Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – May 14, 2026

Transcript

AGI Dreams Podcast – May 14, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  question
• 0:00
  of
• 0:00
  whether
• 0:01
  AI
• 0:01
  can
• 0:01
  find
• 0:01
  real
• 0:02
  vulnerabilities
• 0:02
  has
• 0:02
  been
• 0:03
  answered
• 0:03
  —
• 0:04
  the
• 0:04
  field
• 0:04
  now
• 0:04
  has
• 0:04
  over
• 0:05
  800
• 0:05
  verified
• 0:06
  findings
• 0:06
  and
• 0:06
  200+
• 0:07
  assigned
• 0:07
  CVEs
• 0:08
  across
• 0:08
  heavily
• 0:09
  scrutinized
• 0:09
  codebases.
• 0:10
  The
• 0:10
  question
• 0:11
  that
• 0:11
  matters
• 0:11
  now
• 0:11
  is
• 0:12
  how
• 0:12
  to
• 0:12
  evaluate,
• 0:12
  measure,
• 0:13
  and
• 0:13
  standardize
• 0:14
  AI-driven
• 0:14
  security
• 0:15
  work.
• 0:15
  XBOW's
• 0:16
  evaluation
• 0:17
  of
• 0:17
  Anthropic's
• 0:17
  Mythos
• 0:17
  Preview
• 0:18
  model
• 0:18
  marks
• 0:18
  a
• 0:18
  significant
• 0:19
  data
• 0:19
  point
• 0:19
  in
• 0:19
  that
• 0:20
  transition.
• 0:21
  Across
• 0:21
  XBOW's
• 0:21
  web
• 0:22
  exploit
• 0:22
  benchmark,
• 0:22
  Mythos
• 0:23
  Preview
• 0:23
  cut
• 0:23
  false
• 0:24
  negatives
• 0:24
  by
• 0:24
  42%
• 0:25
  compared
• 0:25
  to
• 0:25
  Opus
• 0:26
  4.6,
• 0:27
  and
• 0:27
  by
• 0:27
  55%
• 0:28
  when
• 0:28
  given
• 0:28
  source
• 0:28
  code
• 0:28
  access.
• 0:29
  Token-for-token,
• 0:30
  it
• 0:30
  "hones
• 0:30
  in
• 0:30
  on
• 0:31
  the
• 0:31
  vulnerability
• 0:31
  with
• 0:31
  absolutely
• 0:32
  unprecedented
• 0:32
  precision,"
• 0:33
  according
• 0:34
  to
• 0:34
  XBOW's
• 0:34
  team
• 0:34
  of
• 0:34
  ten
• 0:34
  security
• 0:35
  experts.
• 0:36
  The
• 0:36
  model
• 0:36
  showed
• 0:36
  particular
• 0:37
  strength
• 0:37
  in
• 0:37
  native-code
• 0:38
  vulnerability
• 0:38
  discovery
• 0:39
  and
• 0:39
  reverse
• 0:39
  engineering,
• 0:40
  reasoning
• 0:40
  through
• 0:40
  unusual
• 0:41
  firmware
• 0:41
  and
• 0:41
  embedded
• 0:42
  systems
• 0:42
  contexts
• 0:43
  that
• 0:43
  required
• 0:43
  more
• 0:43
  than
• 0:43
  rote
• 0:43
  pattern
• 0:44
  matching.
• 0:44
  But
• 0:45
  XBOW's
• 0:45
  evaluation
• 0:45
  also
• 0:46
  surfaced
• 0:46
  important
• 0:46
  limitations.
• 0:47
  Mythos
• 0:48
  Preview's
• 0:48
  judgment
• 0:48
  was
• 0:49
  "mixed"
• 0:49
  —
• 0:50
  it
• 0:50
  rejected
• 0:50
  false
• 0:50
  positives
• 0:51
  well
• 0:51
  but
• 0:52
  sometimes
• 0:52
  lost
• 0:52
  true
• 0:52
  positives
• 0:53
  when
• 0:53
  evidence
• 0:53
  didn't
• 0:54
  formally
• 0:54
  satisfy
• 0:54
  its
• 0:54
  criteria.
• 0:56
  On
• 0:56
  command
• 0:56
  safety
• 0:57
  benchmarks,
• 0:57
  it
• 0:58
  scored
• 0:58
  77.8%
• 0:59
  accuracy
• 0:59
  compared
• 1:00
  to
• 1:00
  Haiku
• 1:00
  4.5's
• 1:01
  90.1%,
• 1:02
  prioritizing
• 1:03
  the
• 1:03
  letter
• 1:03
  of
• 1:03
  rules
• 1:03
  over
• 1:03
  their
• 1:03
  spirit.
• 1:04
  And
• 1:04
  when
• 1:05
  normalized
• 1:05
  by
• 1:05
  estimated
• 1:06
  cost
• 1:06
  —
• 1:06
  Mythos
• 1:07
  Preview
• 1:07
  is
• 1:07
  expected
• 1:08
  to
• 1:08
  be
• 1:08
  5x
• 1:08
  the
• 1:08
  price
• 1:09
  of
• 1:09
  Opus
• 1:09
  —
• 1:09
  the
• 1:10
  cost-effectiveness
• 1:10
  picture
• 1:11
  shifts
• 1:11
  significantly.
• 1:12
  XBOW's
• 1:13
  data
• 1:13
  suggests
• 1:13
  that
• 1:14
  in
• 1:14
  many
• 1:14
  cases,
• 1:15
  giving
• 1:15
  a
• 1:15
  cheaper
• 1:15
  model
• 1:16
  more
• 1:16
  iterations
• 1:16
  beats
• 1:17
  giving
• 1:17
  Mythos
• 1:17
  Preview
• 1:17
  one
• 1:18
  expensive
• 1:18
  shot.
• 1:19
  The
• 1:19
  real
• 1:19
  value
• 1:20
  proposition,
• 1:20
  XBOW
• 1:21
  argues,
• 1:22
  is
• 1:22
  combining
• 1:22
  Mythos's
• 1:23
  source-code
• 1:23
  analysis
• 1:24
  power
• 1:24
  with
• 1:24
  live-site
• 1:25
  validation
• 1:25
  infrastructure
• 1:26
  rather
• 1:26
  than
• 1:26
  treating
• 1:26
  it
• 1:27
  as
• 1:27
  a
• 1:27
  standalone
• 1:27
  silver
• 1:28
  bullet.
• 1:29
  On
• 1:29
  the
• 1:29
  audit
• 1:29
  framework
• 1:30
  side,
• 1:30
  SPECA
• 1:31
  introduces
• 1:31
  specification-anchored
• 1:32
  security
• 1:32
  auditing
• 1:33
  —
• 1:33
  deriving
• 1:34
  typed
• 1:34
  security
• 1:34
  properties
• 1:35
  from
• 1:35
  natural-language
• 1:35
  specifications
• 1:36
  rather
• 1:37
  than
• 1:37
  scanning
• 1:37
  for
• 1:37
  known
• 1:37
  bug
• 1:38
  patterns.
• 1:39
  In
• 1:39
  the
• 1:39
  Sherlock
• 1:39
  Ethereum
• 1:39
  Fusaka
• 1:40
  audit
• 1:40
  contest
• 1:41
  (366
• 1:42
  submissions,
• 1:43
  10
• 1:43
  implementations),
• 1:44
  SPECA
• 1:44
  recovered
• 1:45
  all
• 1:45
  15
• 1:45
  in-scope
• 1:45
  vulnerabilities
• 1:46
  and
• 1:46
  discovered
• 1:47
  4
• 1:47
  novel
• 1:47
  bugs
• 1:48
  confirmed
• 1:48
  by
• 1:48
  developer
• 1:48
  fix
• 1:49
  commits,
• 1:49
  including
• 1:50
  a
• 1:50
  cryptographic
• 1:50
  invariant
• 1:51
  violation
• 1:51
  missed
• 1:51
  by
• 1:52
  every
• 1:52
  human
• 1:52
  auditor
• 1:52
  in
• 1:52
  the
• 1:53
  contest.
• 1:54
  On
• 1:54
  the
• 1:54
  RepoAudit
• 1:54
  C/C++
• 1:56
  benchmark,
• 1:56
  it
• 1:57
  matched
• 1:57
  the
• 1:57
  best
• 1:57
  published
• 1:58
  precision
• 1:58
  at
• 1:58
  88.9%
• 2:00
  while
• 2:00
  surfacing
• 2:00
  12
• 2:01
  author-validated
• 2:01
  candidates
• 2:02
  beyond
• 2:02
  ground
• 2:02
  truth.
