Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 23, 2026

Transcript

AGI Dreams Podcast – April 23, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Dominik
• 0:00
  Blain's
• 0:00
  new
• 0:00
  paper,
• 0:01
  "Mythos
• 0:01
  and
• 0:02
  the
• 0:02
  Unverified
• 0:02
  Cage,"
• 0:03
  presents
• 0:03
  COBALT,
• 0:04
  a
• 0:04
  Z3
• 0:04
  SMT-based
• 0:05
  formal
• 0:05
  verification
• 0:06
  engine
• 0:07
  that
• 0:07
  targets
• 0:07
  the
• 0:07
  exact
• 0:07
  class
• 0:08
  of
• 0:08
  vulnerability
• 0:08
  believed
• 0:08
  to
• 0:09
  have
• 0:09
  enabled
• 0:09
  the
• 0:09
  escape
• 0:11
  —
• 0:12
  CWE-190/191
• 0:14
  arithmetic
• 0:14
  overflows
• 0:15
  in
• 0:15
  C/C++
• 0:16
  infrastructure
• 0:17
  code.
• 0:18
  The
• 0:18
  results
• 0:18
  are
• 0:18
  striking.
• 0:19
  COBALT
• 0:19
  identified
• 0:20
  real
• 0:20
  vulnerabilities
• 0:20
  in
• 0:21
  four
• 0:21
  production
• 0:21
  codebases:
• 0:22
  a
• 0:22
  CWE-195
• 0:25
  resource
• 0:26
  ID
• 0:26
  truncation
• 0:27
  in
• 0:27
  NASA's
• 0:27
  Core
• 0:27
  Flight
• 0:28
  Executive
• 0:28
  (the
• 0:28
  framework
• 0:29
  running
• 0:29
  on
• 0:29
  the
• 0:29
  Mars
• 0:29
  Perseverance
• 0:30
  rover),
• 0:31
  a
• 0:31
  signed
• 0:31
  left-shift
• 0:31
  undefined
• 0:32
  behavior
• 0:32
  in
• 0:32
  wolfSSL's
• 0:33
  ML-DSA
• 0:35
  post-quantum
• 0:35
  implementation
• 0:36
  (confirmed
• 0:36
  and
• 0:36
  patched
• 0:37
  same-day
• 0:37
  via
• 0:37
  PR
• 0:37
  #10096),
• 0:40
  a
• 0:40
  CWE-191
• 0:43
  TLV
• 0:43
  length
• 0:43
  underflow
• 0:43
  in
• 0:44
  Eclipse
• 0:44
  Mosquitto's
• 0:44
  MQTT
• 0:45
  broker
• 0:45
  (CVE
• 0:45
  pending),
• 0:46
  and
• 0:46
  a
• 0:46
  CWE-190
• 0:49
  overflow
• 0:49
  bypass
• 0:49
  in
• 0:50
  NASA
• 0:50
  F
• 0:50
  Prime's
• 0:50
  sequencer
• 0:50
  (the
• 0:51
  framework
• 0:51
  behind
• 0:51
  the
• 0:52
  Ingenuity
• 0:52
  helicopter).
• 0:53
  Every
• 0:53
  SAT
• 0:54
  verdict
• 0:54
  produced
• 0:54
  a
• 0:54
  concrete
• 0:55
  exploitation
• 0:55
  witness;
• 0:56
  no
• 0:56
  report
• 0:56
  was
• 0:56
  rejected
• 0:57
  by
• 0:57
  the
• 0:57
  receiving
• 0:57
  security
• 0:58
  team.
• 0:58
  The
• 0:59
  paper's
• 0:59
  four-layer
• 0:59
  containment
• 1:00
  framework
• 1:00
  —
• 1:01
  COBALT
• 1:01
  for
• 1:01
  pre-deployment
• 1:02
  scanning,
• 1:02
  VERDICT
• 1:02
  for
• 1:03
  pre-execution
• 1:03
  constraint
• 1:04
  checking,
• 1:04
  DIRECTIVE-4
• 1:07
  for
• 1:08
  output
• 1:08
  firewalling,
• 1:09
  and
• 1:09
  SENTINEL
• 1:09
  for
• 1:10
  runtime
• 1:10
  monitoring
• 1:10
  —
• 1:11
  maps
• 1:11
  directly
• 1:11
  to
• 1:11
  failure.
• 1:12
  Blain's
• 1:13
  conditional
• 1:13
  argument
• 1:13
  is
• 1:14
  carefully
• 1:14
  hedged:
• 1:15
  Anthropic
• 1:15
  never
• 1:16
  publicly
• 1:16
  confirmed
• 1:16
  the
• 1:16
  escape
• 1:17
  vector,
• 1:17
  and
• 1:18
  secondary
• 1:18
  accounts
• 1:18
  hypothesizing
• 1:19
  a
• 1:19
  CWE-190
• 1:21
  arithmetic
• 1:22
  vulnerability
• 1:22
  remain
• 1:23
  unverified.
• 1:24
  But
• 1:24
  COBALT
• 1:25
  independently
• 1:25
  scanned
• 1:25
  OpenBSD
• 1:26
  HEAD
• 1:26
  and
• 1:27
  flagged
• 1:27
  a
• 1:27
  CWE-195
• 1:30
  signed-cast
• 1:30
  pattern
• 1:30
  at
• 1:31
  the
• 1:31
  exact
• 1:31
  file
• 1:31
  associated
• 1:32
  with
• 1:32
  the
• 1:32
  TCP
• 1:32
  SACK
• 1:33
  finding
• 1:33
  Mythos
• 1:34
  reportedly
• 1:34
  discovered
• 1:34
  —
• 1:35
  a
• 1:36
  27-year-old
• 1:36
  integer
• 1:37
  overflow.
• 1:38
  The
• 1:38
  runtime
• 1:39
  guard
• 1:39
  prototype
• 1:39
  is
• 1:39
  particularly
• 1:40
  promising:
• 1:41
  Z3-derived
• 1:42
  safe
• 1:42
  bounds
• 1:43
  compiled
• 1:43
  once
• 1:43
  at
• 1:43
  deployment,
• 1:44
  enforced
• 1:45
  at
• 1:45
  each
• 1:45
  trust
• 1:45
  boundary
• 1:46
  crossing
• 1:46
  with
• 1:46
  a
• 1:46
  mean
• 1:47
  latency
• 1:47
  of
• 1:47
  87.2
• 1:48
  nanoseconds
• 1:49
  and
• 1:49
  throughput
• 1:49
  of
• 1:49
  11.55
• 1:50
  million
• 1:51
  checks
• 1:51
  per
• 1:51
  second.
• 1:52
  Zero
• 1:53
  false
• 1:53
  positives,
• 1:53
  zero
• 1:54
  false
• 1:54
  negatives
• 1:54
  on
• 1:54
  4,000
• 1:55
  test
• 1:55
  inputs.
• 1:57
  The
• 1:57
  lesson,
• 1:57
  as
• 1:57
  Blain
• 1:58
  frames
• 1:58
  it,
• 1:58
  is
• 1:58
  not
• 1:59
  that
• 1:59
  Mythos
• 1:59
  was
• 1:59
  too
• 1:59
  capable
• 2:00
  to
• 2:00
  contain
• 2:00
  —
• 2:01
  it
• 2:01
  is
• 2:01
  that
• 2:02
  formal
• 2:02
  verification
• 2:03
  of
• 2:03
  containment
• 2:03
  infrastructure
• 2:04
  is
• 2:04
  an
• 2:04
  engineering
• 2:05
  problem
• 2:05
  with
• 2:05
  a
• 2:05
  formal
• 2:06
  solution.
• 2:07
  The
• 2:07
  community
• 2:08
  is
• 2:08
  not
• 2:08
  waiting
• 2:08
  for
• 2:09
  Anthropic
• 2:09
  to
• 2:09
  open-source
• 2:10
  anything.
• 2:11
  Kye
• 2:11
  Gomez's
• 2:12
  OpenMythos
• 2:12
  project
• 2:13
  offers
• 2:13
  a
• 2:13
  theoretical
• 2:14
  reconstruction
• 2:14
  of
• 2:14
  what
• 2:15
  Mythos
• 2:15
  might
• 2:15
  be
• 2:15
  architecturally:
• 2:16
  a
• 2:16
  Recurrent-Depth
• 2:17
  Transformer
• 2:17
  with
• 2:17
  a
• 2:18
  Prelude-Recurrence-Coda
• 2:18
  structure,
• 2:19
  switchable
• 2:20
  MLA/GQA
• 2:22
  attention,
• 2:22
  and
• 2:22
  sparse
• 2:23
  MoE
• 2:23
  feed-forward
• 2:23
  layers.
