other
• 8:21
  side
• 8:21
  of
• 8:21
  the
• 8:21
  cost
• 8:21
  curve:
• 8:22
  prefill.
• 8:23
  A
• 8:23
  community
• 8:24
  developer
• 8:24
  built
• 8:24
  projectors
• 8:25
  for
• 8:25
  Qwen3.8
• 8:28
  Flash
• 8:28
  Next
• 8:29
  on
• 8:29
  a
• 8:29
  2x
• 8:30
  R9700
• 8:31
  rig
• 8:31
  that
• 8:31
  approximate
• 8:32
  later-layer
• 8:32
  key/value
• 8:33
  computation
• 8:33
  —
• 8:35
  ridge-regression
• 8:36
  maps
• 8:36
  predicting
• 8:36
  layer
• 8:37
  inputs
• 8:37
  while
• 8:37
  the
• 8:37
  actual
• 8:38
  weights
• 8:38
  compute
• 8:38
  keys
• 8:39
  and
• 8:39
  values
• 8:39
  —
• 8:40
  for
• 8:40
  1.85x
• 8:41
  prefill
• 8:41
  speedup
• 8:42
  (1700
• 8:43
  to
• 8:43
  3150
• 8:44
  t/s)
• 8:45
  at
• 8:45
  +8%
• 8:46
  perplexity
• 8:46
  from
• 8:46
  layer
• 8:47
  12,
• 8:47
  down
• 8:48
  to
• 8:48
  1.45x
• 8:49
  at
• 8:49
  +2.6%
• 8:51
  from
• 8:51
  layer
• 8:51
  24;
• 8:52
  a
• 8:52
  multi-layer
• 8:53
  variant
• 8:53
  manages
• 8:54
  1.55x
• 8:55
  at
• 8:55
  +2%.
• 8:57
  The
• 8:57
  author
• 8:57
  is
• 8:57
  candid:
• 8:58
  it
• 8:58
  is
• 8:58
  like
• 8:58
  MTP
• 8:59
  for
• 8:59
  prefill
• 9:00
  instead
• 9:00
  of
• 9:00
  token
• 9:00
  generation
• 9:01
  —
• 9:02
  "except
• 9:02
  it's
• 9:02
  not
• 9:02
  lossless"
• 9:03
  —
• 9:03
  the
• 9:03
  error
• 9:04
  is
• 9:04
  ingested
• 9:04
  by
• 9:04
  the
• 9:04
  model.
• 9:05
  Version
• 9:06
  1
• 9:06
  is
• 9:06
  stable
• 9:06
  on
• 9:07
  Hugging
• 9:07
  Face;
• 9:08
  version
• 9:08
  2
• 9:09
  sits
• 9:09
  behind
• 9:09
  a
• 9:09
  config
• 9:10
  flag,
• 9:10
  explicitly
• 9:11
  not
• 9:11
  confirmed
• 9:11
  stable
• 9:12
  .
• 9:14
  The
• 9:15
  runtime
• 9:15
  surface
• 9:15
  keeps
• 9:16
  widening.
• 9:16
  Nestrilabs'
• 9:17
  virtio-nvgpu
• 9:18
  gives
• 9:18
  a
• 9:18
  KVM
• 9:19
  guest
• 9:19
  near-native
• 9:20
  NVIDIA
• 9:20
  GPU
• 9:20
  access
• 9:21
  by
• 9:21
  forwarding
• 9:22
  the
• 9:22
  kernel
• 9:22
  driver's
• 9:22
  ioctls
• 9:23
  between
• 9:23
  guest
• 9:23
  and
• 9:24
  host
• 9:24
  at
• 9:24
  the
• 9:24
  driver-ABI
• 9:25
  level,
• 9:25
  so
• 9:26
  the
• 9:26
  guest
• 9:26
  runs
• 9:26
  NVIDIA's
• 9:27
  own
• 9:27
  unmodified
• 9:28
  user-mode
• 9:28
  drivers.
• 9:29
  Unlike
• 9:29
  VFIO
• 9:30
  passthrough,
• 9:30
  which
• 9:31
  dedicates
• 9:31
  a
• 9:31
  whole
• 9:31
  GPU
• 9:32
  to
• 9:32
  one
• 9:32
  VM,
• 9:33
  it
• 9:33
  shares
• 9:33
  a
• 9:33
  card:
• 9:34
  four
• 9:34
  simultaneous
• 9:35
  guests
• 9:35
  on
• 9:35
  a
• 9:35
  single
• 9:36
  RTX
• 9:36
  3060
• 9:37
  each
• 9:38
  held
• 9:38
  25.57–26.49
• 9:41
  fps
• 9:41
  (103.7
• 9:42
  combined
• 9:43
  versus
• 9:43
  102.9
• 9:44
  for
• 9:44
  one
• 9:44
  guest),
• 9:45
  all
• 9:45
  encoding
• 9:46
  H.264
• 9:47
  via
• 9:48
  NVENC
• 9:48
  at
• 9:48
  once.
• 9:49
  Unlike
• 9:49
  API-translation
• 9:50
  schemes
• 9:50
  like
• 9:51
  virtio-gpu
• 9:51
  with
• 9:51
  Venus,
• 9:52
  at
• 9:52
  ~2,000
• 9:53
  boundary
• 9:53
  crossings
• 9:54
  per
• 9:54
  frame,
• 9:54
  it
• 9:54
  crosses
• 9:55
  per
• 9:55
  ioctl
• 9:56
  —
• 9:57
  13,792
• 9:58
  messages
• 9:58
  over
• 9:58
  813,691
• 10:01
  frames,
• 10:01
  one
• 10:01
  per
• 10:02
  59
• 10:02
  frames
• 10:02
  —
• 10:03
  and
• 10:03
  guests
• 10:03
  hit
• 10:04
  98–100%
• 10:05
  of
• 10:05
  native
• 10:06
  frame
• 10:06
  times
• 10:06
  above
• 10:06
  ~2
• 10:07
  ms
• 10:07
  .
