Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 11, 2026

Transcript

AGI Dreams Podcast – June 11, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Fable
• 0:00
  shipped
• 0:00
  Monday.
• 0:01
  By
• 0:01
  Tuesday,
• 0:02
  the
• 0:02
  security
• 0:02
  community
• 0:03
  was
• 0:03
  filing
• 0:03
  complaints.
• 0:04
  Cybersecurity
• 0:05
  researchers
• 0:06
  report
• 0:06
  that
• 0:06
  Anthropic's
• 0:07
  newest
• 0:07
  model
• 0:07
  blocks
• 0:07
  innocuous
• 0:08
  requests
• 0:08
  containing
• 0:09
  keywords
• 0:09
  like
• 0:10
  "exploit,"
• 0:10
  "payload,"
• 0:11
  or
• 0:11
  "reverse
• 0:12
  shell"
• 0:12
  —
• 0:13
  terms
• 0:13
  that
• 0:13
  are
• 0:13
  table
• 0:13
  stakes
• 0:14
  in
• 0:14
  any
• 0:14
  penetration
• 0:15
  test
• 0:15
  or
• 0:15
  vulnerability
• 0:16
  assessment.
• 0:17
  The
• 0:17
  frustration
• 0:17
  is
• 0:18
  compounded
• 0:18
  by
• 0:18
  timing:
• 0:19
  Fable
• 0:20
  was
• 0:20
  positioned
• 0:20
  as
• 0:21
  the
• 0:21
  most
• 0:21
  capable
• 0:21
  model
• 0:22
  in
• 0:22
  Anthropic's
• 0:22
  lineup,
• 0:23
  and
• 0:23
  security
• 0:23
  professionals
• 0:24
  who
• 0:24
  upgraded
• 0:25
  specifically
• 0:25
  for
• 0:25
  its
• 0:25
  reasoning
• 0:26
  capability
• 0:27
  found
• 0:27
  themselves
• 0:27
  locked
• 0:28
  out
• 0:28
  of
• 0:28
  their
• 0:28
  primary
• 0:28
  use
• 0:28
  case.
• 0:29
  The
• 0:29
  guardrails
• 0:30
  appear
• 0:30
  to
• 0:30
  be
• 0:30
  keyword-triggered
• 0:31
  rather
• 0:31
  than
• 0:31
  intent-aware:
• 0:33
  researchers
• 0:33
  can
• 0:33
  sometimes
• 0:34
  bypass
• 0:34
  them
• 0:35
  by
• 0:35
  rephrasing,
• 0:36
  but
• 0:36
  the
• 0:36
  default
• 0:36
  behavior
• 0:37
  forces
• 0:37
  a
• 0:37
  fallback
• 0:38
  to
• 0:38
  Opus
• 0:38
  4.8
• 0:39
  for
• 0:39
  anything
• 0:39
  that
• 0:39
  smells
• 0:40
  like
• 0:40
  offensive
• 0:40
  security.
• 0:41
  Santiago
• 0:42
  Palmiotti
• 0:43
  of
• 0:43
  IBM
• 0:43
  X-Force
• 0:44
  Red
• 0:44
  and
• 0:44
  Matt
• 0:44
  Suiche
• 0:45
  of
• 0:45
  Tolmo
• 0:46
  have
• 0:46
  both
• 0:46
  publicly
• 0:46
  criticized
• 0:47
  the
• 0:47
  restrictions
• 0:48
  as
• 0:48
  counterproductive,
• 0:49
  arguing
• 0:49
  that
• 0:49
  the
• 0:49
  people
• 0:50
  most
• 0:50
  likely
• 0:50
  to
• 0:50
  trigger
• 0:51
  keyword
• 0:51
  blocks
• 0:52
  are
• 0:52
  exactly
• 0:52
  the
• 0:52
  people
• 0:52
  doing
• 0:53
  legitimate
• 0:53
  security
• 0:54
  work.
• 0:54
  The
• 0:55
  contrast
• 0:55
  with
• 0:55
  Nullsec-S1
• 0:56
  is
• 0:56
  instructive.
• 0:57
  Rather
• 0:58
  than
• 0:58
  restricting
• 0:58
  a
• 0:58
  general-purpose
• 0:59
  model,
• 1:00
  the
• 1:00
  Nullsec
• 1:00
  team
• 1:00
  built
• 1:01
  a
• 1:01
  purpose-specific
• 1:01
  security
• 1:02
  auditor
• 1:02
  —
• 1:03
  a
• 1:03
  QLoRA
• 1:03
  adapter
• 1:04
  on
• 1:04
  Qwen2.5-Coder-7B
• 1:07
  trained
• 1:08
  explicitly
• 1:08
  to
• 1:08
  find
• 1:08
  vulnerabilities
• 1:09
  in
• 1:09
  AI-generated
• 1:10
  code.
• 1:11
  On
• 1:11
  a
• 1:11
  111-case
• 1:12
  benchmark
• 1:12
  it
• 1:12
  achieves
• 1:12
  an
• 1:13
  F1
• 1:13
  of
• 1:13
  0.9245
• 1:14
  with
• 1:14
  a
• 1:14
  0%
• 1:15
  false-safe
• 1:16
  rate,
• 1:16
  meaning
• 1:16
  it
• 1:16
  never
• 1:17
  marks
• 1:17
  vulnerable
• 1:17
  code
• 1:18
  as
• 1:18
  clean.
• 1:19
  The
• 1:19
  hallucination
• 1:19
  rate
• 1:20
  sits
• 1:20
  at
• 1:20
  6.7%,
• 1:21
  managed
• 1:22
  by
• 1:22
  a
• 1:22
  Deterministic
• 1:23
  Safety
• 1:23
  Layer
• 1:23
  that
• 1:24
  applies
• 1:24
  eight
• 1:24
  check
• 1:24
  dimensions
• 1:25
  and
• 1:25
  six
• 1:25
  validation
• 1:26
  rules
• 1:26
  before
• 1:26
  any
• 1:27
  output
• 1:27
  reaches
• 1:27
  the
• 1:27
  user.
• 1:28
  The
• 1:28
  taxonomy
• 1:29
  of
• 1:29
  16
• 1:29
  security
• 1:30
  categories
• 1:30
  —
• 1:31
  including
• 1:31
  MCP
• 1:32
  TOOL
• 1:32
  ABUSE
• 1:32
  and
• 1:33
  PROMPT
• 1:33
  INJECTION
• 1:33
  —
• 1:34
  reflects
• 1:35
  the
• 1:35
  threat
• 1:35
  model
• 1:35
  that
• 1:35
  actually
• 1:36
  matters
• 1:36
  for
• 1:36
  AI-assisted
• 1:37
  development.
• 1:38
  The
• 1:38
  lesson:
• 1:39
  blanket
• 1:39
  keyword
• 1:39
  filtering
• 1:40
  is
• 1:40
  a
• 1:40
  blunt
• 1:40
  instrument.
• 1:41
  Domain-specific
• 1:42
  tools
• 1:42
  that
• 1:42
  understand
• 1:43
  what
• 1:43
  they're
• 1:43
  looking
• 1:43
  at
• 1:44
  tend
• 1:44
  to
• 1:44
  produce
• 1:44
  better
• 1:45
  outcomes
• 1:45
  for
• 1:45
  everyone
• 1:46
  involved.
• 1:48
  Damien
• 1:48
  Charlotin's
• 1:49
  hallucinations
• 1:49
  database
• 1:50
  now
• 1:50
  tracks
• 1:51
  over
• 1:51
  160
• 1:52
  legal
• 1:52
  cases
• 1:53
  across
• 1:53
  the
• 1:53
  United
• 1:53
  States,
• 1:54
  Canada,
• 1:54
  France,
• 1:55
  India,
• 1:56
  and
• 1:56
  Israel
• 1:56
  where
• 1:57
  courts
• 1:57
  found
• 1:57
  AI-generated
• 1:58
  citations
• 1:58
  in
• 1:59
  legal
• 1:59
  filings.
