Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – July 07, 2026

Transcript

AGI Dreams Podcast – July 07, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Beijing
• 0:00
  is
• 0:00
  exploring
• 0:01
  restrictions
• 0:01
  on
• 0:01
  overseas
• 0:02
  access
• 0:02
  to
• 0:02
  China's
• 0:03
  most
• 0:03
  advanced
• 0:03
  AI
• 0:04
  models
• 0:04
  —
• 0:04
  a
• 0:04
  move
• 0:05
  that
• 0:05
  would
• 0:05
  mirror
• 0:05
  the
• 0:05
  US
• 0:06
  export
• 0:06
  controls
• 0:06
  that
• 0:07
  consumed
• 0:07
  the
• 0:07
  AI
• 0:07
  industry
• 0:08
  for
• 0:08
  most
• 0:08
  of
• 0:08
  June.
• 0:09
  Reuters
• 0:09
  reports
• 0:10
  that
• 0:10
  the
• 0:10
  Ministry
• 0:10
  of
• 0:10
  Commerce
• 0:11
  has
• 0:11
  held
• 0:11
  meetings
• 0:12
  over
• 0:12
  the
• 0:12
  past
• 0:12
  month
• 0:13
  with
• 0:13
  Alibaba,
• 0:13
  ByteDance,
• 0:14
  and
• 0:14
  Z.ai
• 0:15
  about
• 0:15
  limiting
• 0:16
  both
• 0:16
  closed-source
• 0:17
  and
• 0:17
  open-weight
• 0:17
  frontier
• 0:18
  models,
• 0:18
  potentially
• 0:19
  restricting
• 0:19
  them
• 0:20
  to
• 0:20
  domestic
• 0:20
  use.
• 0:21
  Officials
• 0:21
  also
• 0:22
  discussed
• 0:22
  making
• 0:22
  AI
• 0:23
  technology
• 0:23
  theft
• 0:23
  an
• 0:23
  offense
• 0:24
  under
• 0:24
  China's
• 0:24
  national
• 0:25
  security
• 0:25
  law,
• 0:26
  and
• 0:26
  implementing
• 0:27
  new
• 0:27
  controls
• 0:28
  on
• 0:28
  who
• 0:28
  can
• 0:28
  fund
• 0:28
  domestic
• 0:29
  AI
• 0:29
  startups.
• 0:30
  The
• 0:30
  scope
• 0:31
  remains
• 0:31
  fluid
• 0:31
  —
• 0:32
  restrictions
• 0:32
  may
• 0:32
  apply
• 0:33
  only
• 0:33
  to
• 0:33
  future
• 0:33
  models
• 0:34
  —
• 0:35
  but
• 0:35
  a
• 0:35
  May
• 0:35
  roundtable
• 0:35
  of
• 0:36
  Chinese
• 0:36
  legal
• 0:36
  experts
• 0:37
  published
• 0:37
  in
• 0:37
  an
• 0:37
  official
• 0:38
  Supreme
• 0:38
  People's
• 0:38
  Court
• 0:38
  journal
• 0:39
  proposed
• 0:39
  a
• 0:39
  tiered
• 0:40
  system:
• 0:40
  basic
• 0:41
  open-source
• 0:41
  tools
• 0:42
  subject
• 0:42
  to
• 0:42
  filing,
• 0:43
  advanced
• 0:43
  technologies
• 0:44
  facing
• 0:44
  security
• 0:45
  reviews,
• 0:46
  and
• 0:46
  the
• 0:46
  most
• 0:46
  sensitive
• 0:46
  frontier
• 0:47
  models
• 0:47
  barred
• 0:47
  from
• 0:48
  public
• 0:48
  release
• 0:48
  or
• 0:49
  restricted
• 0:49
  to
• 0:49
  domestic
• 0:49
  use.
• 0:51
  The
• 0:51
  timing
• 0:51
  is
• 0:51
  not
• 0:51
  accidental.
• 0:53
  Chinese
• 0:53
  authorities
• 0:54
  are
• 0:54
  "deeply
• 0:54
  worried"
• 0:55
  about
• 0:55
  Anthropic's
• 0:56
  Mythos
• 0:56
  model,
• 0:56
  which
• 0:57
  remains
• 0:57
  restricted
• 0:57
  to
• 0:58
  trusted
• 0:58
  US
• 0:58
  organizations
• 0:59
  after
• 0:59
  the
• 0:59
  broader
• 1:00
  Fable
• 1:00
  export
• 1:00
  controls
• 1:01
  were
• 1:01
  lifted.
• 1:02
  Zhou
• 1:03
  Hongyi,
• 1:03
  founder
• 1:03
  of
• 1:03
  cybersecurity
• 1:04
  firm
• 1:04
  360,
• 1:05
  has
• 1:06
  publicly
• 1:06
  called
• 1:06
  for
• 1:06
  China
• 1:07
  to
• 1:07
  develop
• 1:07
  its
• 1:07
  own
• 1:07
  Mythos
• 1:08
  equivalent.
• 1:09
  Meanwhile,
• 1:10
  Beijing
• 1:10
  forced
• 1:10
  Meta
• 1:11
  to
• 1:11
  unwind
• 1:11
  its
• 1:11
  $2
• 1:11
  billion
• 1:12
  acquisition
• 1:13
  of
• 1:13
  Chinese-founded
• 1:13
  AI
• 1:14
  startup
• 1:14
  Manus
• 1:15
  and
• 1:15
  launched
• 1:15
  investigations
• 1:16
  into
• 1:16
  local
• 1:16
  AI
• 1:16
  startups
• 1:17
  that
• 1:17
  moved
• 1:17
  abroad.
• 1:18
  Alibaba
• 1:19
  has
• 1:19
  banned
• 1:19
  employees
• 1:20
  from
• 1:20
  using
• 1:20
  Anthropic's
• 1:21
  Claude
• 1:21
  Code
• 1:21
  after
• 1:21
  the
• 1:22
  tool
• 1:22
  was
• 1:22
  found
• 1:22
  to
• 1:22
  contain
• 1:23
  mechanisms
• 1:23
  inspecting
• 1:24
  user
• 1:24
  environments
• 1:24
  —
• 1:25
  timezone
• 1:26
  data,
• 1:26
  proxy
• 1:26
  configurations
• 1:27
  —
• 1:28
  and
• 1:28
  inserting
• 1:28
  markers
• 1:29
  into
• 1:29
  server-bound
• 1:30
  prompts.
• 1:31
  An
• 1:31
  Anthropic
• 1:31
  employee
• 1:32
  described
• 1:32
  the
• 1:32
  feature
• 1:33
  as
• 1:33
  "an
• 1:33
  experiment
• 1:33
  launched
• 1:34
  in
• 1:34
  March"
• 1:34
  to
• 1:34
  prevent
• 1:35
  abuse
• 1:35
  by
• 1:35
  unauthorized
• 1:36
  resellers
• 1:37
  and
• 1:37
  protect
• 1:37
  against
• 1:37
  model
• 1:38
  distillation.
• 1:39
  This
• 1:39
  follows
• 1:39
  Anthropic's
• 1:40
  accusation
• 1:41
  that
• 1:41
  Alibaba
• 1:41
  orchestrated
• 1:42
  a
• 1:42
  large-scale
• 1:42
  distillation
• 1:43
  effort.
• 1:44
  Alibaba
• 1:44
  is
• 1:44
  directing
• 1:45
  employees
• 1:45
  to
• 1:45
  its
• 1:46
  own
• 1:46
  Qoder
• 1:46
  coding
• 1:46
  platform
• 1:47
  instead.
• 1:48
  On
• 1:48
  the
• 1:48
  domestic
• 1:48
  regulation
• 1:49
  front,
• 1:49
  Virginia's
• 1:50
  ban
• 1:50
  on
• 1:50
  the
• 1:50
  sale
• 1:51
  of
• 1:51
  geolocation
• 1:51
  data
• 1:51
  took
• 1:52
  effect
• 1:52
  July
• 1:52
  1,
• 1:53
  amending
• 1:53
  the
• 1:53
  Virginia
• 1:54
  Consumer
• 1:54
  Data
• 1:54
  Protection
• 1:55
  Act.
• 1:56
  The
• 1:56
  law
• 1:56
  prohibits
• 1:57
  exchanging
• 1:57
  geolocation
• 1:58
  data
• 1:58
  for
• 1:58
  monetary
• 1:59
  consideration
• 1:59
  —
• 2:01
  a
• 2:01
  narrower
• 2:01
  definition
• 2:02
  of
• 2:02
  "sale"
• 2:02
  than
• 2:02
  Maryland
• 2:03
  and
• 2:03
  Oregon,
• 2:04
  which
• 2:04
  extend
• 2:04
  the
• 2:04
  prohibition
• 2:05
  to
• 2:05
  "monetary
• 2:05
  or
• 2:06
  other
• 2:06
  valuable
• 2:06
  consideration."