• 2:24
  The
• 2:25
  hypothesis
• 2:25
  —
• 2:26
  supported
• 2:26
  by
• 2:26
  Parcae
• 2:27
  scaling
• 2:27
  laws
• 2:28
  showing
• 2:28
  a
• 2:28
  770M-parameter
• 2:30
  looped
• 2:30
  model
• 2:31
  matching
• 2:31
  a
• 2:31
  1.3B
• 2:32
  fixed-depth
• 2:32
  transformer
• 2:33
  —
• 2:34
  is
• 2:34
  that
• 2:34
  Mythos
• 2:34
  achieves
• 2:35
  its
• 2:35
  capability
• 2:35
  through
• 2:35
  loop
• 2:36
  depth
• 2:36
  rather
• 2:37
  than
• 2:37
  raw
• 2:37
  parameter
• 2:37
  count,
• 2:38
  with
• 2:38
  each
• 2:39
  iteration
• 2:39
  functioning
• 2:40
  as
• 2:40
  an
• 2:40
  implicit
• 2:40
  chain-of-thought
• 2:41
  step
• 2:41
  in
• 2:41
  continuous
• 2:42
  latent
• 2:42
  space.
• 2:43
  Hugging
• 2:44
  Face's
• 2:44
  cybersecurity
• 2:45
  team
• 2:45
  argues
• 2:45
  the
• 2:46
  broader
• 2:46
  openness
• 2:47
  is
• 2:47
  a
• 2:47
  structural
• 2:48
  advantage
• 2:48
  for
• 2:48
  defense.
• 2:50
  Their
• 2:50
  analysis
• 2:50
  points
• 2:51
  out
• 2:51
  that
• 2:51
  AI
• 2:51
  systems
• 2:52
  are
• 2:52
  increasingly
• 2:53
  capable
• 2:53
  of
• 2:53
  reverse-engineering
• 2:54
  stripped
• 2:54
  binaries,
• 2:55
  eroding
• 2:56
  proprietary
• 2:56
  obscurity
• 2:57
  as
• 2:57
  a
• 2:57
  security
• 2:58
  strategy.
• 2:59
  When
• 3:00
  companies
• 3:00
  adopt
• 3:00
  AI
• 3:01
  coding
• 3:01
  tools
• 3:01
  under
• 3:02
  wrong
• 3:02
  incentives
• 3:02
  —
• 3:03
  measuring
• 3:03
  engineers
• 3:04
  by
• 3:04
  feature
• 3:05
  volume
• 3:05
  rather
• 3:05
  than
• 3:05
  code
• 3:06
  quality
• 3:06
  —
• 3:07
  they
• 3:07
  introduce
• 3:07
  more
• 3:08
  vulnerabilities
• 3:08
  into
• 3:08
  closed
• 3:09
  codebases
• 3:10
  where
• 3:10
  only
• 3:10
  one
• 3:10
  organization
• 3:11
  can
• 3:11
  find
• 3:11
  and
• 3:12
  fix
• 3:12
  them.
• 3:13
  Open
• 3:14
  ecosystems
• 3:14
  distribute
• 3:15
  the
• 3:15
  detection-verification-coordination-patching
• 3:17
  pipeline
• 3:17
  across
• 3:18
  communities
• 3:19
  rather
• 3:19
  than
• 3:19
  centralizing
• 3:20
  it
• 3:20
  behind
• 3:20
  a
• 3:20
  single
• 3:21
  point
• 3:21
  of
• 3:21
  failure.
• 3:22
  On
• 3:22
  the
• 3:22
  offensive
• 3:23
  OSINT
• 3:23
  side,
• 3:24
  Robin
• 3:24
  offers
• 3:24
  an
• 3:24
  AI-powered
• 3:25
  dark
• 3:25
  web
• 3:26
  investigation
• 3:26
  tool
• 3:27
  that
• 3:27
  routes
• 3:27
  queries
• 3:27
  through
• 3:27
  Tor,
• 3:28
  scrapes
• 3:29
  onion
• 3:29
  search
• 3:29
  engines,
• 3:30
  and
• 3:30
  uses
• 3:30
  LLMs
• 3:31
  (OpenAI,
• 3:32
  Claude,
• 3:32
  Gemini,
• 3:33
  or
• 3:33
  local
• 3:33
  via
• 3:34
  Ollama)
• 3:34
  to
• 3:34
  refine
• 3:35
  queries
• 3:35
  and
• 3:35
  summarize
• 3:36
  findings.
• 3:37
  It
• 3:37
  is
• 3:37
  exactly
• 3:38
  the
• 3:38
  kind
• 3:38
  of
• 3:38
  semi-autonomous
• 3:39
  defensive
• 3:39
  tool
• 3:40
  the
• 3:40
  Hugging
• 3:40
  Face
• 3:40
  team
• 3:41
  envisions
• 3:41
  —
• 3:42
  modular,
• 3:42
  auditable,
• 3:43
  and
• 3:43
  extensible.
• 3:48
  The
• 3:49
  hype
• 3:49
  around
• 3:49
  autonomous
• 3:50
  AI
• 3:50
  agents
• 3:50
  took
• 3:50
  a
• 3:51
  cold
• 3:51
  shower
• 3:51
  this
• 3:51
  week.
• 3:52
  BankerToolBench,
• 3:53
  a
• 3:53
  benchmark
• 3:54
  designed
• 3:54
  with
• 3:55
  actual
• 3:55
  investment
• 3:56
  banking
• 3:56
  practitioners,
• 3:57
  tested
• 3:57
  frontier
• 3:58
  agents
• 3:58
  on
• 3:58
  real
• 3:59
  analyst
• 3:59
  workflows
• 3:59
  —
• 4:00
  multi-file
• 4:01
  deliverables
• 4:01
  involving
• 4:02
  financial
• 4:02
  models,
• 4:03
  slides,
• 4:03
  reports,
• 4:04
  and
• 4:04
  internal
• 4:04
  consistency
• 4:05
  checks
• 4:05
  that
• 4:05
  represent
• 4:06
  hours
• 4:06
  of
• 4:06
  human
• 4:06
  work.
• 4:07
  The
• 4:07
  result:
• 4:08
  even
• 4:08
  the
• 4:08
  best
• 4:08
  agents
• 4:09
  produced
• 4:09
  zero
• 4:09
  percent
• 4:10
  of
• 4:10
  outputs
• 4:10
  that
• 4:10
  bankers
• 4:11
  considered
• 4:11
  client-ready.
• 4:12
  Not
• 4:12
  low.
• 4:13
  Zero.
• 4:14
  The
• 4:14
  failure
• 4:14
  modes
• 4:15
  are
• 4:15
  systemic:
• 4:16
  cross-artifact
• 4:16
  consistency
• 4:17
  collapses
• 4:17
  (numbers
• 4:18
  that
• 4:18
  do
• 4:18
  not
• 4:18
  match
• 4:19
  across
• 4:19
  slides,
• 4:20
  models,
• 4:20
  and
• 4:21
  outputs),
• 4:21
  formula
• 4:22
  and
• 4:22
  code
• 4:22
  errors
• 4:23
  at
• 4:23
  a
• 4:23
  41%
• 4:24
  rate,
• 4:24
  reasoning
• 4:25
  and
• 4:25
  logic
• 4:25
  failures
• 4:26
  at
• 4:26
  27%,
• 4:27
  and
• 4:27
  outright
• 4:28
  fabrication
• 4:28
  when
• 4:28
  the
• 4:28
  system
• 4:29
  gets
• 4:29
  stuck
• 4:29
  —
• 4:30
  inventing
• 4:30
  numbers
• 4:31
  and
• 4:31
  presenting
• 4:31
  them
• 4:31
  as
• 4:31
  sourced
• 4:32
  data.
• 4:32
  Stuart
• 4:33
  Winter-Tear's
• 4:33
  commentary
• 4:34
  cuts
• 4:34
  to
• 4:34
  the
• 4:34
  core:
• 4:35
  the
• 4:35
  current
• 4:35
  narrative
• 4:35
  about
• 4:35
  agents
• 4:36
  rests
• 4:36
  on
• 4:36
  the
• 4:36
  assumption
• 4:37
  that
• 4:37
  junior
• 4:37
  professional
• 4:38
  work
• 4:38
  is
• 4:38
  thin,
• 4:38
  repetitive
• 4:39
  admin
• 4:39
  waiting
• 4:40
  to
• 4:40
  be
• 4:40
  automated.