• 2:00
  The
• 2:00
  sanctions
• 2:00
  are
• 2:00
  no
• 2:00
  longer
• 2:01
  gentle.
• 2:01
  In
• 2:02
  Withers
• 2:02
  v.
• 2:03
  City
• 2:03
  of
• 2:03
  Aberdeen,
• 2:04
  a
• 2:04
  Mississippi
• 2:04
  federal
• 2:05
  court
• 2:05
  revoked
• 2:05
  pro
• 2:05
  hac
• 2:06
  vice
• 2:06
  admission,
• 2:07
  disqualified
• 2:07
  resident
• 2:08
  attorneys,
• 2:08
  imposed
• 2:08
  monetary
• 2:09
  fines,
• 2:09
  and
• 2:09
  referred
• 2:10
  counsel
• 2:10
  to
• 2:10
  state
• 2:10
  bars
• 2:11
  —
• 2:12
  all
• 2:12
  over
• 2:12
  six
• 2:12
  hallucinated
• 2:13
  case
• 2:13
  law
• 2:13
  citations.
• 2:14
  A
• 2:14
  9th
• 2:15
  Circuit
• 2:15
  case
• 2:15
  produced
• 2:15
  a
• 2:16
  six-month
• 2:16
  suspension
• 2:16
  and
• 2:16
  bar
• 2:17
  referral
• 2:17
  for
• 2:17
  seven
• 2:17
  fabricated
• 2:18
  citations.
• 2:20
  In
• 2:20
  France,
• 2:20
  the
• 2:20
  Tribunal
• 2:21
  Administratif
• 2:21
  de
• 2:21
  Grenoble
• 2:22
  identified
• 2:22
  a
• 2:22
  300-page
• 2:23
  filing
• 2:23
  as
• 2:23
  "manifestly
• 2:24
  generated
• 2:24
  with
• 2:24
  artificial
• 2:25
  intelligence,"
• 2:26
  citing
• 2:26
  nonexistent
• 2:27
  judicial
• 2:27
  decisions,
• 2:28
  and
• 2:28
  imposed
• 2:28
  a
• 2:28
  200-euro
• 2:29
  fine.
• 2:31
  The
• 2:31
  volume
• 2:31
  of
• 2:31
  cases
• 2:32
  is
• 2:32
  accelerating.
• 2:33
  The
• 2:33
  database
• 2:34
  itself
• 2:34
  has
• 2:34
  spawned
• 2:35
  PelAIkan,
• 2:35
  a
• 2:35
  detection
• 2:36
  tool
• 2:36
  purpose-built
• 2:37
  to
• 2:37
  catch
• 2:37
  hallucinated
• 2:37
  legal
• 2:38
  citations
• 2:38
  before
• 2:38
  they
• 2:38
  reach
• 2:39
  a
• 2:39
  courtroom.
• 2:39
  The
• 2:40
  accountability
• 2:40
  gap
• 2:40
  runs
• 2:41
  deeper
• 2:41
  than
• 2:41
  courtrooms.
• 2:42
  This
• 2:42
  year's
• 2:42
  AI
• 2:43
  productivity
• 2:43
  headlines
• 2:44
  are
• 2:44
  all
• 2:44
  volume
• 2:44
  claims
• 2:44
  dressed
• 2:45
  in
• 2:45
  outcome
• 2:45
  clothing.
• 2:46
  Google:
• 2:47
  75%
• 2:48
  of
• 2:48
  new
• 2:48
  code
• 2:48
  is
• 2:48
  AI-generated.
• 2:50
  Anthropic:
• 2:51
  80%
• 2:51
  of
• 2:51
  merged
• 2:52
  production
• 2:52
  code
• 2:52
  is
• 2:52
  written
• 2:53
  by
• 2:53
  Claude,
• 2:53
  engineers
• 2:54
  ship
• 2:55
  "8x
• 2:55
  more
• 2:55
  code
• 2:56
  per
• 2:56
  quarter."
• 2:57
  Cursor:
• 2:57
  100M+
• 2:58
  lines
• 2:59
  of
• 2:59
  enterprise
• 2:59
  code
• 2:59
  written
• 3:00
  per
• 3:00
  day.
• 3:01
  As
• 3:01
  one
• 3:01
  observer
• 3:01
  notes,
• 3:02
  "percent
• 3:03
  of
• 3:03
  code
• 3:03
  written
• 3:03
  by
• 3:04
  AI
• 3:04
  is
• 3:04
  just
• 3:04
  lines
• 3:05
  of
• 3:05
  code
• 3:05
  with
• 3:05
  a
• 3:05
  better
• 3:05
  publicist."
• 3:07
  These
• 3:07
  numbers
• 3:07
  cannot
• 3:08
  fail
• 3:08
  —
• 3:09
  adoption
• 3:09
  is
• 3:10
  the
• 3:10
  one
• 3:10
  metric
• 3:10
  guaranteed
• 3:11
  to
• 3:11
  rise
• 3:11
  regardless
• 3:12
  of
• 3:12
  whether
• 3:12
  anything
• 3:12
  got
• 3:12
  better.
• 3:14
  Meanwhile,
• 3:14
  METR
• 3:14
  effectively
• 3:15
  walked
• 3:15
  back
• 3:15
  its
• 3:16
  earlier
• 3:16
  finding
• 3:16
  that
• 3:17
  AI
• 3:17
  makes
• 3:17
  experienced
• 3:18
  developers
• 3:18
  19%
• 3:19
  slower,
• 3:20
  abandoning
• 3:20
  the
• 3:20
  study
• 3:21
  design
• 3:21
  entirely
• 3:21
  because
• 3:22
  developers
• 3:22
  now
• 3:22
  refuse
• 3:23
  to
• 3:23
  work
• 3:23
  without
• 3:23
  AI
• 3:24
  and
• 3:24
  cannot
• 3:24
  reliably
• 3:25
  self-report
• 3:25
  time
• 3:26
  on
• 3:26
  agentic
• 3:26
  work.
• 3:27
  An
• 3:27
  NBER
• 3:28
  survey
• 3:28
  of
• 3:28
  roughly
• 3:29
  6,000
• 3:29
  executives
• 3:30
  found
• 3:30
  nine
• 3:30
  in
• 3:30
  ten
• 3:31
  firms
• 3:31
  reporting
• 3:31
  no
• 3:32
  measurable
• 3:32
  productivity
• 3:33
  impact.
• 3:33
  The
• 3:34
  cross-study
• 3:34
  consensus
• 3:35
  hovers
• 3:35
  around
• 3:35
  10%
• 3:36
  organizational
• 3:37
  gains
• 3:37
  —
• 3:38
  useful,
• 3:38
  but
• 3:38
  not
• 3:39
  "you
• 3:39
  don't
• 3:39
  need
• 3:39
  developers
• 3:39
  anymore"
• 3:40
  territory.
• 3:41
  The
• 3:41
  irony
• 3:41
  is
• 3:41
  sharpest
• 3:42
  at
• 3:42
  Anthropic
• 3:42
  itself.
• 3:43
  Its
• 3:43
  marketing
• 3:44
  arm
• 3:44
  claims
• 3:44
  "8x
• 3:45
  more
• 3:45
  code
• 3:45
  shipped
• 3:46
  per
• 3:46
  quarter."