• 10:08
  Caveats
• 10:08
  as
• 10:09
  plain
• 10:09
  as
• 10:09
  the
• 10:09
  numbers:
• 10:10
  CUDA
• 10:10
  allocation
• 10:11
  and
• 10:11
  zero-copy
• 10:11
  interop
• 10:12
  are
• 10:12
  untested
• 10:13
  beyond
• 10:13
  enumeration,
• 10:14
  there
• 10:14
  is
• 10:14
  no
• 10:14
  IOMMU
• 10:15
  boundary,
• 10:16
  and
• 10:16
  the
• 10:16
  host
• 10:16
  driver
• 10:17
  sits
• 10:17
  in
• 10:17
  the
• 10:17
  trusted
• 10:17
  computing
• 10:18
  base
• 10:18
  —
• 10:19
  "attack-surface
• 10:19
  reduction,
• 10:20
  not
• 10:20
  hardware
• 10:21
  isolation."
• 10:22
  Promising,
• 10:22
  but
• 10:23
  early.
• 10:23
  Model
• 10:24
  portability
• 10:24
  advances
• 10:25
  on
• 10:25
  stranger
• 10:25
  vectors:
• 10:26
  Parakeet.java
• 10:27
  ports
• 10:27
  NVIDIA's
• 10:28
  Parakeet
• 10:28
  ASR
• 10:29
  family
• 10:29
  to
• 10:29
  run
• 10:30
  end-to-end
• 10:30
  on
• 10:30
  the
• 10:30
  JVM,
• 10:31
  competitive
• 10:32
  with
• 10:32
  native
• 10:32
  engines
• 10:33
  on
• 10:33
  CPUs
• 10:33
  —
• 10:34
  Steve
• 10:34
  Jobs'
• 10:34
  15-minute
• 10:35
  Stanford
• 10:35
  speech
• 10:36
  transcribes
• 10:36
  in
• 10:36
  about
• 10:37
  15
• 10:37
  seconds
• 10:37
  on
• 10:38
  a
• 10:38
  laptop
• 10:38
  with
• 10:38
  the
• 10:38
  110M
• 10:39
  Q4
• 10:40
  K
• 10:40
  model,
• 10:40
  streaming
• 10:41
  included,
• 10:41
  Apache
• 10:42
  2.0
• 10:42
  .
• 10:43
  A
• 10:43
  421M-parameter
• 10:45
  Laya
• 10:45
  model
• 10:46
  plays
• 10:46
  Flappy
• 10:46
  Bird
• 10:46
  on
• 10:47
  a
• 10:47
  12th-gen
• 10:47
  i7
• 10:48
  desktop
• 10:48
  CPU
• 10:49
  via
• 10:49
  OpenVINO
• 10:50
  INT8
• 10:50
  —
• 10:51
  a
• 10:51
  commenter's
• 10:52
  tip:
• 10:53
  pin
• 10:53
  to
• 10:53
  P-cores
• 10:54
  and
• 10:54
  dump
• 10:54
  the
• 10:54
  execution
• 10:55
  graph
• 10:55
  to
• 10:55
  confirm
• 10:55
  the
• 10:56
  matmuls
• 10:56
  actually
• 10:57
  went
• 10:57
  int8
• 10:57
  .
• 10:58
  And
• 10:58
  Qwen
• 10:58
  pushed
• 10:58
  its
• 10:59
  image
• 10:59
  line
• 10:59
  forward
• 10:59
  with
• 11:00
  the
• 11:00
  Qwen
• 11:00
  Image
• 11:00
  2.1
• 11:00
  release
• 11:01
  .
• 11:04
  Once
• 11:05
  models
• 11:05
  run
• 11:05
  anywhere,
• 11:06
  organizing
• 11:06
  them
• 11:07
  becomes
• 11:07
  the
• 11:07
  product.
• 11:08
  Paperclip,
• 11:09
  an
• 11:09
  MIT-licensed
• 11:10
  Node.js
• 11:12
  server
• 11:12
  with
• 11:12
  a
• 11:12
  React
• 11:12
  UI,
• 11:13
  runs
• 11:13
  a
• 11:13
  "company"
• 11:14
  of
• 11:14
  AI
• 11:14
  agents:
• 11:15
  org
• 11:15
  charts,
• 11:15
  roles
• 11:16
  —
• 11:16
  CEO,
• 11:17
  CTO,
• 11:18
  engineers,
• 11:19
  designers,
• 11:19
  marketers,
• 11:20
  any
• 11:21
  bot
• 11:21
  from
• 11:21
  any
• 11:21
  provider
• 11:21
  —
• 11:22
  budgets,
• 11:23
  and
• 11:23
  governance.
• 11:24
  "If
• 11:24
  OpenClaw
• 11:24
  is
• 11:24
  an
• 11:25
  employee,
• 11:25
  Paperclip
• 11:26
  is
• 11:26
  the
• 11:26
  company";
• 11:27
  "manage
• 11:27
  business
• 11:27
  goals,
• 11:28
  not
• 11:28
  pull
• 11:28
  requests."
• 11:30
  Twelve
• 11:30
  subsystems
• 11:31
  cover
• 11:31
  work,
• 11:31
  heartbeats,
• 11:32
  governance,
• 11:32
  budgets,
• 11:32
  routines,
• 11:33
  and
• 11:33
  secrets,
• 11:33
  with
• 11:34
  adapters
• 11:34
  for
• 11:34
  Claude
• 11:34
  Code,
• 11:35
  Codex,
• 11:35
  CLI
• 11:36
  agents,
• 11:36
  and
• 11:36
  HTTP
• 11:37
  bots
• 11:37
  —
• 11:38
  "if
• 11:38
  it
• 11:38
  can
• 11:38
  receive
• 11:38
  a
• 11:38
  heartbeat,
• 11:39
  it's
• 11:39
  hired"
• 11:39
  —
• 11:40
  and
• 11:40
  overspending
• 11:40
  automatically
• 11:41
  pauses
• 11:41
  agents
• 11:41
  .