• 4:41
  This
• 4:41
  benchmark
• 4:41
  reveals
• 4:42
  that
• 4:42
  "a
• 4:42
  lot
• 4:42
  of
• 4:42
  that
• 4:42
  layer
• 4:43
  is
• 4:43
  stitching,
• 4:43
  checking,
• 4:44
  reconciling,
• 4:44
  and
• 4:44
  keeping
• 4:45
  the
• 4:45
  whole
• 4:45
  thing
• 4:45
  upright
• 4:46
  when
• 4:46
  multiple
• 4:46
  moving
• 4:46
  parts
• 4:47
  have
• 4:47
  to
• 4:47
  agree
• 4:47
  with
• 4:47
  each
• 4:47
  other
• 4:48
  and
• 4:48
  still
• 4:48
  make
• 4:48
  sense
• 4:48
  to
• 4:48
  someone
• 4:49
  senior,
• 4:49
  busy,
• 4:49
  and
• 4:50
  unforgiving."
• 4:50
  The
• 4:51
  trust
• 4:51
  problem
• 4:51
  extends
• 4:51
  beyond
• 4:52
  agent
• 4:52
  capability
• 4:53
  into
• 4:53
  the
• 4:53
  infrastructure
• 4:53
  that
• 4:53
  serves
• 4:54
  them.
• 4:54
  Kimi's
• 4:55
  vendor
• 4:55
  verifier,
• 4:55
  released
• 4:56
  alongside
• 4:56
  the
• 4:56
  K2.6
• 4:58
  model,
• 4:58
  addresses
• 4:59
  a
• 4:59
  growing
• 4:59
  crisis
• 4:59
  in
• 4:59
  the
• 5:00
  open-weight
• 5:00
  ecosystem:
• 5:01
  inference
• 5:01
  providers
• 5:02
  silently
• 5:02
  swapping
• 5:02
  weaker
• 5:03
  or
• 5:03
  over-quantized
• 5:03
  models,
• 5:04
  producing
• 5:04
  anomalous
• 5:05
  benchmark
• 5:05
  scores
• 5:06
  that
• 5:06
  users
• 5:06
  cannot
• 5:06
  distinguish
• 5:07
  from
• 5:07
  genuine
• 5:07
  model
• 5:07
  deficiencies.
• 5:09
  Kimi's
• 5:09
  approach
• 5:09
  is
• 5:10
  surgical
• 5:10
  —
• 5:11
  a
• 5:11
  six-test
• 5:11
  suite
• 5:11
  covering
• 5:12
  parameter
• 5:12
  enforcement,
• 5:13
  multimodal
• 5:13
  smoke
• 5:14
  tests,
• 5:14
  vision
• 5:15
  preprocessing
• 5:15
  validation,
• 5:16
  long-output
• 5:17
  stress
• 5:17
  tests
• 5:18
  (exposing
• 5:18
  KV
• 5:18
  cache
• 5:19
  bugs
• 5:19
  that
• 5:19
  short
• 5:19
  benchmarks
• 5:20
  hide),
• 5:20
  and
• 5:20
  tool-calling
• 5:21
  consistency.
• 5:22
  The
• 5:22
  team
• 5:22
  embeds
• 5:22
  directly
• 5:23
  with
• 5:23
  vLLM/SGLang
• 5:27
  communities
• 5:27
  to
• 5:27
  fix
• 5:28
  root
• 5:28
  causes
• 5:28
  rather
• 5:28
  than
• 5:28
  just
• 5:29
  detect
• 5:29
  symptoms,
• 5:30
  and
• 5:30
  plans
• 5:30
  a
• 5:30
  public
• 5:30
  vendor
• 5:31
  accuracy
• 5:31
  leaderboard.
• 5:32
  The
• 5:32
  uncomfortable
• 5:33
  truth
• 5:33
  they
• 5:33
  surfaced:
• 5:34
  "The
• 5:34
  more
• 5:34
  open
• 5:35
  the
• 5:35
  weights
• 5:35
  are,
• 5:35
  and
• 5:35
  the
• 5:35
  more
• 5:36
  diverse
• 5:36
  the
• 5:36
  deployment
• 5:36
  channels
• 5:37
  become,
• 5:37
  the
• 5:38
  less
• 5:38
  controllable
• 5:38
  the
• 5:38
  quality
• 5:39
  becomes."
• 5:40
  Denis
• 5:41
  O'Brien
• 5:41
  of
• 5:41
  AIX
• 5:42
  offers
• 5:42
  a
• 5:42
  different
• 5:42
  architectural
• 5:43
  critique
• 5:43
  entirely.
• 5:44
  His
• 5:44
  Seed
• 5:45
  IQ
• 5:45
  framework
• 5:45
  argues
• 5:46
  that
• 5:46
  the
• 5:46
  fundamental
• 5:46
  problem
• 5:47
  with
• 5:47
  LLM-based
• 5:48
  and
• 5:48
  RL+DL
• 5:49
  multi-agent
• 5:49
  systems
• 5:50
  is
• 5:51
  that
• 5:51
  coordination
• 5:51
  is
• 5:51
  message-driven,
• 5:52
  scaling
• 5:53
  quadratically
• 5:53
  as
• 5:54
  agents
• 5:54
  exchange
• 5:54
  prompts,
• 5:55
  state,
• 5:55
  and
• 5:55
  rewards
• 5:56
  across
• 5:56
  expanding
• 5:57
  interaction
• 5:57
  graphs.
• 5:58
  Seed
• 5:58
  IQ's
• 5:59
  alternative:
• 6:00
  each
• 6:00
  agent
• 6:00
  maintains
• 6:01
  its
• 6:01
  own
• 6:01
  internal
• 6:02
  world
• 6:02
  model,
• 6:03
  and
• 6:03
  global
• 6:03
  coherence
• 6:04
  emerges
• 6:04
  because
• 6:04
  those
• 6:05
  models
• 6:05
  evolve
• 6:05
  under
• 6:06
  the
• 6:06
  same
• 6:06
  viability
• 6:06
  bounds
• 6:07
  rather
• 6:07
  than
• 6:07
  through
• 6:07
  explicit
• 6:08
  negotiation.
• 6:09
  Early
• 6:10
  alignment
• 6:10
  is
• 6:10
  O(n),
• 6:11
  but
• 6:11
  through
• 6:11
  resonance
• 6:12
  scaling
• 6:12
  the
• 6:12
  interaction
• 6:13
  surface
• 6:13
  compresses
• 6:14
  to
• 6:14
  O(1)
• 6:14
  —
• 6:15
  effectively
• 6:15
  zero
• 6:16
  communication
• 6:16
  in
• 6:16
  the
• 6:16
  dominant
• 6:17
  regime.
• 6:18
  Whether
• 6:18
  the
• 6:18
  claims
• 6:18
  hold
• 6:19
  at
• 6:19
  production
• 6:19
  scale
• 6:19
  remains
• 6:20
  to
• 6:20
  be
• 6:20
  seen,
• 6:21
  but
• 6:21
  the
• 6:21
  diagnosis
• 6:22
  of
• 6:22
  the
• 6:22
  coordination
• 6:22
  bottleneck
• 6:23
  aligns
• 6:23
  with
• 6:24
  what
• 6:24
  BankerToolBench's
• 6:24
  failures
• 6:25
  expose:
• 6:26
  the
• 6:26
  problem
• 6:27
  is
• 6:27
  architectural,
• 6:28
  not
• 6:28
  just
• 6:28
  a
• 6:28
  matter
• 6:29
  of
• 6:29
  better
• 6:29
  prompts.
• 6:33
  Anyone
• 6:34
  who
• 6:34
  has
• 6:34
  used
• 6:34
  a
• 6:34
  coding
• 6:35
  agent
• 6:35
  knows
• 6:35
  the
• 6:35
  frustration:
• 6:36
  you
• 6:36
  ask
• 6:37
  it
• 6:37
  to
• 6:37
  fix
• 6:37
  a
• 6:37
  single
• 6:37
  off-by-one
• 6:38
  error,
• 6:39
  and
• 6:39
  it
• 6:39
  rewrites
• 6:39
  the
• 6:39
  entire
• 6:40
  function,
• 6:41
  adds
• 6:41
  validation,
• 6:42
  renames
• 6:43
  variables,
• 6:43
  and
• 6:44
  introduces
• 6:44
  a
• 6:44
  helper
• 6:44
  method.
