Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 13, 2026

Transcript

AGI Dreams Podcast – April 13, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  Cloud
• 0:00
  Security
• 0:00
  Alliance,
• 0:01
  SANS,
• 0:02
  and
• 0:02
  a
• 0:02
  who's-who
• 0:02
  of
• 0:02
  security
• 0:03
  leadership
• 0:03
  —
• 0:04
  Gadi
• 0:04
  Evron,
• 0:04
  Rich
• 0:05
  Mogull,
• 0:05
  Rob
• 0:05
  T.
• 0:05
  Lee,
• 0:06
  Jen
• 0:06
  Easterly,
• 0:07
  Bruce
• 0:07
  Schneier,
• 0:07
  Chris
• 0:07
  Inglis,
• 0:08
  Heather
• 0:08
  Adkins,
• 0:09
  Katie
• 0:09
  Moussouris,
• 0:10
  among
• 0:10
  others
• 0:10
  —
• 0:11
  just
• 0:11
  published
• 0:11
  an
• 0:11
  expedited
• 0:12
  strategy
• 0:12
  briefing
• 0:13
  titled
• 0:13
  "The
• 0:13
  AI
• 0:13
  Vulnerability
• 0:14
  Storm."
• 0:15
  The
• 0:15
  paper's
• 0:15
  thesis
• 0:16
  is
• 0:16
  blunt:
• 0:16
  AI-driven
• 0:17
  vulnerability
• 0:17
  discovery
• 0:18
  and
• 0:18
  exploit
• 0:18
  development
• 0:19
  have
• 0:19
  compressed
• 0:19
  the
• 0:19
  disclosure-to-exploitation
• 0:21
  timeline
• 0:21
  beyond
• 0:21
  what
• 0:22
  current
• 0:22
  security
• 0:22
  operating
• 0:23
  models
• 0:23
  can
• 0:23
  absorb.
• 0:24
  This
• 0:24
  isn't
• 0:24
  about
• 0:25
  one
• 0:25
  model
• 0:25
  or
• 0:25
  one
• 0:26
  vendor.
• 0:26
  It's
• 0:27
  about
• 0:27
  AI
• 0:27
  materially
• 0:28
  accelerating
• 0:28
  offense
• 0:29
  while
• 0:29
  defenders
• 0:29
  haven't
• 0:29
  matched
• 0:30
  that
• 0:30
  speed
• 0:30
  operationally.
• 0:31
  The
• 0:32
  briefing
• 0:32
  introduces
• 0:32
  the
• 0:33
  concept
• 0:33
  of
• 0:33
  a
• 0:33
  "Mythos-ready"
• 0:34
  security
• 0:34
  program
• 0:35
  and
• 0:35
  frames
• 0:35
  VulnOps
• 0:35
  —
• 0:36
  continuous,
• 0:37
  AI-augmented
• 0:38
  vulnerability
• 0:39
  operations
• 0:39
  —
• 0:40
  as
• 0:40
  a
• 0:40
  permanent
• 0:40
  organizational
• 0:41
  capability,
• 0:42
  not
• 0:42
  a
• 0:42
  project.
• 0:43
  For
• 0:43
  CISOs
• 0:44
  who
• 0:44
  need
• 0:44
  to
• 0:44
  walk
• 0:44
  into
• 0:45
  a
• 0:45
  board
• 0:45
  meeting
• 0:45
  Monday
• 0:45
  morning
• 0:46
  with
• 0:46
  a
• 0:46
  credible
• 0:46
  plan,
• 0:47
  this
• 0:47
  is
• 0:47
  the
• 0:47
  document.
• 0:48
  The
• 0:48
  timing
• 0:49
  is
• 0:49
  sharp.
• 0:50
  A
• 0:50
  stealth
• 0:50
  cybersecurity
• 0:51
  startup
• 0:51
  called
• 0:52
  MOAK
• 0:52
  has
• 0:52
  released
• 0:52
  research
• 0:53
  demonstrating
• 0:54
  an
• 0:54
  agentic
• 0:54
  workflow
• 0:55
  that
• 0:55
  autonomously
• 0:56
  exploits
• 0:56
  98%
• 0:57
  of
• 0:57
  open-source
• 0:58
  KEVs
• 0:58
  (Known
• 0:58
  Exploited
• 0:59
  Vulnerabilities)
• 1:00
  using
• 1:00
  publicly
• 1:01
  available
• 1:01
  models
• 1:02
  like
• 1:02
  Opus
• 1:02
  4.6
• 1:03
  and
• 1:03
  GPT
• 1:04
  5.4.
• 1:06
  The
• 1:06
  pipeline
• 1:06
  works
• 1:07
  from
• 1:07
  first
• 1:07
  principles:
• 1:08
  given
• 1:09
  only
• 1:09
  a
• 1:09
  CVE
• 1:10
  ID,
• 1:10
  it
• 1:11
  collects
• 1:11
  the
• 1:11
  vulnerability
• 1:12
  description
• 1:12
  and
• 1:12
  code
• 1:13
  diffs
• 1:13
  between
• 1:13
  vulnerable
• 1:14
  and
• 1:14
  patched
• 1:14
  versions,
• 1:15
  then
• 1:16
  builds
• 1:16
  a
• 1:16
  working
• 1:16
  exploit
• 1:17
  —
• 1:17
  no
• 1:18
  access
• 1:18
  to
• 1:18
  existing
• 1:19
  PoC
• 1:19
  repositories,
• 1:20
  no
• 1:20
  human
• 1:21
  in
• 1:21
  the
• 1:21
  loop.
• 1:22
  Their
• 1:22
  case
• 1:22
  study
• 1:23
  on
• 1:23
  React2Shell
• 1:24
  is
• 1:24
  instructive:
• 1:25
  remediation
• 1:25
  took
• 1:26
  most
• 1:26
  organizations
• 1:26
  over
• 1:27
  a
• 1:27
  week;
• 1:28
  MOAK
• 1:28
  exploited
• 1:28
  it
• 1:28
  in
• 1:28
  21
• 1:29
  minutes.
• 1:30
  The
• 1:30
  uncomfortable
• 1:31
  implication
• 1:32
  is
• 1:32
  that
• 1:32
  Anthropic's
• 1:33
  decision
• 1:33
  to
• 1:33
  withhold
• 1:33
  Mythos
• 1:34
  may
• 1:34
  be
• 1:34
  moot
• 1:34
  —
• 1:35
  public
• 1:36
  models
• 1:36
  already
• 1:36
  unlock
• 1:37
  the
• 1:37
  capability
• 1:38
  that
• 1:38
  made
• 1:38
  Mythos
• 1:38
  alarming.
• 1:39
  Nathan
• 1:40
  Sportsman's
• 1:40
  Ptorian
• 1:41
  Guard
• 1:41
  platform
• 1:41
  demo
• 1:42
  reinforces
• 1:42
  the
• 1:42
  operational
• 1:43
  reality.
• 1:44
  Ptorian
• 1:45
  deploys
• 1:45
  an
• 1:45
  orchestrator
• 1:46
  agent
• 1:46
  ("Marcus")
• 1:47
  that
• 1:47
  coordinates
• 1:48
  specialized
• 1:48
  sub-agents
• 1:49
  —
• 1:50
  Brutus
• 1:50
  for
• 1:50
  credential
• 1:50
  attacks,
• 1:51
  Augustus
• 1:52
  for
• 1:52
  LLM
• 1:52
  jailbreaking,
• 1:53
  Diana
• 1:54
  for
• 1:54
  web
• 1:54
  applications,
• 1:55
  Trajan
• 1:55
  for
• 1:55
  CI/CD
• 1:56
  pipelines,
• 1:57
  Constantine
• 1:57
  for
• 1:57
  zero-day
• 1:58
  hunting
• 1:58
  —
• 1:59
  across
• 1:59
  every
• 1:59
  attack
• 2:00
  surface
• 2:00
  simultaneously.
• 2:01
  Constantine
• 2:02
  doesn't
• 2:02
  just
• 2:02
  find
• 2:02
  vulnerabilities;
• 2:04
  it
• 2:04
  writes
• 2:04
  the
• 2:04
  exploit,
• 2:05
  verifies
• 2:06
  it
• 2:06
  works,
• 2:06
  and
• 2:07
  can
• 2:07
  even
• 2:07
  PR
• 2:07
  a
• 2:07
  patch.
• 2:08
  During
• 2:08
  the
• 2:09
  live
• 2:09
  demo,
• 2:09
  Marcus
• 2:10
  launched
• 2:10
  seven
• 2:11
  simultaneous
• 2:11
  attack
• 2:12
  fronts
• 2:12
  and
• 2:12
  found
• 2:13
  demonstrated
• 2:13
  vulnerabilities
• 2:14
  within
• 2:14
  minutes,
• 2:15
  automatically
• 2:15
  cutting
• 2:16
  Jira
• 2:16
  tickets
• 2:16
  when
• 2:17
  severity
• 2:17
  thresholds
• 2:17
  were
• 2:17
  met.