• 11:42
  The
• 11:42
  editor's
• 11:43
  video
• 11:43
  pick
• 11:43
  puts
• 11:44
  that
• 11:44
  architecture
• 11:44
  through
• 11:44
  a
• 11:45
  real
• 11:45
  workload:
• 11:46
  NetworkChuck,
• 11:47
  with
• 11:47
  Paperclip
• 11:47
  creator
• 11:48
  Doda
• 11:48
  making
• 11:48
  his
• 11:49
  first
• 11:49
  on-camera
• 11:49
  appearance,
• 11:50
  built
• 11:50
  an
• 11:50
  all-AI
• 11:51
  IT
• 11:51
  department
• 11:51
  —
• 11:52
  Dumbledore
• 11:52
  as
• 11:52
  CEO
• 11:53
  on
• 11:53
  Claude
• 11:53
  Code,
• 11:54
  Ron
• 11:54
  as
• 11:54
  CTO
• 11:55
  on
• 11:55
  Hermes,
• 11:55
  Mad-Eye
• 11:55
  Moody
• 11:56
  on
• 11:56
  security
• 11:56
  via
• 11:56
  Codex,
• 11:57
  plus
• 11:58
  helpdesk
• 11:58
  agents
• 11:58
  on
• 11:58
  local
• 11:59
  models
• 11:59
  —
• 12:00
  on
• 12:00
  a
• 12:00
  genuine
• 12:01
  studio
• 12:01
  mystery:
• 12:02
  every
• 12:02
  toilet
• 12:02
  flush
• 12:03
  knocked
• 12:03
  everyone
• 12:03
  off
• 12:03
  the
• 12:03
  NAS.
• 12:04
  The
• 12:05
  agents
• 12:05
  traced
• 12:05
  the
• 12:05
  MikroTik
• 12:06
  switch
• 12:06
  dropping
• 12:06
  four
• 12:07
  fiber
• 12:07
  runs
• 12:07
  within
• 12:08
  two
• 12:08
  seconds
• 12:08
  of
• 12:08
  each
• 12:08
  flush,
• 12:09
  36
• 12:10
  times
• 12:10
  that
• 12:10
  week;
• 12:11
  port
• 12:11
  5
• 12:12
  had
• 12:12
  logged
• 12:12
  20,052
• 12:13
  link-downs
• 12:14
  since
• 12:14
  boot
• 12:14
  versus
• 12:14
  123
• 12:15
  on
• 12:15
  a
• 12:15
  healthy
• 12:16
  port.
• 12:17
  Root
• 12:17
  cause:
• 12:18
  cheap
• 12:18
  third-party
• 12:19
  Amazon
• 12:19
  SFPs
• 12:20
  from
• 12:20
  one
• 12:20
  batch
• 12:21
  with
• 12:21
  a
• 12:21
  50%
• 12:21
  in-service
• 12:22
  failure
• 12:22
  rate;
• 12:23
  genuine
• 12:23
  transceivers
• 12:24
  fixed
• 12:24
  it
• 12:24
  —
• 12:25
  a
• 12:25
  real
• 12:25
  investigation,
• 12:26
  artifacts
• 12:27
  and
• 12:27
  escalation
• 12:27
  chains
• 12:28
  included,
• 12:29
  that
• 12:29
  reached
• 12:29
  the
• 12:29
  human
• 12:29
  only
• 12:30
  when
• 12:30
  the
• 12:30
  agents
• 12:31
  were
• 12:31
  stumped
• 12:31
  .
• 12:32
  Agent
• 12:32
  companies
• 12:33
  need
• 12:33
  institutional
• 12:34
  memory,
• 12:34
  which
• 12:34
  Asymptote
• 12:35
  Labs'
• 12:35
  Beacon
• 12:35
  supplies:
• 12:37
  an
• 12:37
  MIT-licensed
• 12:38
  layer
• 12:38
  capturing
• 12:39
  session
• 12:39
  history
• 12:39
  across
• 12:40
  Claude
• 12:40
  Code,
• 12:41
  Cursor,
• 12:41
  Codex,
• 12:42
  OpenCode,
• 12:43
  Cline,
• 12:43
  and
• 12:43
  20-plus
• 12:44
  harnesses,
• 12:45
  normalized
• 12:46
  into
• 12:46
  a
• 12:46
  common
• 12:46
  OpenTelemetry
• 12:47
  event
• 12:47
  model
• 12:47
  —
• 12:48
  prompts,
• 12:49
  tool
• 12:49
  calls,
• 12:50
  approvals,
• 12:51
  Model
• 12:51
  Context
• 12:52
  Protocol
• 12:52
  (MCP)
• 12:53
  interactions,
• 12:54
  token
• 12:54
  usage.
• 12:55
  The
• 12:56
  loop
• 12:56
  —
• 12:56
  run,
• 12:57
  capture,
• 12:57
  evaluate,
• 12:58
  extract,
• 12:59
  review,
• 12:59
  reuse
• 13:00
  —
• 13:01
  turns
• 13:01
  a
• 13:01
  debugging
• 13:02
  path
• 13:02
  that
• 13:02
  fixed
• 13:02
  an
• 13:03
  obscure
• 13:03
  issue
• 13:03
  into
• 13:04
  knowledge
• 13:04
  the
• 13:04
  next
• 13:04
  agent
• 13:05
  starts
• 13:05
  with.