• 6:46
  A
• 6:46
  rigorous
• 6:46
  new
• 6:47
  study
• 6:47
  quantifies
• 6:47
  this
• 6:48
  phenomenon
• 6:48
  across
• 6:48
  21
• 6:49
  frontier
• 6:49
  models
• 6:50
  using
• 6:50
  400
• 6:51
  programmatically
• 6:51
  corrupted
• 6:52
  problems
• 6:52
  from
• 6:52
  HumanEval+,
• 6:54
  where
• 6:54
  every
• 6:54
  ground-truth
• 6:55
  fix
• 6:55
  is
• 6:55
  a
• 6:55
  single-token
• 6:56
  reversal.
• 6:57
  GPT-5.4
• 6:58
  over-edits
• 6:59
  the
• 6:59
  most,
• 7:00
  with
• 7:00
  a
• 7:00
  token-level
• 7:01
  Levenshtein
• 7:01
  distance
• 7:01
  of
• 7:02
  0.39
• 7:03
  in
• 7:03
  reasoning
• 7:03
  mode
• 7:04
  and
• 7:04
  added
• 7:04
  cognitive
• 7:05
  complexity
• 7:05
  of
• 7:05
  2.31
• 7:06
  —
• 7:07
  while
• 7:07
  simultaneously
• 7:08
  posting
• 7:08
  one
• 7:08
  of
• 7:08
  the
• 7:08
  weakest
• 7:09
  Pass@1
• 7:09
  scores
• 7:10
  (0.723).
• 7:12
  Claude
• 7:12
  Opus
• 7:12
  4.6
• 7:13
  achieves
• 7:14
  the
• 7:14
  highest
• 7:14
  correctness
• 7:15
  (0.912)
• 7:16
  with
• 7:16
  the
• 7:17
  smallest
• 7:17
  diffs
• 7:17
  (Levenshtein
• 7:18
  0.06),
• 7:20
  making
• 7:20
  it
• 7:20
  the
• 7:20
  only
• 7:20
  frontier
• 7:21
  model
• 7:21
  where
• 7:21
  reasoning
• 7:22
  mode
• 7:22
  actually
• 7:22
  reduces
• 7:23
  over-editing
• 7:24
  relative
• 7:24
  to
• 7:24
  its
• 7:24
  non-reasoning
• 7:25
  variant.
• 7:26
  The
• 7:26
  study's
• 7:27
  training
• 7:27
  experiments
• 7:28
  are
• 7:28
  where
• 7:28
  it
• 7:28
  gets
• 7:28
  interesting.
• 7:29
  Supervised
• 7:30
  fine-tuning
• 7:30
  on
• 7:31
  minimal
• 7:31
  edits
• 7:31
  looks
• 7:31
  perfect
• 7:32
  in-domain
• 7:33
  but
• 7:33
  collapses
• 7:33
  entirely
• 7:34
  on
• 7:34
  out-of-domain
• 7:35
  corruptions
• 7:35
  —
• 7:36
  Pass@1
• 7:37
  drops
• 7:37
  to
• 7:37
  0.458
• 7:39
  as
• 7:39
  the
• 7:39
  model
• 7:39
  memorizes
• 7:40
  specific
• 7:40
  reversals
• 7:41
  rather
• 7:41
  than
• 7:41
  learning
• 7:41
  general
• 7:42
  minimality.
• 7:43
  Reinforcement
• 7:44
  learning
• 7:44
  is
• 7:45
  the
• 7:45
  only
• 7:45
  method
• 7:45
  that
• 7:45
  generalizes
• 7:46
  cleanly,
• 7:47
  improving
• 7:48
  edit
• 7:48
  minimality
• 7:48
  without
• 7:49
  any
• 7:49
  degradation
• 7:50
  on
• 7:50
  LiveCodeBench,
• 7:51
  consistent
• 7:52
  with
• 7:52
  the
• 7:52
  broader
• 7:52
  finding
• 7:53
  that
• 7:53
  SFT
• 7:53
  memorizes
• 7:54
  while
• 7:54
  RL
• 7:54
  generalizes.
• 7:56
  LoRA
• 7:57
  at
• 7:57
  rank
• 7:57
  64
• 7:58
  nearly
• 7:58
  matches
• 7:58
  full
• 7:58
  RL,
• 7:59
  suggesting
• 8:00
  that
• 8:00
  for
• 8:00
  style-level
• 8:01
  behavioral
• 8:01
  changes
• 8:02
  where
• 8:02
  underlying
• 8:02
  capability
• 8:03
  already
• 8:03
  exists,
• 8:04
  parameter-efficient
• 8:05
  fine-tuning
• 8:06
  is
• 8:06
  sufficient.
• 8:07
  The
• 8:07
  takeaway
• 8:07
  for
• 8:08
  reasoning
• 8:08
  models:
• 8:09
  their
• 8:09
  default
• 8:09
  behavior
• 8:10
  is
• 8:10
  to
• 8:10
  over-edit,
• 8:11
  but
• 8:11
  explicit
• 8:12
  prompting
• 8:12
  to
• 8:12
  preserve
• 8:12
  original
• 8:13
  code
• 8:13
  dramatically
• 8:14
  narrows
• 8:14
  diffs,
• 8:15
  and
• 8:15
  RL
• 8:15
  can
• 8:16
  bake
• 8:16
  that
• 8:16
  constraint
• 8:16
  in
• 8:16
  permanently.
• 8:18
  The
• 8:18
  quality
• 8:18
  problem
• 8:19
  extends
• 8:19
  beyond
• 8:19
  code
• 8:20
  to
• 8:20
  design.
• 8:21
  An
• 8:21
  analysis
• 8:22
  of
• 8:22
  500
• 8:22
  recent
• 8:23
  Show
• 8:23
  HN
• 8:23
  submissions
• 8:24
  scored
• 8:24
  them
• 8:24
  against
• 8:25
  15
• 8:25
  deterministic
• 8:26
  CSS
• 8:27
  and
• 8:27
  DOM
• 8:27
  patterns
• 8:28
  associated
• 8:28
  with
• 8:28
  AI-generated
• 8:29
  design
• 8:29
  —
• 8:30
  Inter-everywhere
• 8:31
  hero
• 8:31
  headlines,
• 8:32
  colored
• 8:32
  left
• 8:33
  borders
• 8:33
  on
• 8:33
  cards,
• 8:34
  perma-dark
• 8:34
  mode
• 8:35
  with
• 8:35
  medium-grey
• 8:35
  body
• 8:36
  text,
• 8:36
  badge-above-hero
• 8:37
  layouts,
• 8:38
  identical
• 8:38
  feature
• 8:39
  cards
• 8:39
  with
• 8:39
  top
• 8:40
  icons.
• 8:41
  Twenty-one
• 8:41
  percent
• 8:41
  of
• 8:42
  sites
• 8:42
  triggered
• 8:42
  five
• 8:43
  or
• 8:43
  more
• 8:43
  patterns
• 8:43
  (classified
• 8:44
  as
• 8:44
  "high
• 8:44
  AI
• 8:45
  slop"),
• 8:46
  another
• 8:46
  46%
• 8:47
  triggered
• 8:47
  two
• 8:47
  to
• 8:47
  four.
• 8:48
  The
• 8:48
  author
• 8:49
  notes
• 8:49
  this
• 8:49
  is
• 8:49
  not
• 8:49
  necessarily
• 8:50
  bad
• 8:50
  —
• 8:51
  "validating
• 8:51
  a
• 8:51
  business
• 8:52
  idea
• 8:52
  was
• 8:52
  never
• 8:53
  about
• 8:53
  fancy
• 8:53
  design"
• 8:53
  —
• 8:54
  but
• 8:54
  wonders
• 8:55
  whether
• 8:55
  people
• 8:55
  will
• 8:55
  craft
• 8:56
  distinctive
• 8:56
  designs
• 8:56
  to
• 8:57
  stand
• 8:57
  out
• 8:57
  from
• 8:57
  the
• 8:57
  slop,
• 8:58
  or
• 8:58
  whether
• 8:58
  it
• 8:58
  even
• 8:59
  matters
• 8:59
  once
• 8:59
  AI
• 9:00
  agents
• 9:00
  are
• 9:00
  the
• 9:00
  primary
• 9:01
  users
• 9:01
  of
• 9:01
  the
• 9:01
  web.