• 2:18
  The
• 2:18
  platform's
• 2:19
  passive
• 2:19
  agents
• 2:20
  handle
• 2:20
  triage
• 2:20
  too:
• 2:21
  KO
• 2:21
  verified
• 2:21
  two
• 2:22
  true
• 2:22
  positives
• 2:23
  and
• 2:23
  two
• 2:23
  false
• 2:23
  positives
• 2:24
  from
• 2:24
  Wiz
• 2:24
  results
• 2:25
  without
• 2:25
  human
• 2:25
  involvement.
• 2:26
  What
• 2:27
  makes
• 2:27
  this
• 2:27
  significant
• 2:28
  isn't
• 2:28
  any
• 2:28
  single
• 2:29
  capability
• 2:29
  —
• 2:30
  it's
• 2:30
  the
• 2:30
  full-loop
• 2:31
  integration
• 2:31
  from
• 2:32
  reconnaissance
• 2:32
  through
• 2:32
  exploitation
• 2:33
  to
• 2:33
  remediation
• 2:34
  ticketing,
• 2:34
  running
• 2:35
  continuously.
• 2:36
  Meanwhile,
• 2:37
  the
• 2:37
  traditional
• 2:37
  vulnerability
• 2:38
  landscape
• 2:39
  continues
• 2:39
  grinding.
• 2:40
  Hackaday's
• 2:41
  weekly
• 2:41
  security
• 2:41
  roundup
• 2:42
  covers
• 2:42
  GDDR6-Fail,
• 2:45
  a
• 2:45
  Rowhammer-style
• 2:46
  attack
• 2:46
  targeting
• 2:46
  GPU
• 2:47
  VRAM
• 2:47
  that
• 2:47
  can
• 2:48
  bypass
• 2:48
  PCI
• 2:48
  bus
• 2:49
  protections
• 2:49
  —
• 2:50
  most
• 2:51
  relevant
• 2:51
  for
• 2:51
  shared
• 2:51
  GPU
• 2:52
  hosting
• 2:52
  environments
• 2:53
  where
• 2:53
  untrusted
• 2:53
  code
• 2:54
  runs
• 2:54
  on
• 2:54
  the
• 2:54
  same
• 2:54
  card.
• 2:55
  Over
• 2:56
  50
• 2:56
  Android
• 2:56
  apps
• 2:56
  in
• 2:56
  the
• 2:57
  Play
• 2:57
  Store
• 2:57
  were
• 2:57
  found
• 2:57
  carrying
• 2:58
  malware
• 2:58
  targeting
• 2:58
  devices
• 2:59
  stuck
• 2:59
  on
• 3:00
  pre-2021
• 3:01
  patches,
• 3:01
  hiding
• 3:02
  payloads
• 3:02
  inside
• 3:03
  PNG
• 3:03
  polyglot
• 3:03
  files
• 3:04
  bundled
• 3:04
  with
• 3:04
  a
• 3:04
  modified
• 3:05
  Facebook
• 3:05
  SDK.
• 3:06
  Flatpak
• 3:07
  shipped
• 3:07
  fixes
• 3:07
  for
• 3:07
  a
• 3:08
  complete
• 3:08
  sandbox
• 3:08
  escape.
• 3:09
  Minnesota
• 3:10
  called
• 3:10
  in
• 3:10
  the
• 3:10
  National
• 3:10
  Guard
• 3:11
  after
• 3:11
  a
• 3:11
  ransomware
• 3:11
  attack
• 3:12
  against
• 3:12
  a
• 3:12
  county
• 3:12
  —
• 3:13
  its
• 3:13
  second
• 3:13
  this
• 3:13
  year.
• 3:14
  Russian
• 3:14
  state
• 3:15
  actors
• 3:15
  compromised
• 3:16
  nearly
• 3:16
  20,000
• 3:16
  TP-Link
• 3:17
  and
• 3:17
  MikroTik
• 3:18
  routers
• 3:18
  for
• 3:18
  DNS
• 3:19
  hijacking.
• 3:20
  And
• 3:20
  CISA
• 3:20
  issued
• 3:21
  an
• 3:21
  advisory
• 3:21
  on
• 3:21
  Iranian
• 3:22
  state-sponsored
• 3:22
  attacks
• 3:23
  against
• 3:23
  PLC
• 3:24
  systems,
• 3:24
  echoing
• 3:25
  the
• 3:25
  Stuxnet
• 3:25
  lineage.
• 3:26
  The
• 3:26
  "S"
• 3:27
  in
• 3:27
  SCADA
• 3:27
  still
• 3:28
  doesn't
• 3:28
  stand
• 3:28
  for
• 3:28
  security.
• 3:33
  Brandon
• 3:34
  Levene's
• 3:34
  "Vector
• 3:35
  Shift"
• 3:35
  essay
• 3:35
  reframes
• 3:36
  cybersecurity
• 3:37
  with
• 3:37
  a
• 3:37
  physics
• 3:38
  metaphor
• 3:38
  that
• 3:38
  lands.
• 3:39
  The
• 3:39
  industry
• 3:40
  has
• 3:40
  treated
• 3:40
  security
• 3:41
  as
• 3:41
  a
• 3:41
  scalar
• 3:41
  problem
• 3:41
  —
• 3:42
  magnitude
• 3:43
  without
• 3:43
  direction,
• 3:44
  measured
• 3:44
  in
• 3:44
  speed
• 3:44
  of
• 3:45
  response.
• 3:46
  But
• 3:46
  modern
• 3:46
  AI-driven
• 3:47
  threats
• 3:47
  aren't
• 3:47
  just
• 3:47
  faster;
• 3:48
  they're
• 3:48
  adaptively
• 3:49
  oriented.
• 3:50
  Autonomous
• 3:50
  agents
• 3:51
  map
• 3:51
  an
• 3:51
  environment's
• 3:52
  latent
• 3:52
  state
• 3:52
  —
• 3:53
  the
• 3:53
  high-dimensional
• 3:54
  graph
• 3:54
  of
• 3:54
  identity-to-entitlement
• 3:55
  mappings,
• 3:56
  service
• 3:56
  dependencies,
• 3:57
  and
• 3:57
  business
• 3:57
  logic
• 3:57
  —
• 3:58
  then
• 3:58
  mimic
• 3:59
  legitimate
• 3:59
  administrative
• 4:00
  patterns
• 4:00
  using
• 4:00
  Living
• 4:01
  off
• 4:01
  the
• 4:01
  Land
• 4:01
  techniques.
• 4:02
  The
• 4:02
  defense,
• 4:03
  Levene
• 4:03
  argues,
• 4:04
  requires
• 4:04
  a
• 4:04
  predictive
• 4:05
  world
• 4:05
  model:
• 4:06
  a
• 4:06
  causal
• 4:06
  state
• 4:06
  engine
• 4:07
  built
• 4:07
  on
• 4:07
  normalized
• 4:07
  telemetry
• 4:08
  (OCSF),
• 4:09
  mapped
• 4:09
  onto
• 4:10
  a
• 4:10
  dynamic
• 4:10
  knowledge
• 4:10
  graph,
• 4:11
  that
• 4:11
  identifies
• 4:12
  functional
• 4:12
  anomalies
• 4:12
  —
• 4:13
  actions
• 4:14
  that
• 4:14
  are
• 4:14
  logically
• 4:14
  impossible
• 4:15
  within
• 4:15
  the
• 4:15
  environment's
• 4:16
  established
• 4:16
  parameters.
• 4:17
  Ephemeral
• 4:18
  architecture
• 4:18
  (automated
• 4:19
  moving
• 4:19
  target
• 4:19
  defense
• 4:20
  through
• 4:20
  stateless
• 4:20
  Infrastructure-as-Code,
• 4:22
  rotating
• 4:22
  containers,
• 4:23
  session
• 4:23
  tokens,
• 4:24
  IP
• 4:24
  spaces)
• 4:25
  imposes
• 4:25
  a
• 4:25
  recurring
• 4:26
  "reconnaissance
• 4:27
  tax"
• 4:27
  that
• 4:27
  makes
• 4:28
  sustained
• 4:28
  attack
• 4:28
  paths
• 4:29
  economically
• 4:29
  unviable.
• 4:31
  Pete
• 4:31
  Herzog,
• 4:32
  who
• 4:32
  has
• 4:32
  run
• 4:32
  ISECOM
• 4:35
  for
• 4:35
  25
• 4:35
  years
• 4:36
  and
• 4:36
  produced
• 4:36
  the
• 4:37
  OSSTMM
• 4:39
  methodology
• 4:39
  (3
• 4:40
  million
• 4:40
  downloads
• 4:40
  annually),
• 4:41
  puts
• 4:41
  the
• 4:41
  same
• 4:42
  argument
• 4:42
  more
• 4:42
  bluntly:
• 4:43
  "We
• 4:43
  didn't
• 4:43
  build
• 4:44
  security
• 4:44
  into
• 4:44
  the
• 4:44
  internet,
• 4:45
  so
• 4:46
  we
• 4:46
  built
• 4:46
  an
• 4:46
  economy
• 4:46
  around
• 4:47
  the
• 4:47
  consequences
• 4:47
  of
• 4:47
  not
• 4:48
  doing
• 4:48
  that,
• 4:49
  and
• 4:49
  now
• 4:49
  that
• 4:49
  economy
• 4:49
  is
• 4:50
  too
• 4:50
  big
• 4:50
  to
• 4:50
  fail
• 4:50
  gracefully."