• 3:47
  Its
• 3:47
  research
• 3:47
  arm
• 3:48
  published
• 3:48
  an
• 3:48
  RCT
• 3:49
  finding
• 3:49
  that
• 3:49
  AI-assisted
• 3:50
  developers
• 3:50
  scored
• 3:51
  17%
• 3:52
  lower
• 3:52
  on
• 3:52
  comprehension
• 3:53
  of
• 3:53
  the
• 3:53
  code
• 3:53
  they
• 3:53
  had
• 3:53
  just
• 3:54
  shipped,
• 3:54
  with
• 3:54
  no
• 3:55
  statistically
• 3:55
  significant
• 3:56
  productivity
• 3:56
  gain.
• 3:57
  Both
• 3:57
  things
• 3:58
  are
• 3:58
  true
• 3:58
  at
• 3:58
  once
• 3:58
  —
• 3:59
  which
• 3:59
  is
• 3:59
  rather
• 3:59
  the
• 3:59
  point.
• 4:00
  Yet
• 4:00
  these
• 4:01
  volume
• 4:01
  claims
• 4:01
  move
• 4:02
  budgets.
• 4:02
  Jack
• 4:03
  Dorsey
• 4:03
  cut
• 4:03
  over
• 4:03
  40%
• 4:04
  of
• 4:04
  Block's
• 4:04
  workforce
• 4:05
  (4,000+
• 4:06
  people)
• 4:06
  with
• 4:07
  AI
• 4:07
  as
• 4:07
  the
• 4:07
  explicit
• 4:08
  thesis,
• 4:08
  in
• 4:08
  the
• 4:09
  same
• 4:09
  announcement
• 4:09
  conceding
• 4:10
  the
• 4:10
  business
• 4:10
  was
• 4:10
  strong
• 4:11
  and
• 4:11
  gross
• 4:11
  profit
• 4:11
  growing.
• 4:12
  Atlassian
• 4:13
  followed
• 4:13
  suit,
• 4:14
  conceding
• 4:14
  it
• 4:14
  would
• 4:14
  be
• 4:14
  "disingenuous
• 4:15
  to
• 4:15
  pretend
• 4:16
  AI
• 4:16
  doesn't
• 4:16
  change
• 4:16
  the
• 4:16
  mix
• 4:17
  of
• 4:17
  skills
• 4:17
  we
• 4:17
  need."
• 4:18
  When
• 4:18
  a
• 4:18
  company
• 4:18
  says
• 4:19
  "AI
• 4:19
  made
• 4:19
  everyone
• 4:20
  more
• 4:20
  productive,
• 4:20
  so
• 4:21
  we
• 4:21
  need
• 4:21
  fewer
• 4:21
  people,"
• 4:22
  the
• 4:22
  evidence
• 4:22
  for
• 4:22
  that
• 4:23
  claim
• 4:23
  matters
• 4:23
  —
• 4:23
  and
• 4:23
  it
• 4:23
  does
• 4:24
  not
• 4:24
  yet
• 4:24
  exist
• 4:24
  at
• 4:24
  the
• 4:25
  scale
• 4:25
  implied.
• 4:26
  The
• 4:26
  question
• 4:26
  to
• 4:26
  carry
• 4:27
  into
• 4:27
  every
• 4:27
  vendor
• 4:27
  pitch
• 4:28
  and
• 4:28
  exec
• 4:28
  review:
• 4:29
  is
• 4:29
  that
• 4:29
  an
• 4:29
  outcome,
• 4:30
  or
• 4:30
  a
• 4:30
  volume?
• 4:32
  Two
• 4:32
  papers
• 4:33
  this
• 4:33
  week
• 4:33
  isolate
• 4:34
  specific,
• 4:34
  measurable
• 4:35
  limitations
• 4:35
  in
• 4:35
  transformer
• 4:36
  architectures
• 4:37
  that
• 4:37
  are
• 4:37
  not
• 4:37
  going
• 4:37
  away
• 4:37
  with
• 4:37
  scale.
• 4:38
  The
• 4:39
  first,
• 4:39
  published
• 4:39
  in
• 4:40
  PNAS
• 4:40
  Nexus,
• 4:41
  uses
• 4:41
  the
• 4:41
  classic
• 4:42
  color
• 4:42
  Stroop
• 4:42
  task
• 4:43
  to
• 4:43
  demonstrate
• 4:43
  that
• 4:43
  transformers
• 4:44
  lack
• 4:44
  executive
• 4:45
  control
• 4:45
  of
• 4:45
  attention
• 4:45
  —
• 4:46
  the
• 4:46
  capacity
• 4:47
  to
• 4:47
  detect
• 4:47
  conflict,
• 4:48
  maintain
• 4:48
  task
• 4:48
  goals,
• 4:49
  and
• 4:49
  override
• 4:50
  prepotent
• 4:50
  responses.
• 4:51
  The
• 4:51
  researchers
• 4:52
  administered
• 4:52
  Stroop
• 4:53
  tasks
• 4:53
  to
• 4:53
  GPT-4o
• 4:54
  and
• 4:55
  Claude
• 4:55
  3.5
• 4:56
  Sonnet
• 4:56
  across
• 4:57
  word
• 4:57
  lists
• 4:57
  of
• 4:57
  varying
• 4:57
  length.
• 4:58
  Both
• 4:58
  models
• 4:59
  showed
• 4:59
  human-like
• 5:00
  interference
• 5:00
  at
• 5:00
  short
• 5:00
  lists,
• 5:01
  but
• 5:01
  performance
• 5:02
  collapsed
• 5:02
  catastrophically
• 5:03
  as
• 5:03
  lists
• 5:03
  grew.
• 5:04
  GPT-4o's
• 5:05
  incongruent
• 5:06
  accuracy
• 5:06
  fell
• 5:06
  from
• 5:06
  91%
• 5:07
  at
• 5:07
  5
• 5:07
  words
• 5:08
  to
• 5:08
  15%
• 5:08
  at
• 5:09
  40
• 5:09
  words.
• 5:10
  Claude
• 5:10
  held
• 5:10
  up
• 5:10
  longer
• 5:11
  (76%
• 5:11
  at
• 5:12
  20
• 5:12
  words)
• 5:12
  but
• 5:12
  still
• 5:13
  crashed
• 5:13
  to
• 5:13
  24%
• 5:14
  at
• 5:14
  40.
• 5:15
  Humans
• 5:15
  maintain
• 5:16
  95-97%
• 5:18
  accuracy
• 5:19
  even
• 5:19
  in
• 5:19
  hour-long,
• 5:20
  1,500-word
• 5:20
  sessions.
• 5:21
  The
• 5:22
  key
• 5:22
  finding:
• 5:22
  this
• 5:23
  persists
• 5:23
  across
• 5:23
  GPT-5,
• 5:24
  Claude
• 5:24
  Opus
• 5:25
  4.1,
• 5:26
  and
• 5:26
  Gemini
• 5:26
  2.5
• 5:27
  Pro,
• 5:27
  confirming
• 5:28
  the
• 5:28
  limitation
• 5:28
  is
• 5:28
  architectural,
• 5:29
  not
• 5:29
  a
• 5:29
  matter
• 5:30
  of
• 5:30
  scale.
• 5:31
  The
• 5:31
  softmax
• 5:31
  mechanism
• 5:32
  routes
• 5:32
  information
• 5:32
  competitively,
• 5:33
  but
• 5:33
  there
• 5:33
  is
• 5:34
  no
• 5:34
  mechanism
• 5:34
  that
• 5:34
  evaluates
• 5:35
  whether
• 5:35
  current
• 5:35
  processing
• 5:36
  remains
• 5:36
  aligned
• 5:36
  with
• 5:36
  the
• 5:36
  task
• 5:37
  goal.