• 9:02
  For
• 9:02
  those
• 9:02
  scaling
• 9:03
  AI-assisted
• 9:03
  development
• 9:04
  to
• 9:04
  parallel
• 9:04
  workflows,
• 9:06
  a
• 9:06
  detailed
• 9:06
  practitioner
• 9:07
  playbook
• 9:07
  covers
• 9:08
  the
• 9:08
  five
• 9:08
  pillars
• 9:08
  of
• 9:08
  parallel
• 9:09
  agentic
• 9:09
  development:
• 9:10
  issues-as-specs,
• 9:12
  git
• 9:12
  worktrees
• 9:12
  for
• 9:12
  codebase
• 9:13
  isolation,
• 9:14
  plan-build-validate
• 9:15
  loops,
• 9:15
  fresh-context
• 9:16
  PR
• 9:16
  reviews
• 9:17
  (never
• 9:17
  let
• 9:17
  the
• 9:17
  agent
• 9:17
  grade
• 9:18
  its
• 9:18
  own
• 9:18
  homework
• 9:18
  in
• 9:19
  the
• 9:19
  same
• 9:19
  context
• 9:19
  window),
• 9:20
  and
• 9:20
  a
• 9:20
  self-healing
• 9:21
  layer
• 9:21
  that
• 9:21
  evolves
• 9:22
  rules,
• 9:22
  skills,
• 9:23
  and
• 9:23
  workflows
• 9:24
  based
• 9:24
  on
• 9:24
  post-review
• 9:24
  The
• 9:26
  critical
• 9:26
  infrastructure
• 9:27
  includes
• 9:27
  per-worktree
• 9:28
  database
• 9:28
  branching
• 9:29
  (via
• 9:29
  Neon
• 9:29
  or
• 9:29
  local
• 9:30
  SQLite)
• 9:30
  and
• 9:31
  dynamic
• 9:31
  port
• 9:31
  assignment
• 9:32
  to
• 9:32
  prevent
• 9:32
  conflicts
• 9:32
  when
• 9:33
  multiple
• 9:33
  application
• 9:33
  instances
• 9:34
  run
• 9:34
  simultaneously.
• 9:39
  A
• 9:39
  comprehensive
• 9:39
  benchmark
• 9:40
  of
• 9:40
  21
• 9:40
  local
• 9:41
  LLMs
• 9:41
  on
• 9:41
  a
• 9:41
  MacBook
• 9:42
  Air
• 9:42
  M5
• 9:42
  using
• 9:43
  HumanEval+
• 9:43
  reveals
• 9:44
  some
• 9:44
  surprises.
• 9:45
  Qwen
• 9:45
  3.6
• 9:46
  35B-A3B
• 9:47
  dominates
• 9:48
  at
• 9:48
  89.6%
• 9:49
  Pass@1
• 9:50
  while
• 9:50
  running
• 9:50
  at
• 9:50
  16.9
• 9:51
  tok/s
• 9:52
  thanks
• 9:52
  to
• 9:52
  its
• 9:52
  MoE
• 9:53
  architecture
• 9:53
  —
• 9:54
  active
• 9:54
  parameter
• 9:54
  count
• 9:55
  determines
• 9:55
  speed
• 9:55
  while
• 9:55
  total
• 9:56
  parameter
• 9:56
  count
• 9:56
  determines
• 9:57
  quality.
• 9:58
  The
• 9:58
  best
• 9:58
  bang-for-RAM
• 9:59
  goes
• 9:59
  to
• 9:59
  Qwen
• 9:59
  2.5
• 10:00
  Coder
• 10:00
  7B
• 10:00
  at
• 10:00
  84.2%
• 10:01
  in
• 10:02
  just
• 10:02
  4.5
• 10:02
  GB.
• 10:03
  But
• 10:03
  the
• 10:03
  Gemma
• 10:04
  4
• 10:04
  results
• 10:04
  are
• 10:04
  baffling:
• 10:05
  Gemma
• 10:05
  4
• 10:06
  31B
• 10:06
  scores
• 10:07
  31.1%,
• 10:08
  lower
• 10:08
  than
• 10:09
  Llama
• 10:09
  3.2
• 10:10
  1B
• 10:10
  (32.9%)
• 10:11
  and
• 10:11
  drastically
• 10:12
  below
• 10:12
  Gemma
• 10:12
  3
• 10:13
  27B
• 10:14
  (78.7%).
• 10:15
  The
• 10:15
  community
• 10:16
  suspects
• 10:16
  a
• 10:16
  tool-calling
• 10:17
  premature
• 10:17
  stop
• 10:18
  bug
• 10:18
  that
• 10:18
  both
• 10:18
  Google
• 10:18
  and
• 10:18
  llama.cpp
• 10:20
  have
• 10:20
  issued
• 10:20
  partial
• 10:21
  fixes
• 10:21
  for,
• 10:22
  though
• 10:22
  Q4
• 10:22
  K
• 10:23
  M
• 10:23
  quantization
• 10:23
  hitting
• 10:23
  the
• 10:24
  architecture
• 10:24
  harder
• 10:24
  than
• 10:24
  others
• 10:25
  cannot
• 10:25
  be
• 10:25
  ruled
• 10:25
  out.
• 10:26
  At
• 10:26
  the
• 10:26
  edge,
• 10:27
  a
• 10:27
  Gemma
• 10:27
  4
• 10:28
  VLA
• 10:28
  demo
• 10:28
  on
• 10:28
  the
• 10:28
  NVIDIA
• 10:29
  Jetson
• 10:29
  Orin
• 10:29
  Nano
• 10:30
  Super
• 10:30
  shows
• 10:31
  what
• 10:31
  a
• 10:31
  vision-language-action
• 10:32
  pipeline
• 10:32
  looks
• 10:32
  like
• 10:33
  on
• 10:33
  an
• 10:33
  8GB
• 10:33
  embedded
• 10:34
  board:
• 10:34
  speech-to-text
• 10:35
  via
• 10:35
  Parakeet,
• 10:36
  Gemma
• 10:36
  4
• 10:37
  deciding
• 10:37
  autonomously
• 10:38
  whether
• 10:38
  to
• 10:38
  activate
• 10:39
  the
• 10:39
  webcam
• 10:39
  (no
• 10:39
  keyword
• 10:40
  triggers,
• 10:40
  no
• 10:41
  hardcoded
• 10:41
  logic),
• 10:42
  and
• 10:42
  Kokoro
• 10:42
  text-to-speech
• 10:43
  for
• 10:43
  output.
• 10:44
  The
• 10:44
  model
• 10:44
  receives
• 10:45
  a
• 10:45
  single
• 10:45
  tool
• 10:45
  definition
• 10:46
  —
• 10:46
  "take
• 10:47
  a
• 10:47
  photo
• 10:47
  and
• 10:47
  analyze
• 10:48
  what
• 10:48
  is
• 10:48
  visible"
• 10:48
  —
• 10:49
  and
• 10:49
  decides
• 10:49
  on
• 10:49
  its
• 10:50
  own
• 10:50
  when
• 10:50
  visual
• 10:50
  context
• 10:51
  is
• 10:51
  needed.
• 10:52
  It
• 10:52
  runs
• 10:52
  with
• 10:52
  all
• 10:53
  layers
• 10:53
  offloaded
• 10:53
  to
• 10:53
  GPU,
• 10:55
  the
• 10:55
  Q4
• 10:55
  K
• 10:55
  M
• 10:55
  quantization
• 10:56
  sweet
• 10:56
  spot,
• 10:57
  and
• 10:57
  2048
• 10:58
  context.
• 10:59
  Meanwhile,
• 11:00
  the
• 11:00
  extreme
• 11:00
  quantization
• 11:01
  frontier
• 11:01
  keeps
• 11:01
  advancing.
• 11:03
  A
• 11:03
  merged
• 11:03
  PR
• 11:04
  in
• 11:04
  llama.cpp
• 11:05
  optimizes
• 11:06
  the
• 11:06
  x86
• 11:07
  and
• 11:07
  generic
• 11:07
  CPU
• 11:08
  q1
• 11:08
  0
• 11:08
  dot
• 11:09
  product,
• 11:09
  pushing
• 11:10
  1-bit
• 11:10
  inference
• 11:10
  from
• 11:11
  0.3
• 11:11
  to
• 11:11
  1.7
• 11:12
  tok/s
• 11:13
  even
• 11:13
  on
• 11:13
  an
• 11:13
  old
• 11:14
  laptop
• 11:14
  without
• 11:14
  AVX
• 11:15
  support
• 11:15
  —
• 11:16
  with
• 11:16
  Metal,
• 11:16
  Vulkan,
• 11:17
  and
• 11:17
  CUDA
• 11:17
  backends
• 11:18
  also
• 11:18
  supporting
• 11:19
  the
• 11:19
  Bonsai
• 11:19
  1-bit
• 11:19
  format.