• 13:06
  "Your
• 13:06
  Cursor
• 13:06
  sessions
• 13:07
  can
• 13:07
  improve
• 13:07
  Codex";
• 13:08
  forwarding
• 13:09
  to
• 13:09
  Splunk
• 13:09
  or
• 13:10
  Datadog
• 13:10
  is
• 13:10
  optional
• 13:11
  .
• 13:11
  The
• 13:12
  bill
• 13:12
  arrives
• 13:12
  at
• 13:12
  CI.
• 13:14
  Linear
• 13:14
  reworked
• 13:14
  its
• 13:15
  pipeline
• 13:15
  after
• 13:15
  agents
• 13:16
  accelerated
• 13:16
  shipping
• 13:17
  faster
• 13:17
  than
• 13:17
  validation
• 13:18
  could
• 13:18
  absorb:
• 13:19
  every
• 13:19
  PR
• 13:19
  still
• 13:20
  passes
• 13:20
  through
• 13:20
  CI,
• 13:21
  so
• 13:21
  CI
• 13:22
  became
• 13:22
  the
• 13:22
  bottleneck.
• 13:24
  Faster
• 13:24
  third-party
• 13:25
  runners
• 13:25
  (34%
• 13:26
  average
• 13:26
  job
• 13:27
  speedup),
• 13:28
  a
• 13:28
  native
• 13:28
  TypeScript
• 13:29
  compiler
• 13:29
  (73%
• 13:30
  off
• 13:30
  the
• 13:30
  typecheck
• 13:31
  median),
• 13:31
  pure-AST
• 13:32
  lint
• 13:32
  rules
• 13:33
  (68%
• 13:33
  off
• 13:34
  API
• 13:34
  lint),
• 13:35
  capped
• 13:35
  fetch
• 13:35
  depth
• 13:36
  (slowest
• 13:36
  gate
• 13:36
  from
• 13:37
  94
• 13:37
  to
• 13:37
  20
• 13:37
  seconds),
• 13:38
  schema
• 13:39
  snapshots
• 13:39
  instead
• 13:40
  of
• 13:40
  replaying
• 13:40
  migrations,
• 13:41
  batched
• 13:42
  checks
• 13:42
  saving
• 13:42
  ~87,000
• 13:44
  runner-minutes
• 13:44
  monthly,
• 13:45
  and
• 13:45
  eight
• 13:45
  shards
• 13:46
  instead
• 13:46
  of
• 13:46
  four
• 13:46
  cut
• 13:47
  PR
• 13:47
  wait
• 13:47
  from
• 13:47
  over
• 13:47
  six
• 13:48
  minutes
• 13:48
  to
• 13:48
  just
• 13:48
  over
• 13:49
  five
• 13:49
  despite
• 13:50
  a
• 13:50
  year
• 13:50
  of
• 13:50
  test-suite
• 13:51
  growth
• 13:51
  —
• 13:52
  without
• 13:53
  the
• 13:53
  rework,
• 13:53
  today's
• 13:53
  suite
• 13:54
  would
• 13:54
  take
• 13:54
  about
• 13:54
  eleven
• 13:54
  minutes.
• 13:55
  Roughly
• 13:56
  2,000
• 13:56
  tests
• 13:57
  are
• 13:57
  added
• 13:57
  weekly,
• 13:58
  and
• 13:58
  "agents
• 13:58
  now
• 13:58
  write
• 13:59
  the
• 13:59
  majority
• 13:59
  of
• 13:59
  our
• 13:59
  tests"
• 14:00
  .
• 14:02
  When
• 14:03
  generation
• 14:03
  is
• 14:03
  cheap,
• 14:04
  measurement
• 14:04
  becomes
• 14:04
  the
• 14:04
  product.
• 14:05
  FIDES,
• 14:05
  a
• 14:06
  protocol
• 14:06
  posted
• 14:06
  to
• 14:06
  arXiv,
• 14:07
  treats
• 14:07
  an
• 14:07
  LLM-generated
• 14:08
  trading
• 14:08
  strategy
• 14:09
  as
• 14:09
  three
• 14:09
  artifacts
• 14:10
  to
• 14:10
  reconcile:
• 14:11
  the
• 14:11
  stated
• 14:11
  rationale,
• 14:12
  the
• 14:12
  executable
• 14:12
  code,
• 14:13
  and
• 14:13
  the
• 14:13
  track
• 14:13
  record
• 14:13
  that
• 14:13
  code
• 14:14
  produces.
• 14:14
  "Dual
• 14:15
  delivery"
• 14:15
  returns
• 14:16
  both
• 14:16
  prose
• 14:16
  and
• 14:16
  a
• 14:16
  self-contained
• 14:17
  function
• 14:17
  from
• 14:17
  one
• 14:17
  call;
• 14:18
  the
• 14:18
  code
• 14:18
  runs
• 14:19
  in
• 14:19
  a
• 14:19
  sandboxed
• 14:19
  subprocess
• 14:20
  with
• 14:20
  whitelisted
• 14:20
  imports
• 14:21
  against
• 14:21
  a
• 14:21
  lag-one
• 14:22
  out-of-sample
• 14:22
  backtest
• 14:22
  —
• 14:23
  day-t
• 14:23
  positions
• 14:24
  use
• 14:24
  day-(t−1)
• 14:24
  signals
• 14:24
  —
• 14:25
  that
• 14:25
  the
• 14:25
  model
• 14:25
  cannot
• 14:26
  game.