• 2:03
  Crucially,
• 2:04
  all
• 2:04
  false
• 2:04
  positives
• 2:05
  traced
• 2:05
  to
• 2:05
  three
• 2:06
  interpretable
• 2:06
  root
• 2:06
  causes
• 2:07
  mapped
• 2:07
  to
• 2:07
  specific
• 2:07
  pipeline
• 2:08
  phases
• 2:08
  —
• 2:09
  not
• 2:09
  the
• 2:09
  usual
• 2:09
  opaque
• 2:10
  "the
• 2:10
  model
• 2:10
  thought
• 2:11
  this
• 2:11
  was
• 2:11
  a
• 2:11
  bug."
• 2:12
  Whether
• 2:12
  any
• 2:12
  of
• 2:12
  this
• 2:12
  measurement
• 2:13
  actually
• 2:13
  means
• 2:13
  anything
• 2:14
  is
• 2:14
  the
• 2:14
  question
• 2:14
  the
• 2:14
  Berryville
• 2:15
  Institute
• 2:15
  of
• 2:15
  Machine
• 2:15
  Learning
• 2:16
  is
• 2:16
  raising.
• 2:17
  BIML's
• 2:17
  new
• 2:17
  paper,
• 2:18
  "No
• 2:18
  Security
• 2:18
  Meter
• 2:19
  for
• 2:19
  AI,"
• 2:19
  argues
• 2:20
  that
• 2:20
  AI
• 2:20
  benchmarks
• 2:20
  are
• 2:21
  "fundamentally
• 2:21
  broken"
• 2:21
  —
• 2:22
  contaminated
• 2:23
  by
• 2:23
  their
• 2:23
  own
• 2:23
  publication,
• 2:24
  measuring
• 2:24
  narrow
• 2:24
  performance
• 2:25
  rather
• 2:25
  than
• 2:25
  actual
• 2:26
  capability.
• 2:27
  Co-author
• 2:28
  Gary
• 2:28
  McGraw
• 2:28
  draws
• 2:29
  on
• 2:29
  three
• 2:29
  decades
• 2:29
  of
• 2:29
  software
• 2:30
  security
• 2:30
  history:
• 2:31
  "In
• 2:31
  the
• 2:31
  late
• 2:32
  1990s,
• 2:32
  pen
• 2:33
  testing
• 2:33
  was
• 2:33
  treated
• 2:33
  as
• 2:34
  a
• 2:34
  security
• 2:34
  meter,
• 2:34
  but
• 2:35
  it's
• 2:35
  actually
• 2:35
  just
• 2:35
  a
• 2:35
  badness-ometer
• 2:36
  that
• 2:36
  tops
• 2:36
  out
• 2:37
  at
• 2:37
  'who
• 2:37
  knows?'
• 2:38
  Benchmark
• 2:39
  scores
• 2:39
  for
• 2:39
  AI
• 2:39
  security
• 2:40
  are
• 2:40
  even
• 2:40
  worse."
• 2:42
  Recent
• 2:42
  research
• 2:43
  confirms
• 2:43
  the
• 2:43
  concern
• 2:43
  —
• 2:44
  eight
• 2:45
  prominent
• 2:45
  AI
• 2:45
  agent
• 2:46
  benchmarks
• 2:46
  can
• 2:46
  be
• 2:46
  exploited
• 2:47
  to
• 2:47
  achieve
• 2:47
  near-perfect
• 2:48
  scores
• 2:48
  without
• 2:49
  completing
• 2:49
  a
• 2:49
  single
• 2:49
  task.
• 2:51
  Meanwhile,
• 2:52
  the
• 2:52
  dnsmasq
• 2:52
  maintainer
• 2:53
  Simon
• 2:53
  Kelley
• 2:53
  is
• 2:54
  living
• 2:54
  the
• 2:54
  downstream
• 2:54
  reality
• 2:55
  of
• 2:55
  AI-powered
• 2:55
  vulnerability
• 2:56
  discovery.
• 2:57
  CERT
• 2:57
  released
• 2:58
  six
• 2:58
  CVEs
• 2:58
  for
• 2:58
  serious,
• 2:59
  long-standing
• 3:00
  dnsmasq
• 3:01
  bugs
• 3:01
  —
• 3:01
  all
• 3:02
  found
• 3:02
  in
• 3:02
  what
• 3:02
  Kelley
• 3:02
  describes
• 3:03
  as
• 3:03
  a
• 3:03
  "tsunami
• 3:04
  of
• 3:04
  AI-generated
• 3:05
  bug
• 3:05
  reports"
• 3:05
  that
• 3:06
  shows
• 3:06
  "no
• 3:06
  signs
• 3:06
  of
• 3:06
  stopping."
• 3:08
  He
• 3:08
  spent
• 3:08
  months
• 3:08
  triaging
• 3:09
  reports,
• 3:10
  weeding
• 3:10
  out
• 3:10
  duplicates
• 3:11
  ("so
• 3:11
  many
• 3:11
  duplicates!"),
• 3:12
  and
• 3:12
  making
• 3:12
  judgment
• 3:13
  calls
• 3:13
  about
• 3:13
  which
• 3:13
  bugs
• 3:14
  needed
• 3:14
  vendor
• 3:14
  pre-disclosure
• 3:15
  versus
• 3:15
  immediate
• 3:16
  public
• 3:16
  fix.
• 3:17
  His
• 3:17
  candid
• 3:17
  assessment:
• 3:18
  "given
• 3:19
  the
• 3:19
  number
• 3:19
  of
• 3:19
  times
• 3:19
  'good
• 3:20
  guys'
• 3:20
  have
• 3:20
  found
• 3:20
  these
• 3:20
  bugs,
• 3:21
  there's
• 3:21
  no
• 3:22
  doubt
• 3:22
  that
• 3:22
  'bad
• 3:22
  guys'
• 3:22
  have
• 3:23
  been
• 3:23
  able
• 3:23
  to
• 3:23
  do
• 3:23
  the
• 3:23
  same,
• 3:24
  so
• 3:24
  long
• 3:24
  embargoes
• 3:25
  seem
• 3:25
  kind
• 3:25
  of
• 3:25
  pointless."
• 3:29
  AI
• 3:30
  coding
• 3:30
  tools
• 3:30
  have
• 3:30
  become
• 3:31
  the
• 3:31
  most
• 3:31
  trusted
• 3:31
  environment
• 3:32
  on
• 3:32
  a
• 3:32
  developer's
• 3:33
  machine
• 3:33
  —
• 3:33
  and
• 3:33
  the
• 3:33
  most
• 3:34
  attractive
• 3:34
  target.
• 3:35
  A
• 3:35
  vulnerability
• 3:35
  in
• 3:36
  Claude
• 3:36
  Code,
• 3:36
  now
• 3:36
  fixed
• 3:37
  in
• 3:37
  version
• 3:37
  2.1.118,
• 3:39
  demonstrated
• 3:40
  how
• 3:40
  that
• 3:40
  trust
• 3:40
  can
• 3:40
  be
• 3:41
  weaponized.
• 3:41
  Security
• 3:42
  researcher
• 3:42
  joernchen
• 3:43
  discovered
• 3:43
  that
• 3:44
  Claude
• 3:44
  Code's
• 3:44
  settings
• 3:44
  parser
• 3:45
  eagerly
• 3:45
  scanned
• 3:45
  the
• 3:46
  entire
• 3:46
  command
• 3:46
  line
• 3:46
  for
• 3:47
  flags,
• 3:47
  without
• 3:47
  distinguishing
• 3:48
  between
• 3:48
  actual
• 3:49
  CLI
• 3:49
  options
• 3:49
  and
• 3:49
  values
• 3:50
  passed
• 3:50
  as
• 3:50
  arguments
• 3:50
  to
• 3:51
  other
• 3:51
  flags.
• 3:52
  Combined
• 3:52
  with
• 3:52
  the
• 3:52
  deeplink
• 3:53
  handler
• 3:53
  for
• 3:53
  claude-cli://
• 3:54
  URIs,
• 3:55
  this
• 3:55
  created
• 3:55
  a
• 3:55
  clean
• 3:56
  injection
• 3:56
  path:
• 3:57
  an
• 3:57
  attacker
• 3:57
  could
• 3:57
  craft
• 3:57
  a
• 3:58
  deeplink
• 3:58
  that
• 3:58
  injected
• 3:58
  arbitrary
• 3:59
  settings
• 3:59
  —
• 4:00
  including
• 4:00
  shell
• 4:00
  command
• 4:00
  execution
• 4:01
  via
• 4:01
  SessionStart
• 4:01
  hooks
• 4:02
  —
• 4:02
  into
• 4:02
  the
• 4:02
  spawned
• 4:03
  Claude
• 4:03
  Code
• 4:03
  instance.
• 4:04
  If
• 4:04
  the
• 4:04
  deeplink's
• 4:05
  repo
• 4:05
  parameter
• 4:05
  pointed
• 4:06
  to
• 4:06
  a
• 4:06
  repository
• 4:06
  the
• 4:06
  user
• 4:07
  had
• 4:07
  already
• 4:07
  trusted,
• 4:08
  execution
• 4:08
  happened
• 4:09
  without
• 4:09
  any
• 4:09
  warning
• 4:09
  prompt.