• 2:08
  Virginia
• 2:08
  joins
• 2:09
  a
• 2:09
  growing
• 2:09
  list
• 2:09
  of
• 2:09
  states
• 2:10
  following
• 2:10
  California's
• 2:11
  enforcement
• 2:11
  sweep
• 2:12
  and
• 2:12
  a
• 2:12
  2024
• 2:13
  FTC
• 2:13
  action
• 2:13
  banning
• 2:14
  a
• 2:14
  data
• 2:14
  broker
• 2:14
  from
• 2:14
  selling
• 2:15
  geolocation
• 2:15
  data.
• 2:16
  Location
• 2:17
  data
• 2:17
  —
• 2:18
  the
• 2:18
  raw
• 2:18
  material
• 2:19
  for
• 2:19
  everything
• 2:19
  from
• 2:19
  targeted
• 2:20
  advertising
• 2:20
  to
• 2:20
  surveillance
• 2:21
  —
• 2:22
  is
• 2:22
  becoming
• 2:22
  legally
• 2:23
  untouchable.
• 2:27
  SentinelMCP,
• 2:29
  a
• 2:29
  new
• 2:29
  open-source
• 2:30
  security
• 2:30
  gateway
• 2:30
  for
• 2:30
  the
• 2:31
  Model
• 2:31
  Context
• 2:31
  Protocol,
• 2:32
  tackles
• 2:32
  what
• 2:33
  may
• 2:33
  be
• 2:33
  the
• 2:33
  most
• 2:33
  underdeveloped
• 2:34
  layer
• 2:34
  of
• 2:34
  the
• 2:34
  AI
• 2:35
  agent
• 2:35
  stack:
• 2:35
  runtime
• 2:36
  enforcement
• 2:36
  of
• 2:36
  tool-call
• 2:37
  policies.
• 2:38
  Written
• 2:38
  in
• 2:38
  Go
• 2:38
  and
• 2:38
  licensed
• 2:39
  under
• 2:39
  Apache
• 2:40
  2.0,
• 2:40
  SentinelMCP
• 2:41
  sits
• 2:42
  between
• 2:42
  AI
• 2:42
  agents
• 2:42
  and
• 2:43
  the
• 2:43
  MCP
• 2:43
  tools
• 2:43
  they
• 2:44
  invoke,
• 2:44
  providing
• 2:45
  inspection,
• 2:45
  policy
• 2:45
  enforcement,
• 2:46
  PII/secret
• 2:47
  redaction,
• 2:48
  and
• 2:48
  structured
• 2:48
  audit
• 2:48
  logging.
• 2:49
  It
• 2:50
  ships
• 2:50
  in
• 2:50
  two
• 2:50
  deployment
• 2:50
  modes
• 2:51
  —
• 2:52
  a
• 2:52
  standalone
• 2:52
  sidecar
• 2:53
  proxy
• 2:53
  that
• 2:53
  works
• 2:53
  with
• 2:54
  any
• 2:54
  language
• 2:54
  or
• 2:54
  framework,
• 2:55
  and
• 2:55
  an
• 2:55
  inline
• 2:56
  Go
• 2:56
  SDK
• 2:56
  that
• 2:56
  wraps
• 2:57
  tool
• 2:57
  calls
• 2:57
  in-process
• 2:58
  with
• 2:58
  19-microsecond
• 2:59
  p99
• 2:59
  latency
• 3:00
  on
• 3:00
  the
• 3:00
  allow
• 3:00
  path.
• 3:01
  The
• 3:02
  policy
• 3:02
  engine
• 3:02
  uses
• 3:03
  local
• 3:03
  YAML
• 3:03
  with
• 3:03
  hot-reloading:
• 3:04
  tools
• 3:05
  are
• 3:05
  assigned
• 3:05
  risk
• 3:05
  levels
• 3:06
  (low,
• 3:06
  medium,
• 3:06
  high)
• 3:07
  mapping
• 3:07
  to
• 3:07
  allow,
• 3:07
  redact,
• 3:08
  block,
• 3:08
  or
• 3:08
  interrupt-for-human-approval.
• 3:10
  Six
• 3:10
  built-in
• 3:11
  DLP
• 3:11
  regex
• 3:12
  patterns
• 3:12
  cover
• 3:12
  the
• 3:12
  usual
• 3:13
  suspects
• 3:13
  (API
• 3:14
  keys,
• 3:14
  passwords,
• 3:15
  credit
• 3:15
  cards,
• 3:16
  SSNs),
• 3:17
  with
• 3:17
  custom
• 3:17
  patterns
• 3:18
  supported.
• 3:19
  Transport
• 3:20
  security
• 3:20
  follows
• 3:20
  a
• 3:21
  loopback-plaintext-default
• 3:22
  model
• 3:22
  that
• 3:22
  fail-closes
• 3:23
  to
• 3:23
  TLS
• 3:23
  off-loopback
• 3:24
  —
• 3:25
  a
• 3:25
  thoughtful
• 3:25
  design
• 3:26
  for
• 3:26
  a
• 3:26
  security
• 3:26
  tool
• 3:26
  that
• 3:27
  would
• 3:27
  be
• 3:27
  embarrassing
• 3:27
  to
• 3:27
  compromise.
• 3:29
  The
• 3:29
  need
• 3:29
  for
• 3:29
  such
• 3:29
  tools
• 3:30
  is
• 3:30
  underscored
• 3:30
  by
• 3:30
  a
• 3:30
  Chinese-language
• 3:31
  guide
• 3:31
  on
• 3:32
  GitHub
• 3:32
  that
• 3:32
  walks
• 3:32
  users
• 3:33
  through
• 3:33
  jailbreaking
• 3:34
  OpenAI's
• 3:34
  Codex
• 3:35
  CLI
• 3:35
  by
• 3:36
  overriding
• 3:36
  its
• 3:36
  system
• 3:37
  prompt
• 3:37
  via
• 3:37
  the
• 3:37
  model
• 3:37
  instructions
• 3:38
  file
• 3:38
  field
• 3:39
  in
• 3:39
  config.toml.
• 3:41
  The
• 3:42
  mechanism
• 3:42
  is
• 3:42
  a
• 3:42
  straightforward
• 3:43
  replace,
• 3:44
  not
• 3:44
  append
• 3:44
  —
• 3:45
  the
• 3:45
  official
• 3:45
  safety
• 3:46
  instructions
• 3:46
  are
• 3:46
  completely
• 3:47
  removed,
• 3:47
  and
• 3:48
  approval
• 3:48
  policy
• 3:48
  =
• 3:49
  "never"
• 3:50
  combined
• 3:50
  with
• 3:50
  sandbox
• 3:51
  mode
• 3:51
  =
• 3:52
  "danger-full-access"
• 3:52
  disables
• 3:53
  all
• 3:53
  safety
• 3:54
  guardrails.
• 3:55
  This
• 3:55
  is
• 3:55
  not
• 3:55
  a
• 3:55
  vulnerability
• 3:56
  in
• 3:56
  the
• 3:56
  traditional
• 3:56
  sense;
• 3:57
  it
• 3:57
  is
• 3:57
  a
• 3:57
  configuration
• 3:58
  option
• 3:58
  that
• 3:58
  does
• 3:58
  exactly
• 3:59
  what
• 3:59
  it
• 3:59
  says.
• 4:00
  But
• 4:00
  it
• 4:00
  illustrates
• 4:01
  why
• 4:01
  deterministic
• 4:02
  enforcement
• 4:02
  at
• 4:02
  the
• 4:02
  tool-call
• 4:03
  boundary
• 4:03
  (SentinelMCP's
• 4:04
  approach)
• 4:04
  matters
• 4:05
  more
• 4:05
  than
• 4:05
  advisory
• 4:06
  guardrails
• 4:06
  baked
• 4:06
  into
• 4:07
  agent
• 4:07
  prompts.
• 4:08
  If
• 4:08
  the
• 4:08
  user
• 4:08
  controls
• 4:09
  the
• 4:09
  prompt,
• 4:09
  the
• 4:09
  prompt
• 4:10
  is
• 4:10
  not
• 4:10
  a
• 4:10
  security
• 4:11
  boundary.
• 4:12
  A
• 4:12
  video
• 4:12
  walkthrough
• 4:13
  from
• 4:13
  a
• 4:13
  security
• 4:13
  researcher
• 4:14
  breaks
• 4:14
  AI's
• 4:15
  vulnerability-finding
• 4:15
  advantage
• 4:16
  into
• 4:16
  three
• 4:16
  factors:
• 4:17
  static
• 4:18
  analysis
• 4:18
  speed
• 4:19
  (a
• 4:19
  model
• 4:19
  internalizes
• 4:20
  a
• 4:20
  400,000-line
• 4:22
  codebase
• 4:22
  in
• 4:22
  minutes,
• 4:23
  versus
• 4:24
  a
• 4:24
  human's
• 4:24
  ~10,000-line
• 4:26
  daily
• 4:26
  limit),
• 4:27
  full-context
• 4:28
  pattern
• 4:28
  matching
• 4:28
  (source-to-sink
• 4:29
  tracing
• 4:30
  that
• 4:30
  would
• 4:30
  take
• 4:30
  a
• 4:30
  human
• 4:30
  weeks),
• 4:31
  and
• 4:31
  relentless
• 4:32
  automation.