• 14:26
  Three
• 14:27
  concordance
• 14:27
  gaps
• 14:27
  are
• 14:27
  scored:
• 14:28
  does
• 14:28
  the
• 14:28
  code
• 14:29
  implement
• 14:29
  the
• 14:29
  stated
• 14:29
  rules,
• 14:30
  does
• 14:30
  it
• 14:30
  execute
• 14:31
  as
• 14:31
  written
• 14:31
  without
• 14:31
  look-ahead,
• 14:32
  does
• 14:32
  the
• 14:32
  claimed
• 14:32
  edge
• 14:32
  survive
• 14:33
  out-of-sample
• 14:33
  .
• 14:34
  The
• 14:34
  findings
• 14:34
  are
• 14:34
  unflattering:
• 14:35
  across
• 14:35
  40
• 14:36
  strategies
• 14:36
  on
• 14:36
  eight
• 14:36
  liquid
• 14:37
  US
• 14:37
  ETFs
• 14:37
  and
• 14:38
  four
• 14:38
  models,
• 14:39
  concordance
• 14:39
  did
• 14:39
  not
• 14:39
  predict
• 14:40
  profit
• 14:40
  —
• 14:41
  the
• 14:41
  most
• 14:41
  speech-code-consistent
• 14:42
  model
• 14:42
  scored
• 14:43
  a
• 14:43
  perfect
• 14:43
  say-do
• 14:43
  gap
• 14:44
  and
• 14:44
  still
• 14:44
  lost
• 14:44
  to
• 14:44
  buy-and-hold;
• 14:45
  only
• 14:46
  2
• 14:46
  of
• 14:46
  40
• 14:46
  strategies
• 14:47
  beat
• 14:47
  it
• 14:47
  (both
• 14:47
  on
• 14:47
  TLT),
• 14:48
  and
• 14:48
  a
• 14:48
  plain
• 14:49
  SMA(50,200)
• 14:50
  crossover
• 14:50
  beat
• 14:51
  every
• 14:51
  model's
• 14:51
  mean
• 14:51
  Sharpe,
• 14:52
  robust
• 14:52
  to
• 14:53
  an
• 14:53
  alternate
• 14:53
  window
• 14:53
  and
• 14:53
  costs.
• 14:54
  "Concordance
• 14:54
  is
• 14:55
  thus
• 14:55
  necessary
• 14:55
  bookkeeping,
• 14:56
  not
• 14:56
  evidence
• 14:57
  of
• 14:57
  edge."
• 14:58
  Self-assessment
• 14:58
  was
• 14:59
  badly
• 14:59
  calibrated
• 14:59
  —
• 15:00
  most
• 15:00
  strategies
• 15:01
  claimed
• 15:01
  to
• 15:01
  beat
• 15:01
  buy-and-hold,
• 15:02
  exactly
• 15:02
  one
• 15:03
  did
• 15:03
  —
• 15:03
  and
• 15:03
  swapping
• 15:04
  the
• 15:04
  LLM
• 15:04
  judge
• 15:05
  for
• 15:05
  a
• 15:05
  second
• 15:05
  model
• 15:05
  flipped
• 15:06
  verdicts
• 15:06
  on
• 15:06
  over
• 15:06
  half
• 15:07
  the
• 15:07
  items.
• 15:08
  Contrastive
• 15:08
  Language
• 15:09
  Models
• 15:09
  build
• 15:09
  that
• 15:09
  discipline
• 15:10
  into
• 15:10
  architecture:
• 15:11
  a
• 15:11
  frozen
• 15:12
  Qwen3-8B
• 15:13
  backbone
• 15:13
  with
• 15:13
  20M-parameter
• 15:14
  projection
• 15:15
  heads
• 15:15
  for
• 15:15
  states
• 15:15
  and
• 15:15
  actions
• 15:16
  (the
• 15:17
  CLM-8B
• 15:18
  head
• 15:18
  is
• 15:18
  75
• 15:19
  MB),
• 15:20
  trained
• 15:20
  with
• 15:20
  bidirectional
• 15:21
  InfoNCE
• 15:21
  loss
• 15:21
  plus
• 15:22
  mid-training
• 15:22
  hard
• 15:22
  negatives
• 15:23
  over
• 15:23
  ~60M
• 15:24
  question-answer
• 15:24
  pairs
• 15:25
  and
• 15:25
  a
• 15:25
  million
• 15:26
  agentic
• 15:26
  trajectories.
• 15:27
  The
• 15:27
  staging
• 15:28
  matters:
• 15:29
  pre-training
• 15:29
  alone
• 15:30
  reached
• 15:30
  52.1%
• 15:31
  top-1
• 15:31
  on
• 15:32
  held-out
• 15:32
  questions,
• 15:33
  hard
• 15:33
  negatives
• 15:34
  lifted
• 15:34
  it
• 15:34
  to
• 15:34
  69.2%,
• 15:36
  and
• 15:36
  training
• 15:37
  on
• 15:37
  them
• 15:37
  from
• 15:37
  the
• 15:37
  start
• 15:37
  peaked
• 15:38
  seven
• 15:38
  points
• 15:38
  lower.
• 15:39
  Instead
• 15:39
  of
• 15:40
  generating
• 15:40
  text,
• 15:41
  the
• 15:41
  model
• 15:41
  scores
• 15:42
  how
• 15:42
  well
• 15:42
  candidate
• 15:42
  actions
• 15:43
  align
• 15:43
  with
• 15:43
  a
• 15:43
  state
• 15:43
  —
• 15:44
  ranking
• 15:45
  best-of-N
• 15:45
  answers,
• 15:45
  routing
• 15:46
  tools
• 15:46
  —
• 15:47
  and
• 15:47
  as
• 15:47
  a
• 15:47
  verifier
• 15:48
  over
• 15:48
  sampled
• 15:48
  solutions
• 15:49
  it
• 15:49
  sets
• 15:49
  87.6%
• 15:50
  on
• 15:51
  Terminal-Bench
• 15:51
  2.1
• 15:52
  and
• 15:52
  81.6%
• 15:53
  on
• 15:53
  DeepSWE,
• 15:54
  4.1–5.7x
• 15:57
  faster
• 15:57
  than
• 15:57
  the
• 15:57
  Jev
• 15:58
  baseline
• 15:58
  .