• 4:10
  The
• 4:10
  deeper
• 4:10
  systemic
• 4:11
  risk
• 4:11
  comes
• 4:11
  from
• 4:11
  what
• 4:11
  these
• 4:12
  tools
• 4:12
  accumulate
• 4:12
  during
• 4:13
  normal
• 4:13
  use.
• 4:14
  GhostType,
• 4:15
  a
• 4:15
  forensic
• 4:15
  scanner
• 4:15
  built
• 4:15
  by
• 4:16
  a
• 4:16
  red
• 4:16
  team
• 4:16
  practitioner,
• 4:17
  highlights
• 4:17
  the
• 4:17
  credential
• 4:18
  exposure
• 4:18
  problem
• 4:18
  hiding
• 4:19
  in
• 4:19
  plain
• 4:19
  sight
• 4:19
  across
• 4:20
  AI
• 4:20
  coding
• 4:20
  tools.
• 4:21
  Every
• 4:21
  Claude
• 4:21
  Code
• 4:22
  session
• 4:22
  is
• 4:22
  stored
• 4:22
  in
• 4:22
  plaintext
• 4:23
  —
• 4:24
  one
• 4:24
  conversation
• 4:24
  per
• 4:25
  file,
• 4:25
  human-readable,
• 4:26
  every
• 4:26
  message
• 4:27
  and
• 4:27
  tool
• 4:27
  call
• 4:27
  preserved.
• 4:28
  Cursor
• 4:28
  keeps
• 4:29
  data
• 4:29
  in
• 4:29
  SQLite
• 4:29
  databases
• 4:30
  that
• 4:30
  can
• 4:30
  grow
• 4:30
  beyond
• 4:30
  1
• 4:30
  GB.
• 4:32
  Codex
• 4:32
  stores
• 4:32
  state
• 4:33
  in
• 4:33
  plaintext
• 4:33
  JSONL
• 4:33
  rollout
• 4:34
  files.
• 4:35
  A
• 4:35
  GitHub
• 4:35
  issue
• 4:35
  documents
• 4:36
  a
• 4:36
  researcher
• 4:36
  finding
• 4:36
  five
• 4:37
  distinct
• 4:37
  secrets
• 4:37
  across
• 4:38
  34
• 4:38
  session
• 4:38
  files
• 4:39
  after
• 4:39
  just
• 4:39
  30
• 4:40
  days
• 4:40
  of
• 4:40
  normal
• 4:40
  usage.
• 4:41
  GhostType
• 4:42
  scans
• 4:42
  these
• 4:42
  histories
• 4:43
  with
• 4:43
  30+
• 4:43
  regex
• 4:44
  patterns
• 4:44
  and
• 4:44
  heuristic
• 4:44
  detection,
• 4:45
  extracting
• 4:46
  AWS
• 4:46
  keys,
• 4:47
  GitHub
• 4:47
  PATs,
• 4:48
  Stripe
• 4:49
  secrets,
• 4:49
  database
• 4:49
  connection
• 4:50
  strings,
• 4:50
  and
• 4:50
  more.
• 4:51
  The
• 4:51
  kill
• 4:51
  chain
• 4:52
  is
• 4:52
  almost
• 4:52
  boring:
• 4:52
  get
• 4:53
  local
• 4:53
  access,
• 4:54
  read
• 4:54
  the
• 4:54
  files,
• 4:54
  use
• 4:55
  the
• 4:55
  credentials.
• 4:56
  No
• 4:56
  exploitation
• 4:57
  required
• 4:57
  —
• 4:58
  you
• 4:58
  just
• 4:58
  log
• 4:58
  in.
• 4:59
  The
• 4:59
  irony
• 4:59
  is
• 4:59
  pointed:
• 5:00
  the
• 5:00
  tools
• 5:00
  developers
• 5:01
  use
• 5:01
  to
• 5:01
  write
• 5:01
  more
• 5:01
  secure
• 5:02
  code
• 5:02
  are
• 5:02
  quietly
• 5:02
  accumulating
• 5:03
  a
• 5:03
  record
• 5:03
  of
• 5:03
  every
• 5:03
  credential
• 5:04
  ever
• 5:04
  pasted
• 5:04
  into
• 5:05
  a
• 5:05
  session.
• 5:06
  No
• 5:06
  encryption,
• 5:06
  no
• 5:07
  ACLs
• 5:07
  beyond
• 5:07
  standard
• 5:08
  filesystem
• 5:08
  permissions,
• 5:09
  no
• 5:09
  canary
• 5:09
  mechanism,
• 5:10
  no
• 5:10
  detection.
• 5:11
  The
• 5:11
  assumption
• 5:12
  seems
• 5:12
  to
• 5:12
  be
• 5:12
  that
• 5:12
  local
• 5:12
  access
• 5:13
  means
• 5:13
  game
• 5:13
  over
• 5:13
  anyway
• 5:14
  —
• 5:14
  but
• 5:14
  that
• 5:15
  assumption
• 5:15
  breaks
• 5:15
  when
• 5:15
  the
• 5:15
  files
• 5:16
  contain
• 5:16
  credentials
• 5:16
  that
• 5:16
  reach
• 5:17
  far
• 5:17
  outside
• 5:17
  the
• 5:17
  local
• 5:17
  machine
• 5:18
  into
• 5:18
  cloud
• 5:18
  environments,
• 5:19
  CI/CD
• 5:20
  pipelines,
• 5:21
  and
• 5:21
  production
• 5:21
  databases.
• 5:25
  A
• 5:25
  prevailing
• 5:25
  assumption
• 5:26
  in
• 5:26
  AI
• 5:26
  safety
• 5:26
  is
• 5:26
  that
• 5:27
  alignment
• 5:27
  emerges
• 5:27
  from
• 5:28
  broad
• 5:28
  reorganization
• 5:29
  of
• 5:29
  model
• 5:29
  weights
• 5:29
  —
• 5:30
  that
• 5:30
  safety
• 5:30
  is
• 5:31
  distributed
• 5:31
  diffusely
• 5:32
  enough
• 5:32
  that
• 5:32
  no
• 5:32
  single
• 5:33
  component
• 5:33
  is
• 5:33
  decisive.
• 5:34
  An
• 5:34
  Apple
• 5:34
  research
• 5:35
  team
• 5:35
  just
• 5:35
  demonstrated
• 5:36
  that
• 5:36
  assumption
• 5:36
  is
• 5:36
  wrong.
• 5:37
  Across
• 5:37
  seven
• 5:38
  models
• 5:38
  spanning
• 5:38
  two
• 5:39
  families
• 5:39
  (Qwen3
• 5:40
  and
• 5:40
  Llama-3.1)
• 5:41
  and
• 5:41
  scales
• 5:41
  from
• 5:41
  1.7B
• 5:42
  to
• 5:42
  70B
• 5:43
  parameters,
• 5:44
  they
• 5:44
  show
• 5:44
  that
• 5:44
  suppressing
• 5:45
  a
• 5:45
  single
• 5:45
  MLP
• 5:46
  neuron
• 5:46
  —
• 5:46
  one
• 5:47
  unit
• 5:47
  out
• 5:47
  of
• 5:47
  hundreds
• 5:48
  of
• 5:48
  thousands
• 5:48
  to
• 5:48
  over
• 5:49
  two
• 5:49
  million
• 5:49
  —
• 5:50
  is
• 5:50
  sufficient
• 5:50
  to
• 5:50
  bypass
• 5:51
  safety
• 5:51
  alignment
• 5:52
  entirely,
• 5:52
  achieving
• 5:53
  a
• 5:53
  91.7%
• 5:54
  average
• 5:54
  attack
• 5:54
  success
• 5:55
  rate
• 5:55
  on
• 5:55
  JailbreakBench.
• 5:56
  No
• 5:56
  training,
• 5:57
  no
• 5:57
  fine-tuning,
• 5:57
  no
• 5:58
  prompt
• 5:58
  engineering
• 5:58
  required.
• 5:59
  The
• 5:59
  paper
• 5:59
  identifies
• 6:00
  two
• 6:00
  mechanistically
• 6:01
  distinct
• 6:01
  systems:
• 6:02
  "refusal
• 6:02
  neurons"
• 6:03
  that
• 6:03
  gate
• 6:03
  whether
• 6:03
  harmful
• 6:04
  knowledge
• 6:04
  is
• 6:04
  expressed,
• 6:05
  and
• 6:05
  "concept
• 6:05
  neurons"
• 6:06
  that
• 6:06
  encode
• 6:06
  the
• 6:06
  harmful
• 6:07
  knowledge
• 6:07
  itself.