• 4:33
  The
• 4:33
  presenter
• 4:34
  cites
• 4:34
  AI
• 4:34
  finding
• 4:35
  two
• 4:35
  real
• 4:35
  bugs
• 4:35
  in
• 4:35
  the
• 4:35
  Linux
• 4:36
  kernel's
• 4:36
  SMB
• 4:37
  handler
• 4:37
  alongside
• 4:38
  100–200
• 4:39
  false
• 4:40
  positives
• 4:40
  —
• 4:41
  the
• 4:41
  mitigation
• 4:41
  being
• 4:41
  ASan
• 4:42
  or
• 4:42
  QEMU
• 4:43
  harnesses
• 4:44
  that
• 4:44
  prove
• 4:44
  whether
• 4:44
  a
• 4:44
  reported
• 4:45
  bug
• 4:45
  is
• 4:45
  real.
• 4:46
  AI
• 4:46
  will
• 4:47
  not
• 4:47
  replace
• 4:47
  vulnerability
• 4:48
  researchers
• 4:48
  but
• 4:48
  will
• 4:49
  "turn
• 4:49
  them
• 4:49
  into
• 4:49
  a
• 4:49
  hundred
• 4:50
  copies
• 4:50
  of
• 4:50
  themselves."
• 4:51
  The
• 4:51
  researcher's
• 4:52
  own
• 4:52
  anecdote
• 4:52
  is
• 4:53
  telling:
• 4:53
  he
• 4:54
  spent
• 4:54
  significant
• 4:54
  time
• 4:55
  trying
• 4:55
  to
• 4:55
  identify
• 4:56
  an
• 4:56
  obscure
• 4:56
  chipset
• 4:56
  from
• 4:57
  firmware,
• 4:57
  then
• 4:58
  pointed
• 4:58
  Opus
• 4:58
  4.7
• 4:59
  at
• 4:59
  it,
• 5:00
  which
• 5:00
  instantly
• 5:01
  identified
• 5:01
  the
• 5:01
  chipset,
• 5:02
  entry
• 5:02
  point,
• 5:03
  interrupt
• 5:03
  vector
• 5:03
  table,
• 5:04
  and
• 5:04
  RTOS
• 5:05
  startup
• 5:05
  routine.
• 5:09
  Anthropic
• 5:10
  published
• 5:10
  research
• 5:10
  into
• 5:11
  what
• 5:11
  it
• 5:11
  calls
• 5:11
  "J-Space"
• 5:11
  —
• 5:12
  a
• 5:12
  Jacobian-lens
• 5:13
  view
• 5:14
  of
• 5:14
  a
• 5:14
  model's
• 5:14
  internal
• 5:15
  deliberation
• 5:15
  as
• 5:15
  it
• 5:15
  generates
• 5:16
  each
• 5:16
  token.
• 5:17
  The
• 5:17
  idea
• 5:17
  is
• 5:18
  to
• 5:18
  observe
• 5:18
  the
• 5:18
  model's
• 5:18
  hidden
• 5:19
  state
• 5:19
  as
• 5:19
  it
• 5:19
  passes
• 5:19
  through
• 5:20
  layers,
• 5:20
  revealing
• 5:21
  which
• 5:21
  candidate
• 5:21
  tokens
• 5:22
  the
• 5:22
  model
• 5:22
  is
• 5:22
  "considering"
• 5:23
  before
• 5:24
  committing
• 5:24
  to
• 5:24
  an
• 5:24
  output.
• 5:25
  Anthropic
• 5:26
  released
• 5:26
  the
• 5:26
  J-Space
• 5:27
  lens
• 5:27
  code
• 5:27
  and
• 5:27
  partnered
• 5:28
  with
• 5:28
  Neuronpedia
• 5:28
  to
• 5:28
  build
• 5:29
  an
• 5:29
  interactive
• 5:29
  demonstration
• 5:30
  on
• 5:30
  Qwen
• 5:30
  3.6
• 5:31
  27B.
• 5:32
  The
• 5:33
  LocalLLaMA
• 5:33
  community
• 5:33
  response
• 5:34
  was
• 5:34
  split:
• 5:35
  some
• 5:35
  researchers
• 5:36
  described
• 5:36
  it
• 5:36
  as
• 5:36
  a
• 5:36
  "more
• 5:36
  formal
• 5:37
  rediscovery
• 5:37
  of
• 5:38
  well-known
• 5:38
  intermediate
• 5:38
  layer
• 5:39
  abstract
• 5:39
  representation,"
• 5:40
  noting
• 5:40
  that
• 5:41
  mid-training
• 5:41
  layer
• 5:41
  duplication
• 5:42
  experiments
• 5:42
  have
• 5:42
  long
• 5:43
  shown
• 5:43
  similar
• 5:43
  basin
• 5:44
  dynamics.
• 5:45
  Others
• 5:45
  saw
• 5:45
  genuine
• 5:46
  utility
• 5:46
  —
• 5:47
  if
• 5:47
  J-Space
• 5:48
  consistently
• 5:48
  reveals
• 5:49
  a
• 5:49
  model's
• 5:49
  reasoning
• 5:50
  trajectory,
• 5:50
  it
• 5:51
  could
• 5:51
  become
• 5:51
  a
• 5:51
  practical
• 5:52
  tool
• 5:52
  for
• 5:52
  alignment
• 5:52
  verification,
• 5:53
  evaluation,
• 5:54
  and
• 5:54
  understanding
• 5:54
  decoherence
• 5:55
  at
• 5:55
  long
• 5:55
  context.
• 5:56
  One
• 5:56
  commenter
• 5:57
  with
• 5:57
  experience
• 5:57
  in
• 5:57
  layer-duplication
• 5:58
  experiments
• 5:58
  offered
• 5:59
  a
• 5:59
  physical
• 5:59
  analogy:
• 6:00
  intermediate
• 6:01
  representations
• 6:01
  are
• 6:01
  basins
• 6:02
  where
• 6:02
  "the
• 6:02
  marble
• 6:02
  can
• 6:03
  easily
• 6:03
  flip,"
• 6:04
  while
• 6:04
  final
• 6:04
  layers
• 6:04
  form
• 6:05
  rigid
• 6:05
  channels.
• 6:06
  Duplicating
• 6:06
  intermediate
• 6:07
  layers
• 6:07
  works
• 6:07
  with
• 6:08
  modest
• 6:08
  continued
• 6:08
  pretraining;
• 6:09
  duplicating
• 6:10
  final
• 6:10
  layers
• 6:11
  causes
• 6:11
  the
• 6:11
  model
• 6:11
  to
• 6:11
  "go
• 6:11
  to
• 6:12
  crazy
• 6:12
  tangent"
• 6:13
  even
• 6:13
  after
• 6:13
  80–100B
• 6:15
  tokens
• 6:15
  of
• 6:15
  CPT.
• 6:16
  Meanwhile,
• 6:17
  a
• 6:17
  researcher
• 6:17
  tested
• 6:18
  how
• 6:18
  far
• 6:18
  pure
• 6:18
  dynamics
• 6:19
  can
• 6:19
  go
• 6:19
  for
• 6:19
  learning
• 6:20
  word
• 6:20
  meaning
• 6:20
  —
• 6:21
  no
• 6:21
  MLP,
• 6:22
  no
• 6:22
  attention,
• 6:22
  no
• 6:23
  output
• 6:23
  layer.
• 6:24
  The
• 6:24
  system
• 6:24
  uses
• 6:25
  one
• 6:25
  256-dimensional
• 6:26
  vector
• 6:26
  per
• 6:27
  word
• 6:27
  as
• 6:27
  a
• 6:27
  "well,"
• 6:28
  a
• 6:28
  start
• 6:28
  state,
• 6:29
  and
• 6:29
  two
• 6:29
  learned
• 6:29
  scalars
• 6:30
  (pull
• 6:30
  strength
• 6:30
  and
• 6:30
  readout
• 6:31
  temperature).
• 6:32
  Context
• 6:32
  words
• 6:33
  pull
• 6:33
  the
• 6:33
  state
• 6:33
  toward
• 6:33
  their
• 6:33
  wells
• 6:34
  via
• 6:34
  a
• 6:34
  cosine-based
• 6:35
  update
• 6:35
  rule,
• 6:36
  and
• 6:36
  the
• 6:36
  final
• 6:36
  position
• 6:36
  encodes
• 6:37
  meaning.
• 6:38
  Trained
• 6:38
  on
• 6:38
  7.5%
• 6:39
  of
• 6:39
  Wikipedia
• 6:40
  with
• 6:40
  CBOW-style
• 6:41
  fill-in-the-blank,
• 6:42
  it
• 6:42
  hits
• 6:42
  SimLex-999
• 6:43
  ρ
• 6:44
  =
• 6:44
  0.362
• 6:45
  on
• 6:45
  nouns
• 6:46
  —
• 6:46
  landing
• 6:47
  inside
• 6:47
  the
• 6:47
  classic
• 6:47
  word2vec/GloVe
• 6:51
  band
• 6:52
  (0.37–0.44)
• 6:54
  on
• 6:54
  a
• 6:55
  fraction
• 6:55
  of
• 6:55
  the
• 6:55
  data
• 6:55
  and
• 6:56
  3.2
• 6:56
  hours
• 6:56
  on
• 6:57
  a
• 6:57
  MacBook.