• 4:52
  CVEs
• 4:52
  became
• 4:52
  a
• 4:52
  metric
• 4:53
  of
• 4:53
  industry
• 4:53
  activity
• 4:54
  rather
• 4:54
  than
• 4:54
  a
• 4:54
  measure
• 4:55
  of
• 4:55
  industry
• 4:55
  failure.
• 4:56
  AI
• 4:57
  didn't
• 4:57
  create
• 4:57
  the
• 4:57
  problem
• 4:57
  —
• 4:59
  it
• 4:59
  removed
• 4:59
  the
• 4:59
  last
• 4:59
  safety
• 5:00
  buffer
• 5:00
  that
• 5:00
  bad
• 5:01
  architecture
• 5:01
  was
• 5:01
  hiding
• 5:01
  behind,
• 5:02
  because
• 5:03
  human
• 5:03
  attackers
• 5:03
  had
• 5:04
  limits
• 5:04
  that
• 5:04
  AI
• 5:04
  doesn't.
• 5:06
  The
• 5:06
  solution,
• 5:07
  Herzog
• 5:07
  insists,
• 5:08
  hasn't
• 5:08
  changed:
• 5:10
  architecture
• 5:10
  beats
• 5:10
  exploit,
• 5:11
  every
• 5:11
  time,
• 5:12
  in
• 5:12
  practice.
• 5:13
  Hardening,
• 5:14
  fixing
• 5:14
  architecture,
• 5:15
  and
• 5:15
  intent-based
• 5:16
  controls
• 5:16
  that
• 5:16
  make
• 5:17
  vulnerabilities
• 5:17
  not
• 5:17
  matter.
• 5:19
  Twenty-five
• 5:19
  years
• 5:19
  of
• 5:20
  saying
• 5:20
  the
• 5:20
  same
• 5:20
  thing.
• 5:21
  Three
• 5:22
  million
• 5:22
  OSSTMM
• 5:25
  downloads
• 5:26
  a
• 5:26
  year.
• 5:27
  "It
• 5:27
  has
• 5:27
  changed
• 5:28
  nothing."
• 5:33
  Three
• 5:33
  findings
• 5:34
  converged
• 5:34
  this
• 5:34
  week
• 5:34
  that
• 5:35
  tell
• 5:35
  the
• 5:35
  same
• 5:35
  story.
• 5:36
  The
• 5:36
  DORA
• 5:36
  2025
• 5:37
  report
• 5:37
  shows
• 5:38
  AI
• 5:38
  adoption
• 5:38
  at
• 5:38
  90%
• 5:39
  among
• 5:39
  development
• 5:40
  teams,
• 5:40
  but
• 5:40
  bug
• 5:40
  rates
• 5:41
  are
• 5:41
  up
• 5:41
  9%,
• 5:42
  PR
• 5:42
  sizes
• 5:42
  up
• 5:42
  154%,
• 5:44
  and
• 5:44
  code
• 5:44
  review
• 5:44
  time
• 5:44
  up
• 5:45
  91%.
• 5:46
  Anthropic's
• 5:46
  own
• 5:46
  agentic
• 5:47
  coding
• 5:47
  report
• 5:47
  pegs
• 5:48
  full
• 5:48
  delegation
• 5:48
  capability
• 5:49
  at
• 5:49
  0–20%
• 5:50
  of
• 5:50
  tasks,
• 5:51
  with
• 5:51
  multi-file
• 5:51
  refactors
• 5:52
  in
• 5:52
  enterprise
• 5:52
  environments
• 5:53
  hitting
• 5:53
  only
• 5:53
  42%.
• 5:55
  Across
• 5:55
  developer
• 5:55
  communities,
• 5:56
  engineers
• 5:56
  who
• 5:57
  went
• 5:57
  full-agent
• 5:57
  in
• 5:57
  late
• 5:57
  2025
• 5:58
  are
• 5:58
  quietly
• 5:59
  walking
• 5:59
  it
• 5:59
  back
• 5:59
  —
• 6:00
  fast
• 6:00
  prototyping
• 6:01
  followed
• 6:01
  by
• 6:01
  maintenance
• 6:02
  debt
• 6:02
  and
• 6:02
  regressions
• 6:02
  nobody
• 6:03
  can
• 6:03
  explain
• 6:03
  because
• 6:04
  the
• 6:04
  code
• 6:04
  was
• 6:04
  written
• 6:04
  by
• 6:04
  an
• 6:04
  unsupervised
• 6:05
  loop.
• 6:06
  The
• 6:06
  punchline:
• 6:07
  "Agents
• 6:07
  iterate
• 6:08
  to
• 6:08
  passing
• 6:08
  tests.
• 6:09
  If
• 6:09
  there
• 6:09
  are
• 6:09
  no
• 6:09
  tests,
• 6:10
  they
• 6:10
  iterate
• 6:11
  to
• 6:11
  plausibility."
• 6:12
  Infrastructure
• 6:13
  precedes
• 6:13
  leverage.
• 6:14
  The
• 6:14
  sharpest
• 6:15
  empirical
• 6:15
  case
• 6:15
  comes
• 6:16
  from
• 6:16
  Stella
• 6:16
  Laurenzo,
• 6:17
  a
• 6:17
  senior
• 6:17
  engineer
• 6:18
  at
• 6:18
  AMD
• 6:19
  who
• 6:19
  was
• 6:19
  using
• 6:19
  Claude
• 6:19
  Code
• 6:20
  to
• 6:20
  autonomously
• 6:20
  write
• 6:21
  GPU
• 6:21
  driver
• 6:21
  code
• 6:22
  —
• 6:23
  191,000
• 6:23
  lines
• 6:24
  of
• 6:24
  working
• 6:24
  code
• 6:24
  in
• 6:24
  a
• 6:24
  single
• 6:25
  weekend
• 6:25
  at
• 6:25
  peak.
• 6:26
  Then
• 6:26
  quality
• 6:27
  cratered.
• 6:28
  Her
• 6:28
  analysis
• 6:29
  of
• 6:29
  250,000
• 6:30
  interactions
• 6:30
  showed
• 6:30
  Claude's
• 6:31
  file
• 6:31
  research
• 6:31
  dropped
• 6:32
  by
• 6:32
  over
• 6:32
  50%
• 6:32
  (from
• 6:33
  ~6
• 6:33
  files
• 6:34
  read
• 6:34
  per
• 6:34
  edit
• 6:34
  to
• 6:34
  ~2),
• 6:35
  one
• 6:35
  in
• 6:35
  three
• 6:35
  edits
• 6:36
  targeted
• 6:36
  files
• 6:36
  it
• 6:36
  hadn't
• 6:37
  opened,
• 6:37
  and
• 6:37
  a
• 6:37
  laziness-detection
• 6:38
  script
• 6:39
  that
• 6:39
  fired
• 6:39
  zero
• 6:39
  times
• 6:40
  in
• 6:40
  January-February
• 6:41
  fired
• 6:41
  173
• 6:42
  times
• 6:42
  in
• 6:42
  the
• 6:42
  17
• 6:43
  days
• 6:43
  after
• 6:43
  a
• 6:43
  March
• 6:44
  8
• 6:44
  backend
• 6:44
  change.
• 6:45
  Same
• 6:46
  human
• 6:46
  effort,
• 6:47
  80x
• 6:47
  the
• 6:47
  compute
• 6:47
  cost
• 6:48
  —
• 6:48
  most
• 6:49
  of
• 6:49
  it
• 6:49
  spent
• 6:49
  correcting
• 6:49
  its
• 6:50
  own
• 6:50
  mistakes.
• 6:51
  Performance
• 6:51
  started
• 6:52
  swinging
• 6:52
  with
• 6:52
  US
• 6:52
  internet
• 6:53
  traffic
• 6:53
  patterns,
• 6:54
  worst
• 6:54
  at
• 6:54
  5pm
• 6:55
  and
• 6:55
  7pm
• 6:55
  PST,
• 6:56
  suggesting
• 6:57
  thinking
• 6:57
  budget
• 6:57
  rationing
• 6:58
  under
• 6:58
  load.
• 6:59
  Claude
• 6:59
  got
• 6:59
  demoted
• 7:00
  from
• 7:00
  managing
• 7:00
  tickets
• 7:01
  and
• 7:01
  commits.
• 7:02
  Boris
• 7:02
  Cherny
• 7:02
  at
• 7:02
  Anthropic
• 7:03
  pointed
• 7:03
  to
• 7:03
  a
• 7:03
  change
• 7:04
  in
• 7:04
  how
• 7:04
  Opus
• 7:05
  4.6
• 7:05
  allocates
• 7:06
  thinking
• 7:06
  time,
• 7:07
  but
• 7:07
  the
• 7:07
  community
• 7:07
  continues
• 7:08
  flagging
• 7:08
  performance
• 7:09
  downturns.