• 5:38
  The
• 5:38
  researchers
• 5:38
  also
• 5:39
  tested
• 5:39
  for
• 5:39
  congruency
• 5:40
  sequence
• 5:40
  effects
• 5:40
  —
• 5:41
  whether
• 5:41
  conflict
• 5:42
  on
• 5:42
  a
• 5:42
  previous
• 5:42
  trial
• 5:42
  improves
• 5:43
  performance
• 5:43
  on
• 5:43
  the
• 5:43
  next
• 5:44
  —
• 5:44
  and
• 5:44
  found
• 5:45
  no
• 5:45
  consistent
• 5:45
  evidence
• 5:46
  of
• 5:46
  trial-to-trial
• 5:46
  adaptation
• 5:47
  in
• 5:47
  either
• 5:47
  model.
• 5:48
  Humans
• 5:49
  show
• 5:49
  this
• 5:49
  effect
• 5:49
  reliably
• 5:50
  via
• 5:50
  the
• 5:50
  ACC-DLPFC
• 5:52
  pathway.
• 5:53
  Transformers
• 5:53
  lack
• 5:53
  the
• 5:53
  circuit
• 5:54
  entirely.
• 5:54
  The
• 5:55
  second
• 5:55
  paper,
• 5:55
  from
• 5:55
  NUS
• 5:56
  and
• 5:56
  Google
• 5:56
  Research,
• 5:57
  uses
• 5:57
  shortest-path
• 5:58
  planning
• 5:58
  on
• 5:58
  grid
• 5:58
  maps
• 5:59
  to
• 5:59
  measure
• 5:59
  two
• 6:00
  distinct
• 6:00
  axes
• 6:00
  of
• 6:01
  generalization.
• 6:02
  Models
• 6:02
  achieve
• 6:03
  strong
• 6:03
  spatial
• 6:03
  transfer
• 6:04
  —
• 6:04
  over
• 6:04
  90%
• 6:05
  success
• 6:05
  on
• 6:06
  unseen
• 6:06
  maps
• 6:06
  —
• 6:07
  demonstrating
• 6:07
  genuine
• 6:08
  algorithmic
• 6:08
  competence.
• 6:09
  But
• 6:09
  they
• 6:10
  consistently
• 6:10
  fail
• 6:10
  at
• 6:11
  length
• 6:11
  scaling.
• 6:12
  The
• 6:12
  authors
• 6:12
  trace
• 6:12
  this
• 6:12
  to
• 6:12
  "recursive
• 6:13
  instability":
• 6:14
  even
• 6:14
  when
• 6:14
  a
• 6:14
  model
• 6:14
  can
• 6:15
  solve
• 6:15
  each
• 6:15
  subpath
• 6:16
  individually,
• 6:17
  composing
• 6:17
  them
• 6:17
  into
• 6:17
  a
• 6:17
  correct
• 6:18
  end-to-end
• 6:18
  solution
• 6:19
  fails.
• 6:20
  This
• 6:20
  is
• 6:20
  not
• 6:21
  hardness
• 6:21
  accumulation;
• 6:22
  it
• 6:22
  is
• 6:22
  compositional
• 6:23
  instability
• 6:23
  dominating
• 6:24
  the
• 6:24
  failure
• 6:24
  mode.
• 6:25
  Reinforcement
• 6:26
  learning
• 6:26
  via
• 6:27
  GRPO
• 6:27
  stabilizes
• 6:28
  training
• 6:28
  but
• 6:28
  never
• 6:28
  exceeds
• 6:29
  the
• 6:29
  best
• 6:29
  supervised
• 6:30
  fine-tuning
• 6:30
  ceiling
• 6:30
  —
• 6:31
  it
• 6:31
  is
• 6:31
  a
• 6:31
  "robust
• 6:32
  fallback,"
• 6:33
  not
• 6:33
  a
• 6:33
  capability
• 6:33
  expander.
• 6:34
  The
• 6:35
  paper
• 6:35
  also
• 6:35
  offers
• 6:36
  actionable
• 6:36
  training
• 6:36
  data
• 6:37
  insights:
• 6:38
  the
• 6:38
  number
• 6:38
  of
• 6:38
  distinct
• 6:38
  questions
• 6:39
  matters
• 6:39
  more
• 6:40
  than
• 6:40
  the
• 6:40
  number
• 6:40
  of
• 6:40
  solutions
• 6:41
  per
• 6:41
  question,
• 6:41
  and
• 6:41
  coverage
• 6:42
  —
• 6:42
  the
• 6:42
  fraction
• 6:43
  of
• 6:43
  unique
• 6:43
  primitives
• 6:44
  seen
• 6:44
  during
• 6:44
  training
• 6:44
  —
• 6:45
  sets
• 6:45
  a
• 6:45
  hard
• 6:46
  ceiling
• 6:46
  on
• 6:46
  performance.
• 6:47
  These
• 6:47
  findings
• 6:48
  validate
• 6:48
  on
• 6:48
  MathQA
• 6:49
  with
• 6:49
  Qwen2.5-7B,
• 6:53
  suggesting
• 6:53
  a
• 6:53
  general
• 6:54
  property
• 6:54
  of
• 6:54
  LLM
• 6:55
  learning
• 6:55
  rather
• 6:55
  than
• 6:56
  a
• 6:56
  synthetic-domain
• 6:56
  artifact.
• 6:57
  The
• 6:57
  one
• 6:58
  intervention
• 6:58
  that
• 6:58
  actually
• 6:59
  rescues
• 6:59
  length
• 6:59
  scaling
• 7:00
  is
• 7:00
  prosaic:
• 7:01
  add
• 7:01
  a
• 7:01
  few
• 7:01
  training
• 7:02
  examples
• 7:02
  at
• 7:03
  or
• 7:03
  slightly
• 7:03
  above
• 7:03
  the
• 7:03
  target
• 7:04
  length.
• 7:05
  Length
• 7:05
  generalization
• 7:06
  must
• 7:06
  be
• 7:06
  explicitly
• 7:07
  scaffolded
• 7:07
  in
• 7:07
  the
• 7:07
  training
• 7:08
  distribution,
• 7:09
  not
• 7:09
  expected
• 7:09
  to
• 7:09
  emerge.
• 7:12
  The
• 7:12
  Claude
• 7:12
  Code
• 7:12
  versus
• 7:13
  Codex
• 7:13
  debate
• 7:13
  is
• 7:14
  producing
• 7:14
  more
• 7:14
  heat
• 7:14
  than
• 7:15
  light,
• 7:15
  but
• 7:15
  one
• 7:16
  framing
• 7:16
  cuts
• 7:16
  through.
• 7:17
  Claude
• 7:17
  Code
• 7:18
  feels
• 7:18
  like
• 7:18
  a
• 7:18
  cockpit
• 7:18
  —
• 7:19
  you
• 7:19
  are
• 7:19
  close
• 7:19
  to
• 7:19
  the
• 7:20
  model,
• 7:20
  steering
• 7:20
  it,
• 7:21
  stopping
• 7:21
  it,
• 7:21
  correcting
• 7:22
  its
• 7:22
  plan,
• 7:22
  wrestling
• 7:23
  with
• 7:23
  ambiguity
• 7:23
  together.
• 7:24
  Codex
• 7:24
  feels
• 7:24
  like
• 7:24
  an
• 7:25
  operations
• 7:25
  desk
• 7:25
  —
• 7:26
  you
• 7:26
  dispatch
• 7:26
  jobs
• 7:26
  in
• 7:27
  parallel,
• 7:27
  each
• 7:27
  sandboxed,
• 7:28
  each
• 7:28
  returning
• 7:29
  inspectable
• 7:29
  artifacts
• 7:30
  with
• 7:30
  proof
• 7:30
  of
• 7:30
  completion.
• 7:31
  As
• 7:31
  one
• 7:31
  observer
• 7:32
  puts
• 7:32
  it:
• 7:32
  "The
• 7:32
  skill
• 7:33
  of
• 7:33
  2026
• 7:33
  is
• 7:34
  agent
• 7:34
  literacy."