• 11:20
  In
• 11:20
  the
• 11:21
  pre-registered
• 11:21
  experiment
• 11:22
  category,
• 11:22
  the
• 11:23
  T3
• 11:23
  (Toroidal
• 11:24
  Tesseract
• 11:24
  Transformer)
• 11:25
  v3.5
• 11:27
  project
• 11:27
  is
• 11:27
  attempting
• 11:27
  to
• 11:28
  apply
• 11:28
  a
• 11:28
  proprietary
• 11:29
  architecture
• 11:29
  to
• 11:30
  Google's
• 11:30
  Gemma
• 11:30
  3
• 11:31
  270M
• 11:32
  weights
• 11:32
  with
• 11:32
  5
• 11:32
  billion
• 11:33
  tokens
• 11:33
  of
• 11:33
  continued
• 11:34
  training.
• 11:35
  The
• 11:35
  claim
• 11:35
  is
• 11:35
  architectural,
• 11:36
  not
• 11:37
  data-compute:
• 11:37
  that
• 11:37
  the
• 11:37
  T3
• 11:38
  ecology
• 11:38
  mechanism
• 11:39
  (0.003%
• 11:41
  of
• 11:41
  parameters
• 11:42
  absorbing
• 11:42
  7.2%
• 11:43
  of
• 11:44
  gradient
• 11:44
  norm
• 11:45
  at
• 11:45
  normalized
• 11:45
  pressure
• 11:46
  2,463x)
• 11:48
  will
• 11:48
  cross
• 11:48
  the
• 11:49
  released
• 11:49
  Gemma
• 11:49
  reasoning
• 11:50
  benchmark
• 11:50
  composite
• 11:51
  before
• 11:51
  75%
• 11:52
  of
• 11:52
  training.
• 11:53
  Prior
• 11:54
  GPT-2
• 11:54
  Medium
• 11:55
  experiments
• 11:55
  showed
• 11:55
  the
• 11:56
  ecology
• 11:56
  is
• 11:56
  load-bearing
• 11:57
  —
• 11:58
  ablating
• 11:58
  the
• 11:58
  sigma
• 11:59
  clamp
• 11:59
  dropped
• 11:59
  ARC-Easy
• 12:00
  by
• 12:00
  7.7
• 12:01
  percentage
• 12:01
  points.
• 12:02
  The
• 12:02
  protocol
• 12:03
  is
• 12:03
  frozen,
• 12:03
  SHA-256
• 12:06
  hashed,
• 12:07
  and
• 12:07
  publicly
• 12:07
  verifiable,
• 12:08
  with
• 12:08
  a
• 12:08
  live
• 12:08
  dashboard
• 12:09
  tracking
• 12:09
  perplexity
• 12:10
  trajectory.
• 12:11
  OpenAI
• 12:12
  quietly
• 12:12
  released
• 12:13
  a
• 12:13
  privacy
• 12:13
  filter
• 12:13
  model
• 12:14
  that
• 12:14
  runs
• 12:14
  entirely
• 12:15
  in-browser
• 12:15
  via
• 12:15
  WebGPU
• 12:16
  —
• 12:17
  a
• 12:17
  notable
• 12:17
  shift
• 12:17
  from
• 12:17
  a
• 12:18
  company
• 12:18
  that
• 12:18
  has
• 12:18
  historically
• 12:19
  kept
• 12:19
  everything
• 12:19
  server-side.
• 12:21
  On
• 12:21
  the
• 12:21
  image
• 12:21
  generation
• 12:22
  front,
• 12:22
  Qwen
• 12:22
  Image
• 12:23
  Edit
• 12:23
  2511
• 12:24
  is
• 12:24
  now
• 12:24
  servable
• 12:25
  in-browser
• 12:25
  with
• 12:25
  generation
• 12:26
  times
• 12:26
  down
• 12:26
  to
• 12:27
  10
• 12:27
  seconds
• 12:27
  on
• 12:27
  an
• 12:27
  L4
• 12:28
  GPU
• 12:28
  ,
• 12:29
  while
• 12:29
  OpenAI
• 12:30
  launched
• 12:30
  ChatGPT
• 12:31
  Images
• 12:31
  2.0
• 12:32
  as
• 12:32
  its
• 12:32
  next-generation
• 12:33
  image
• 12:33
  creation
• 12:34
  system
• 12:34
  .
• 12:39
  Uber
• 12:40
  has
• 12:40
  rewritten
• 12:40
  its
• 12:40
  ledger
• 12:41
  systems
• 12:41
  five
• 12:42
  times
• 12:42
  in
• 12:42
  ten
• 12:42
  years,
• 12:43
  and
• 12:43
  at
• 12:43
  least
• 12:43
  one
• 12:44
  rewrite
• 12:44
  —
• 12:44
  the
• 12:44
  one
• 12:45
  built
• 12:45
  on
• 12:45
  DynamoDB
• 12:46
  —
• 12:47
  could
• 12:47
  have
• 12:47
  been
• 12:47
  avoided
• 12:47
  with
• 12:48
  napkin
• 12:48
  math.
• 12:49
  In
• 12:49
  2017,
• 12:50
  Uber
• 12:50
  launched
• 12:50
  its
• 12:51
  payment
• 12:51
  ledger
• 12:51
  on
• 12:51
  DynamoDB,
• 12:53
  a
• 12:53
  consumption-priced
• 12:53
  database
• 12:54
  where
• 12:54
  you
• 12:54
  pay
• 12:54
  for
• 12:54
  every
• 12:55
  read
• 12:55
  and
• 12:55
  every
• 12:55
  write.
• 12:56
  With
• 12:56
  11
• 12:57
  million
• 12:57
  trips
• 12:57
  per
• 12:57
  day
• 12:57
  generating
• 12:58
  roughly
• 12:58
  10
• 12:58
  ledger
• 12:59
  entries
• 12:59
  each
• 12:59
  at
• 12:59
  5
• 13:00
  write
• 13:00
  capacity
• 13:00
  units
• 13:01
  per
• 13:01
  entry,
• 13:02
  the
• 13:02
  write
• 13:02
  costs
• 13:02
  alone
• 13:02
  reached
• 13:03
  $250,000
• 13:04
  annually.
• 13:05
  With
• 13:05
  3x
• 13:05
  annual
• 13:06
  growth,
• 13:06
  that
• 13:06
  became
• 13:06
  $2.25
• 13:07
  million
• 13:07
  per
• 13:08
  year
• 13:08
  by
• 13:08
  year
• 13:08
  three.
• 13:09
  Adding
• 13:10
  storage
• 13:10
  costs
• 13:11
  for
• 13:11
  the
• 13:11
  1.2
• 13:11
  petabytes
• 13:12
  accumulated
• 13:12
  by
• 13:12
  2020
• 13:13
  at
• 13:13
  $0.25
• 13:14
  per
• 13:14
  gigabyte
• 13:15
  brings
• 13:15
  the
• 13:15
  cumulative
• 13:16
  bill
• 13:16
  to
• 13:16
  approximately
• 13:17
  $8
• 13:17
  million
• 13:17
  —
• 13:18
  for
• 13:18
  a
• 13:19
  ledger
• 13:19
  that
• 13:19
  did
• 13:19
  not
• 13:19
  need
• 13:19
  to
• 13:19
  be
• 13:19
  on
• 13:20
  DynamoDB
• 13:20
  in
• 13:20
  the
• 13:20
  first
• 13:20
  place.
• 13:21
  The
• 13:21
  critical
• 13:22
  distinction:
• 13:22
  DynamoDB
• 13:23
  is
• 13:23
  excellent
• 13:24
  for
• 13:24
  payments
• 13:24
  (independent
• 13:25
  transactions,
• 13:26
  causal
• 13:26
  consistency
• 13:27
  sufficient)
• 13:27
  but
• 13:28
  fundamentally
• 13:28
  wrong
• 13:28
  for
• 13:29
  ledgers
• 13:29
  (which
• 13:29
  require
• 13:29
  global
• 13:30
  consistency
• 13:30
  across
• 13:31
  the
• 13:31
  full
• 13:31
  scope
• 13:31
  of
• 13:31
  financial
• 13:32
  state).
• 13:32
  The
• 13:33
  real
• 13:33
  indictment
• 13:33
  is
• 13:33
  not
• 13:33
  the
• 13:33
  technical
• 13:34
  mistake
• 13:34
  but
• 13:35
  the
• 13:35
  incentive
• 13:35
  structure.
• 13:36
  Every
• 13:36
  rewrite
• 13:37
  was
• 13:37
  someone's
• 13:37
  promotion
• 13:37
  project.