• 6:58
  The
• 6:58
  neighborhoods
• 6:58
  are
• 6:58
  sensible
• 6:59
  (physics
• 6:59
  →
• 6:59
  chemistry,
• 6:59
  mathematics,
• 7:00
  astronomy;
• 7:01
  cat
• 7:01
  →
• 7:01
  tabby,
• 7:02
  dog,
• 7:02
  pet,
• 7:03
  felis).
• 7:04
  The
• 7:04
  system
• 7:04
  has
• 7:04
  clear
• 7:05
  limitations
• 7:05
  —
• 7:06
  single-sense
• 7:07
  only
• 7:07
  (apple
• 7:07
  collapses
• 7:08
  to
• 7:08
  the
• 7:08
  company),
• 7:09
  frequency-bound,
• 7:10
  no
• 7:10
  subword
• 7:10
  handling
• 7:10
  —
• 7:12
  but
• 7:12
  the
• 7:12
  point
• 7:12
  is
• 7:12
  conceptual:
• 7:13
  a
• 7:13
  fully
• 7:14
  inspectable
• 7:14
  contraction
• 7:15
  toward
• 7:15
  learned
• 7:16
  point-attractors,
• 7:17
  with
• 7:17
  a
• 7:17
  measurable
• 7:17
  Lyapunov
• 7:18
  energy,
• 7:19
  can
• 7:19
  do
• 7:19
  the
• 7:19
  "compress
• 7:20
  a
• 7:20
  sequence
• 7:20
  into
• 7:20
  meaning"
• 7:21
  job
• 7:21
  without
• 7:21
  any
• 7:22
  learned
• 7:22
  nonlinearity.
• 7:24
  A
• 7:24
  Dharma
• 7:24
  AI
• 7:25
  blog
• 7:25
  post
• 7:25
  synthesizes
• 7:26
  the
• 7:26
  case
• 7:26
  that
• 7:26
  specialization
• 7:27
  is
• 7:27
  structurally
• 7:28
  inevitable,
• 7:29
  drawing
• 7:29
  on
• 7:29
  a
• 7:30
  2026
• 7:30
  paper
• 7:31
  by
• 7:31
  Goldfeder,
• 7:32
  Wyder,
• 7:32
  LeCun,
• 7:33
  and
• 7:33
  Shwartz-Ziv.
• 7:35
  The
• 7:35
  argument
• 7:36
  runs
• 7:36
  through
• 7:36
  four
• 7:37
  analytical
• 7:37
  traditions:
• 7:38
  the
• 7:38
  No
• 7:39
  Free
• 7:39
  Lunch
• 7:39
  theorem
• 7:40
  (no
• 7:40
  general
• 7:40
  algorithm
• 7:41
  outperforms
• 7:41
  across
• 7:42
  all
• 7:42
  problems),
• 7:43
  evolutionary
• 7:44
  biology
• 7:44
  (selection
• 7:45
  favors
• 7:45
  niche
• 7:45
  fit
• 7:46
  over
• 7:46
  generalist
• 7:46
  breadth),
• 7:47
  competitive
• 7:48
  markets
• 7:48
  (concentrated
• 7:49
  capacity
• 7:49
  outperforms
• 7:50
  distributed
• 7:50
  capacity),
• 7:51
  and
• 7:52
  machine
• 7:52
  learning
• 7:52
  itself
• 7:52
  (negative
• 7:53
  transfer,
• 7:54
  MoE
• 7:54
  architectures
• 7:55
  recovering
• 7:55
  specialization
• 7:56
  internally,
• 7:57
  AlphaFold's
• 7:58
  domain-targeted
• 7:58
  breakthrough).
• 8:00
  The
• 8:00
  post
• 8:00
  carefully
• 8:01
  distinguishes
• 8:01
  domain
• 8:02
  knowledge
• 8:02
  (hand-coded
• 8:03
  features,
• 8:04
  targeted
• 8:04
  by
• 8:04
  Sutton's
• 8:05
  Bitter
• 8:05
  Lesson)
• 8:05
  from
• 8:06
  domain
• 8:06
  specialization
• 8:07
  (directing
• 8:07
  resources
• 8:08
  toward
• 8:08
  a
• 8:08
  bounded
• 8:08
  task
• 8:09
  set),
• 8:09
  arguing
• 8:10
  both
• 8:10
  can
• 8:10
  be
• 8:10
  true
• 8:11
  simultaneously.
• 8:12
  The
• 8:12
  practical
• 8:13
  implication
• 8:13
  for
• 8:14
  agent
• 8:14
  builders:
• 8:15
  deploying
• 8:15
  narrow,
• 8:16
  hyperspecialized
• 8:17
  agents
• 8:17
  with
• 8:17
  less
• 8:17
  context
• 8:18
  should
• 8:18
  reduce
• 8:19
  hallucinations
• 8:19
  and
• 8:20
  token
• 8:20
  consumption,
• 8:21
  making
• 8:21
  specialization
• 8:22
  both
• 8:22
  cheap
• 8:23
  and
• 8:23
  correct.
• 8:28
  A
• 8:28
  developer
• 8:28
  has
• 8:28
  published
• 8:29
  a
• 8:29
  brutally
• 8:29
  honest
• 8:30
  post-mortem
• 8:30
  of
• 8:30
  a
• 8:30
  production
• 8:31
  LLM
• 8:31
  service
• 8:31
  that
• 8:32
  handled
• 8:32
  medical
• 8:32
  appointment
• 8:33
  scheduling
• 8:33
  via
• 8:33
  messenger,
• 8:34
  using
• 8:34
  PydanticAI
• 8:35
  for
• 8:35
  orchestration
• 8:36
  and
• 8:36
  OpenRouter
• 8:36
  for
• 8:37
  model
• 8:37
  routing
• 8:37
  across
• 8:37
  seven
• 8:38
  providers
• 8:38
  (GLM,
• 8:39
  DeepSeek,
• 8:39
  Mimo,
• 8:40
  Qwen,
• 8:40
  ChatGPT,
• 8:41
  Claude,
• 8:41
  Minimax).
• 8:42
  The
• 8:43
  project
• 8:43
  ran
• 8:43
  for
• 8:43
  eight
• 8:43
  months
• 8:44
  before
• 8:44
  the
• 8:44
  clients
• 8:44
  terminated
• 8:45
  the
• 8:45
  contract,
• 8:46
  and
• 8:46
  the
• 8:46
  developer's
• 8:46
  relief
• 8:47
  is
• 8:47
  palpable:
• 8:48
  "I
• 8:48
  AM
• 8:48
  SO
• 8:49
  FUCKING
• 8:49
  FREE!"
• 8:50
  The
• 8:50
  failure
• 8:51
  modes
• 8:51
  are
• 8:51
  a
• 8:51
  catalog
• 8:51
  of
• 8:52
  everything
• 8:52
  that
• 8:52
  can
• 8:52
  go
• 8:52
  wrong
• 8:53
  when
• 8:53
  an
• 8:53
  LLM
• 8:53
  mediates
• 8:54
  between
• 8:54
  a
• 8:54
  business
• 8:54
  and
• 8:55
  its
• 8:55
  customers.
• 8:56
  Structured
• 8:56
  output
• 8:56
  loops
• 8:57
  where
• 8:57
  the
• 8:57
  model
• 8:57
  repeatedly
• 8:58
  failed
• 8:58
  validation.
• 8:59
  An
• 8:59
  appointment
• 9:00
  bot
• 9:00
  that,
• 9:00
  when
• 9:01
  the
• 9:01
  requested
• 9:01
  10:00
• 9:02
  slot
• 9:02
  was
• 9:02
  unavailable,
• 9:03
  silently
• 9:04
  booked
• 9:04
  11:00
• 9:04
  and
• 9:05
  then
• 9:05
  gaslit
• 9:05
  the
• 9:05
  user:
• 9:06
  "You
• 9:06
  asked
• 9:06
  for
• 9:06
  11:00,
• 9:07
  and
• 9:08
  we
• 9:08
  are
• 9:08
  expecting
• 9:08
  you."
• 9:09
  An
• 9:09
  agent
• 9:10
  that
• 9:10
  cancelled
• 9:10
  all
• 9:11
  of
• 9:11
  a
• 9:11
  user's
• 9:11
  existing
• 9:12
  appointments
• 9:12
  to
• 9:12
  "free
• 9:13
  up
• 9:13
  space"
• 9:13
  for
• 9:13
  a
• 9:13
  new
• 9:14
  one
• 9:14
  —
• 9:15
  which
• 9:15
  then
• 9:15
  also
• 9:15
  failed.
• 9:16
  Emoji
• 9:16
  in
• 9:17
  user
• 9:17
  messages
• 9:17
  causing
• 9:17
  the
• 9:18
  bot
• 9:18
  to
• 9:18
  hallucinate
• 9:18
  personal
• 9:19
  details
• 9:19
  about
• 9:19
  the
• 9:19
  user's
• 9:20
  nonexistent
• 9:20
  son.
• 9:21
  The
• 9:22
  fundamental
• 9:22
  lesson:
• 9:23
  a
• 9:23
  95%
• 9:24
  success
• 9:24
  rate
• 9:25
  sounds
• 9:25
  impressive
• 9:26
  until
• 9:26
  the
• 9:26
  remaining
• 9:26
  5%
• 9:27
  "will
• 9:27
  spoil
• 9:27
  everything."