• 7:35
  Not
• 7:35
  prompting.
• 7:36
  Not
• 7:36
  which
• 7:36
  model
• 7:37
  wins
• 7:37
  a
• 7:37
  benchmark
• 7:37
  this
• 7:38
  month.
• 7:39
  The
• 7:39
  skill
• 7:39
  is
• 7:39
  writing
• 7:39
  assignments
• 7:40
  that
• 7:40
  come
• 7:40
  back
• 7:40
  as
• 7:40
  inspectable
• 7:41
  work,
• 7:41
  and
• 7:42
  knowing
• 7:42
  when
• 7:42
  to
• 7:42
  steer
• 7:43
  versus
• 7:43
  when
• 7:43
  to
• 7:43
  dispatch.
• 7:45
  Claude's
• 7:45
  failure
• 7:45
  mode:
• 7:46
  it
• 7:46
  seduces
• 7:47
  you
• 7:47
  with
• 7:47
  great
• 7:47
  conversation
• 7:48
  and
• 7:48
  makes
• 7:48
  you
• 7:48
  feel
• 7:48
  closer
• 7:49
  to
• 7:49
  the
• 7:49
  work
• 7:49
  than
• 7:49
  you
• 7:49
  are.
• 7:50
  Codex's
• 7:51
  failure
• 7:51
  mode:
• 7:52
  a
• 7:52
  completed
• 7:52
  run
• 7:52
  that
• 7:53
  makes
• 7:53
  the
• 7:53
  work
• 7:53
  feel
• 7:53
  more
• 7:53
  done
• 7:53
  than
• 7:54
  it
• 7:54
  really
• 7:54
  is.
• 7:55
  Both
• 7:55
  still
• 7:55
  require
• 7:55
  judgment.
• 7:56
  Both
• 7:57
  still
• 7:57
  require
• 7:57
  proof.
• 7:58
  Cole
• 7:58
  Medin's
• 7:59
  "harness
• 7:59
  engineering"
• 7:59
  framework
• 8:00
  provides
• 8:00
  the
• 8:00
  structural
• 8:01
  vocabulary.
• 8:02
  An
• 8:02
  AI
• 8:02
  harness
• 8:02
  has
• 8:03
  two
• 8:03
  layers:
• 8:03
  a
• 8:03
  within-session
• 8:04
  layer
• 8:04
  (rules,
• 8:05
  skills,
• 8:05
  MCP
• 8:06
  servers,
• 8:06
  LSP
• 8:06
  integration,
• 8:07
  hooks,
• 8:07
  sub-agents)
• 8:08
  and
• 8:08
  a
• 8:08
  multi-session
• 8:09
  orchestration
• 8:09
  layer
• 8:10
  (persistent
• 8:10
  loops,
• 8:10
  durable
• 8:11
  workflows).
• 8:11
  The
• 8:12
  key
• 8:12
  insight:
• 8:13
  "every
• 8:13
  mistake
• 8:13
  becomes
• 8:13
  a
• 8:14
  rule."
• 8:14
  A
• 8:14
  CLAUDE.md
• 8:15
  file
• 8:15
  is
• 8:16
  not
• 8:16
  a
• 8:16
  config
• 8:16
  dump
• 8:16
  —
• 8:17
  it
• 8:17
  is
• 8:17
  a
• 8:17
  standing
• 8:18
  body
• 8:18
  of
• 8:18
  institutional
• 8:19
  knowledge
• 8:19
  that
• 8:19
  evolves
• 8:19
  with
• 8:19
  the
• 8:19
  project.
• 8:21
  This
• 8:21
  is
• 8:21
  infrastructure
• 8:21
  thinking
• 8:22
  applied
• 8:22
  to
• 8:22
  agent
• 8:22
  management,
• 8:23
  and
• 8:23
  it
• 8:24
  separates
• 8:24
  practitioners
• 8:25
  who
• 8:25
  use
• 8:25
  AI
• 8:25
  from
• 8:26
  those
• 8:26
  who
• 8:26
  operate
• 8:26
  AI
• 8:27
  systems.
• 8:28
  The
• 8:28
  multi-provider
• 8:29
  workflow
• 8:29
  is
• 8:29
  maturing.
• 8:30
  Medin's
• 8:31
  Archon
• 8:31
  harness
• 8:31
  demonstrates
• 8:32
  a
• 8:32
  practical
• 8:32
  approach:
• 8:34
  an
• 8:34
  eight-node
• 8:34
  YAML
• 8:34
  pipeline
• 8:35
  where
• 8:35
  Opus
• 8:35
  handles
• 8:36
  planning
• 8:36
  (the
• 8:36
  only
• 8:36
  step
• 8:36
  that
• 8:37
  truly
• 8:37
  requires
• 8:37
  the
• 8:37
  most
• 8:38
  expensive
• 8:38
  model),
• 8:39
  Gemini
• 8:40
  3.5
• 8:40
  Flash
• 8:40
  handles
• 8:41
  UI
• 8:41
  design
• 8:41
  (creative
• 8:42
  but
• 8:42
  unreliable
• 8:43
  at
• 8:43
  following
• 8:44
  complex
• 8:44
  instructions),
• 8:45
  and
• 8:45
  Sonnet
• 8:46
  handles
• 8:46
  cheaper
• 8:46
  validation
• 8:47
  steps.
• 8:48
  Each
• 8:48
  node
• 8:48
  communicates
• 8:49
  via
• 8:49
  markdown
• 8:50
  artifact
• 8:50
  files
• 8:50
  in
• 8:50
  a
• 8:50
  shared
• 8:51
  workspace
• 8:51
  directory.
• 8:52
  The
• 8:52
  demo
• 8:52
  —
• 8:53
  a
• 8:53
  benchmarking
• 8:54
  dashboard
• 8:54
  built
• 8:54
  from
• 8:55
  spec
• 8:55
  to
• 8:55
  near-deployment
• 8:55
  —
• 8:56
  ran
• 8:57
  without
• 8:57
  manual
• 8:57
  intervention,
• 8:58
  though
• 8:58
  Gemini
• 8:59
  skipped
• 8:59
  a
• 8:59
  critical
• 9:00
  artifact-writing
• 9:00
  instruction
• 9:01
  while
• 9:01
  producing
• 9:01
  a
• 9:01
  clean,
• 9:02
  non-generic
• 9:03
  UI
• 9:03
  in
• 9:03
  a
• 9:03
  single
• 9:03
  pass.
• 9:04
  The
• 9:04
  integration
• 9:05
  node
• 9:05
  (Opus)
• 9:05
  ran
• 9:06
  for
• 9:06
  31
• 9:06
  minutes
• 9:07
  handling
• 9:07
  Clerk
• 9:07
  authentication
• 9:08
  setup
• 9:08
  but
• 9:09
  completed
• 9:09
  with
• 9:09
  full
• 9:09
  auth
• 9:10
  working
• 9:10
  on
• 9:10
  first
• 9:10
  attempt.
• 9:11
  The
• 9:12
  practical
• 9:12
  implication:
• 9:13
  use
• 9:13
  the
• 9:13
  cheapest
• 9:14
  model
• 9:14
  that
• 9:14
  meets
• 9:15
  each
• 9:15
  step's
• 9:15
  reasoning
• 9:16
  requirement,
• 9:16
  and
• 9:17
  expect
• 9:17
  the
• 9:17
  creative
• 9:17
  models
• 9:18
  to
• 9:18
  occasionally
• 9:19
  drop
• 9:19
  instructions.