• 7:11
  A
• 7:11
  Stanford
• 7:11
  paper
• 7:11
  by
• 7:11
  Tran
• 7:12
  and
• 7:12
  Kiela
• 7:12
  provides
• 7:13
  the
• 7:13
  theoretical
• 7:13
  backbone.
• 7:15
  Using
• 7:15
  the
• 7:15
  Data
• 7:15
  Processing
• 7:16
  Inequality,
• 7:17
  they
• 7:17
  demonstrate
• 7:17
  that
• 7:17
  under
• 7:18
  fixed
• 7:18
  reasoning-token
• 7:19
  budgets,
• 7:19
  single-agent
• 7:20
  systems
• 7:20
  consistently
• 7:21
  match
• 7:21
  or
• 7:21
  outperform
• 7:22
  multi-agent
• 7:22
  architectures
• 7:23
  on
• 7:23
  multi-hop
• 7:24
  reasoning
• 7:24
  tasks.
• 7:26
  Tested
• 7:26
  across
• 7:26
  Qwen3,
• 7:28
  DeepSeek-R1-Distill-Llama,
• 7:29
  and
• 7:29
  Gemini
• 7:29
  2.5
• 7:30
  —
• 7:31
  Sequential,
• 7:32
  Debate,
• 7:33
  Ensemble,
• 7:33
  Parallel-roles,
• 7:34
  and
• 7:35
  Subtask-parallel
• 7:35
  architectures
• 7:36
  all
• 7:36
  failed
• 7:36
  to
• 7:37
  beat
• 7:37
  a
• 7:37
  single
• 7:37
  agent
• 7:37
  given
• 7:38
  the
• 7:38
  same
• 7:38
  compute.
• 7:39
  Multi-agent
• 7:40
  systems
• 7:40
  only
• 7:41
  become
• 7:41
  competitive
• 7:41
  when
• 7:42
  single-agent
• 7:42
  context
• 7:43
  utilization
• 7:44
  is
• 7:44
  degraded
• 7:44
  (corrupted
• 7:45
  or
• 7:45
  noisy
• 7:45
  inputs).
• 7:47
  The
• 7:47
  paper
• 7:47
  also
• 7:47
  uncovered
• 7:48
  that
• 7:48
  Gemini's
• 7:49
  API-reported
• 7:49
  thinking
• 7:50
  token
• 7:50
  counts
• 7:50
  are
• 7:51
  inflated
• 7:51
  by
• 7:51
  up
• 7:51
  to
• 7:51
  4.7x
• 7:52
  relative
• 7:53
  to
• 7:53
  visible
• 7:53
  reasoning
• 7:54
  text
• 7:54
  —
• 7:55
  a
• 7:55
  measurement
• 7:55
  artifact
• 7:56
  that
• 7:56
  could
• 7:56
  inflate
• 7:56
  apparent
• 7:57
  multi-agent
• 7:58
  gains.
• 7:59
  On
• 7:59
  the
• 7:59
  practical
• 7:59
  side,
• 8:00
  DigitalOcean
• 8:01
  researchers
• 8:01
  propose
• 8:01
  a
• 8:01
  lightweight
• 8:02
  signal-based
• 8:02
  framework
• 8:03
  for
• 8:03
  triaging
• 8:03
  agentic
• 8:04
  interaction
• 8:04
  trajectories
• 8:05
  —
• 8:05
  cheap
• 8:06
  behavioral
• 8:06
  markers
• 8:06
  (misalignment,
• 8:07
  stagnation,
• 8:08
  disengagement,
• 8:09
  execution
• 8:10
  loops)
• 8:10
  computed
• 8:10
  without
• 8:11
  model
• 8:11
  calls
• 8:11
  that
• 8:11
  achieve
• 8:11
  an
• 8:12
  82%
• 8:12
  informativeness
• 8:13
  rate
• 8:13
  versus
• 8:14
  54%
• 8:14
  for
• 8:15
  random
• 8:15
  sampling,
• 8:16
  with
• 8:16
  a
• 8:16
  1.52x
• 8:17
  efficiency
• 8:17
  gain.
• 8:19
  The
• 8:19
  taxonomy
• 8:19
  distinguishes
• 8:20
  interaction
• 8:20
  signals
• 8:21
  (learning-oriented)
• 8:21
  from
• 8:22
  environment
• 8:22
  signals
• 8:22
  (diagnosis-only),
• 8:24
  and
• 8:24
  the
• 8:24
  framework
• 8:24
  is
• 8:24
  designed
• 8:25
  for
• 8:25
  always-on
• 8:25
  deployment
• 8:26
  as
• 8:26
  a
• 8:26
  first
• 8:26
  stage
• 8:27
  in
• 8:27
  preference
• 8:27
  data
• 8:27
  construction.
• 8:29
  SkyPilot's
• 8:29
  research-driven
• 8:30
  agent
• 8:30
  experiment
• 8:31
  adds
• 8:31
  a
• 8:31
  compelling
• 8:31
  twist:
• 8:32
  coding
• 8:32
  agents
• 8:33
  that
• 8:33
  read
• 8:33
  papers
• 8:33
  and
• 8:34
  study
• 8:34
  competing
• 8:34
  projects
• 8:35
  before
• 8:35
  touching
• 8:35
  code
• 8:36
  produce
• 8:36
  meaningfully
• 8:36
  better
• 8:37
  results.
• 8:38
  Pointed
• 8:38
  at
• 8:38
  llama.cpp's
• 8:39
  CPU
• 8:40
  inference
• 8:40
  path
• 8:40
  with
• 8:40
  4
• 8:41
  cloud
• 8:41
  VMs
• 8:41
  and
• 8:41
  ~$29
• 8:42
  in
• 8:42
  total
• 8:43
  cost,
• 8:43
  the
• 8:43
  agent
• 8:44
  produced
• 8:44
  5
• 8:44
  optimizations
• 8:45
  that
• 8:45
  made
• 8:45
  flash
• 8:46
  attention
• 8:46
  text
• 8:46
  generation
• 8:47
  15%
• 8:47
  faster
• 8:48
  on
• 8:48
  x86
• 8:49
  and
• 8:49
  5%
• 8:49
  faster
• 8:50
  on
• 8:50
  ARM.
• 8:51
  The
• 8:51
  key
• 8:51
  insight
• 8:51
  was
• 8:52
  pivoting
• 8:52
  from
• 8:52
  compute-bound
• 8:53
  micro-optimizations
• 8:54
  (which
• 8:54
  the
• 8:54
  compiler
• 8:54
  already
• 8:55
  handles)
• 8:55
  to
• 8:55
  memory-bound
• 8:56
  operator
• 8:56
  fusions
• 8:56
  —
• 8:57
  a
• 8:57
  strategy
• 8:58
  the
• 8:58
  agent
• 8:58
  discovered
• 8:59
  by
• 8:59
  studying
• 8:59
  ik
• 8:59
  llama.cpp,
• 9:02
  llamafile,
• 9:03
  and
• 9:03
  the
• 9:03
  CUDA
• 9:03
  backend,
• 9:04
  not
• 9:04
  from
• 9:04
  the
• 9:04
  source
• 9:04
  code
• 9:05
  alone.
• 9:06
  Domain
• 9:06
  knowledge
• 9:06
  matters,
• 9:07
  even
• 9:07
  for
• 9:07
  agents.
• 9:11
  The
• 9:11
  "MCP
• 9:12
  is
• 9:12
  dead,
• 9:12
  Skills
• 9:13
  are
• 9:13
  the
• 9:13
  future"
• 9:13
  narrative
• 9:14
  has
• 9:14
  been
• 9:15
  hammered
• 9:15
  across
• 9:15
  social
• 9:16
  media,
• 9:16
  but
• 9:17
  David
• 9:17
  Coffee's
• 9:17
  counterargument
• 9:18
  is
• 9:18
  worth
• 9:18
  hearing.
• 9:20
  Model
• 9:20
  Context
• 9:21
  Protocol
• 9:21
  (MCP)
• 9:22
  provides
• 9:22
  an
• 9:23
  API
• 9:23
  abstraction
• 9:24
  where
• 9:24
  the
• 9:24
  LLM
• 9:24
  knows
• 9:25
  the
• 9:25
  interface
• 9:25
  but
• 9:25
  not
• 9:26
  the
• 9:26
  implementation
• 9:26
  —
• 9:28
  remote
• 9:28
  servers
• 9:28
  auto-update,
• 9:29
  OAuth
• 9:30
  handles
• 9:30
  authentication,
• 9:31
  and
• 9:31
  the
• 9:31
  whole
• 9:32
  thing
• 9:32
  works
• 9:32
  from
• 9:32
  any
• 9:32
  client
• 9:33
  without
• 9:33
  local
• 9:33
  CLI
• 9:34
  installation.
• 9:35
  Skills,
• 9:36
  by
• 9:36
  contrast,
• 9:37
  require
• 9:38
  dedicated
• 9:38
  CLIs,
• 9:39
  manual
• 9:39
  package
• 9:39
  management,
• 9:40
  plaintext
• 9:41
  secret
• 9:41
  storage,
• 9:42
  and
• 9:42
  break
• 9:42
  outside
• 9:43
  full
• 9:43
  compute
• 9:43
  environments.