• 13:39
  When
• 13:39
  the
• 13:39
  DynamoDB
• 13:39
  costs
• 13:40
  became
• 13:40
  prohibitive,
• 13:41
  Uber
• 13:41
  migrated
• 13:42
  to
• 13:42
  an
• 13:42
  internal
• 13:42
  database
• 13:43
  (DocStore)
• 13:44
  that
• 13:44
  lacked
• 13:44
  change
• 13:45
  data
• 13:45
  capture
• 13:45
  —
• 13:46
  so
• 13:46
  they
• 13:46
  had
• 13:46
  to
• 13:46
  build
• 13:46
  a
• 13:47
  streaming
• 13:47
  framework
• 13:47
  on
• 13:47
  top
• 13:48
  of
• 13:48
  it.
• 13:48
  And
• 13:49
  when
• 13:49
  AWS
• 13:49
  invited
• 13:50
  Uber
• 13:50
  to
• 13:50
  present
• 13:50
  at
• 13:50
  re:Invent
• 13:51
  2019,
• 13:52
  they
• 13:52
  said
• 13:53
  yes,
• 13:53
  turning
• 13:54
  an
• 13:54
  atrocious
• 13:54
  decision
• 13:54
  into
• 13:55
  a
• 13:55
  case
• 13:55
  study.
• 13:56
  ByteByteGo
• 13:57
  later
• 13:57
  praised
• 13:57
  "the
• 13:57
  cost
• 13:58
  savings
• 13:58
  from
• 13:58
  this
• 13:58
  migration"
• 13:59
  —
• 14:00
  savings
• 14:00
  from
• 14:01
  migrating
• 14:01
  away
• 14:01
  from
• 14:01
  a
• 14:01
  database
• 14:02
  that
• 14:02
  should
• 14:02
  never
• 14:02
  have
• 14:02
  been
• 14:03
  chosen.
• 14:04
  The
• 14:04
  author's
• 14:04
  framing
• 14:05
  is
• 14:05
  brutal:
• 14:06
  "the
• 14:06
  technological
• 14:07
  equivalent
• 14:07
  of
• 14:07
  an
• 14:07
  arsonist
• 14:08
  writing
• 14:08
  a
• 14:08
  fire
• 14:08
  safety
• 14:09
  manual."
• 14:10
  The
• 14:10
  lesson
• 14:10
  applies
• 14:10
  far
• 14:11
  beyond
• 14:11
  Uber:
• 14:12
  if
• 14:12
  you
• 14:12
  are
• 14:12
  building
• 14:12
  a
• 14:12
  system
• 14:13
  that
• 14:13
  makes
• 14:13
  the
• 14:13
  economics
• 14:14
  of
• 14:14
  your
• 14:14
  company
• 14:14
  impossible,
• 14:16
  you
• 14:16
  are
• 14:16
  better
• 14:16
  off
• 14:16
  not
• 14:16
  building
• 14:17
  it.
• 14:18
  Focusing
• 14:18
  solely
• 14:19
  on
• 14:19
  technical
• 14:19
  requirements
• 14:20
  without
• 14:20
  modeling
• 14:21
  costs
• 14:21
  is
• 14:22
  a
• 14:22
  disservice
• 14:22
  to
• 14:22
  the
• 14:22
  business.
• 14:23
  In
• 14:24
  the
• 14:24
  AI
• 14:24
  tooling
• 14:24
  space,
• 14:25
  Claude's
• 14:25
  new
• 14:25
  co-work
• 14:26
  live
• 14:26
  artifacts
• 14:27
  feature
• 14:27
  takes
• 14:27
  a
• 14:27
  different
• 14:28
  approach
• 14:28
  to
• 14:28
  the
• 14:28
  dashboard
• 14:29
  problem:
• 14:30
  rather
• 14:30
  than
• 14:30
  building
• 14:31
  dashboards
• 14:31
  that
• 14:31
  become
• 14:32
  stale
• 14:32
  snapshots,
• 14:33
  live
• 14:33
  artifacts
• 14:34
  connect
• 14:34
  directly
• 14:35
  to
• 14:35
  apps
• 14:35
  and
• 14:35
  files,
• 14:36
  refreshing
• 14:37
  with
• 14:37
  current
• 14:37
  data
• 14:37
  every
• 14:38
  time
• 14:38
  they
• 14:38
  are
• 14:38
  opened.
• 14:39
  The
• 14:39
  pitch
• 14:40
  is
• 14:40
  operational
• 14:40
  —
• 14:41
  a
• 14:41
  lead
• 14:41
  tracker
• 14:42
  that
• 14:42
  deduplicates
• 14:42
  across
• 14:43
  sources
• 14:43
  and
• 14:43
  shows
• 14:44
  who
• 14:44
  to
• 14:44
  follow
• 14:44
  up
• 14:44
  with
• 14:44
  today,
• 14:45
  a
• 14:46
  content
• 14:46
  planner
• 14:46
  that
• 14:46
  surfaces
• 14:47
  top
• 14:47
  posts
• 14:48
  and
• 14:48
  coverage
• 14:48
  gaps,
• 14:49
  client
• 14:49
  reports
• 14:50
  that
• 14:50
  stay
• 14:50
  current
• 14:50
  without
• 14:51
  manual
• 14:51
  assembly.
• 14:56
  Shopify's
• 14:57
  CTO
• 14:57
  Muel
• 14:58
  Parkin
• 14:58
  revealed
• 14:58
  the
• 14:58
  scale
• 14:59
  of
• 14:59
  internal
• 14:59
  AI
• 15:00
  adoption:
• 15:00
  daily
• 15:01
  active
• 15:01
  usage
• 15:02
  of
• 15:02
  AI
• 15:02
  tools
• 15:02
  approaches
• 15:03
  100%
• 15:03
  of
• 15:04
  employees,
• 15:04
  with
• 15:04
  CLI-based
• 15:05
  and
• 15:05
  headless
• 15:06
  tools
• 15:06
  growing
• 15:07
  fastest
• 15:07
  while
• 15:07
  IDE-integrated
• 15:08
  tools
• 15:08
  like
• 15:09
  GitHub
• 15:09
  Copilot
• 15:09
  plateau.
• 15:10
  The
• 15:10
  company
• 15:11
  funds
• 15:11
  unlimited
• 15:12
  tokens
• 15:12
  for
• 15:12
  everyone
• 15:12
  but
• 15:13
  enforces
• 15:13
  a
• 15:13
  floor
• 15:13
  rather
• 15:14
  than
• 15:14
  a
• 15:14
  ceiling
• 15:14
  —
• 15:15
  discouraging
• 15:15
  anything
• 15:16
  less
• 15:16
  capable
• 15:16
  than
• 15:16
  Opus
• 15:17
  Token
• 15:18
  consumption
• 15:19
  distribution
• 15:19
  is
• 15:19
  increasingly
• 15:20
  skewed,
• 15:21
  with
• 15:21
  the
• 15:21
  top
• 15:21
  10th
• 15:21
  percentile
• 15:22
  growing
• 15:22
  faster
• 15:23
  than
• 15:23
  the
• 15:23
  median,
• 15:24
  raising
• 15:24
  questions
• 15:25
  about
• 15:25
  whether
• 15:25
  this
• 15:25
  represents
• 15:26
  power-user
• 15:26
  efficiency
• 15:27
  or
• 15:27
  diminishing
• 15:27
  returns
• 15:28
  for
• 15:28
  the
• 15:28
  majority.
• 15:29
  On
• 15:29
  the
• 15:29
  Jensen
• 15:30
  Huang
• 15:30
  "$100K
• 15:31
  of
• 15:31
  tokens
• 15:32
  per
• 15:32
  $200K
• 15:33
  engineer"
• 15:33
  thesis,
• 15:34
  Parkin
• 15:35
  agrees
• 15:35
  directionally
• 15:36
  but
• 15:36
  warns
• 15:36
  that
• 15:36
  the
• 15:36
  anti-pattern
• 15:37
  is
• 15:37
  running
• 15:37
  too
• 15:37
  many
• 15:38
  agents
• 15:38
  in
• 15:38
  parallel
• 15:39
  without
• 15:39
  communication
• 15:39
  —
• 15:40
  burning
• 15:40
  tokens
• 15:41
  inefficiently.
• 15:42
  The
• 15:42
  right
• 15:42
  approach:
• 15:43
  critique
• 15:43
  loops
• 15:44
  where
• 15:44
  one
• 15:44
  agent
• 15:44
  generates
• 15:45
  and
• 15:45
  another
• 15:45
  (ideally
• 15:45
  a
• 15:45
  different
• 15:46
  model)
• 15:46
  reviews,
• 15:47
  accepting
• 15:47
  higher
• 15:48
  latency
• 15:48
  for
• 15:48
  dramatically
• 15:49
  better
• 15:49
  code
• 15:49
  quality.