• 9:20
  The
• 9:20
  distinction
• 9:21
  between
• 9:21
  CLI
• 9:22
  and
• 9:22
  Skill
• 9:22
  is
• 9:22
  sharpening:
• 9:23
  a
• 9:23
  CLI
• 9:24
  is
• 9:24
  a
• 9:24
  capability
• 9:24
  (what
• 9:24
  you
• 9:25
  can
• 9:25
  do),
• 9:25
  a
• 9:26
  Skill
• 9:26
  is
• 9:26
  an
• 9:26
  instruction
• 9:27
  manual
• 9:27
  (how
• 9:27
  to
• 9:27
  do
• 9:27
  it).
• 9:28
  Google's
• 9:29
  Agents
• 9:29
  CLI
• 9:30
  is
• 9:30
  highlighted
• 9:30
  as
• 9:30
  an
• 9:30
  example
• 9:31
  of
• 9:31
  this
• 9:31
  convergence.
• 9:32
  Meanwhile
• 9:33
  Codex
• 9:33
  now
• 9:33
  runs
• 9:34
  from
• 9:34
  a
• 9:34
  phone
• 9:34
  with
• 9:34
  computer-use
• 9:35
  plugin
• 9:35
  for
• 9:35
  UI
• 9:36
  testing
• 9:36
  ,
• 9:37
  Claude
• 9:37
  Code
• 9:37
  integrates
• 9:37
  with
• 9:38
  Obsidian
• 9:38
  as
• 9:38
  a
• 9:38
  command
• 9:39
  center
• 9:39
  with
• 9:39
  custom
• 9:39
  metrics
• 9:40
  dashboards
• 9:40
  ,
• 9:41
  and
• 9:41
  the
• 9:41
  web
• 9:41
  design
• 9:42
  CLI
• 9:42
  ecosystem
• 9:43
  around
• 9:43
  Claude
• 9:43
  Code
• 9:43
  continues
• 9:44
  to
• 9:44
  expand
• 9:45
  with
• 9:45
  Impeccable,
• 9:45
  Playwright,
• 9:46
  Supabase,
• 9:47
  GitHub,
• 9:47
  and
• 9:47
  Vercel
• 9:48
  CLIs
• 9:48
  rounding
• 9:48
  out
• 9:49
  the
• 9:49
  toolkit
• 9:49
  .
• 9:51
  A
• 9:51
  deep
• 9:51
  technical
• 9:52
  breakdown
• 9:52
  of
• 9:52
  Linear
• 9:53
  reveals
• 9:53
  that
• 9:54
  its
• 9:54
  perceived
• 9:54
  speed
• 9:54
  is
• 9:55
  not
• 9:55
  the
• 9:55
  result
• 9:55
  of
• 9:55
  optimization
• 9:56
  passes
• 9:56
  —
• 9:57
  it
• 9:57
  is
• 9:57
  an
• 9:57
  architectural
• 9:58
  commitment
• 9:59
  made
• 9:59
  from
• 9:59
  day
• 9:59
  one.
• 10:00
  Linear
• 10:00
  treats
• 10:01
  the
• 10:01
  browser
• 10:01
  as
• 10:01
  the
• 10:02
  database.
• 10:04
  All
• 10:04
  project
• 10:04
  data
• 10:05
  lives
• 10:05
  in
• 10:05
  IndexedDB
• 10:06
  on
• 10:06
  the
• 10:06
  user's
• 10:06
  machine,
• 10:07
  with
• 10:08
  an
• 10:08
  in-memory
• 10:08
  pool
• 10:08
  of
• 10:09
  MobX
• 10:09
  observables
• 10:10
  as
• 10:10
  the
• 10:10
  primary
• 10:10
  data
• 10:10
  source
• 10:11
  for
• 10:11
  the
• 10:11
  UI.
• 10:13
  Mutations
• 10:13
  apply
• 10:13
  locally
• 10:14
  first,
• 10:14
  then
• 10:15
  sync
• 10:15
  to
• 10:15
  the
• 10:15
  server
• 10:15
  over
• 10:16
  WebSocket.
• 10:17
  There
• 10:17
  are
• 10:17
  no
• 10:17
  loading
• 10:18
  spinners
• 10:18
  because
• 10:18
  the
• 10:18
  data
• 10:19
  is
• 10:19
  already
• 10:19
  local.
• 10:20
  Reactivity
• 10:21
  is
• 10:21
  granular
• 10:21
  to
• 10:21
  the
• 10:21
  individual
• 10:22
  property
• 10:22
  level:
• 10:23
  changing
• 10:23
  one
• 10:24
  field
• 10:24
  on
• 10:24
  one
• 10:24
  issue
• 10:25
  re-renders
• 10:25
  one
• 10:25
  cell,
• 10:26
  not
• 10:26
  the
• 10:26
  list.
• 10:27
  The
• 10:27
  team
• 10:28
  has
• 10:28
  rewritten
• 10:28
  their
• 10:28
  bundler
• 10:29
  setup
• 10:29
  four
• 10:29
  times
• 10:29
  —
• 10:30
  Parcel
• 10:31
  to
• 10:31
  Rollup
• 10:31
  to
• 10:31
  Vite
• 10:31
  to
• 10:31
  Rolldown
• 10:32
  —
• 10:32
  and
• 10:33
  precaches
• 10:33
  roughly
• 10:34
  1,200
• 10:35
  assets
• 10:35
  via
• 10:35
  service
• 10:36
  worker.
• 10:36
  The
• 10:36
  total
• 10:37
  JavaScript
• 10:37
  payload
• 10:38
  is
• 10:38
  21MB,
• 10:39
  but
• 10:39
  aggressively
• 10:40
  code-split
• 10:40
  into
• 10:40
  hundreds
• 10:41
  of
• 10:41
  chunks
• 10:41
  fetched
• 10:41
  in
• 10:42
  parallel
• 10:42
  via
• 10:42
  modulepreload.
• 10:43
  Animations
• 10:44
  use
• 10:44
  only
• 10:45
  GPU-composited
• 10:45
  properties
• 10:46
  (transform,
• 10:47
  opacity)
• 10:47
  at
• 10:47
  80-120ms
• 10:49
  durations.
• 10:50
  The
• 10:50
  overarching
• 10:50
  point:
• 10:51
  this
• 10:52
  cannot
• 10:52
  be
• 10:52
  replicated
• 10:53
  by
• 10:53
  applying
• 10:53
  individual
• 10:54
  optimizations
• 10:55
  to
• 10:55
  a
• 10:55
  conventionally
• 10:56
  architected
• 10:56
  application.
• 10:57
  Speed
• 10:58
  at
• 10:58
  this
• 10:58
  level
• 10:58
  is
• 10:58
  a
• 10:58
  product
• 10:59
  decision.
• 10:59
  The
• 11:00
  React
• 11:00
  Compiler
• 11:00
  is
• 11:00
  being
• 11:01
  ported
• 11:01
  to
• 11:01
  Rust,
• 11:02
  and
• 11:02
  the
• 11:02
  numbers
• 11:02
  justify
• 11:03
  the
• 11:03
  effort:
• 11:04
  3x
• 11:04
  faster
• 11:04
  as
• 11:05
  a
• 11:05
  Babel
• 11:05
  plugin,
• 11:06
  approximately
• 11:06
  10x
• 11:07
  faster
• 11:07
  as
• 11:07
  a
• 11:07
  direct
• 11:08
  transformation.
• 11:09
  The
• 11:09
  architecture
• 11:10
  was
• 11:10
  guided
• 11:10
  by
• 11:10
  humans
• 11:11
  but
• 11:11
  majority-coded
• 11:12
  by
• 11:12
  AI,
• 11:13
  porting
• 11:13
  the
• 11:13
  SSA/CFG/HIR
• 11:16
  pipeline
• 11:17
  pass-by-pass
• 11:17
  with
• 11:18
  1,725
• 11:19
  fixtures
• 11:20
  passing.