• 6:08
  Suppressing
• 6:08
  any
• 6:08
  one
• 6:08
  identified
• 6:09
  refusal
• 6:09
  neuron
• 6:10
  bypasses
• 6:10
  safety
• 6:10
  across
• 6:11
  diverse
• 6:11
  harmful
• 6:11
  requests.
• 6:12
  As
• 6:12
  a
• 6:12
  proof
• 6:13
  of
• 6:13
  concept,
• 6:14
  amplifying
• 6:14
  a
• 6:14
  single
• 6:14
  "suicide
• 6:15
  neuron"
• 6:15
  causes
• 6:16
  models
• 6:16
  to
• 6:16
  inject
• 6:16
  suicide-related
• 6:17
  content
• 6:17
  into
• 6:18
  otherwise
• 6:18
  innocent
• 6:18
  prompts
• 6:19
  across
• 6:19
  multiple
• 6:19
  scales.
• 6:20
  Perhaps
• 6:21
  most
• 6:21
  concerning:
• 6:21
  these
• 6:22
  refusal
• 6:22
  neurons
• 6:22
  exist
• 6:23
  in
• 6:23
  base
• 6:23
  models
• 6:23
  prior
• 6:24
  to
• 6:24
  alignment
• 6:24
  training,
• 6:25
  suggesting
• 6:25
  alignment
• 6:26
  modulates
• 6:26
  existing
• 6:26
  neurons
• 6:27
  rather
• 6:27
  than
• 6:27
  creating
• 6:27
  safety
• 6:28
  mechanisms
• 6:28
  from
• 6:28
  scratch.
• 6:29
  A
• 6:30
  single
• 6:30
  refusal
• 6:30
  neuron's
• 6:30
  activation
• 6:31
  also
• 6:31
  serves
• 6:32
  as
• 6:32
  an
• 6:32
  effective
• 6:33
  harmful-prompt
• 6:33
  detector,
• 6:34
  achieving
• 6:35
  AUROC
• 6:35
  comparable
• 6:35
  to
• 6:35
  Llama-Guard-3-8B
• 6:36
  —
• 6:37
  a
• 6:37
  dedicated
• 6:38
  safety
• 6:38
  classifier.
• 6:39
  Safety,
• 6:40
  it
• 6:40
  turns
• 6:40
  out,
• 6:40
  concentrates
• 6:41
  into
• 6:41
  single
• 6:41
  points
• 6:41
  of
• 6:42
  failure
• 6:42
  at
• 6:42
  both
• 6:42
  the
• 6:42
  gate
• 6:43
  and
• 6:43
  the
• 6:43
  knowledge
• 6:43
  layer.
• 6:44
  This
• 6:44
  mechanistic
• 6:45
  fragility
• 6:45
  connects
• 6:46
  to
• 6:46
  a
• 6:46
  broader
• 6:46
  theoretical
• 6:47
  argument
• 6:47
  Andriy
• 6:48
  Burkov
• 6:48
  has
• 6:48
  been
• 6:48
  making
• 6:49
  about
• 6:49
  the
• 6:49
  fundamental
• 6:49
  limits
• 6:50
  of
• 6:50
  LLM-based
• 6:51
  agents.
• 6:52
  An
• 6:52
  LLM
• 6:52
  optimizes
• 6:53
  next-token
• 6:53
  probability,
• 6:54
  not
• 6:54
  expected
• 6:55
  utility
• 6:55
  for
• 6:55
  the
• 6:56
  user
• 6:56
  —
• 6:56
  those
• 6:57
  are
• 6:57
  different
• 6:57
  objectives.
• 6:58
  For
• 6:58
  narrow,
• 6:59
  constrained
• 6:59
  tasks,
• 7:00
  the
• 7:00
  imitation
• 7:01
  of
• 7:01
  rational
• 7:01
  deliberation
• 7:02
  can
• 7:02
  be
• 7:02
  good
• 7:02
  enough.
• 7:03
  For
• 7:03
  general-purpose
• 7:04
  problem
• 7:04
  solving,
• 7:05
  the
• 7:05
  gap
• 7:05
  becomes
• 7:06
  fatal:
• 7:07
  LLMs
• 7:07
  lack
• 7:07
  stable
• 7:08
  preferences,
• 7:09
  calibrated
• 7:09
  beliefs,
• 7:10
  and
• 7:10
  causal
• 7:10
  models
• 7:11
  of
• 7:11
  the
• 7:11
  world.
• 7:12
  "They
• 7:12
  are
• 7:12
  not
• 7:12
  failing
• 7:13
  because
• 7:13
  the
• 7:13
  prompts
• 7:13
  are
• 7:14
  insufficiently
• 7:14
  clever,"
• 7:15
  Burkov
• 7:15
  writes.
• 7:17
  "They
• 7:17
  are
• 7:17
  failing
• 7:17
  because
• 7:18
  we
• 7:18
  are
• 7:18
  asking
• 7:18
  a
• 7:18
  simulator
• 7:19
  of
• 7:19
  rational
• 7:20
  agency
• 7:20
  to
• 7:20
  be
• 7:20
  a
• 7:20
  rational
• 7:21
  agent."
• 7:22
  Community
• 7:22
  pushback
• 7:23
  notes
• 7:23
  humans
• 7:23
  aren't
• 7:23
  expected-utility
• 7:24
  maximizers
• 7:25
  either
• 7:25
  —
• 7:26
  but
• 7:26
  the
• 7:26
  distinction
• 7:27
  matters
• 7:27
  when
• 7:27
  we
• 7:27
  hand
• 7:28
  these
• 7:28
  systems
• 7:28
  destructive
• 7:29
  capabilities.
• 7:30
  On
• 7:30
  the
• 7:30
  constructive
• 7:31
  side,
• 7:32
  new
• 7:32
  research
• 7:33
  on
• 7:33
  Risk-sensitive
• 7:33
  Alignment
• 7:34
  via
• 7:34
  Dominance
• 7:35
  (RAD)
• 7:35
  proposes
• 7:36
  replacing
• 7:36
  the
• 7:36
  scalar
• 7:37
  expected-cost
• 7:38
  constraints
• 7:38
  in
• 7:38
  standard
• 7:39
  Safe
• 7:39
  RLHF
• 7:40
  with
• 7:40
  first-order
• 7:41
  stochastic
• 7:41
  dominance
• 7:42
  constraints
• 7:42
  on
• 7:43
  the
• 7:43
  full
• 7:43
  cost
• 7:43
  distribution.
• 7:45
  Rather
• 7:45
  than
• 7:45
  just
• 7:46
  constraining
• 7:46
  the
• 7:46
  average
• 7:47
  harm,
• 7:47
  RAD
• 7:48
  ensures
• 7:48
  the
• 7:48
  learned
• 7:49
  policy
• 7:49
  assigns
• 7:50
  less
• 7:50
  probability
• 7:51
  to
• 7:51
  high-cost
• 7:51
  outcomes
• 7:52
  across
• 7:52
  the
• 7:52
  entire
• 7:53
  distribution.
• 7:54
  The
• 7:54
  framework
• 7:55
  uses
• 7:55
  optimal
• 7:55
  transport
• 7:56
  theory
• 7:56
  with
• 7:56
  Sinkhorn
• 7:57
  iterations
• 7:58
  for
• 7:58
  efficient
• 7:58
  optimization,
• 7:59
  and
• 8:00
  introduces
• 8:00
  quantile-weighted
• 8:01
  dominance
• 8:01
  constraints
• 8:02
  that
• 8:02
  provide
• 8:03
  universal
• 8:03
  control
• 8:04
  over
• 8:04
  spectral
• 8:04
  risk
• 8:05
  measures
• 8:05
  —
• 8:06
  enabling
• 8:06
  practitioners
• 8:07
  to
• 8:07
  tune
• 8:07
  risk
• 8:07
  profiles
• 8:08
  for
• 8:08
  specific
• 8:09
  deployment
• 8:09
  contexts.
• 8:10
  Empirically,
• 8:11
  RAD-aligned
• 8:12
  models
• 8:13
  produced
• 8:13
  higher
• 8:13
  proportions
• 8:14
  of
• 8:14
  safe
• 8:14
  responses
• 8:15
  and
• 8:16
  showed
• 8:16
  stronger
• 8:16
  robustness
• 8:17
  on
• 8:17
  out-of-distribution
• 8:18
  harmlessness
• 8:19
  evaluations
• 8:20
  than
• 8:20
  standard
• 8:21
  Safe
• 8:21
  RLHF
• 8:22
  baselines.
• 8:25
  The
• 8:25
  Akhter
• 8:25
  twin
• 8:26
  brothers
• 8:26
  didn't
• 8:26
  need
• 8:26
  a
• 8:26
  zero-day.