• 15:58
  Frontier
• 15:59
  pricing
• 15:59
  gets
• 15:59
  the
• 15:59
  same
• 16:00
  receipts
• 16:00
  treatment.
• 16:01
  An
• 16:01
  independent
• 16:02
  harness
• 16:02
  ran
• 16:02
  25
• 16:03
  graded
• 16:03
  tasks
• 16:04
  —
• 16:04
  23
• 16:05
  across
• 16:05
  three
• 16:05
  difficulty
• 16:06
  tiers
• 16:06
  plus
• 16:07
  two
• 16:07
  built
• 16:07
  to
• 16:07
  probe
• 16:08
  Fable
• 16:08
  5.1's
• 16:09
  supposed
• 16:09
  long-context
• 16:10
  edge
• 16:10
  —
• 16:11
  with
• 16:12
  automatic
• 16:12
  graders
• 16:12
  (exact
• 16:13
  answers,
• 16:14
  SQL
• 16:14
  against
• 16:14
  a
• 16:14
  real
• 16:15
  database,
• 16:15
  hidden
• 16:16
  unit
• 16:16
  tests)
• 16:17
  and
• 16:17
  every
• 16:17
  failure
• 16:18
  hand-checked.
• 16:19
  Opus
• 16:19
  5.5
• 16:20
  hit
• 16:20
  100%
• 16:21
  at
• 16:21
  roughly
• 16:21
  $0.09
• 16:22
  estimated
• 16:22
  per
• 16:23
  passed
• 16:23
  task
• 16:23
  and
• 16:23
  27
• 16:24
  seconds
• 16:24
  average
• 16:25
  —
• 16:25
  fastest
• 16:26
  and
• 16:26
  cheapest
• 16:26
  of
• 16:26
  the
• 16:27
  perfect
• 16:27
  scorers;
• 16:29
  Fable
• 16:29
  5.1
• 16:30
  matched
• 16:30
  accuracy
• 16:31
  at
• 16:31
  ~$0.26,
• 16:33
  Sonnet
• 16:33
  5
• 16:33
  at
• 16:33
  ~$0.11
• 16:34
  but
• 16:34
  61
• 16:35
  seconds;
• 16:36
  Haiku
• 16:37
  4.5
• 16:38
  failed
• 16:38
  silently,
• 16:40
  wrong
• 16:40
  three-for-three
• 16:41
  on
• 16:41
  a
• 16:41
  counting
• 16:41
  problem
• 16:42
  (429,
• 16:43
  429,
• 16:45
  233
• 16:46
  against
• 16:47
  a
• 16:47
  correct
• 16:47
  6,205).
• 16:50
  The
• 16:50
  frontier
• 16:50
  tier
• 16:50
  separated
• 16:51
  them:
• 16:52
  Sonnet
• 16:52
  5
• 16:52
  burned
• 16:53
  ~3.5x
• 16:54
  the
• 16:54
  tokens
• 16:54
  (~218k
• 16:56
  versus
• 16:56
  ~61k),
• 16:58
  canceling
• 16:58
  its
• 16:59
  lower
• 16:59
  per-token
• 16:59
  price;
• 17:00
  on
• 17:00
  the
• 17:00
  two
• 17:01
  Fable-edge
• 17:01
  tasks,
• 17:02
  Opus
• 17:02
  5.5
• 17:03
  at
• 17:03
  low
• 17:03
  effort
• 17:04
  passed
• 17:04
  3/3
• 17:05
  at
• 17:05
  roughly
• 17:05
  a
• 17:05
  third
• 17:05
  of
• 17:05
  Fable's
• 17:06
  estimated
• 17:06
  cost.
• 17:07
  Caveats
• 17:08
  as
• 17:08
  plain
• 17:08
  as
• 17:09
  the
• 17:09
  rankings:
• 17:09
  small
• 17:10
  samples,
• 17:10
  one
• 17:11
  account,
• 17:12
  a
• 17:12
  CLI
• 17:12
  harness,
• 17:13
  first-day
• 17:14
  timing
• 17:14
  .
• 17:17
  At
• 17:18
  the
• 17:18
  bottom
• 17:18
  of
• 17:18
  the
• 17:18
  stack,
• 17:19
  the
• 17:19
  day's
• 17:19
  best
• 17:20
  read
• 17:20
  dissects
• 17:21
  what
• 17:21
  trusted
• 17:21
  execution
• 17:22
  environments
• 17:22
  actually
• 17:22
  prove.
• 17:23
  The
• 17:23
  explainer
• 17:24
  at
• 17:24
  tee.public.computer
• 17:25
  walks
• 17:25
  Meta's
• 17:26
  Muse
• 17:26
  agent
• 17:26
  —
• 17:27
  which
• 17:27
  books
• 17:27
  travel,
• 17:28
  fills
• 17:28
  forms,
• 17:28
  and
• 17:29
  sends
• 17:29
  messages,
• 17:29
  each
• 17:30
  instance
• 17:30
  in
• 17:30
  its
• 17:30
  own
• 17:30
  VM
• 17:31
  on
• 17:31
  Meta
• 17:31
  servers
• 17:31
  —
• 17:32
  through
• 17:32
  the
• 17:32
  attestation
• 17:33
  machinery
• 17:33
  a
• 17:33
  verifier
• 17:34
  needs:
• 17:34
  the
• 17:35
  CPU
• 17:35
  hashes
• 17:35
  the
• 17:35
  initial
• 17:36
  image
• 17:36
  (firmware,
• 17:37
  kernel,
• 17:37
  initrd,
• 17:38
  command
• 17:38
  line)
• 17:38
  into
• 17:39
  a
• 17:39
  48-byte
• 17:40
  SHA-384
• 17:42
  launch
• 17:42
  measurement
• 17:42
  fixed
• 17:42
  before
• 17:43
  the
• 17:43
  first
• 17:43
  instruction
• 17:43
  runs,
• 17:44
  then
• 17:45
  signs
• 17:45
  a
• 17:45
  report
• 17:46
  carrying
• 17:46
  that
• 17:46
  measurement,
• 17:47
  the
• 17:47
  TCB
• 17:48
  version,
• 17:48
  launch
• 17:48
  policy,
• 17:49
  and
• 17:50
  a
• 17:50
  verifier
• 17:50
  nonce,
• 17:51
  under
• 17:51
  a
• 17:51
  per-chip
• 17:52
  key
• 17:52
  chaining
• 17:52
  to
• 17:52
  AMD's
• 17:53
  root.