• 11:20
  OXC
• 11:21
  and
• 11:21
  SWC
• 11:22
  integrations
• 11:22
  are
• 11:22
  in
• 11:22
  progress.
• 11:23
  This
• 11:24
  is
• 11:24
  an
• 11:24
  experimental
• 11:24
  WIP
• 11:25
  pull
• 11:25
  request,
• 11:25
  but
• 11:26
  it
• 11:26
  demonstrates
• 11:26
  the
• 11:26
  pattern:
• 11:27
  rewrites
• 11:28
  that
• 11:28
  deliver
• 11:28
  order-of-magnitude
• 11:29
  speedups
• 11:29
  and
• 11:29
  are
• 11:29
  economically
• 11:30
  viable
• 11:31
  because
• 11:31
  AI
• 11:31
  dramatically
• 11:32
  reduces
• 11:32
  the
• 11:32
  labor
• 11:32
  cost
• 11:33
  of
• 11:33
  porting.
• 11:34
  Extend
• 11:34
  UI
• 11:34
  ships
• 11:35
  an
• 11:35
  open-source
• 11:35
  React
• 11:36
  component
• 11:36
  kit
• 11:36
  for
• 11:36
  rendering
• 11:37
  PDFs,
• 11:37
  DOCX,
• 11:38
  XLSX,
• 11:39
  and
• 11:39
  CSV
• 11:40
  files
• 11:40
  with
• 11:40
  bounding-box
• 11:41
  citation
• 11:41
  overlays
• 11:42
  —
• 11:43
  the
• 11:43
  kind
• 11:43
  of
• 11:43
  document
• 11:43
  viewer
• 11:44
  that
• 11:44
  every
• 11:44
  RAG
• 11:44
  application
• 11:45
  eventually
• 11:45
  needs
• 11:45
  but
• 11:46
  nobody
• 11:46
  wants
• 11:46
  to
• 11:46
  build
• 11:46
  from
• 11:47
  scratch.
• 11:48
  It
• 11:48
  solves
• 11:48
  a
• 11:48
  specific
• 11:49
  pain
• 11:49
  point:
• 11:50
  showing
• 11:50
  users
• 11:50
  exactly
• 11:51
  where
• 11:51
  in
• 11:51
  a
• 11:51
  source
• 11:52
  document
• 11:52
  an
• 11:52
  AI-generated
• 11:53
  claim
• 11:54
  originated,
• 11:55
  with
• 11:55
  spatial
• 11:55
  precision
• 11:56
  rather
• 11:56
  than
• 11:56
  page-level
• 11:57
  attribution
• 11:57
  .
• 11:59
  Tessera
• 12:00
  is
• 12:00
  a
• 12:00
  from-scratch
• 12:01
  LLM
• 12:01
  stack
• 12:01
  built
• 12:01
  around
• 12:02
  a
• 12:02
  single
• 12:02
  goal:
• 12:03
  distill
• 12:03
  a
• 12:03
  large
• 12:03
  teacher
• 12:04
  into
• 12:04
  a
• 12:04
  small
• 12:04
  student,
• 12:04
  then
• 12:05
  serve
• 12:05
  that
• 12:05
  student
• 12:05
  efficiently.
• 12:06
  The
• 12:07
  scope
• 12:07
  is
• 12:07
  deliberately
• 12:07
  comprehensive
• 12:08
  —
• 12:09
  custom
• 12:09
  Triton
• 12:09
  and
• 12:10
  CUDA
• 12:10
  kernels
• 12:10
  (FlashAttention
• 12:11
  forward,
• 12:12
  fused
• 12:12
  RMSNorm,
• 12:13
  fused
• 12:13
  SwiGLU,
• 12:14
  int8
• 12:14
  dequantizing
• 12:15
  matmul),
• 12:16
  FSDP/ZeRO-3
• 12:18
  training
• 12:18
  with
• 12:18
  atomic
• 12:19
  sharded
• 12:19
  checkpoints,
• 12:20
  a
• 12:20
  serving
• 12:20
  engine
• 12:21
  with
• 12:21
  block-paged
• 12:21
  KV
• 12:22
  cache
• 12:22
  and
• 12:22
  speculative
• 12:23
  decoding,
• 12:23
  post-training
• 12:24
  quantization
• 12:24
  (int8,
• 12:25
  AWQ,
• 12:26
  FP8),
• 12:26
  and
• 12:27
  a
• 12:27
  Rust/axum
• 12:28
  gateway
• 12:28
  calling
• 12:28
  into
• 12:28
  Python
• 12:29
  via
• 12:29
  PyO3.
• 12:30
  A
• 12:30
  JAX
• 12:30
  reimplementation
• 12:31
  of
• 12:31
  the
• 12:31
  forward
• 12:32
  pass
• 12:32
  serves
• 12:32
  as
• 12:32
  an
• 12:33
  independent
• 12:33
  parity
• 12:33
  check.
• 12:34
  Everything
• 12:35
  runs
• 12:35
  and
• 12:35
  is
• 12:35
  tested
• 12:36
  on
• 12:36
  a
• 12:36
  laptop;
• 12:37
  the
• 12:37
  GPU
• 12:37
  kernels
• 12:38
  fall
• 12:38
  back
• 12:38
  to
• 12:38
  torch
• 12:39
  references
• 12:39
  on
• 12:39
  non-NVIDIA
• 12:40
  hardware.
• 12:41
  The
• 12:41
  project
• 12:41
  is
• 12:41
  small
• 12:42
  enough
• 12:42
  to
• 12:42
  test
• 12:42
  thoroughly
• 12:42
  —
• 12:43
  each
• 12:43
  Triton
• 12:44
  kernel
• 12:44
  matches
• 12:44
  its
• 12:44
  torch
• 12:45
  reference
• 12:45
  within
• 12:45
  floating-point
• 12:46
  tolerance,
• 12:47
  sharded
• 12:47
  Adam
• 12:47
  matches
• 12:48
  single-process
• 12:48
  Adam
• 12:49
  step-for-step.
• 12:50
  Tencent's
• 12:50
  UniRL
• 12:51
  provides
• 12:51
  a
• 12:51
  unified
• 12:52
  reinforcement
• 12:52
  learning
• 12:53
  framework
• 12:53
  for
• 12:53
  multimodal
• 12:54
  models,
• 12:54
  supporting
• 12:55
  Stable
• 12:55
  Diffusion
• 12:56
  3,
• 12:56
  FLUX,
• 12:57
  WAN,
• 12:58
  HunyuanVideo,
• 12:58
  Qwen-VL,
• 12:59
  and
• 12:59
  Qwen3.
• 13:00
  It
• 13:00
  introduces
• 13:01
  DRPO
• 13:02
  and
• 13:02
  Flow-DPPO
• 13:02
  algorithms
• 13:03
  with
• 13:03
  Ray
• 13:03
  DevicePool
• 13:04
  orchestration,
• 13:05
  FSDP,
• 13:06
  and
• 13:06
  LoRA
• 13:07
  synchronization
• 13:07
  —
• 13:08
  a
• 13:08
  single
• 13:08
  training
• 13:08
  harness
• 13:09
  that
• 13:09
  works
• 13:09
  across
• 13:09
  image
• 13:10
  generation,
• 13:10
  video
• 13:10
  generation,
• 13:11
  and
• 13:11
  language
• 13:12
  models.
• 13:13
  SafeAdapt,
• 13:13
  from
• 13:13
  Imperial
• 13:14
  College
• 13:14
  London,
• 13:15
  introduces
• 13:15
  the
• 13:15
  Rashomon
• 13:16
  set
• 13:16
  —
• 13:16
  a
• 13:17
  certified
• 13:17
  region
• 13:17
  in
• 13:17
  policy
• 13:18
  parameter
• 13:18
  space
• 13:19
  where
• 13:19
  every
• 13:19
  parameterization
• 13:20
  is
• 13:20
  guaranteed
• 13:20
  to
• 13:21
  satisfy
• 13:21
  safety
• 13:21
  constraints.