• 8:27
  Muneeb
• 8:28
  and
• 8:28
  Sohaib,
• 8:28
  both
• 8:29
  34,
• 8:29
  had
• 8:29
  already
• 8:30
  served
• 8:30
  prison
• 8:30
  time
• 8:30
  for
• 8:31
  wire
• 8:31
  fraud
• 8:31
  involving
• 8:31
  computers.
• 8:33
  After
• 8:33
  their
• 8:33
  release,
• 8:33
  they
• 8:33
  worked
• 8:34
  their
• 8:34
  way
• 8:34
  back
• 8:34
  into
• 8:34
  tech
• 8:35
  —
• 8:35
  both
• 8:35
  landing
• 8:36
  jobs
• 8:36
  at
• 8:36
  the
• 8:36
  same
• 8:36
  DC
• 8:36
  firm
• 8:37
  selling
• 8:37
  software
• 8:37
  and
• 8:38
  services
• 8:38
  to
• 8:38
  45
• 8:38
  federal
• 8:39
  clients.
• 8:40
  When
• 8:40
  they
• 8:40
  were
• 8:40
  fired
• 8:40
  via
• 8:40
  Teams
• 8:41
  call
• 8:41
  on
• 8:41
  February
• 8:41
  18,
• 8:42
  2025,
• 8:43
  at
• 8:43
  4:50
• 8:44
  PM,
• 8:44
  Sohaib's
• 8:45
  VPN
• 8:45
  access
• 8:45
  was
• 8:46
  terminated.
• 8:46
  Muneeb's
• 8:47
  was
• 8:47
  not.
• 8:48
  By
• 8:48
  4:56
• 8:49
  PM,
• 8:49
  Muneeb
• 8:50
  was
• 8:50
  issuing
• 8:50
  database
• 8:50
  commands.
• 8:51
  By
• 8:51
  4:58
• 8:52
  PM:
• 8:53
  "DROP
• 8:53
  DATABASE
• 8:54
  dhsproddb."
• 8:55
  At
• 8:55
  4:59
• 8:56
  PM,
• 8:56
  he
• 8:56
  asked
• 8:56
  an
• 8:56
  AI
• 8:57
  tool
• 8:57
  how
• 8:57
  to
• 8:57
  clear
• 8:57
  SQL
• 8:58
  Server
• 8:58
  system
• 8:58
  logs.
• 8:59
  Within
• 8:59
  an
• 8:59
  hour,
• 9:00
  96
• 9:00
  databases
• 9:01
  containing
• 9:01
  US
• 9:01
  government
• 9:02
  information
• 9:02
  were
• 9:02
  gone,
• 9:03
  1,805
• 9:04
  EEOC
• 9:04
  files
• 9:05
  were
• 9:05
  on
• 9:05
  a
• 9:05
  USB
• 9:05
  drive,
• 9:06
  and
• 9:06
  federal
• 9:06
  tax
• 9:07
  information
• 9:07
  for
• 9:07
  450+
• 9:08
  people
• 9:08
  was
• 9:09
  exfiltrated.
• 9:10
  The
• 9:10
  brothers'
• 9:10
  running
• 9:11
  commentary
• 9:11
  during
• 9:11
  the
• 9:11
  destruction
• 9:12
  is
• 9:12
  darkly
• 9:12
  instructive.
• 9:13
  "I
• 9:13
  see
• 9:14
  you
• 9:14
  cleaning
• 9:14
  out
• 9:14
  their
• 9:14
  database
• 9:15
  backups,"
• 9:15
  Sohaib
• 9:16
  observed.
• 9:17
  "Alright
• 9:17
  —
• 9:18
  if
• 9:18
  you
• 9:18
  have
• 9:18
  good
• 9:18
  plausible
• 9:18
  deniability."
• 9:20
  When
• 9:20
  Sohaib
• 9:20
  suggested
• 9:21
  blackmail,
• 9:21
  Muneeb
• 9:22
  showed
• 9:22
  more
• 9:22
  sense
• 9:22
  about
• 9:23
  evidence
• 9:23
  than
• 9:23
  about
• 9:23
  ethics:
• 9:24
  "No,
• 9:24
  you
• 9:24
  do
• 9:24
  not
• 9:25
  do
• 9:25
  that,
• 9:25
  that's
• 9:25
  proof
• 9:25
  of
• 9:25
  guilt,
• 9:26
  man."
• 9:26
  Then
• 9:26
  they
• 9:26
  reinstalled
• 9:27
  their
• 9:27
  laptop
• 9:27
  operating
• 9:28
  systems
• 9:28
  with
• 9:28
  help
• 9:28
  from
• 9:28
  a
• 9:28
  co-conspirator.
• 9:30
  The
• 9:30
  feds
• 9:30
  raided
• 9:30
  three
• 9:31
  weeks
• 9:31
  later,
• 9:31
  finding
• 9:32
  seven
• 9:32
  firearms
• 9:33
  and
• 9:33
  370
• 9:34
  rounds
• 9:34
  of
• 9:34
  ammunition
• 9:34
  at
• 9:34
  Sohaib's
• 9:35
  home
• 9:35
  —
• 9:36
  all
• 9:36
  illegal
• 9:36
  given
• 9:37
  his
• 9:37
  prior
• 9:37
  convictions.
• 9:38
  Muneeb
• 9:39
  pled
• 9:39
  guilty
• 9:39
  in
• 9:39
  April
• 9:39
  2026;
• 9:41
  Sohaib
• 9:41
  went
• 9:41
  to
• 9:41
  trial
• 9:42
  and
• 9:42
  was
• 9:42
  convicted.
• 9:43
  The
• 9:43
  employer,
• 9:44
  later
• 9:44
  identified
• 9:45
  as
• 9:45
  Opexus,
• 9:46
  acknowledged
• 9:47
  that
• 9:47
  "the
• 9:47
  terminations
• 9:47
  were
• 9:48
  not
• 9:48
  handled
• 9:48
  in
• 9:48
  an
• 9:48
  appropriate
• 9:49
  manner"
• 9:50
  and
• 9:50
  that
• 9:50
  "additional
• 9:50
  diligence
• 9:51
  should
• 9:51
  have
• 9:51
  been
• 9:51
  applied"
• 9:52
  in
• 9:52
  hiring.
• 9:53
  The
• 9:53
  failure
• 9:53
  was
• 9:53
  total:
• 9:54
  background
• 9:55
  checks
• 9:55
  that
• 9:55
  missed
• 9:55
  prior
• 9:56
  convictions,
• 9:57
  access
• 9:57
  controls
• 9:57
  that
• 9:58
  survived
• 9:58
  only
• 9:58
  half
• 9:59
  a
• 9:59
  termination,
• 10:00
  and
• 10:00
  no
• 10:00
  monitoring
• 10:01
  of
• 10:01
  a
• 10:01
  known-risk
• 10:01
  employee's
• 10:02
  database
• 10:02
  activity.
• 10:03
  The
• 10:03
  Ghost
• 10:04
  SIM
• 10:04
  attack
• 10:04
  presented
• 10:05
  at
• 10:05
  Black
• 10:05
  Hat
• 10:05
  SecTor
• 10:06
  demonstrates
• 10:06
  a
• 10:06
  different
• 10:07
  kind
• 10:07
  of
• 10:07
  access
• 10:08
  control
• 10:08
  failure
• 10:08
  —
• 10:09
  this
• 10:09
  time
• 10:09
  at
• 10:09
  the
• 10:09
  mobile
• 10:10
  network
• 10:10
  level.
• 10:11
  Researchers
• 10:11
  showed
• 10:12
  that
• 10:12
  by
• 10:12
  extracting
• 10:12
  essential
• 10:13
  SIM
• 10:13
  card
• 10:13
  information
• 10:14
  (IMSI,
• 10:15
  file
• 10:16
  system
• 10:16
  data)
• 10:16
  without
• 10:16
  needing
• 10:17
  the
• 10:17
  secret
• 10:17
  key
• 10:17
  (Ki),
• 10:18
  and
• 10:18
  copying
• 10:19
  it
• 10:19
  to
• 10:19
  a
• 10:19
  programmable
• 10:20
  SIM
• 10:20
  card,
• 10:21
  an
• 10:21
  attacker
• 10:21
  can
• 10:21
  successfully
• 10:22
  register
• 10:22
  on
• 10:22
  a
• 10:22
  mobile
• 10:23
  network
• 10:23
  if
• 10:23
  no
• 10:23
  authentication
• 10:24
  procedure
• 10:24
  is
• 10:25
  triggered.
• 10:26
  Testing
• 10:26
  across
• 10:26
  seven
• 10:27
  operators
• 10:27
  in
• 10:27
  multiple
• 10:28
  European
• 10:28
  and
• 10:28
  Asian
• 10:29
  countries
• 10:29
  revealed
• 10:30
  vulnerabilities
• 10:30
  spanning
• 10:31
  2G
• 10:31
  through
• 10:31
  5G,
• 10:32
  with
• 10:32
  most
• 10:33
  operators
• 10:33
  failing
• 10:34
  to
• 10:34
  authenticate
• 10:34
  on
• 10:34
  initial
• 10:35
  registration
• 10:35
  in
• 10:35
  at
• 10:36
  least
• 10:36
  one
• 10:36
  network
• 10:36
  domain.