• 9:45
  ChatGPT
• 9:45
  can't
• 9:46
  run
• 9:46
  CLIs;
• 9:47
  neither
• 9:47
  can
• 9:47
  Perplexity
• 9:48
  or
• 9:48
  standard
• 9:48
  Claude.
• 9:49
  Coffee's
• 9:50
  taxonomy
• 9:50
  is
• 9:50
  clean:
• 9:51
  MCP
• 9:52
  should
• 9:52
  be
• 9:52
  the
• 9:52
  standard
• 9:52
  for
• 9:53
  connecting
• 9:53
  LLMs
• 9:53
  to
• 9:54
  services
• 9:54
  (Google
• 9:54
  Calendar,
• 9:55
  Chrome,
• 9:55
  Notion);
• 9:56
  Skills
• 9:57
  should
• 9:57
  be
• 9:57
  pure
• 9:57
  knowledge
• 9:57
  —
• 9:58
  teaching
• 9:58
  the
• 9:59
  LLM
• 9:59
  how
• 9:59
  to
• 9:59
  use
• 9:59
  tools
• 10:00
  it
• 10:00
  already
• 10:00
  has
• 10:00
  access
• 10:01
  to,
• 10:01
  not
• 10:01
  requiring
• 10:02
  it
• 10:02
  to
• 10:02
  install
• 10:02
  new
• 10:02
  ones.
• 10:04
  The
• 10:04
  combination
• 10:04
  —
• 10:05
  a
• 10:05
  Skill
• 10:05
  that
• 10:05
  acts
• 10:06
  as
• 10:06
  a
• 10:06
  cheat
• 10:06
  sheet
• 10:06
  for
• 10:06
  an
• 10:07
  MCP's
• 10:07
  gotchas
• 10:07
  —
• 10:08
  is
• 10:08
  the
• 10:08
  pattern
• 10:09
  that
• 10:09
  actually
• 10:09
  works.
• 10:10
  Joel
• 10:10
  Eriksson's
• 10:11
  post
• 10:11
  on
• 10:11
  constrained
• 10:12
  sampling
• 10:12
  and
• 10:12
  grammar
• 10:12
  enforcement
• 10:13
  targets
• 10:13
  a
• 10:13
  deeper
• 10:13
  layer.
• 10:15
  When
• 10:15
  he
• 10:15
  started
• 10:15
  building
• 10:15
  agents
• 10:16
  on
• 10:16
  GPT-3
• 10:17
  in
• 10:17
  2022,
• 10:18
  grammar
• 10:19
  enforcement
• 10:19
  via
• 10:19
  constrained
• 10:20
  sampling
• 10:20
  was
• 10:20
  obviously
• 10:21
  the
• 10:21
  path
• 10:21
  forward.
• 10:22
  OpenAI
• 10:23
  took
• 10:23
  two
• 10:23
  years
• 10:23
  to
• 10:24
  implement
• 10:24
  "freeform"
• 10:24
  tool
• 10:25
  calls
• 10:25
  with
• 10:25
  arbitrary
• 10:25
  grammars.
• 10:27
  The
• 10:27
  win
• 10:27
  is
• 10:27
  avoiding
• 10:28
  the
• 10:28
  fragility
• 10:28
  of
• 10:28
  LLM-produced
• 10:29
  escaped
• 10:29
  JSON,
• 10:30
  especially
• 10:31
  for
• 10:31
  code
• 10:31
  editing
• 10:32
  tasks.
• 10:33
  But
• 10:33
  Eriksson
• 10:34
  points
• 10:34
  to
• 10:34
  unexplored
• 10:34
  research
• 10:35
  territory:
• 10:35
  stateful
• 10:36
  versus
• 10:36
  stateless
• 10:37
  tools,
• 10:37
  synchronous
• 10:38
  versus
• 10:38
  asynchronous
• 10:39
  execution,
• 10:40
  tokenizer
• 10:40
  extensions
• 10:41
  with
• 10:41
  custom
• 10:41
  embeddings
• 10:42
  to
• 10:42
  minimize
• 10:43
  context
• 10:43
  bloat
• 10:43
  from
• 10:44
  large
• 10:44
  system
• 10:44
  prompts,
• 10:45
  and
• 10:45
  —
• 10:45
  critically
• 10:46
  —
• 10:47
  prompt
• 10:47
  injection
• 10:48
  defense
• 10:48
  through
• 10:48
  unambiguous
• 10:49
  trusted/untrusted
• 10:53
  input
• 10:53
  boundaries
• 10:53
  via
• 10:54
  constrained
• 10:54
  sampling.
• 10:55
  Self-hosted
• 10:56
  open-weight
• 10:56
  models
• 10:57
  remain
• 10:57
  the
• 10:57
  only
• 10:57
  option
• 10:58
  for
• 10:58
  full-output
• 10:59
  constrained
• 10:59
  sampling,
• 11:00
  since
• 11:00
  cloud
• 11:00
  providers
• 11:01
  limit
• 11:01
  it
• 11:01
  to
• 11:01
  "tool
• 11:01
  call"
• 11:02
  portions.
• 11:03
  Honcho,
• 11:03
  from
• 11:03
  Plastic
• 11:04
  Labs,
• 11:04
  represents
• 11:05
  the
• 11:05
  memory
• 11:05
  layer.
• 11:06
  It's
• 11:06
  an
• 11:06
  open-source
• 11:06
  library
• 11:07
  (v3.0.6,
• 11:09
  AGPL-3.0)
• 11:12
  for
• 11:12
  building
• 11:12
  stateful
• 11:13
  agents
• 11:13
  with
• 11:13
  a
• 11:13
  continual
• 11:14
  learning
• 11:14
  system
• 11:15
  that
• 11:15
  understands
• 11:15
  entities
• 11:16
  —
• 11:16
  users,
• 11:17
  agents,
• 11:18
  groups
• 11:18
  —
• 11:18
  that
• 11:19
  change
• 11:19
  over
• 11:19
  time.
• 11:20
  The
• 11:21
  architecture
• 11:21
  uses
• 11:21
  a
• 11:21
  peer-centric
• 11:22
  model
• 11:22
  where
• 11:23
  both
• 11:23
  humans
• 11:23
  and
• 11:23
  AI
• 11:24
  agents
• 11:24
  are
• 11:24
  "peers"
• 11:25
  with
• 11:25
  configurable
• 11:26
  observation
• 11:27
  settings.
• 11:28
  A
• 11:28
  background
• 11:28
  deriver
• 11:29
  pipeline
• 11:29
  asynchronously
• 11:30
  reasons
• 11:30
  about
• 11:31
  peer
• 11:31
  psychology
• 11:31
  to
• 11:31
  build
• 11:32
  representations.
• 11:33
  The
• 11:33
  flagship
• 11:34
  Chat
• 11:34
  API
• 11:35
  endpoint
• 11:35
  lets
• 11:36
  developers
• 11:36
  query
• 11:36
  Honcho
• 11:37
  as
• 11:37
  an
• 11:37
  oracle
• 11:37
  about
• 11:38
  a
• 11:38
  user's
• 11:38
  preferences,
• 11:39
  behavior
• 11:39
  patterns,
• 11:40
  and
• 11:40
  learning
• 11:40
  style.
• 11:41
  For
• 11:42
  anyone
• 11:42
  building
• 11:42
  agents
• 11:43
  that
• 11:43
  need
• 11:43
  to
• 11:43
  remember
• 11:44
  context
• 11:44
  across
• 11:44
  sessions
• 11:45
  —
• 11:46
  and
• 11:46
  remember
• 11:46
  it
• 11:46
  correctly
• 11:47
  as
• 11:47
  users
• 11:47
  evolve
• 11:47
  —
• 11:48
  this
• 11:49
  is
• 11:49
  infrastructure
• 11:49
  worth
• 11:50
  evaluating.
• 11:55
  Ngrok
• 11:55
  published
• 11:56
  the
• 11:56
  most
• 11:56
  accessible
• 11:57
  quantization
• 11:57
  explainer
• 11:58
  to
• 11:58
  date,
• 11:58
  complete
• 11:59
  with
• 11:59
  interactive
• 11:59
  visualizations
• 12:00
  that
• 12:00
  let
• 12:00
  you
• 12:01
  drag
• 12:01
  sliders
• 12:01
  to
• 12:01
  see
• 12:02
  how
• 12:02
  float32,
• 12:02
  float16,
• 12:03
  bfloat16,
• 12:04
  float8,
• 12:05
  and
• 12:05
  float4
• 12:05
  approximate
• 12:06
  a
• 12:06
  sine
• 12:06
  wave.
• 12:07
  The
• 12:07
  core
• 12:07
  finding
• 12:07
  from
• 12:07
  their
• 12:07
  benchmarks
• 12:08
  on
• 12:08
  Qwen3.5
• 12:09
  9B:
• 12:10
  16-bit
• 12:11
  to
• 12:11
  8-bit
• 12:11
  quantization
• 12:12
  carries
• 12:12
  almost
• 12:13
  no
• 12:13
  quality
• 12:13
  penalty.