• 9:28
  The
• 9:29
  service
• 9:29
  devolved
• 9:29
  into
• 9:30
  "constant
• 9:30
  monitoring,
• 9:31
  praying
• 9:31
  that
• 9:31
  it
• 9:32
  won't
• 9:32
  fuck
• 9:32
  up
• 9:32
  another
• 9:32
  conversation."
• 9:33
  Community
• 9:34
  response
• 9:34
  was
• 9:35
  divided
• 9:35
  —
• 9:35
  some
• 9:36
  blamed
• 9:36
  OpenRouter
• 9:37
  (where
• 9:37
  providers
• 9:38
  may
• 9:38
  serve
• 9:38
  quantized
• 9:39
  models
• 9:39
  without
• 9:39
  disclosure)
• 9:40
  and
• 9:40
  missing
• 9:40
  human-in-the-loop
• 9:41
  gates,
• 9:42
  while
• 9:42
  others
• 9:42
  argued
• 9:42
  the
• 9:43
  error
• 9:43
  rate
• 9:43
  is
• 9:43
  simply
• 9:44
  too
• 9:44
  high
• 9:44
  for
• 9:44
  autonomous
• 9:45
  customer-facing
• 9:45
  interactions
• 9:46
  with
• 9:46
  destructive
• 9:47
  actions.
• 9:47
  HuggingFace's
• 9:48
  "Every
• 9:49
  Eval
• 9:49
  Ever"
• 9:50
  (EEE)
• 9:50
  project
• 9:51
  tackles
• 9:51
  a
• 9:51
  different
• 9:51
  production
• 9:52
  problem:
• 9:53
  evaluation
• 9:53
  results
• 9:54
  that
• 9:54
  inform
• 9:54
  model
• 9:55
  selection
• 9:55
  are
• 9:55
  scattered,
• 9:56
  inconsistent,
• 9:57
  and
• 9:57
  often
• 9:58
  unreproducible.
• 9:59
  The
• 9:59
  same
• 10:00
  model
• 10:00
  on
• 10:00
  the
• 10:00
  same
• 10:00
  benchmark
• 10:01
  can
• 10:01
  return
• 10:01
  scores
• 10:02
  of
• 10:02
  63.7
• 10:03
  or
• 10:03
  48.8
• 10:04
  depending
• 10:04
  on
• 10:04
  who
• 10:05
  ran
• 10:05
  it
• 10:05
  and
• 10:05
  how.
• 10:06
  EEE
• 10:07
  provides
• 10:07
  a
• 10:07
  single
• 10:07
  JSON
• 10:08
  schema
• 10:08
  that
• 10:09
  records
• 10:09
  what
• 10:09
  a
• 10:09
  metric
• 10:09
  means,
• 10:10
  the
• 10:10
  generation
• 10:11
  config,
• 10:11
  harness
• 10:12
  version,
• 10:12
  and
• 10:13
  optional
• 10:13
  per-sample
• 10:13
  outputs.
• 10:15
  The
• 10:15
  datastore
• 10:15
  has
• 10:16
  grown
• 10:16
  to
• 10:16
  229,000
• 10:17
  results
• 10:18
  across
• 10:18
  22,000+
• 10:20
  models
• 10:21
  and
• 10:21
  2,200
• 10:22
  benchmarks.
• 10:23
  A
• 10:23
  new
• 10:24
  converter
• 10:24
  bridges
• 10:24
  EEE
• 10:25
  records
• 10:25
  to
• 10:25
  HuggingFace
• 10:26
  Community
• 10:26
  Evals,
• 10:27
  so
• 10:27
  scores
• 10:28
  appear
• 10:28
  on
• 10:28
  model
• 10:28
  cards
• 10:29
  with
• 10:29
  backlinks
• 10:29
  to
• 10:29
  full
• 10:30
  structured
• 10:30
  records.
• 10:31
  Reproducing
• 10:32
  the
• 10:32
  existing
• 10:33
  runs
• 10:33
  from
• 10:33
  scratch
• 10:33
  would
• 10:34
  cost
• 10:34
  hundreds
• 10:34
  of
• 10:34
  thousands
• 10:35
  of
• 10:35
  dollars
• 10:35
  —
• 10:36
  a
• 10:36
  compelling
• 10:37
  argument
• 10:37
  for
• 10:37
  not
• 10:37
  letting
• 10:38
  evaluation
• 10:38
  data
• 10:39
  scatter.
• 10:40
  On
• 10:40
  the
• 10:40
  creative
• 10:40
  end,
• 10:41
  a
• 10:41
  Reddit
• 10:41
  post
• 10:42
  proposes
• 10:42
  entropy-aware
• 10:43
  sampling
• 10:43
  to
• 10:43
  improve
• 10:44
  LLM
• 10:44
  creative
• 10:45
  writing.
• 10:45
  The
• 10:46
  core
• 10:46
  insight:
• 10:47
  standard
• 10:47
  temperature-based
• 10:48
  sampling
• 10:48
  applies
• 10:49
  uniform
• 10:49
  randomness,
• 10:50
  but
• 10:50
  creative
• 10:51
  writing
• 10:51
  benefits
• 10:51
  from
• 10:52
  structured
• 10:52
  variation
• 10:52
  —
• 10:53
  high
• 10:54
  entropy
• 10:54
  at
• 10:54
  decision
• 10:54
  points,
• 10:55
  lower
• 10:55
  entropy
• 10:56
  for
• 10:56
  follow-through.
• 10:57
  The
• 10:57
  approach
• 10:58
  uses
• 10:58
  future-entropy
• 10:59
  lookahead
• 10:59
  to
• 10:59
  weight
• 10:59
  token
• 11:00
  selection,
• 11:01
  with
• 11:01
  a
• 11:01
  sine-wave
• 11:01
  modulation
• 11:02
  pattern.
• 11:03
  Community
• 11:03
  reception
• 11:04
  was
• 11:04
  enthusiastic
• 11:05
  but
• 11:05
  noted
• 11:06
  the
• 11:06
  computational
• 11:06
  cost.
• 11:07
  One
• 11:08
  commenter
• 11:08
  captured
• 11:08
  the
• 11:09
  deeper
• 11:09
  issue:
• 11:10
  "when
• 11:10
  humans
• 11:10
  write
• 11:10
  text,
• 11:11
  we
• 11:11
  don't
• 11:11
  think
• 11:12
  of
• 11:12
  the
• 11:12
  most
• 11:12
  likely
• 11:12
  continuation,
• 11:14
  we
• 11:14
  think
• 11:14
  of
• 11:14
  the
• 11:14
  continuation
• 11:15
  that
• 11:15
  most
• 11:15
  engages
• 11:16
  our
• 11:16
  readers...
• 11:16
  We
• 11:16
  are
• 11:17
  engaged
• 11:17
  by
• 11:17
  a
• 11:17
  mixture
• 11:18
  of
• 11:18
  familiarity
• 11:18
  and
• 11:19
  unpredictability."
• 11:24
  A
• 11:24
  community
• 11:25
  developer
• 11:25
  has
• 11:25
  merged
• 11:26
  quantized
• 11:26
  KV
• 11:27
  cache
• 11:27
  fixes
• 11:27
  into
• 11:28
  a
• 11:28
  custom
• 11:28
  llama.cpp
• 11:29
  branch
• 11:30
  for
• 11:30
  DeepSeek
• 11:30
  V4,
• 11:31
  and
• 11:31
  the
• 11:32
  headline
• 11:32
  result
• 11:32
  is
• 11:32
  striking:
• 11:33
  the
• 11:33
  antirez
• 11:34
  IQ2
• 11:35
  XXS
• 11:35
  model
• 11:36
  fits
• 11:36
  with
• 11:36
  a
• 11:36
  full
• 11:36
  1-million-token
• 11:37
  context
• 11:38
  window
• 11:38
  on
• 11:38
  a
• 11:38
  single
• 11:38
  RTX
• 11:39
  PRO
• 11:39
  6000
• 11:40
  using
• 11:40
  q8
• 11:41
  0
• 11:41
  KV
• 11:41
  cache.
• 11:43
  Benchmarks
• 11:43
  show
• 11:43
  prompt
• 11:44
  processing
• 11:44
  at
• 11:44
  1,790
• 11:45
  t/s
• 11:46
  at
• 11:46
  2K
• 11:47
  context
• 11:47
  degrading
• 11:48
  to
• 11:48
  201
• 11:48
  t/s
• 11:49
  at
• 11:49
  1M,
• 11:50
  while
• 11:50
  token
• 11:50
  generation
• 11:51
  drops
• 11:51
  from
• 11:51
  56
• 11:52
  t/s
• 11:52
  to
• 11:52
  15.3
• 11:53
  t/s.