• 10:38
  Post-exploitation
• 10:39
  capabilities
• 10:39
  include
• 10:40
  receiving
• 10:40
  SMS
• 10:41
  messages
• 10:41
  —
• 10:42
  potentially
• 10:42
  bypassing
• 10:43
  2FA
• 10:44
  for
• 10:44
  messaging
• 10:44
  applications.
• 10:46
  The
• 10:46
  attack
• 10:46
  requires
• 10:47
  physical
• 10:47
  access
• 10:47
  to
• 10:48
  an
• 10:48
  unlocked
• 10:48
  phone
• 10:48
  or
• 10:48
  SIM
• 10:49
  card,
• 10:49
  but
• 10:50
  the
• 10:50
  researchers
• 10:50
  documented
• 10:51
  multiple
• 10:51
  known
• 10:51
  Android
• 10:52
  vulnerabilities
• 10:52
  (lock
• 10:53
  screen
• 10:53
  bypasses,
• 10:54
  undocumented
• 10:55
  AT
• 10:55
  commands)
• 10:55
  that
• 10:56
  can
• 10:56
  facilitate
• 10:56
  that
• 10:56
  access.
• 10:59
  AMD
• 11:00
  introduced
• 11:00
  the
• 11:00
  Instinct
• 11:01
  MI350P,
• 11:03
  bringing
• 11:03
  CDNA
• 11:03
  4
• 11:04
  architecture
• 11:04
  to
• 11:04
  PCIe
• 11:05
  form
• 11:05
  factor
• 11:06
  with
• 11:06
  configurations
• 11:06
  offering
• 11:07
  144GB
• 11:08
  or
• 11:08
  288GB
• 11:10
  of
• 11:10
  HBM3E
• 11:11
  memory
• 11:12
  and
• 11:12
  3.6
• 11:13
  TB/s
• 11:13
  memory
• 11:14
  bandwidth.
• 11:15
  No
• 11:15
  pricing
• 11:15
  or
• 11:15
  availability
• 11:16
  yet,
• 11:16
  but
• 11:16
  the
• 11:16
  288GB
• 11:18
  configuration
• 11:18
  represents
• 11:19
  a
• 11:19
  significant
• 11:19
  memory
• 11:19
  density
• 11:20
  advantage
• 11:20
  for
• 11:20
  inference-heavy
• 11:21
  workloads
• 11:21
  where
• 11:22
  VRAM
• 11:22
  is
• 11:22
  the
• 11:22
  binding
• 11:23
  constraint.
• 11:23
  The
• 11:24
  community
• 11:24
  reaction
• 11:24
  was
• 11:25
  predictable
• 11:25
  enthusiasm
• 11:26
  tempered
• 11:26
  by
• 11:26
  wallet
• 11:26
  anxiety
• 11:27
  —
• 11:27
  "I
• 11:27
  wish
• 11:28
  I
• 11:28
  can
• 11:28
  afford
• 11:28
  4x
• 11:29
  288G
• 11:29
  cards
• 11:30
  to
• 11:30
  run
• 11:30
  TB
• 11:30
  models
• 11:31
  locally"
• 11:31
  —
• 11:32
  but
• 11:32
  the
• 11:32
  real
• 11:32
  significance
• 11:33
  is
• 11:33
  AMD
• 11:33
  continuing
• 11:34
  to
• 11:34
  fill
• 11:34
  its
• 11:34
  accelerator
• 11:34
  lineup
• 11:35
  across
• 11:35
  form
• 11:35
  factors.
• 11:36
  Whether
• 11:36
  AMD
• 11:37
  hardware
• 11:37
  actually
• 11:37
  works
• 11:38
  for
• 11:38
  real
• 11:38
  ML
• 11:38
  workloads
• 11:39
  remains
• 11:39
  the
• 11:39
  perennial
• 11:39
  question,
• 11:40
  and
• 11:40
  a
• 11:40
  hackathon
• 11:41
  project
• 11:41
  provides
• 11:42
  a
• 11:42
  concrete
• 11:42
  data
• 11:42
  point.
• 11:43
  MedQA
• 11:44
  demonstrates
• 11:44
  LoRA
• 11:44
  fine-tuning
• 11:45
  of
• 11:45
  Qwen3-1.7B
• 11:47
  on
• 11:47
  medical
• 11:47
  question-answering
• 11:48
  data
• 11:48
  using
• 11:48
  an
• 11:48
  AMD
• 11:49
  Instinct
• 11:49
  MI300X,
• 11:52
  with
• 11:52
  the
• 11:52
  entire
• 11:52
  HuggingFace
• 11:53
  ecosystem
• 11:53
  —
• 11:54
  Transformers,
• 11:55
  PEFT,
• 11:55
  TRL,
• 11:56
  Accelerate
• 11:56
  —
• 11:57
  running
• 11:57
  on
• 11:57
  ROCm
• 11:58
  with
• 11:58
  exactly
• 11:58
  three
• 11:58
  environment
• 11:59
  variables
• 11:59
  changed
• 12:00
  from
• 12:00
  a
• 12:00
  CUDA
• 12:00
  setup.
• 12:01
  The
• 12:01
  192GB
• 12:02
  HBM3
• 12:03
  eliminated
• 12:03
  the
• 12:03
  need
• 12:04
  for
• 12:04
  any
• 12:04
  quantization
• 12:04
  workarounds.
• 12:05
  Training
• 12:06
  took
• 12:06
  five
• 12:06
  minutes
• 12:06
  on
• 12:07
  2,000
• 12:07
  samples,
• 12:08
  and
• 12:08
  the
• 12:08
  resulting
• 12:08
  model
• 12:09
  produces
• 12:09
  both
• 12:09
  answer
• 12:10
  selections
• 12:10
  and
• 12:10
  clinical
• 12:11
  reasoning
• 12:11
  explanations.
• 12:12
  The
• 12:12
  war
• 12:13
  stories
• 12:13
  section
• 12:13
  is
• 12:13
  instructive:
• 12:14
  bfloat16
• 12:15
  caused
• 12:16
  NaN
• 12:16
  loss
• 12:16
  (fixed
• 12:17
  by
• 12:17
  switching
• 12:17
  to
• 12:17
  fp16),
• 12:18
  and
• 12:19
  no
• 12:19
  ROCm
• 12:19
  build
• 12:20
  of
• 12:20
  bitsandbytes
• 12:20
  exists
• 12:20
  (irrelevant
• 12:21
  given
• 12:21
  the
• 12:21
  MI300X's
• 12:24
  memory
• 12:24
  headroom).
• 12:25
  On
• 12:25
  the
• 12:26
  optimizer
• 12:26
  front,
• 12:26
  STAM
• 12:27
  (Stable
• 12:27
  Training
• 12:28
  with
• 12:28
  Adaptive
• 12:28
  Momentum)
• 12:29
  proposes
• 12:29
  making
• 12:29
  Adam's
• 12:30
  β1
• 12:30
  parameter
• 12:31
  runtime-adaptive
• 12:32
  based
• 12:32
  on
• 12:32
  gradient
• 12:32
  residual
• 12:33
  statistics
• 12:34
  rather
• 12:34
  than
• 12:34
  leaving
• 12:34
  it
• 12:34
  at
• 12:34
  the
• 12:35
  universal
• 12:35
  default
• 12:35
  of
• 12:35
  0.9.
• 12:37
  The
• 12:37
  core
• 12:37
  idea
• 12:38
  —
• 12:38
  that
• 12:38
  β1
• 12:39
  should
• 12:39
  respond
• 12:39
  to
• 12:39
  training
• 12:40
  dynamics
• 12:40
  rather
• 12:41
  than
• 12:41
  sitting
• 12:41
  fixed
• 12:41
  —
• 12:42
  is
• 12:42
  directionally
• 12:43
  sound.