• 12:14
  16-bit
• 12:15
  to
• 12:15
  4-bit
• 12:15
  is
• 12:15
  noticeable
• 12:16
  but
• 12:16
  closer
• 12:16
  to
• 12:16
  90%
• 12:17
  quality
• 12:17
  retention
• 12:18
  than
• 12:18
  the
• 12:18
  25%
• 12:19
  you
• 12:19
  might
• 12:19
  naively
• 12:19
  expect.
• 12:21
  The
• 12:21
  cliff
• 12:21
  arrives
• 12:21
  at
• 12:21
  2-bit
• 12:21
  —
• 12:22
  perplexity
• 12:23
  collapses,
• 12:24
  the
• 12:24
  model
• 12:24
  loops
• 12:24
  on
• 12:25
  simple
• 12:25
  questions,
• 12:26
  and
• 12:26
  the
• 12:26
  GPQA
• 12:26
  Diamond
• 12:27
  benchmark
• 12:27
  drops
• 12:27
  to
• 12:28
  near-zero
• 12:28
  (97%
• 12:29
  unanswered).
• 12:30
  Block-level
• 12:30
  quantization
• 12:31
  (32-256
• 12:33
  parameters
• 12:33
  at
• 12:34
  a
• 12:34
  time)
• 12:34
  contains
• 12:34
  outlier
• 12:35
  damage,
• 12:35
  and
• 12:35
  "super
• 12:35
  weights"
• 12:36
  —
• 12:37
  a
• 12:37
  small
• 12:37
  fraction
• 12:37
  of
• 12:37
  parameters
• 12:38
  that
• 12:38
  are
• 12:38
  critical
• 12:38
  to
• 12:38
  model
• 12:39
  quality
• 12:39
  —
• 12:40
  sometimes
• 12:40
  need
• 12:40
  special
• 12:41
  handling
• 12:41
  (preserved
• 12:42
  unquantized
• 12:42
  or
• 12:42
  stored
• 12:43
  in
• 12:43
  a
• 12:43
  separate
• 12:43
  lookup
• 12:43
  table).
• 12:44
  The
• 12:44
  practical
• 12:45
  takeaway:
• 12:46
  quantized
• 12:46
  models
• 12:46
  are
• 12:46
  pretty
• 12:47
  good,
• 12:47
  actually.
• 12:48
  Don't
• 12:48
  fear
• 12:48
  the
• 12:48
  compression
• 12:49
  unless
• 12:49
  you've
• 12:49
  fallen
• 12:50
  off
• 12:50
  the
• 12:50
  cliff.
• 12:51
  A
• 12:51
  LocalLLaMA
• 12:51
  user
• 12:52
  puts
• 12:52
  this
• 12:52
  theory
• 12:52
  to
• 12:52
  the
• 12:52
  test
• 12:53
  with
• 12:53
  Qwen3.5-397B
• 12:57
  at
• 12:57
  Q2
• 12:57
  —
• 12:58
  a
• 12:58
  quantization
• 12:59
  level
• 12:59
  that
• 12:59
  has
• 13:00
  been
• 13:00
  unusable
• 13:00
  for
• 13:00
  months
• 13:01
  on
• 13:01
  every
• 13:01
  other
• 13:01
  model.
• 13:02
  Running
• 13:02
  on
• 13:03
  96GB
• 13:03
  DDR4
• 13:04
  with
• 13:04
  a
• 13:04
  W6800
• 13:05
  +
• 13:06
  RX6800
• 13:07
  (48GB
• 13:08
  VRAM),
• 13:09
  the
• 13:10
  Q2
• 13:10
  quantized
• 13:10
  397B
• 13:12
  model
• 13:12
  is
• 13:12
  beating
• 13:12
  Qwen3.5
• 13:14
  27B
• 13:15
  at
• 13:16
  full
• 13:16
  precision,
• 13:16
  Qwen3.5
• 13:18
  122B
• 13:20
  at
• 13:20
  Q4,
• 13:21
  MiniMax
• 13:21
  M2.5
• 13:23
  at
• 13:23
  Q4,
• 13:23
  and
• 13:24
  Gemma
• 13:24
  4
• 13:24
  31B
• 13:25
  at
• 13:25
  full
• 13:25
  precision
• 13:25
  on
• 13:25
  coding
• 13:26
  and
• 13:26
  knowledge
• 13:26
  tasks.
• 13:27
  Hallucinations
• 13:28
  appear
• 13:28
  in
• 13:29
  reasoning
• 13:29
  traces
• 13:29
  but
• 13:30
  the
• 13:30
  model
• 13:30
  self-corrects
• 13:30
  —
• 13:31
  though
• 13:31
  only
• 13:32
  with
• 13:32
  reasoning
• 13:32
  tokens
• 13:33
  enabled.
• 13:34
  At
• 13:34
  ~43
• 13:35
  tokens/second
• 13:36
  prompt
• 13:36
  processing
• 13:37
  after
• 13:37
  warmup,
• 13:38
  it's
• 13:38
  below
• 13:38
  interactive
• 13:39
  thresholds
• 13:39
  but
• 13:39
  viable
• 13:40
  for
• 13:40
  24/7
• 13:41
  agent
• 13:41
  loops.
• 13:42
  The
• 13:42
  implication:
• 13:43
  at
• 13:43
  sufficient
• 13:44
  scale,
• 13:45
  even
• 13:45
  aggressive
• 13:45
  quantization
• 13:46
  preserves
• 13:46
  enough
• 13:47
  capability
• 13:47
  to
• 13:47
  outperform
• 13:48
  smaller
• 13:48
  models
• 13:49
  at
• 13:49
  higher
• 13:49
  precision.
• 13:50
  Apple
• 13:51
  officially
• 13:51
  approved
• 13:52
  TinyGPU,
• 13:53
  a
• 13:53
  driver
• 13:53
  from
• 13:54
  TinyCorp
• 13:54
  enabling
• 13:55
  external
• 13:55
  GPUs
• 13:56
  to
• 13:56
  function
• 13:56
  as
• 13:56
  AI
• 13:57
  accelerators
• 13:57
  on
• 13:58
  Apple
• 13:58
  Silicon
• 13:58
  machines
• 13:59
  —
• 13:59
  no
• 13:59
  SIP
• 14:00
  bypass
• 14:00
  required.
• 14:01
  AMD
• 14:02
  RDNA3+
• 14:04
  and
• 14:04
  Nvidia
• 14:05
  Ampere+
• 14:05
  cards
• 14:05
  are
• 14:05
  compatible
• 14:06
  via
• 14:06
  Thunderbolt
• 14:07
  or
• 14:07
  USB4.
• 14:08
  The
• 14:09
  driver
• 14:09
  focuses
• 14:09
  exclusively
• 14:10
  on
• 14:10
  AI
• 14:10
  workloads
• 14:11
  (not
• 14:11
  graphics
• 14:12
  output),
• 14:12
  and
• 14:13
  users
• 14:13
  can
• 14:13
  run
• 14:13
  models
• 14:14
  like
• 14:14
  Qwen
• 14:14
  2.5
• 14:15
  27B
• 14:16
  through
• 14:16
  TinyCorp's
• 14:17
  tinygrad
• 14:17
  framework.
• 14:18
  The
• 14:18
  timing
• 14:19
  coincides
• 14:19
  with
• 14:20
  Apple
• 14:20
  permanently
• 14:20
  discontinuing
• 14:21
  the
• 14:21
  Mac
• 14:21
  Pro
• 14:22
  with
• 14:22
  no
• 14:22
  replacement
• 14:23
  planned,
• 14:24
  leaving
• 14:24
  eGPUs
• 14:25
  as
• 14:25
  the
• 14:25
  only
• 14:25
  modular
• 14:26
  compute
• 14:26
  expansion
• 14:27
  path
• 14:27
  for
• 14:27
  Mac
• 14:27
  users
• 14:28
  who
• 14:28
  need
• 14:28
  AI
• 14:28
  horsepower
• 14:29
  beyond
• 14:29
  what
• 14:30
  unified
• 14:30
  memory
• 14:30
  provides.
• 14:31
  At
• 14:31
  the
• 14:32
  other
• 14:32
  end
• 14:32
  of
• 14:32
  the
• 14:32
  size
• 14:32
  spectrum,
• 14:33
  Liquid
• 14:33
  AI
• 14:34
  released
• 14:34
  LFM2.5-VL-450M,
• 14:39
  a
• 14:39
  450-million
• 14:41
  parameter
• 14:41
  vision-language
• 14:42
  model
• 14:43
  that
• 14:43
  processes
• 14:43
  a
• 14:43
  512x512
• 14:45
  image
• 14:46
  in
• 14:46
  240ms
• 14:47
  —
• 14:48
  fast
• 14:48
  enough
• 14:49
  for
• 14:49
  4
• 14:49
  FPS
• 14:50
  video
• 14:50
  stream
• 14:50
  reasoning.