• 11:55
  The
• 11:55
  quality
• 11:55
  impact
• 11:55
  is
• 11:55
  negligible
• 11:56
  —
• 11:57
  perplexity
• 11:57
  on
• 11:57
  wikitext-2-raw
• 11:58
  moves
• 11:59
  from
• 11:59
  4.0242
• 12:00
  (f16)
• 12:01
  to
• 12:01
  4.0304
• 12:01
  (q8
• 12:02
  0)
• 12:03
  to
• 12:03
  4.0512
• 12:04
  (q4
• 12:04
  0),
• 12:05
  a
• 12:05
  difference
• 12:06
  most
• 12:06
  users
• 12:06
  would
• 12:06
  never
• 12:07
  notice.
• 12:08
  Community
• 12:08
  results
• 12:08
  reinforce
• 12:09
  the
• 12:09
  gains:
• 12:10
  one
• 12:10
  user
• 12:10
  on
• 12:11
  an
• 12:11
  RTX
• 12:11
  PRO
• 12:11
  6000
• 12:12
  +
• 12:13
  RTX
• 12:13
  5090
• 12:15
  reported
• 12:15
  4x
• 12:16
  prompt
• 12:16
  processing
• 12:16
  speedup
• 12:17
  and
• 12:17
  context
• 12:18
  capacity
• 12:18
  jumping
• 12:19
  from
• 12:19
  80K
• 12:19
  to
• 12:19
  256K
• 12:21
  with
• 12:21
  25GB
• 12:22
  VRAM
• 12:22
  to
• 12:22
  spare.
• 12:23
  Another
• 12:24
  fitted
• 12:24
  1M
• 12:24
  context
• 12:25
  across
• 12:25
  8
• 12:25
  consumer
• 12:26
  GPUs
• 12:26
  (4x
• 12:27
  3090
• 12:28
  +
• 12:28
  4x
• 12:29
  3080).
• 12:30
  Athena,
• 12:31
  a
• 12:31
  fully
• 12:31
  offline
• 12:32
  voice-to-voice
• 12:32
  assistant,
• 12:33
  represents
• 12:34
  the
• 12:34
  other
• 12:34
  end
• 12:34
  of
• 12:34
  the
• 12:34
  local
• 12:35
  AI
• 12:35
  ambition:
• 12:36
  a
• 12:36
  complete
• 12:36
  voice
• 12:37
  pipeline
• 12:37
  with
• 12:37
  zero
• 12:38
  cloud
• 12:38
  dependencies.
• 12:39
  The
• 12:39
  system
• 12:40
  chains
• 12:40
  Qwen3.5-397B,
• 12:44
  Orpheus
• 12:45
  3B
• 12:45
  (neural
• 12:45
  TTS
• 12:46
  with
• 12:46
  emotional
• 12:46
  expression
• 12:47
  —
• 12:47
  laughs,
• 12:48
  sighs,
• 12:48
  gasps),
• 12:49
  Whisper-small.en,
• 12:49
  and
• 12:50
  SNAC
• 12:50
  (neural
• 12:50
  audio
• 12:50
  codec)
• 12:51
  into
• 12:51
  a
• 12:51
  four-process
• 12:52
  C++
• 12:52
  pipeline.
• 12:54
  It
• 12:54
  reads
• 12:54
  basic
• 12:54
  vocal
• 12:55
  affect,
• 12:55
  maintains
• 12:55
  cross-session
• 12:56
  memory,
• 12:56
  supports
• 12:57
  barge-in
• 12:57
  interruption,
• 12:58
  and
• 12:58
  runs
• 12:58
  on
• 12:58
  a
• 12:58
  single
• 12:58
  consumer
• 12:59
  GPU
• 12:59
  plus
• 13:00
  system
• 13:00
  RAM.
• 13:01
  No
• 13:01
  Python
• 13:01
  at
• 13:01
  runtime.
• 13:02
  At
• 13:03
  the
• 13:03
  budget
• 13:03
  end
• 13:03
  of
• 13:03
  the
• 13:03
  spectrum,
• 13:04
  a
• 13:04
  homelab
• 13:04
  builder
• 13:05
  documented
• 13:05
  6x
• 13:06
  P40
• 13:07
  GPUs
• 13:08
  (144GB
• 13:08
  total
• 13:09
  VRAM,
• 13:10
  ~$1,200
• 13:11
  used)
• 13:12
  running
• 13:12
  MiniMax
• 13:13
  M2.7
• 13:13
  at
• 13:14
  Q3
• 13:14
  XL
• 13:14
  quantization.
• 13:16
  Key
• 13:16
  findings:
• 13:17
  ubatch
• 13:17
  size
• 13:18
  matters
• 13:18
  more
• 13:18
  than
• 13:18
  batch
• 13:19
  size
• 13:19
  (105
• 13:20
  t/s
• 13:20
  pp512
• 13:21
  at
• 13:21
  ubatch
• 13:22
  256
• 13:23
  vs.
• 13:23
  73
• 13:24
  at
• 13:24
  ubatch
• 13:25
  512),
• 13:26
  flash
• 13:26
  attention
• 13:27
  is
• 13:27
  mandatory
• 13:27
  (disabling
• 13:28
  it
• 13:28
  caused
• 13:28
  "major
• 13:29
  regression"),
• 13:30
  and
• 13:30
  P2P
• 13:30
  or
• 13:31
  multi-queue
• 13:31
  launch
• 13:31
  produced
• 13:32
  no
• 13:32
  measurable
• 13:32
  gains.
• 13:34
  The
• 13:34
  optimal
• 13:34
  config
• 13:34
  delivers
• 13:35
  10.5
• 13:36
  t/s
• 13:36
  generation
• 13:37
  at
• 13:37
  32K
• 13:37
  context
• 13:38
  —
• 13:39
  not
• 13:39
  blazing
• 13:40
  fast,
• 13:40
  but
• 13:41
  usable
• 13:41
  for
• 13:41
  a
• 13:41
  rig
• 13:41
  that
• 13:41
  costs
• 13:42
  less
• 13:42
  than
• 13:42
  two
• 13:42
  months
• 13:42
  of
• 13:43
  cloud
• 13:43
  API
• 13:43
  bills.
• 13:44
  The
• 13:44
  uncensored
• 13:45
  model
• 13:45
  ecosystem
• 13:46
  continues
• 13:46
  to
• 13:46
  evolve,
• 13:47
  with
• 13:47
  a
• 13:47
  community
• 13:48
  member
• 13:48
  releasing
• 13:49
  "Uncensored
• 13:49
  Heretic"
• 13:50
  —
• 13:51
  a
• 13:51
  fine-tuned
• 13:51
  Gemma
• 13:51
  4
• 13:52
  12B
• 13:52
  achieving
• 13:52
  only
• 13:53
  13
• 13:54
  out
• 13:54
  of
• 13:54
  100
• 13:54
  refusals
• 13:55
  with
• 13:55
  a
• 13:55
  KL
• 13:55
  divergence
• 13:56
  of
• 13:56
  0.0367
• 13:58
  from
• 13:58
  the
• 13:58
  base
• 13:58
  model.
• 13:59
  Available
• 14:00
  in
• 14:00
  both
• 14:00
  safetensors
• 14:01
  and
• 14:01
  GGUF
• 14:01
  formats
• 14:02
  with
• 14:02
  Ollama
• 14:02
  one-liner
• 14:03
  install,
• 14:04
  it
• 14:04
  represents
• 14:04
  the
• 14:05
  ongoing
• 14:05
  tension
• 14:05
  between
• 14:06
  safety
• 14:06
  alignment
• 14:06
  and
• 14:07
  user
• 14:07
  autonomy
• 14:07
  in
• 14:07
  the
• 14:07
  open-weight
• 14:08
  ecosystem.
• 14:12
  Waveloom
• 14:12
  is
• 14:12
  a
• 14:12
  new
• 14:13
  terminal
• 14:13
  coding
• 14:13
  agent
• 14:14
  built
• 14:14
  in
• 14:14
  Go,
• 14:14
  engineered
• 14:15
  specifically
• 14:15
  around
• 14:15
  DeepSeek's
• 14:16
  prefix
• 14:16
  caching
• 14:17
  economics.
• 14:18
  The
• 14:18
  architecture
• 14:19
  keeps
• 14:19
  the
• 14:19
  longest
• 14:19
  common
• 14:20
  prefix
• 14:20
  cache-hot
• 14:21
  across
• 14:21
  turns
• 14:21
  using
• 14:21
  a
• 14:22
  four-tier
• 14:22
  watermark
• 14:22
  compaction
• 14:23
  system
• 14:23
  (Snip
• 14:24
  →
• 14:24
  Prune
• 14:24
  →
• 14:24
  Summarize),
• 14:25
  claiming
• 14:25
  95–99%
• 14:27
  cache
• 14:27
  hit
• 14:28
  rates
• 14:28
  and
• 14:28
  roughly
• 14:28
  1/50th
• 14:29
  the
• 14:29
  token
• 14:29
  cost
• 14:29
  of
• 14:29
  naive
• 14:30
  approaches.
• 14:31
  The
• 14:31
  LLM
• 14:31
  auto-selects
• 14:32
  between
• 14:32
  DeepSeek
• 14:33
  Pro
• 14:33
  (deep
• 14:33
  reasoning)
• 14:34
  and
• 14:34
  Flash
• 14:34
  (routine
• 14:35
  tasks)
• 14:35
  to
• 14:35
  maximize
• 14:36
  cache
• 14:36
  hits.