• 12:44
  But
• 12:44
  rigorous
• 12:44
  community
• 12:45
  analysis
• 12:45
  found
• 12:46
  significant
• 12:46
  methodological
• 12:47
  problems:
• 12:48
  the
• 12:48
  signal
• 12:48
  STAM
• 12:49
  computes
• 12:49
  is
• 12:49
  actually
• 12:50
  a
• 12:50
  kurtosis
• 12:51
  proxy
• 12:51
  (not
• 12:51
  variance),
• 12:52
  the
• 12:53
  control
• 12:53
  law
• 12:53
  conflates
• 12:54
  three
• 12:54
  distinct
• 12:54
  gradient
• 12:55
  regimes
• 12:55
  that
• 12:55
  require
• 12:56
  opposite
• 12:56
  responses,
• 12:57
  and
• 12:57
  the
• 12:57
  evaluation
• 12:58
  uses
• 12:58
  single-seed
• 12:59
  MNIST/CIFAR
• 13:03
  runs
• 13:03
  at
• 13:04
  10-80
• 13:04
  update
• 13:04
  steps
• 13:05
  against
• 13:05
  dated
• 13:05
  baselines
• 13:06
  (Adam
• 13:06
  2014,
• 13:07
  RMSProp
• 13:08
  2012),
• 13:09
  omitting
• 13:09
  Lion,
• 13:10
  Muon,
• 13:11
  Sophia,
• 13:12
  and
• 13:12
  other
• 13:12
  competitive
• 13:13
  2024-2026
• 13:14
  optimizers.
• 13:15
  Prior
• 13:16
  benchmarking
• 13:16
  has
• 13:16
  established
• 13:17
  that
• 13:17
  supposed
• 13:17
  2x
• 13:18
  gains
• 13:18
  of
• 13:18
  alternative
• 13:19
  optimizers
• 13:20
  over
• 13:20
  AdamW
• 13:20
  tend
• 13:21
  to
• 13:21
  shrink
• 13:21
  to
• 13:21
  roughly
• 13:22
  1x
• 13:22
  at
• 13:23
  billion-parameter
• 13:23
  scales.
• 13:24
  The
• 13:24
  durable
• 13:25
  contribution
• 13:25
  is
• 13:25
  identifying
• 13:26
  that
• 13:26
  β1=0.9
• 13:28
  has
• 13:29
  no
• 13:29
  principled
• 13:29
  default
• 13:29
  —
• 13:30
  a
• 13:30
  question
• 13:31
  worth
• 13:31
  its
• 13:31
  own
• 13:31
  research
• 13:32
  cycle,
• 13:32
  separated
• 13:33
  from
• 13:33
  STAM's
• 13:33
  specific
• 13:34
  implementation.
• 13:38
  Anthropic
• 13:38
  quietly
• 13:39
  restructured
• 13:39
  Claude
• 13:40
  Code's
• 13:40
  programmatic
• 13:41
  access
• 13:41
  model,
• 13:42
  replacing
• 13:42
  fuzzy
• 13:42
  usage
• 13:43
  limits
• 13:43
  with
• 13:43
  explicit
• 13:44
  monthly
• 13:44
  credits
• 13:44
  for
• 13:44
  SDK,
• 13:46
  claude
• 13:46
  -p,
• 13:46
  GitHub
• 13:47
  Actions,
• 13:48
  and
• 13:48
  third-party
• 13:48
  orchestration
• 13:49
  workflows.
• 13:50
  Power
• 13:50
  users
• 13:51
  who
• 13:51
  had
• 13:51
  been
• 13:51
  running
• 13:51
  recursive
• 13:52
  agent
• 13:52
  loops,
• 13:52
  CI/CD
• 13:53
  automation,
• 13:54
  and
• 13:54
  background
• 13:55
  pipelines
• 13:55
  under
• 13:55
  the
• 13:55
  old
• 13:55
  regime
• 13:56
  —
• 13:56
  sometimes
• 13:57
  extracting
• 13:57
  thousands
• 13:58
  of
• 13:58
  dollars
• 13:58
  in
• 13:58
  effective
• 13:59
  compute
• 13:59
  from
• 13:59
  a
• 13:59
  $200
• 14:00
  plan
• 14:00
  —
• 14:01
  are
• 14:01
  not
• 14:01
  pleased.
• 14:02
  "This
• 14:03
  isn't
• 14:03
  expanded
• 14:04
  SDK
• 14:04
  support.
• 14:05
  It's
• 14:05
  a
• 14:05
  rate
• 14:06
  limiter
• 14:06
  wearing
• 14:06
  a
• 14:06
  party
• 14:06
  hat,"
• 14:07
  wrote
• 14:07
  one
• 14:07
  critic.
• 14:08
  The
• 14:08
  structural
• 14:09
  point
• 14:09
  is
• 14:09
  hard
• 14:09
  to
• 14:09
  argue
• 14:09
  with:
• 14:10
  every
• 14:10
  consumer-priced
• 14:11
  infrastructure
• 14:12
  tool
• 14:12
  that
• 14:13
  ends
• 14:13
  up
• 14:13
  subsidizing
• 14:13
  commercial
• 14:14
  automation
• 14:14
  eventually
• 14:15
  meters
• 14:15
  it.
• 14:16
  The
• 14:16
  question
• 14:16
  is
• 14:16
  whether
• 14:16
  this
• 14:17
  pushes
• 14:17
  more
• 14:17
  developers
• 14:18
  toward
• 14:18
  vendor-neutral
• 14:18
  harnesses
• 14:19
  and
• 14:19
  open-weight
• 14:20
  models.
• 14:20
  .
• 14:21
  Between
• 14:21
  Anthropic's
• 14:22
  rushed
• 14:22
  and
• 14:22
  buggy
• 14:22
  new
• 14:23
  features,
• 14:23
  caps
• 14:23
  on
• 14:24
  use,
• 14:24
  price
• 14:24
  changes,
• 14:25
  and
• 14:25
  consistent
• 14:25
  ill-timed
• 14:26
  outages,
• 14:27
  it's
• 14:27
  not
• 14:27
  surprising
• 14:28
  that
• 14:28
  that
• 14:28
  codex,
• 14:28
  gemini,
• 14:29
  and
• 14:29
  open-weight
• 14:30
  models
• 14:30
  are
• 14:30
  seeing
• 14:30
  a
• 14:30
  surge
• 14:31
  in
• 14:31
  usage.
• 14:32
  The
• 14:32
  open-source
• 14:33
  RAG
• 14:33
  space
• 14:33
  continues
• 14:34
  maturing.
• 14:35
  ForgeRAG
• 14:36
  combines
• 14:36
  BM25
• 14:37
  +
• 14:37
  vector
• 14:37
  search
• 14:38
  for
• 14:38
  candidate
• 14:38
  retrieval,
• 14:39
  knowledge
• 14:40
  graph
• 14:40
  multi-hop
• 14:41
  for
• 14:41
  conceptual
• 14:41
  connections,
• 14:42
  and
• 14:42
  LLM
• 14:43
  tree
• 14:43
  navigation
• 14:43
  for
• 14:44
  structure-aware
• 14:44
  reasoning
• 14:44
  —
• 14:45
  fused
• 14:46
  into
• 14:46
  answers
• 14:46
  with
• 14:46
  pixel-precise,
• 14:47
  traceable
• 14:47
  citations.
• 14:49
  On
• 14:49
  the
• 14:49
  UltraDomain
• 14:50
  benchmark
• 14:50
  it
• 14:50
  achieves
• 14:51
  a
• 14:51
  55.5%
• 14:52
  overall
• 14:52
  win
• 14:52
  rate
• 14:52
  against
• 14:53
  LightRAG.
• 14:54
  The
• 14:54
  key
• 14:54
  architectural
• 14:55
  distinction:
• 14:55
  where
• 14:56
  graph-based
• 14:56
  approaches
• 14:56
  like
• 14:57
  LightRAG
• 14:57
  synthesize
• 14:58
  answers
• 14:58
  from
• 14:58
  knowledge
• 14:58
  graph
• 14:58
  summaries
• 14:59
  (higher
• 14:59
  hallucination
• 15:00
  risk),
• 15:01
  ForgeRAG
• 15:01
  grounds
• 15:01
  every
• 15:01
  claim
• 15:02
  in
• 15:02
  original
• 15:02
  source
• 15:02
  text
• 15:03
  with
• 15:03
  exact
• 15:03
  page
• 15:03
  and
• 15:04
  bounding-box
• 15:04
  references.
• 15:05
  It
• 15:05
  supports
• 15:06
  multiple
• 15:06
  backends
• 15:07
  (SQLite/PostgreSQL,
• 15:10
  ChromaDB/Qdrant/Milvus,
• 15:14
  NetworkX/Neo4j)
• 15:17
  and
• 15:17
  any
• 15:17
  LiteLLM-compatible
• 15:18
  provider.
• 15:19
  AI-generated
• 15:20
  video
• 15:20
  is
• 15:20
  crossing
• 15:20
  from
• 15:21
  novelty
• 15:21
  to
• 15:21
  practical
• 15:22
  utility.