• 17:54
  The
• 17:54
  verifier
• 17:54
  checks
• 17:54
  the
• 17:54
  chain,
• 17:55
  the
• 17:55
  nonce,
• 17:56
  and
• 17:56
  that
• 17:56
  the
• 17:56
  measurement
• 17:56
  matches
• 17:57
  a
• 17:57
  reproducible,
• 17:57
  approved
• 17:58
  build
• 17:58
  .
• 17:59
  The
• 17:59
  hard
• 17:59
  lessons
• 18:00
  earn
• 18:00
  the
• 18:00
  read:
• 18:01
  a
• 18:01
  valid
• 18:01
  vendor
• 18:02
  signature
• 18:02
  proves
• 18:02
  the
• 18:03
  platform,
• 18:04
  not
• 18:04
  the
• 18:04
  image
• 18:04
  —
• 18:06
  an
• 18:06
  attacker
• 18:06
  can
• 18:06
  boot
• 18:07
  a
• 18:07
  modified
• 18:07
  image
• 18:07
  on
• 18:08
  a
• 18:08
  real
• 18:08
  chip
• 18:08
  and
• 18:08
  receive
• 18:09
  a
• 18:09
  genuinely
• 18:09
  signed
• 18:10
  report
• 18:10
  with
• 18:10
  a
• 18:10
  different
• 18:10
  measurement,
• 18:12
  so
• 18:12
  verifiers
• 18:12
  must
• 18:13
  enforce
• 18:13
  whitelists
• 18:13
  of
• 18:14
  approved
• 18:14
  measurements,
• 18:15
  multi-signer
• 18:16
  policies,
• 18:17
  and
• 18:17
  append-only
• 18:17
  transparency
• 18:18
  logs.
• 18:19
  Debug-enabled
• 18:20
  launches
• 18:20
  sign
• 18:20
  validly
• 18:21
  but
• 18:21
  stay
• 18:22
  readable;
• 18:23
  SEV-SNP's
• 18:26
  measurement
• 18:26
  never
• 18:27
  covers
• 18:27
  post-launch
• 18:28
  changes
• 18:28
  like
• 18:28
  attached
• 18:29
  disks
• 18:29
  or
• 18:29
  downloaded
• 18:30
  binaries,
• 18:31
  unlike
• 18:31
  TDX's
• 18:31
  runtime
• 18:32
  measurement
• 18:32
  registers;
• 18:33
  and
• 18:33
  every
• 18:34
  release
• 18:34
  brings
• 18:34
  a
• 18:34
  new
• 18:34
  measurement,
• 18:35
  so
• 18:35
  a
• 18:35
  mismatch
• 18:36
  may
• 18:36
  be
• 18:36
  a
• 18:36
  mistake
• 18:36
  —
• 18:37
  "refuse
• 18:38
  first,
• 18:38
  then
• 18:38
  investigate."
• 18:40
  The
• 18:40
  sharpest
• 18:40
  line:
• 18:41
  "a
• 18:41
  measurement
• 18:41
  identifies
• 18:42
  code;
• 18:43
  it
• 18:43
  doesn't
• 18:43
  vouch
• 18:43
  for
• 18:43
  it."
• 18:44
  Meta's
• 18:45
  claim
• 18:45
  that
• 18:45
  agent
• 18:46
  data
• 18:46
  is
• 18:46
  "encrypted
• 18:47
  with
• 18:47
  a
• 18:47
  key
• 18:47
  only
• 18:47
  they
• 18:47
  hold,
• 18:48
  so
• 18:48
  not
• 18:48
  even
• 18:49
  Meta
• 18:49
  can
• 18:49
  access
• 18:49
  it"
• 18:50
  remains
• 18:50
  an
• 18:50
  assertion
• 18:51
  awaiting
• 18:51
  the
• 18:51
  external
• 18:52
  audit
• 18:52
  planned
• 18:52
  later
• 18:52
  this
• 18:53
  year.
• 18:54
  Hardware
• 18:54
  randomness
• 18:55
  is
• 18:55
  having
• 18:55
  its
• 18:55
  own
• 18:55
  epistemic
• 18:56
  moment.
• 18:57
  A
• 18:57
  flat
• 18:57
  assembler
• 18:58
  forum
• 18:58
  thread
• 18:58
  reports
• 18:59
  that
• 18:59
  AMD's
• 18:59
  RDRAND
• 19:02
  and
• 19:02
  RDSEED
• 19:04
  instructions
• 19:04
  never
• 19:05
  produce
• 19:05
  a
• 19:05
  zero
• 19:05
  spanning
• 19:06
  the
• 19:06
  full
• 19:06
  requested
• 19:06
  register
• 19:07
  width
• 19:07
  —
• 19:08
  found
• 19:08
  by
• 19:08
  a
• 19:08
  Brazilian
• 19:09
  assembly
• 19:09
  programmer
• 19:10
  charting
• 19:10
  the
• 19:10
  16-bit
• 19:11
  number
• 19:11
  space:
• 19:12
  the
• 19:12
  zero
• 19:12
  bucket
• 19:13
  never
• 19:13
  moved
• 19:13
  on
• 19:13
  AMD
• 19:14
  hosts
• 19:14
  while
• 19:14
  Intel
• 19:15
  machines
• 19:15
  filled
• 19:15
  it
• 19:15
  normally.