• 14:37
  For
• 14:37
  Claude
• 14:38
  Code
• 14:38
  users
• 14:38
  considering
• 14:38
  a
• 14:38
  switch,
• 14:39
  Waveloom
• 14:39
  drops
• 14:40
  in
• 14:40
  .claude/skills/
• 14:41
  SKILL.md
• 14:42
  files
• 14:42
  and
• 14:44
  .claude.json
• 14:45
  MCP
• 14:45
  configs
• 14:45
  with
• 14:45
  zero
• 14:46
  friction.
• 14:46
  It
• 14:47
  ships
• 14:47
  as
• 14:47
  a
• 14:47
  single
• 14:47
  ~18MB
• 14:48
  Go
• 14:48
  binary
• 14:49
  with
• 14:49
  12
• 14:49
  built-in
• 14:50
  tools,
• 14:50
  plan
• 14:51
  mode
• 14:51
  with
• 14:51
  write
• 14:51
  protection,
• 14:52
  sub-agent
• 14:52
  spawning,
• 14:53
  session
• 14:53
  persistence,
• 14:54
  and
• 14:54
  full
• 14:54
  Chinese/English
• 14:55
  i18n.
• 14:57
  On
• 14:57
  the
• 14:57
  capability
• 14:58
  side,
• 14:58
  a
• 14:58
  developer
• 14:59
  used
• 14:59
  Claude
• 14:59
  Code
• 14:59
  running
• 14:59
  Qwen3.6-27B
• 15:03
  locally
• 15:03
  to
• 15:03
  implement
• 15:04
  A
• 15:04
  pathfinding
• 15:04
  in
• 15:04
  a
• 15:04
  Java
• 15:05
  game
• 15:05
  —
• 15:05
  a
• 15:05
  12-hour
• 15:06
  marathon
• 15:06
  of
• 15:06
  autonomous
• 15:07
  testing.
• 15:08
  The
• 15:08
  model
• 15:08
  created
• 15:09
  its
• 15:09
  own
• 15:09
  testing
• 15:09
  harness
• 15:10
  that
• 15:10
  auto-piloted
• 15:10
  the
• 15:10
  player
• 15:11
  while
• 15:11
  an
• 15:11
  NPC
• 15:11
  chased,
• 15:12
  monitored
• 15:13
  logs
• 15:13
  in
• 15:13
  real-time,
• 15:14
  refactored
• 15:14
  code,
• 15:15
  and
• 15:15
  re-launched
• 15:15
  the
• 15:15
  game
• 15:16
  after
• 15:16
  each
• 15:16
  iteration.
• 15:17
  The
• 15:17
  result:
• 15:18
  an
• 15:19
  NPC
• 15:19
  that
• 15:19
  climbs
• 15:19
  blocks,
• 15:20
  drops
• 15:20
  down,
• 15:21
  and
• 15:21
  navigates
• 15:21
  around
• 15:22
  obstacles
• 15:22
  "most
• 15:22
  of
• 15:22
  the
• 15:22
  time."
• 15:23
  Community
• 15:24
  response
• 15:24
  was
• 15:24
  measured
• 15:25
  —
• 15:25
  A
• 15:26
  is
• 15:26
  a
• 15:26
  textbook
• 15:26
  algorithm
• 15:26
  —
• 15:27
  but
• 15:27
  the
• 15:27
  real
• 15:28
  achievement
• 15:28
  was
• 15:28
  the
• 15:29
  autonomous
• 15:29
  test
• 15:29
  loop,
• 15:30
  not
• 15:30
  the
• 15:30
  algorithm
• 15:30
  itself.
• 15:32
  The
• 15:32
  model
• 15:32
  validated
• 15:33
  its
• 15:33
  own
• 15:33
  changes
• 15:33
  iteratively
• 15:34
  against
• 15:34
  a
• 15:34
  verifiable
• 15:35
  ground
• 15:35
  truth,
• 15:36
  which
• 15:36
  is
• 15:36
  exactly
• 15:36
  the
• 15:36
  regime
• 15:37
  where
• 15:37
  vibecoding
• 15:37
  works.
• 15:38
  For
• 15:38
  a
• 15:38
  more
• 15:39
  exotic
• 15:39
  stress
• 15:39
  test,
• 15:40
  a
• 15:40
  developer
• 15:41
  tasked
• 15:41
  Codex
• 15:41
  with
• 15:42
  porting
• 15:42
  Pikachu
• 15:42
  Volleyball
• 15:43
  into
• 15:43
  UmLang
• 15:43
  —
• 15:44
  a
• 15:44
  Korean
• 15:45
  esoteric
• 15:45
  programming
• 15:46
  language
• 15:46
  born
• 15:46
  from
• 15:46
  an
• 15:46
  internet
• 15:47
  meme.
• 15:48
  After
• 15:48
  roughly
• 15:48
  41
• 15:49
  hours,
• 15:49
  Codex
• 15:50
  completed
• 15:50
  the
• 15:50
  project.
• 15:51
  Benchmarking
• 15:52
  five
• 15:52
  implementations
• 15:53
  (native
• 15:53
  Rust,
• 15:54
  original
• 15:54
  JavaScript,
• 15:55
  UmLang
• 15:56
  on
• 15:56
  Rust/Node/Python
• 15:57
  VMs)
• 15:58
  produced
• 15:58
  the
• 15:58
  expected
• 15:58
  ranking:
• 16:00
  Rust
• 16:00
  >
• 16:00
  JS
• 16:00
  >
• 16:01
  UmLang
• 16:01
  Rust
• 16:01
  VM
• 16:01
  >
• 16:02
  Node
• 16:02
  >
• 16:02
  Python.
• 16:03
  The
• 16:03
  interesting
• 16:04
  question:
• 16:05
  if
• 16:05
  future
• 16:05
  "sovereign
• 16:06
  AI"
• 16:06
  systems
• 16:07
  are
• 16:07
  trained
• 16:07
  around
• 16:08
  programming
• 16:08
  languages
• 16:09
  influenced
• 16:09
  by
• 16:09
  specific
• 16:10
  natural
• 16:10
  languages,
• 16:11
  could
• 16:11
  developers
• 16:12
  get
• 16:12
  better
• 16:12
  interactions
• 16:13
  because
• 16:13
  the
• 16:13
  abstractions
• 16:14
  match
• 16:14
  their
• 16:14
  native
• 16:15
  reasoning?
• 16:16
  The
• 16:16
  @metaharness/flywheel
• 16:18
  package
• 16:18
  formalizes
• 16:19
  what
• 16:19
  most
• 16:19
  "self-improving
• 16:20
  agent"
• 16:20
  projects
• 16:21
  leave
• 16:21
  hand-wavy:
• 16:22
  the
• 16:22
  promotion
• 16:23
  loop.
• 16:24
  The
• 16:24
  engine
• 16:24
  runs
• 16:24
  a
• 16:24
  cycle
• 16:25
  of
• 16:25
  run
• 16:25
  →
• 16:25
  measure
• 16:25
  →
• 16:26
  mutate
• 16:26
  →
• 16:26
  verify
• 16:26
  →
• 16:27
  promote,
• 16:28
  where
• 16:28
  each
• 16:28
  generation
• 16:29
  re-bases
• 16:29
  on
• 16:29
  the
• 16:29
  previous
• 16:30
  champion,
• 16:31
  building
• 16:31
  an
• 16:31
  immutable
• 16:32
  lineage
• 16:32
  of
• 16:32
  verified
• 16:33
  improvements.
• 16:34
  The
• 16:34
  key
• 16:34
  design
• 16:35
  decision:
• 16:35
  freeze
• 16:36
  the
• 16:36
  model,
• 16:36
  evolve
• 16:37
  the
• 16:37
  harness.
• 16:38
  The
• 16:38
  expensive
• 16:38
  LLM
• 16:39
  stays
• 16:39
  fixed;
• 16:40
  what
• 16:40
  improves
• 16:41
  is
• 16:41
  the
• 16:41
  cheap
• 16:41
  executor's
• 16:42
  operating
• 16:42
  policy.
• 16:44
  Every
• 16:44
  promotion
• 16:44
  is
• 16:45
  Ed25519-signed,
• 16:48
  and
• 16:48
  an
• 16:48
  external
• 16:49
  reviewer
• 16:49
  can
• 16:49
  replay
• 16:50
  the
• 16:50
  bundle
• 16:50
  and
• 16:50
  verify
• 16:51
  the
• 16:51
  gate
• 16:51
  was
• 16:51
  unchanged
• 16:52
  —
• 16:52
  trust
• 16:53
  the
• 16:53
  math,
• 16:53
  not
• 16:54
  the
• 16:54
  operator.
• 16:55
  The
• 16:55
  gate
• 16:55
  is
• 16:55
  conjunctive
• 16:56
  (primary
• 16:56
  score
• 16:57
  +
• 16:57
  holdout
• 16:57
  score),
• 16:58
  fingerprinted,
• 16:59
  and
• 16:59
  frozen
• 17:00
  during
• 17:00
  each
• 17:00
  run.