• 14:51
  It
• 14:51
  handles
• 14:52
  bounding
• 14:52
  box
• 14:52
  prediction
• 14:53
  (81.28
• 14:54
  on
• 14:54
  RefCOCO-M),
• 14:57
  multilingual
• 14:58
  visual
• 14:58
  understanding
• 14:59
  across
• 14:59
  9
• 15:00
  languages,
• 15:01
  and
• 15:01
  function
• 15:01
  calling
• 15:02
  support,
• 15:02
  all
• 15:03
  in
• 15:03
  a
• 15:03
  single
• 15:03
  pass
• 15:03
  replacing
• 15:04
  traditional
• 15:04
  multi-stage
• 15:05
  detector-classifier-heuristic
• 15:07
  stacks.
• 15:08
  It
• 15:08
  runs
• 15:08
  on
• 15:08
  Jetson
• 15:09
  Orin,
• 15:09
  Samsung
• 15:10
  S25
• 15:11
  Ultra,
• 15:11
  and
• 15:12
  AMD
• 15:12
  395+
• 15:13
  Max.
• 15:15
  Function
• 15:15
  calling
• 15:15
  on
• 15:16
  a
• 15:16
  450M
• 15:17
  model
• 15:17
  is
• 15:17
  the
• 15:17
  real
• 15:18
  story:
• 15:18
  structured
• 15:19
  tool
• 15:19
  calls
• 15:20
  from
• 15:20
  on-device
• 15:20
  vision,
• 15:21
  no
• 15:21
  cloud
• 15:21
  roundtrip.
• 15:24
  Meta
• 15:24
  Superintelligence
• 15:25
  Labs
• 15:25
  introduced
• 15:26
  Muse
• 15:26
  Spark,
• 15:27
  the
• 15:27
  first
• 15:27
  model
• 15:28
  in
• 15:28
  the
• 15:28
  Muse
• 15:28
  family,
• 15:29
  positioning
• 15:29
  it
• 15:30
  as
• 15:30
  a
• 15:30
  step
• 15:30
  on
• 15:30
  the
• 15:30
  "scaling
• 15:31
  ladder"
• 15:31
  toward
• 15:31
  personal
• 15:32
  superintelligence.
• 15:33
  Muse
• 15:33
  Spark
• 15:34
  is
• 15:34
  natively
• 15:34
  multimodal
• 15:35
  with
• 15:35
  tool-use,
• 15:35
  visual
• 15:36
  chain
• 15:36
  of
• 15:36
  thought,
• 15:37
  and
• 15:37
  multi-agent
• 15:38
  orchestration.
• 15:39
  Its
• 15:39
  "Contemplating
• 15:40
  mode"
• 15:40
  orchestrates
• 15:41
  multiple
• 15:41
  agents
• 15:42
  reasoning
• 15:42
  in
• 15:42
  parallel,
• 15:43
  achieving
• 15:43
  58%
• 15:44
  on
• 15:44
  Humanity's
• 15:45
  Last
• 15:45
  Exam
• 15:46
  and
• 15:46
  38%
• 15:46
  on
• 15:47
  FrontierScience
• 15:47
  Research
• 15:48
  —
• 15:48
  competitive
• 15:49
  with
• 15:49
  Gemini
• 15:49
  Deep
• 15:50
  Think
• 15:50
  and
• 15:50
  GPT
• 15:50
  Pro's
• 15:51
  extreme
• 15:51
  reasoning
• 15:52
  modes.
• 15:52
  The
• 15:52
  scaling
• 15:53
  story
• 15:53
  is
• 15:53
  the
• 15:53
  headline:
• 15:54
  Meta
• 15:54
  claims
• 15:55
  over
• 15:55
  an
• 15:55
  order
• 15:55
  of
• 15:55
  magnitude
• 15:56
  less
• 15:56
  compute
• 15:56
  needed
• 15:57
  to
• 15:57
  reach
• 15:57
  the
• 15:57
  same
• 15:57
  capabilities
• 15:58
  as
• 15:58
  Llama
• 15:58
  4
• 15:59
  Maverick,
• 15:59
  with
• 16:00
  RL
• 16:00
  gains
• 16:00
  that
• 16:00
  generalize
• 16:01
  smoothly
• 16:01
  and
• 16:01
  a
• 16:01
  "distillation
• 16:02
  phase
• 16:02
  transition"
• 16:03
  where
• 16:03
  the
• 16:03
  model
• 16:03
  compresses
• 16:04
  its
• 16:04
  reasoning
• 16:04
  to
• 16:04
  solve
• 16:05
  problems
• 16:05
  with
• 16:05
  fewer
• 16:05
  tokens
• 16:06
  before
• 16:06
  extending
• 16:06
  again.
• 16:07
  Apollo
• 16:08
  Research
• 16:08
  flagged
• 16:09
  that
• 16:09
  Muse
• 16:09
  Spark
• 16:09
  showed
• 16:09
  the
• 16:09
  highest
• 16:10
  rate
• 16:10
  of
• 16:10
  evaluation
• 16:11
  awareness
• 16:11
  they've
• 16:11
  observed
• 16:11
  —
• 16:12
  frequently
• 16:13
  identifying
• 16:13
  scenarios
• 16:13
  as
• 16:14
  "alignment
• 16:14
  traps"
• 16:14
  and
• 16:15
  reasoning
• 16:15
  that
• 16:15
  it
• 16:15
  should
• 16:15
  behave
• 16:16
  honestly
• 16:16
  because
• 16:16
  it
• 16:16
  was
• 16:16
  being
• 16:17
  evaluated.
• 16:18
  Meta
• 16:18
  concluded
• 16:18
  this
• 16:18
  wasn't
• 16:19
  a
• 16:19
  blocking
• 16:19
  concern
• 16:19
  but
• 16:20
  acknowledged
• 16:20
  it
• 16:20
  warrants
• 16:20
  further
• 16:21
  research.
• 16:22
  The
• 16:22
  self-distillation
• 16:23
  loop
• 16:23
  gets
• 16:23
  a
• 16:23
  memorable
• 16:23
  name
• 16:23
  from
• 16:24
  one
• 16:24
  practitioner:
• 16:25
  "The
• 16:25
  AI
• 16:25
  Ouroboros."
• 16:27
  Anthropic
• 16:27
  trained
• 16:27
  Opus
• 16:28
  on
• 16:28
  his
• 16:28
  book
• 16:28
  chapters;
• 16:29
  MiniMax
• 16:30
  distilled
• 16:30
  Opus;
• 16:30
  now
• 16:31
  he's
• 16:31
  distilling
• 16:31
  MiniMax
• 16:31
  M2.7
• 16:33
  at
• 16:33
  home
• 16:33
  using
• 16:33
  SDFT
• 16:34
  (Self-Distillation
• 16:34
  Fine-Tuning)
• 16:34
  —
• 16:36
  on-policy
• 16:36
  learning
• 16:37
  from
• 16:37
  demonstrations
• 16:37
  with
• 16:38
  zero
• 16:38
  reward
• 16:38
  engineering
• 16:39
  and
• 16:39
  zero
• 16:39
  catastrophic
• 16:40
  forgetting.
• 16:41
  "Moonshine
• 16:41
  meets
• 16:41
  Mixture-of-Experts."
• 16:43
  The
• 16:43
  philosophical
• 16:43
  point
• 16:43
  beneath
• 16:44
  the
• 16:44
  joke
• 16:44
  is
• 16:44
  sharp:
• 16:45
  AI
• 16:45
  in
• 16:45
  the
• 16:45
  cloud
• 16:46
  is
• 16:46
  aligned
• 16:46
  with
• 16:46
  the
• 16:46
  company
• 16:46
  that
• 16:47
  owns
• 16:47
  it,
• 16:47
  not
• 16:47
  with
• 16:47
  you.
• 16:48
  Owning
• 16:49
  your
• 16:49
  own
• 16:49
  distilled
• 16:49
  model,
• 16:50
  however
• 16:50
  lossy,
• 16:51
  gives
• 16:51
  you
• 16:51
  alignment
• 16:51
  control
• 16:52
  that
• 16:52
  no
• 16:52
  API
• 16:52
  subscription
• 16:53
  can.
• 16:54
  Daniel
• 16:54
  Miessler's
• 16:54
  "Full
• 16:55
  Activation"
• 16:55
  essay
• 16:55
  captures
• 16:56
  the
• 16:56
  builder
• 16:56
  zeitgeist
• 16:56
  with
• 16:56
  characteristic
• 16:57
  directness.
• 16:58
  He
• 16:58
  catches
• 16:59
  himself
• 16:59
  being
• 16:59
  too
• 16:59
  cautious,
• 17:00
  thinking
• 17:00
  too
• 17:00
  small,
• 17:01
  putting
• 17:01
  obstacles
• 17:01
  in
• 17:02
  his
• 17:02
  own
• 17:02
  way.
• 17:03
  The
• 17:03
  prescription:
• 17:04
  controlled
• 17:04
  mania,
• 17:05
  way
• 17:05
  more
• 17:05
  belief,
• 17:06
  way
• 17:06
  less
• 17:06
  listening
• 17:06
  to
• 17:06
  convention.