• 19:17
  Corroborated
• 19:17
  so
• 19:17
  far:
• 19:18
  nine
• 19:18
  days
• 19:19
  of
• 19:19
  testing
• 19:19
  across
• 19:19
  two
• 19:20
  AMD
• 19:20
  systems
• 19:20
  produced
• 19:21
  "not
• 19:21
  a
• 19:21
  single
• 19:21
  0,"
• 19:22
  zeros
• 19:23
  do
• 19:23
  appear
• 19:23
  in
• 19:23
  the
• 19:23
  low
• 19:23
  16
• 19:24
  bits
• 19:24
  of
• 19:24
  wider
• 19:25
  requests,
• 19:25
  and
• 19:25
  AMD
• 19:26
  —
• 19:27
  whose
• 19:27
  first
• 19:27
  reply
• 19:27
  misread
• 19:28
  the
• 19:28
  results
• 19:28
  —
• 19:29
  escalated
• 19:30
  without
• 19:30
  an
• 19:30
  explanation.
• 19:31
  No
• 19:31
  root
• 19:32
  cause
• 19:32
  exists;
• 19:33
  what
• 19:33
  would
• 19:33
  settle
• 19:33
  it
• 19:33
  is
• 19:33
  AMD's
• 19:34
  engineering
• 19:34
  answer
• 19:35
  and
• 19:35
  independent
• 19:35
  reproduction
• 19:36
  on
• 19:36
  more
• 19:36
  silicon.
• 19:37
  Until
• 19:38
  then:
• 19:38
  never
• 19:39
  lean
• 19:39
  on
• 19:39
  RDRAND
• 19:42
  alone;
• 19:43
  mix
• 19:43
  entropy
• 19:43
  sources
• 19:43
  .
• 19:44
  The
• 19:44
  physical
• 19:45
  layer
• 19:45
  also
• 19:45
  gained
• 19:46
  a
• 19:46
  consumer
• 19:46
  detection
• 19:47
  story.
• 19:47
  ZuckOff,
• 19:48
  a
• 19:48
  free
• 19:48
  app
• 19:49
  by
• 19:49
  Polish
• 19:49
  developer
• 19:49
  Pawel
• 19:50
  Szydlowski,
• 19:51
  fingerprints
• 19:51
  the
• 19:51
  Bluetooth
• 19:52
  signals
• 19:52
  of
• 19:52
  smart
• 19:53
  glasses
• 19:53
  —
• 19:54
  Ray-Ban
• 19:54
  Meta,
• 19:54
  Oakley
• 19:55
  Meta,
• 19:55
  Snap
• 19:55
  Spectacles
• 19:56
  —
• 19:56
  and
• 19:56
  reports
• 19:57
  when
• 19:57
  a
• 19:57
  pair
• 19:57
  is
• 19:57
  in
• 19:57
  the
• 19:58
  room,
• 19:58
  over
• 19:58
  5,000
• 19:59
  App
• 19:59
  Store
• 19:59
  downloads
• 20:00
  in
• 20:00
  its
• 20:00
  launch
• 20:00
  month.
• 20:01
  Its
• 20:01
  limits
• 20:02
  are
• 20:02
  honest:
• 20:03
  it
• 20:03
  cannot
• 20:03
  tell
• 20:03
  whether
• 20:04
  the
• 20:04
  glasses
• 20:04
  are
• 20:04
  recording
• 20:05
  or
• 20:05
  who
• 20:05
  wears
• 20:06
  them,
• 20:06
  only
• 20:07
  signal
• 20:07
  strength
• 20:07
  for
• 20:08
  rough
• 20:08
  proximity,
• 20:09
  "turning
• 20:09
  an
• 20:09
  invisible
• 20:10
  problem
• 20:10
  into
• 20:11
  a
• 20:11
  partially
• 20:11
  visible
• 20:11
  one"
• 20:11
  —
• 20:13
  and
• 20:13
  it
• 20:13
  is
• 20:13
  hard
• 20:13
  to
• 20:13
  take
• 20:13
  down
• 20:14
  legally,
• 20:14
  since
• 20:14
  it
• 20:14
  intercepts
• 20:15
  nothing.
• 20:16
  Meta,
• 20:16
  for
• 20:17
  its
• 20:17
  part,
• 20:17
  promised
• 20:18
  a
• 20:18
  July
• 20:18
  update
• 20:18
  to
• 20:19
  detect
• 20:19
  when
• 20:19
  the
• 20:19
  recording
• 20:20
  LED
• 20:20
  has
• 20:20
  been
• 20:20
  "physically
• 20:21
  tampered
• 20:21
  with
• 20:22
  or
• 20:22
  destroyed,"
• 20:23
  with
• 20:23
  CTO
• 20:23
  Andrew
• 20:24
  Bosworth
• 20:24
  insisting
• 20:25
  the
• 20:25
  camera
• 20:25
  was
• 20:25
  designed
• 20:26
  to
• 20:26
  be
• 20:26
  noticed
• 20:26
  —
• 20:27
  awkward,
• 20:28
  beside
• 20:28
  the
• 20:28
  product's
• 20:29
  core
• 20:29
  proposition
• 20:30
  that
• 20:30
  the
• 20:30
  camera
• 20:30
  doesn't
• 20:30
  look
• 20:31
  like
• 20:31
  one
• 20:31
  .