• 15:50
  The
• 15:50
  internal
• 15:50
  tools
• 15:51
  are
• 15:51
  impressive.
• 15:52
  Tangle,
• 15:53
  a
• 15:53
  third-generation
• 15:54
  ML
• 15:54
  experimentation
• 15:55
  platform,
• 15:56
  uses
• 15:56
  content-hash-based
• 15:57
  caching
• 15:58
  so
• 15:58
  that
• 15:58
  if
• 15:58
  multiple
• 15:59
  people
• 15:59
  start
• 15:59
  experiments
• 16:00
  requiring
• 16:00
  the
• 16:00
  same
• 16:00
  data
• 16:01
  preprocessing,
• 16:02
  it
• 16:02
  runs
• 16:02
  only
• 16:02
  once
• 16:02
  —
• 16:03
  creating
• 16:04
  a
• 16:04
  network
• 16:04
  effect
• 16:04
  where
• 16:04
  parallel
• 16:05
  experimentation
• 16:06
  amortizes
• 16:06
  compute
• 16:07
  across
• 16:07
  the
• 16:07
  organization.
• 16:08
  Tangent,
• 16:08
  built
• 16:09
  on
• 16:09
  top
• 16:09
  of
• 16:09
  Tangle,
• 16:10
  implements
• 16:11
  auto-research
• 16:11
  loops
• 16:11
  that
• 16:12
  can
• 16:12
  run
• 16:12
  hundreds
• 16:12
  of
• 16:12
  experiments
• 16:13
  autonomously,
• 16:14
  optimizing
• 16:15
  toward
• 16:15
  a
• 16:15
  loss
• 16:15
  function.
• 16:16
  Parkin
• 16:17
  ran
• 16:17
  400
• 16:17
  experiments
• 16:18
  on
• 16:18
  a
• 16:18
  personal
• 16:18
  project;
• 16:19
  only
• 16:20
  one
• 16:20
  succeeded,
• 16:21
  but
• 16:21
  it
• 16:21
  found
• 16:21
  an
• 16:21
  improvement
• 16:22
  on
• 16:22
  a
• 16:22
  system
• 16:22
  he
• 16:23
  believed
• 16:23
  was
• 16:23
  fully
• 16:23
  optimized.
• 16:24
  SimGym,
• 16:25
  their
• 16:25
  customer
• 16:26
  simulation
• 16:26
  platform,
• 16:27
  took
• 16:28
  nearly
• 16:28
  a
• 16:28
  year
• 16:28
  to
• 16:28
  calibrate
• 16:29
  against
• 16:29
  real
• 16:29
  historical
• 16:30
  data
• 16:30
  —
• 16:31
  training
• 16:31
  agents
• 16:32
  on
• 16:32
  decades
• 16:32
  of
• 16:32
  merchant
• 16:33
  behavior
• 16:33
  to
• 16:33
  achieve
• 16:34
  0.7
• 16:35
  correlation
• 16:35
  with
• 16:35
  add-to-cart
• 16:36
  events.
• 16:37
  The
• 16:37
  key
• 16:37
  differentiator
• 16:38
  from
• 16:38
  generic
• 16:38
  prompt-driven
• 16:39
  simulations:
• 16:40
  without
• 16:40
  historical
• 16:41
  customer
• 16:41
  data,
• 16:42
  agents
• 16:42
  just
• 16:42
  do
• 16:42
  whatever
• 16:43
  you
• 16:43
  prompt
• 16:43
  them
• 16:43
  to
• 16:43
  do,
• 16:44
  making
• 16:44
  the
• 16:44
  results
• 16:45
  meaningless.
• 16:46
  Perhaps
• 16:46
  most
• 16:46
  surprising
• 16:47
  is
• 16:47
  Shopify's
• 16:48
  adoption
• 16:48
  of
• 16:48
  Liquid
• 16:49
  Neural
• 16:49
  Networks
• 16:49
  —
• 16:50
  a
• 16:50
  non-transformer
• 16:51
  architecture
• 16:52
  more
• 16:52
  complex
• 16:52
  than
• 16:53
  state-space
• 16:53
  models
• 16:54
  but
• 16:54
  sub-quadratic
• 16:55
  in
• 16:55
  context
• 16:55
  length.
• 16:56
  They
• 16:56
  run
• 16:56
  a
• 16:57
  300M-parameter
• 16:58
  Liquid
• 16:58
  model
• 16:58
  at
• 16:59
  30
• 16:59
  milliseconds
• 17:00
  end-to-end
• 17:00
  for
• 17:00
  search
• 17:00
  query
• 17:01
  understanding,
• 17:02
  distill
• 17:02
  larger
• 17:03
  models
• 17:03
  into
• 17:03
  Liquid
• 17:03
  for
• 17:04
  offline
• 17:04
  taxonomy
• 17:05
  classification,
• 17:06
  and
• 17:06
  report
• 17:06
  it
• 17:06
  steadily
• 17:07
  taking
• 17:07
  share
• 17:07
  from
• 17:07
  Qwen
• 17:08
  internally.
• 17:09
  Parkin's
• 17:09
  assessment:
• 17:10
  liquid
• 17:10
  models
• 17:10
  are
• 17:11
  the
• 17:11
  only
• 17:11
  non-transformer
• 17:12
  architecture
• 17:12
  he
• 17:12
  has
• 17:13
  found
• 17:13
  genuinely
• 17:13
  competitive,
• 17:15
  especially
• 17:15
  as
• 17:15
  distillation
• 17:16
  targets
• 17:16
  for
• 17:16
  low-latency
• 17:17
  and
• 17:17
  high-throughput
• 17:18
  workloads.
• 17:19
  The
• 17:19
  infrastructure
• 17:20
  financing
• 17:20
  angle
• 17:20
  gets
• 17:21
  a
• 17:21
  different
• 17:21
  treatment
• 17:21
  from
• 17:21
  Hamid
• 17:22
  Dejam,
• 17:22
  CTO
• 17:23
  of
• 17:23
  Vertical
• 17:23
  Data,
• 17:24
  who
• 17:24
  reports
• 17:25
  that
• 17:25
  roughly
• 17:25
  70%
• 17:26
  of
• 17:26
  AI
• 17:26
  deployments
• 17:27
  fail
• 17:27
  to
• 17:27
  monetize
• 17:28
  properly
• 17:28
  —
• 17:29
  not
• 17:29
  because
• 17:29
  the
• 17:29
  technology
• 17:30
  is
• 17:30
  wrong
• 17:31
  but
• 17:31
  because
• 17:31
  organizations
• 17:32
  skip
• 17:32
  the
• 17:32
  fundamental
• 17:33
  step
• 17:33
  of
• 17:33
  identifying
• 17:34
  what
• 17:34
  problem
• 17:34
  they
• 17:34
  are
• 17:34
  solving,
• 17:35
  what
• 17:35
  it
• 17:35
  costs
• 17:36
  to
• 17:36
  not
• 17:36
  solve
• 17:36
  it,
• 17:37
  and
• 17:37
  what
• 17:37
  they
• 17:37
  expect
• 17:38
  the
• 17:38
  solution
• 17:38
  to
• 17:38
  deliver.
• 17:39
  Average
• 17:40
  deal
• 17:40
  size
• 17:40
  for
• 17:40
  GPU
• 17:41
  infrastructure
• 17:41
  financing
• 17:42
  runs
• 17:42
  around
• 17:42
  $125
• 17:43
  million,
• 17:44
  with
• 17:45
  revenue-share
• 17:45
  models
• 17:46
  derisking
• 17:46
  both
• 17:46
  sides.
• 17:47
  His
• 17:48
  advice
• 17:48
  to
• 17:48
  anyone
• 17:48
  considering
• 17:49
  an
• 17:49
  AI
• 17:49
  build:
• 17:50
  understand
• 17:51
  your
• 17:51
  data
• 17:51
  portfolio
• 17:51
  before
• 17:52
  touching
• 17:52
  hardware,
• 17:53
  because
• 17:53
  "garbage
• 17:53
  in,
• 17:54
  garbage
• 17:54
  out"
• 17:55
  is
• 17:55
  not
• 17:55
  a
• 17:55
  cliche
• 17:55
  in
• 17:55
  this
• 17:56
  context
• 17:56
  —
• 17:57
  it
• 17:57
  is
• 17:57
  the
• 17:57
  dominant
• 17:57
  failure
• 17:58
  mode.