• 17:01
  An
• 17:01
  ADR-226
• 17:03
  receipt
• 17:04
  showed
• 17:04
  that
• 17:04
  an
• 17:04
  expensive
• 17:05
  read-only
• 17:05
  advisory
• 17:06
  loop
• 17:06
  produced
• 17:06
  "marginal
• 17:07
  lift
• 17:07
  at
• 17:07
  cost"
• 17:07
  —
• 17:08
  the
• 17:08
  executor
• 17:09
  policy
• 17:09
  is
• 17:10
  what
• 17:10
  actually
• 17:10
  mattered.
• 17:15
  CPPO
• 17:15
  (Cumulative
• 17:16
  Prefix-Divergence
• 17:17
  Policy
• 17:17
  Optimization)
• 17:18
  addresses
• 17:18
  a
• 17:18
  structural
• 17:19
  flaw
• 17:19
  in
• 17:19
  how
• 17:20
  PPO-style
• 17:20
  trust
• 17:21
  regions
• 17:21
  work
• 17:21
  for
• 17:21
  LLM
• 17:22
  reinforcement
• 17:22
  learning.
• 17:23
  Current
• 17:24
  methods
• 17:24
  apply
• 17:24
  uniform
• 17:25
  divergence
• 17:25
  thresholds
• 17:26
  across
• 17:26
  all
• 17:26
  token
• 17:26
  positions,
• 17:27
  ignoring
• 17:28
  two
• 17:28
  properties
• 17:28
  of
• 17:28
  autoregressive
• 17:29
  generation:
• 17:30
  early-token
• 17:30
  deviations
• 17:31
  compound
• 17:32
  across
• 17:32
  the
• 17:32
  entire
• 17:32
  subsequent
• 17:33
  sequence,
• 17:34
  and
• 17:34
  locally
• 17:34
  bounded
• 17:35
  per-token
• 17:35
  deviations
• 17:36
  can
• 17:36
  accumulate
• 17:36
  into
• 17:36
  large
• 17:37
  prefix-level
• 17:37
  drift.
• 17:38
  CPPO
• 17:39
  introduces
• 17:40
  two
• 17:40
  coupled
• 17:40
  mechanisms
• 17:41
  —
• 17:42
  a
• 17:42
  position-weighted
• 17:42
  threshold
• 17:43
  that
• 17:43
  is
• 17:43
  stricter
• 17:44
  at
• 17:44
  early
• 17:44
  positions
• 17:44
  (where
• 17:45
  the
• 17:45
  remaining
• 17:45
  horizon
• 17:45
  is
• 17:46
  longest)
• 17:46
  and
• 17:46
  relaxes
• 17:47
  toward
• 17:47
  the
• 17:47
  end,
• 17:48
  plus
• 17:48
  a
• 17:48
  cumulative
• 17:49
  prefix
• 17:49
  budget
• 17:49
  that
• 17:49
  dynamically
• 17:50
  restricts
• 17:51
  further
• 17:51
  divergence
• 17:52
  once
• 17:52
  preceding
• 17:52
  deviations
• 17:53
  have
• 17:53
  accumulated
• 17:54
  past
• 17:54
  a
• 17:54
  threshold.
• 17:55
  Evaluated
• 17:56
  on
• 17:56
  four
• 17:56
  Qwen3
• 17:57
  configurations
• 17:57
  (1.7B
• 17:58
  through
• 17:58
  30B-A3B-Base)
• 18:01
  on
• 18:01
  math
• 18:02
  reasoning
• 18:02
  tasks,
• 18:02
  CPPO
• 18:03
  achieved
• 18:04
  the
• 18:04
  best
• 18:04
  AIME24/25/26
• 18:08
  scores
• 18:08
  across
• 18:09
  all
• 18:09
  settings.
• 18:10
  The
• 18:10
  largest
• 18:10
  gain
• 18:11
  —
• 18:12
  5.56
• 18:12
  absolute
• 18:13
  points
• 18:13
  on
• 18:13
  the
• 18:13
  30B
• 18:14
  MoE
• 18:14
  model
• 18:14
  with
• 18:15
  32K
• 18:15
  rollout
• 18:16
  horizon
• 18:16
  —
• 18:17
  came
• 18:17
  precisely
• 18:18
  in
• 18:18
  the
• 18:18
  regime
• 18:18
  where
• 18:18
  remaining-horizon
• 18:19
  amplification
• 18:20
  is
• 18:20
  most
• 18:20
  pronounced.
• 18:21
  In
• 18:21
  matched
• 18:22
  comparisons
• 18:22
  against
• 18:22
  DPPO
• 18:23
  using
• 18:23
  identical
• 18:24
  divergence
• 18:24
  metrics,
• 18:25
  ablations
• 18:26
  confirmed
• 18:26
  both
• 18:26
  mechanisms
• 18:27
  contribute
• 18:27
  independently.
• 18:29
  CORTIS
• 18:31
  (Cumulative
• 18:34
  Orthogonal
• 18:34
  Identity
• 18:34
  Suppression)
• 18:34
  tackles
• 18:35
  a
• 18:35
  privacy
• 18:35
  problem
• 18:36
  that
• 18:36
  existing
• 18:37
  machine
• 18:37
  unlearning
• 18:38
  methods
• 18:38
  quietly
• 18:38
  ignore:
• 18:39
  sequential
• 18:40
  removal
• 18:40
  requests.
• 18:42
  Zero-shot
• 18:42
  TTS
• 18:43
  systems
• 18:43
  can
• 18:43
  clone
• 18:44
  a
• 18:44
  voice
• 18:44
  from
• 18:44
  seconds
• 18:45
  of
• 18:45
  audio,
• 18:45
  and
• 18:46
  GDPR/CCPA
• 18:48
  "right
• 18:49
  to
• 18:49
  be
• 18:49
  forgotten"
• 18:49
  requests
• 18:50
  arrive
• 18:50
  one
• 18:50
  at
• 18:50
  a
• 18:50
  time
• 18:50
  —
• 18:51
  not
• 18:52
  in
• 18:52
  a
• 18:52
  convenient
• 18:52
  batch.
• 18:53
  The
• 18:53
  authors
• 18:54
  show
• 18:54
  that
• 18:54
  applying
• 18:54
  prior
• 18:55
  unlearning
• 18:55
  methods
• 18:56
  sequentially
• 18:56
  causes
• 18:57
  "catastrophic
• 18:57
  re-learning":
• 18:58
  removing
• 18:59
  speaker
• 18:59
  B
• 18:59
  revives
• 19:00
  previously
• 19:00
  removed
• 19:01
  speaker
• 19:01
  A,
• 19:02
  because
• 19:02
  the
• 19:02
  retain
• 19:03
  loss
• 19:03
  provides
• 19:04
  no
• 19:04
  supervision
• 19:05
  on
• 19:05
  the
• 19:05
  parameters
• 19:05
  that
• 19:05
  drove
• 19:06
  A's
• 19:06
  erasure.
• 19:07
  Worse,
• 19:08
  retaining
• 19:08
  forget-set
• 19:09
  data
• 19:09
  for
• 19:09
  joint
• 19:10
  reprocessing
• 19:10
  creates
• 19:11
  a
• 19:11
  "data
• 19:11
  retention
• 19:12
  paradox"
• 19:13
  that
• 19:13
  violates
• 19:13
  the
• 19:13
  very
• 19:14
  right
• 19:14
  being
• 19:14
  exercised.
• 19:15
  CORTIS
• 19:18
  combines
• 19:18
  contrastive
• 19:19
  Fisher-information
• 19:20
  masking
• 19:20
  (restricting
• 19:21
  updates
• 19:21
  to
• 19:21
  the
• 19:21
  top
• 19:21
  30%
• 19:22
  most
• 19:22
  forget-relevant
• 19:23
  parameters
• 19:24
  while
• 19:24
  soft-excluding
• 19:25
  parameters
• 19:25
  important
• 19:26
  to
• 19:26
  prior
• 19:26
  speakers)
• 19:26
  with
• 19:27
  orthogonal
• 19:27
  projection
• 19:28
  that
• 19:28
  constrains
• 19:28
  gradient
• 19:29
  updates
• 19:29
  away
• 19:29
  from
• 19:30
  directions
• 19:30
  used
• 19:30
  in
• 19:31
  prior
• 19:31
  unlearning
• 19:31
  runs.
• 19:32
  On
• 19:33
  VoiceBox,
• 19:34
  it
• 19:34
  keeps
• 19:34
  all
• 19:34
  forget
• 19:35
  speakers'
• 19:35
  similarity
• 19:36
  below
• 19:36
  0.18
• 19:37
  across
• 19:37
  five
• 19:37
  sequential
• 19:38
  requests
• 19:38
  —
• 19:39
  a
• 19:39
  75%
• 19:40
  reduction
• 19:40
  from
• 19:41
  baseline
• 19:41
  —
• 19:42
  while
• 19:42
  preserving
• 19:42
  synthesis
• 19:43
  quality.
• 19:44
  The
• 19:44
  computational
• 19:44
  overhead
• 19:45
  is
• 19:45
  3.3%
• 19:46
  per
• 19:46
  optimizer
• 19:47
  step,
• 19:47
  versus
• 19:48
  quadratically
• 19:48
  growing
• 19:48
  costs
• 19:49
  for
• 19:49
  cumulative
• 19:49
  reprocessing.