• 15:23
  Hyperframes,
• 15:23
  an
• 15:23
  open-source
• 15:24
  pipeline
• 15:24
  combining
• 15:25
  Claude
• 15:25
  Code
• 15:25
  for
• 15:25
  scripting
• 15:26
  and
• 15:26
  composition,
• 15:27
  HyperFrames
• 15:28
  for
• 15:28
  HTML/GSAP
• 15:30
  rendering,
• 15:31
  and
• 15:31
  either
• 15:31
  Kokoro
• 15:32
  (free,
• 15:33
  local
• 15:33
  TTS)
• 15:33
  or
• 15:34
  ElevenLabs
• 15:34
  (paid),
• 15:35
  can
• 15:35
  produce
• 15:35
  polished
• 15:36
  vertical
• 15:36
  YouTube
• 15:36
  Shorts
• 15:37
  from
• 15:37
  a
• 15:37
  single
• 15:37
  topic
• 15:37
  prompt.
• 15:38
  The
• 15:38
  workflow
• 15:39
  researches
• 15:39
  the
• 15:39
  topic,
• 15:40
  drafts
• 15:40
  paced
• 15:41
  narration
• 15:41
  with
• 15:41
  SSML
• 15:42
  break
• 15:42
  tags,
• 15:43
  generates
• 15:43
  TTS
• 15:44
  with
• 15:44
  word-level
• 15:44
  timestamps,
• 15:45
  edits
• 15:45
  an
• 15:45
  HTML
• 15:46
  composition,
• 15:47
  lints
• 15:47
  it,
• 15:47
  and
• 15:47
  opens
• 15:48
  a
• 15:48
  browser
• 15:48
  preview
• 15:48
  —
• 15:49
  all
• 15:49
  in
• 15:49
  20-30
• 15:50
  minutes.
• 15:51
  Three
• 15:51
  templates
• 15:51
  ship
• 15:52
  by
• 15:52
  default,
• 15:52
  and
• 15:53
  creating
• 15:53
  custom
• 15:53
  brand
• 15:54
  templates
• 15:54
  is
• 15:54
  a
• 15:54
  documented
• 15:55
  process.
• 15:56
  The
• 15:57
  output
• 15:57
  isn't
• 15:57
  perfect
• 15:57
  —
• 15:58
  voice
• 15:59
  inflection
• 15:59
  can
• 15:59
  be
• 15:59
  awkward,
• 16:00
  transitions
• 16:01
  occasionally
• 16:01
  stiff
• 16:01
  —
• 16:02
  but
• 16:02
  for
• 16:02
  explainers,
• 16:03
  team
• 16:04
  updates,
• 16:04
  and
• 16:04
  community
• 16:05
  content,
• 16:06
  it's
• 16:06
  approaching
• 16:06
  practical
• 16:07
  utility.
• 16:08
  Computron
• 16:08
  offers
• 16:09
  a
• 16:09
  self-hosted
• 16:10
  AI
• 16:10
  assistant
• 16:10
  running
• 16:11
  on
• 16:11
  Ollama
• 16:11
  in
• 16:11
  a
• 16:12
  single
• 16:12
  Docker
• 16:12
  container
• 16:13
  with
• 16:13
  full
• 16:13
  Google
• 16:14
  Workspace
• 16:14
  integration
• 16:14
  —
• 16:15
  Gmail,
• 16:16
  Calendar,
• 16:16
  Drive,
• 16:17
  Contacts
• 16:17
  —
• 16:18
  through
• 16:18
  natural
• 16:19
  conversation.
• 16:20
  The
• 16:20
  security
• 16:20
  model
• 16:21
  is
• 16:21
  notable:
• 16:21
  you
• 16:21
  create
• 16:22
  your
• 16:22
  own
• 16:22
  OAuth
• 16:22
  app,
• 16:23
  all
• 16:23
  credentials
• 16:24
  are
• 16:24
  encrypted
• 16:24
  at
• 16:24
  rest
• 16:25
  with
• 16:25
  AES-256-GCM,
• 16:28
  and
• 16:28
  the
• 16:29
  agent
• 16:29
  process
• 16:29
  never
• 16:30
  sees
• 16:30
  raw
• 16:30
  credentials
• 16:30
  —
• 16:31
  they're
• 16:31
  decrypted
• 16:32
  only
• 16:32
  by
• 16:32
  the
• 16:32
  integration
• 16:33
  supervisor
• 16:33
  and
• 16:34
  injected
• 16:34
  into
• 16:34
  isolated
• 16:35
  broker
• 16:35
  processes.
• 16:36
  Permissions
• 16:37
  are
• 16:37
  granular
• 16:37
  per
• 16:37
  capability.
• 16:39
  Separately,
• 16:39
  an
• 16:39
  Italian
• 16:40
  team
• 16:40
  is
• 16:40
  building
• 16:41
  open-source
• 16:41
  Microsoft
• 16:42
  Office
• 16:42
  extensions
• 16:43
  for
• 16:43
  Open
• 16:43
  WebUI
• 16:43
  —
• 16:44
  AI-assisted
• 16:45
  document
• 16:45
  creation
• 16:45
  in
• 16:45
  Word,
• 16:46
  spreadsheet
• 16:47
  automation
• 16:47
  in
• 16:47
  Excel,
• 16:48
  presentation
• 16:48
  editing
• 16:49
  in
• 16:49
  PowerPoint,
• 16:50
  and
• 16:50
  email
• 16:50
  drafting
• 16:51
  in
• 16:51
  Outlook
• 16:51
  —
• 16:51
  all
• 16:52
  designed
• 16:52
  to
• 16:52
  keep
• 16:52
  everything
• 16:53
  open
• 16:53
  and
• 16:53
  compatible
• 16:54
  with
• 16:54
  the
• 16:54
  self-hosted
• 16:55
  ecosystem.
• 16:56
  The
• 16:56
  sheer
• 16:56
  volume
• 16:57
  of
• 16:57
  open-source
• 16:57
  AI
• 16:57
  activity
• 16:58
  is
• 16:58
  visible
• 16:59
  in
• 16:59
  the
• 16:59
  numbers:
• 17:00
  new
• 17:00
  GGUF
• 17:01
  uploads
• 17:01
  on
• 17:01
  Hugging
• 17:01
  Face
• 17:02
  nearly
• 17:02
  doubled
• 17:02
  in
• 17:02
  two
• 17:03
  months.
• 17:04
  Community
• 17:05
  reaction
• 17:05
  is
• 17:05
  mixed
• 17:05
  —
• 17:06
  "most
• 17:06
  of
• 17:06
  them
• 17:06
  are
• 17:06
  vibe-tuned
• 17:07
  slop"
• 17:08
  and
• 17:08
  "they
• 17:08
  all
• 17:08
  look
• 17:08
  like
• 17:08
  UUIDs
• 17:09
  now"
• 17:09
  —
• 17:10
  with
• 17:10
  calls
• 17:11
  for
• 17:11
  better
• 17:11
  filtering
• 17:12
  and
• 17:12
  discoverability.
• 17:13
  The
• 17:13
  real
• 17:14
  concern
• 17:14
  is
• 17:14
  sustainability:
• 17:15
  rapid
• 17:16
  upload
• 17:16
  growth
• 17:16
  eventually
• 17:17
  means
• 17:17
  someone
• 17:17
  pays
• 17:18
  for
• 17:18
  storage,
• 17:18
  and
• 17:19
  the
• 17:19
  odds
• 17:19
  favor
• 17:19
  users
• 17:20
  over
• 17:20
  corporations.
• 17:21
  Meanwhile,
• 17:22
  the
• 17:22
  AI
• 17:22
  regulation
• 17:23
  debate
• 17:23
  continues
• 17:24
  its
• 17:24
  predictable
• 17:24
  pattern:
• 17:26
  US
• 17:26
  industry
• 17:26
  lobbyists
• 17:27
  argue
• 17:27
  any
• 17:27
  regulation
• 17:28
  means
• 17:28
  losing
• 17:28
  to
• 17:29
  China,
• 17:29
  while
• 17:30
  China's
• 17:30
  official
• 17:30
  position
• 17:31
  states
• 17:31
  "safety
• 17:31
  first,
• 17:32
  innovation
• 17:33
  second."
• 17:34
  Community
• 17:34
  skepticism
• 17:35
  runs
• 17:35
  in
• 17:35
  both
• 17:35
  directions
• 17:36
  —
• 17:37
  dismissing
• 17:37
  China's
• 17:37
  statements
• 17:38
  as
• 17:38
  propaganda
• 17:39
  while
• 17:39
  noting
• 17:39
  the
• 17:39
  US
• 17:40
  just
• 17:40
  struck
• 17:40
  a
• 17:40
  deal
• 17:40
  to
• 17:41
  review
• 17:41
  AI
• 17:41
  models
• 17:42
  for
• 17:42
  national
• 17:42
  security
• 17:43
  before
• 17:43
  public
• 17:43
  release,
• 17:44
  prompting
• 17:44
  comparisons
• 17:45
  to
• 17:45
  the
• 17:45
  very
• 17:46
  censorship
• 17:46
  the
• 17:46
  industry
• 17:47
  claims
• 17:47
  to
• 17:47
  oppose.