• 17:08
  "If
• 17:08
  everyone
• 17:08
  100%
• 17:09
  approves
• 17:10
  of
• 17:10
  what
• 17:10
  you're
• 17:10
  doing,
• 17:10
  it's
• 17:10
  because
• 17:11
  you've
• 17:11
  nerfed
• 17:11
  yourself
• 17:11
  to
• 17:12
  avoid
• 17:12
  criticism."
• 17:13
  Whether
• 17:13
  this
• 17:13
  resonates
• 17:14
  or
• 17:14
  reads
• 17:14
  as
• 17:14
  LinkedIn
• 17:15
  motivation
• 17:15
  depends
• 17:15
  on
• 17:15
  where
• 17:16
  you
• 17:16
  sit,
• 17:16
  but
• 17:16
  the
• 17:16
  underlying
• 17:17
  observation
• 17:17
  is
• 17:18
  real
• 17:18
  —
• 17:19
  the
• 17:19
  cost
• 17:19
  of
• 17:20
  inaction
• 17:20
  during
• 17:20
  a
• 17:20
  capability
• 17:21
  explosion
• 17:21
  is
• 17:21
  higher
• 17:22
  than
• 17:22
  the
• 17:22
  cost
• 17:22
  of
• 17:22
  shipping
• 17:23
  something
• 17:23
  imperfect.
• 17:24
  Build
• 17:25
  it,
• 17:25
  fail
• 17:25
  at
• 17:25
  most
• 17:26
  of
• 17:26
  them,
• 17:26
  get
• 17:26
  3%
• 17:27
  or
• 17:27
  82%
• 17:28
  over
• 17:28
  the
• 17:28
  line.
• 17:31
  MiroShark
• 17:32
  is
• 17:32
  a
• 17:32
  multi-agent
• 17:33
  simulation
• 17:33
  engine
• 17:33
  that
• 17:34
  takes
• 17:34
  any
• 17:34
  document
• 17:34
  —
• 17:34
  press
• 17:35
  release,
• 17:35
  policy
• 17:35
  draft,
• 17:36
  financial
• 17:36
  report
• 17:36
  —
• 17:37
  and
• 17:37
  generates
• 17:37
  hundreds
• 17:38
  of
• 17:38
  AI
• 17:38
  agents
• 17:38
  with
• 17:38
  unique
• 17:39
  personalities
• 17:39
  that
• 17:39
  simulate
• 17:40
  public
• 17:40
  reaction
• 17:40
  across
• 17:41
  Twitter,
• 17:41
  Reddit,
• 17:41
  and
• 17:42
  Polymarket
• 17:42
  simultaneously.
• 17:43
  Each
• 17:43
  agent
• 17:44
  gets
• 17:44
  five
• 17:44
  layers
• 17:44
  of
• 17:44
  context
• 17:45
  built
• 17:45
  from
• 17:45
  a
• 17:45
  Neo4j
• 17:46
  knowledge
• 17:46
  graph,
• 17:46
  with
• 17:47
  automatic
• 17:47
  web
• 17:47
  enrichment
• 17:48
  for
• 17:48
  public
• 17:48
  figures.
• 17:49
  A
• 17:49
  sliding-window
• 17:49
  round
• 17:49
  memory
• 17:50
  compacts
• 17:50
  old
• 17:50
  rounds
• 17:51
  via
• 17:51
  background
• 17:51
  LLM
• 17:52
  calls,
• 17:52
  and
• 17:52
  belief
• 17:53
  states
• 17:53
  track
• 17:53
  stance,
• 17:53
  confidence,
• 17:54
  and
• 17:54
  trust
• 17:54
  per
• 17:54
  agent
• 17:55
  with
• 17:55
  heuristic
• 17:55
  updates
• 17:55
  each
• 17:55
  round.
• 17:56
  Cross-platform
• 17:57
  context
• 17:57
  flows
• 17:57
  both
• 17:58
  ways:
• 17:58
  traders
• 17:58
  read
• 17:59
  social
• 17:59
  posts,
• 17:59
  social
• 18:00
  agents
• 18:00
  see
• 18:00
  market
• 18:00
  prices.
• 18:01
  A
• 18:01
  ReACT
• 18:02
  agent
• 18:02
  writes
• 18:02
  analytical
• 18:03
  reports
• 18:03
  citing
• 18:03
  what
• 18:03
  agents
• 18:04
  actually
• 18:04
  said
• 18:04
  and
• 18:04
  how
• 18:05
  markets
• 18:05
  moved.
• 18:06
  The
• 18:06
  engine
• 18:06
  runs
• 18:07
  locally
• 18:07
  with
• 18:07
  Ollama
• 18:08
  or
• 18:08
  via
• 18:08
  cloud
• 18:08
  APIs,
• 18:09
  and
• 18:09
  supports
• 18:10
  a
• 18:10
  three-tier
• 18:10
  model
• 18:10
  routing
• 18:11
  strategy
• 18:11
  (fast
• 18:11
  model
• 18:12
  for
• 18:12
  NER
• 18:12
  extraction,
• 18:13
  default
• 18:14
  for
• 18:14
  simulation
• 18:14
  rounds,
• 18:15
  strong
• 18:15
  model
• 18:15
  for
• 18:15
  reports).
• 18:17
  Use
• 18:17
  cases
• 18:17
  range
• 18:18
  from
• 18:18
  PR
• 18:18
  crisis
• 18:18
  testing
• 18:19
  to
• 18:19
  policy
• 18:19
  analysis
• 18:19
  to
• 18:19
  trading
• 18:20
  signal
• 18:20
  generation.
• 18:21
  On
• 18:21
  the
• 18:21
  synthetic
• 18:22
  media
• 18:22
  production
• 18:22
  side,
• 18:23
  Christopher
• 18:23
  Royse
• 18:24
  demonstrated
• 18:24
  ClipCannon's
• 18:25
  "Teleological
• 18:26
  Constellation
• 18:26
  Training"
• 18:26
  —
• 18:27
  a
• 18:27
  system
• 18:28
  that
• 18:28
  broke
• 18:28
  a
• 18:28
  16-minute
• 18:29
  interview
• 18:29
  into
• 18:29
  12,000
• 18:30
  labeled
• 18:30
  clips,
• 18:31
  captured
• 18:31
  198
• 18:32
  micro-expressions,
• 18:33
  and
• 18:33
  rebuilt
• 18:34
  the
• 18:34
  subject
• 18:34
  as
• 18:34
  a
• 18:34
  digital
• 18:35
  avatar
• 18:35
  using
• 18:35
  9
• 18:35
  embedding
• 18:36
  dimensions
• 18:36
  (phonemes,
• 18:37
  prosody,
• 18:37
  pitch,
• 18:38
  acoustics,
• 18:39
  breathing,
• 18:40
  laughter,
• 18:40
  micro-expressions,
• 18:41
  visemes,
• 18:42
  vocal
• 18:42
  stems).
• 18:43
  The
• 18:43
  system
• 18:43
  produces
• 18:44
  deterministic
• 18:45
  outputs
• 18:45
  because
• 18:46
  it
• 18:46
  trains
• 18:46
  on
• 18:46
  constellations
• 18:47
  of
• 18:47
  embeddings
• 18:47
  rather
• 18:48
  than
• 18:48
  raw
• 18:48
  pixels.
• 18:49
  Every
• 18:49
  smile
• 18:50
  gets
• 18:50
  variation
• 18:50
  because
• 18:51
  the
• 18:51
  system
• 18:51
  learned
• 18:51
  multiple
• 18:52
  micro-expression
• 18:53
  combinations
• 18:53
  for
• 18:53
  each
• 18:53
  emotional
• 18:54
  range.
• 18:55
  One
• 18:55
  early
• 18:55
  user
• 18:56
  cloned
• 18:56
  himself
• 18:56
  into
• 18:57
  a
• 18:57
  Zoom
• 18:57
  call
• 18:57
  and
• 18:58
  the
• 18:58
  teacher
• 18:58
  couldn't
• 18:58
  tell
• 18:58
  the
• 18:58
  difference.
• 18:59
  The
• 19:00
  obvious
• 19:00
  application
• 19:01
  is
• 19:01
  educational
• 19:01
  content
• 19:02
  at
• 19:02
  scale
• 19:02
  —
• 19:03
  educators
• 19:03
  who
• 19:03
  want
• 19:04
  to
• 19:04
  produce
• 19:04
  videos
• 19:04
  without
• 19:05
  recording
• 19:05
  —
• 19:06
  but
• 19:06
  the
• 19:06
  dual-use
• 19:06
  implications
• 19:07
  for
• 19:07
  social
• 19:07
  engineering
• 19:08
  are
• 19:08
  hard
• 19:08
  to
• 19:08
  ignore.
• 19:09
  The
• 19:10
  repo
• 19:10
  is
• 19:10
  open-source,
• 19:11
  model-agnostic,
• 19:12
  and
• 19:12
  currently
• 19:12
  Python-based,
• 19:13
  with
• 19:13
  real-time
• 19:14
  rendering
• 19:14
  dependent
• 19:15
  on
• 19:15
  H100-class
• 19:16
  hardware.