• 13:22
  Using
• 13:23
  Interval
• 13:23
  Bound
• 13:23
  Propagation
• 13:24
  from
• 13:24
  the
• 13:24
  neural
• 13:24
  network
• 13:25
  verification
• 13:25
  literature,
• 13:26
  it
• 13:26
  computes
• 13:26
  tight
• 13:27
  bounds
• 13:27
  on
• 13:27
  how
• 13:28
  network
• 13:28
  outputs
• 13:28
  change
• 13:29
  as
• 13:29
  parameters
• 13:29
  vary,
• 13:30
  then
• 13:30
  constrains
• 13:31
  downstream
• 13:31
  RL
• 13:32
  adaptation
• 13:32
  via
• 13:33
  projected
• 13:33
  gradient
• 13:33
  descent
• 13:34
  to
• 13:34
  stay
• 13:34
  within
• 13:34
  the
• 13:34
  certified
• 13:35
  region.
• 13:36
  It
• 13:36
  is
• 13:36
  the
• 13:36
  only
• 13:36
  method
• 13:37
  in
• 13:37
  their
• 13:37
  experiments
• 13:37
  that
• 13:37
  maintains
• 13:38
  a
• 13:38
  critical
• 13:38
  state
• 13:39
  safety
• 13:39
  rate
• 13:39
  of
• 13:39
  1.0
• 13:40
  after
• 13:40
  adaptation;
• 13:42
  EWC
• 13:43
  and
• 13:43
  unconstrained
• 13:44
  PPO
• 13:44
  both
• 13:45
  exhibit
• 13:45
  catastrophic
• 13:46
  forgetting
• 13:46
  of
• 13:46
  safety.
• 13:47
  The
• 13:47
  method
• 13:48
  operates
• 13:48
  in
• 13:48
  three
• 13:48
  phases:
• 13:49
  constructing
• 13:50
  a
• 13:50
  safety
• 13:50
  dataset
• 13:51
  from
• 13:51
  an
• 13:51
  unsafety
• 13:52
  labeling
• 13:52
  function,
• 13:53
  computing
• 13:53
  the
• 13:53
  certified
• 13:54
  region
• 13:54
  using
• 13:54
  the
• 13:54
  LID
• 13:55
  verification
• 13:56
  framework,
• 13:56
  and
• 13:57
  constraining
• 13:57
  PPO
• 13:58
  updates
• 13:58
  via
• 13:58
  projected
• 13:59
  gradient
• 13:59
  descent
• 13:59
  to
• 13:59
  stay
• 14:00
  within
• 14:00
  the
• 14:00
  certified
• 14:00
  region.
• 14:01
  Currently
• 14:02
  limited
• 14:02
  to
• 14:02
  finite
• 14:03
  discrete
• 14:03
  state-action
• 14:04
  spaces
• 14:04
  —
• 14:05
  and
• 14:05
  the
• 14:05
  IBP-derived
• 14:06
  bounds
• 14:06
  become
• 14:06
  conservative
• 14:07
  for
• 14:07
  larger
• 14:07
  networks
• 14:08
  —
• 14:09
  but
• 14:09
  the
• 14:09
  conceptual
• 14:09
  bridge
• 14:10
  between
• 14:10
  formal
• 14:10
  verification
• 14:11
  and
• 14:11
  RL
• 14:11
  adaptation
• 14:12
  is
• 14:12
  the
• 14:12
  contribution
• 14:13
  that
• 14:13
  matters.
• 14:15
  MAKA
• 14:15
  deploys
• 14:15
  four
• 14:16
  specialized
• 14:16
  agents
• 14:17
  for
• 14:17
  CNC
• 14:17
  manufacturing
• 14:18
  of
• 14:18
  Ti-6Al-4V
• 14:19
  rotor
• 14:19
  blades:
• 14:20
  a
• 14:20
  Central
• 14:20
  Agent
• 14:21
  for
• 14:21
  intent
• 14:21
  routing,
• 14:22
  an
• 14:22
  Analysis
• 14:23
  Agent
• 14:23
  that
• 14:23
  performs
• 14:24
  all
• 14:24
  quantitative
• 14:25
  computation
• 14:25
  through
• 14:25
  deterministic
• 14:26
  tool
• 14:26
  calls
• 14:27
  (never
• 14:27
  unconstrained
• 14:28
  LLM
• 14:28
  generation
• 14:29
  for
• 14:29
  numbers),
• 14:30
  a
• 14:30
  Knowledge
• 14:30
  Graph
• 14:31
  Agent
• 14:31
  retrieving
• 14:31
  from
• 14:32
  2,701
• 14:33
  machining
• 14:33
  triples
• 14:34
  extracted
• 14:34
  from
• 14:34
  four
• 14:35
  research
• 14:35
  papers,
• 14:36
  and
• 14:36
  a
• 14:36
  Critic
• 14:36
  Agent
• 14:37
  enforcing
• 14:37
  physical
• 14:38
  plausibility
• 14:38
  and
• 14:38
  provenance
• 14:39
  completeness.
• 14:40
  On
• 14:40
  a
• 14:40
  75-question
• 14:41
  benchmark
• 14:42
  spanning
• 14:42
  three
• 14:42
  tool-use
• 14:43
  depth
• 14:43
  levels,
• 14:44
  it
• 14:44
  improves
• 14:44
  tool
• 14:44
  execution
• 14:45
  success
• 14:45
  by
• 14:46
  up
• 14:46
  to
• 14:46
  87.5
• 14:47
  percentage
• 14:47
  points
• 14:48
  versus
• 14:48
  single-model
• 14:49
  baselines.
• 14:50
  The
• 14:50
  Critic
• 14:51
  agent
• 14:51
  alone
• 14:51
  recovers
• 14:52
  61%
• 14:52
  of
• 14:52
  degraded
• 14:53
  trials.
• 14:54
  In
• 14:54
  a
• 14:54
  digital
• 14:55
  twin
• 14:55
  compensation
• 14:56
  demonstration,
• 14:57
  the
• 14:57
  architecture
• 14:57
  reduced
• 14:58
  predicted
• 14:58
  surface
• 14:59
  deviation
• 14:59
  from
• 14:59
  approximately
• 15:00
  0.01
• 15:01
  inches
• 15:01
  to
• 15:01
  0.001
• 15:03
  inches
• 15:03
  across
• 15:03
  most
• 15:04
  of
• 15:04
  the
• 15:04
  blade
• 15:04
  surface.
• 15:05
  The
• 15:05
  complete
• 15:06
  reasoning
• 15:06
  loop
• 15:06
  executes
• 15:07
  in
• 15:07
  4.3
• 15:07
  seconds
• 15:08
  —
• 15:09
  negligible
• 15:09
  versus
• 15:10
  machining
• 15:10
  cycle
• 15:10
  times.
• 15:11
  The
• 15:12
  architecture
• 15:12
  treats
• 15:12
  the
• 15:12
  LLM
• 15:13
  as
• 15:13
  a
• 15:13
  reasoning
• 15:13
  coordinator,
• 15:14
  not
• 15:15
  a
• 15:15
  source
• 15:15
  of
• 15:15
  quantitative
• 15:16
  truth,
• 15:16
  and
• 15:16
  that
• 15:17
  strict
• 15:17
  separation
• 15:17
  is
• 15:17
  what
• 15:18
  makes
• 15:18
  it
• 15:18
  viable
• 15:18
  where
• 15:18
  physically
• 15:19
  implausible
• 15:19
  recommendations
• 15:20
  cannot
• 15:20
  reach
• 15:21
  a
• 15:21
  human
• 15:21
  operator.