All  right.  Thank  you,  everybody. Thank  you  for  coming. Can  everybody  hear  me  okay  in  the  back? Is  Zoom  okay?  Yeah.  All  right. Then  we  can  get  started. As  I  said,  my  name  is  Anya, and  I'm  an  assistant  professor  in  psychology, and  my  background  is  mainly  in cognitive  neuroscience  with  recent  force  into  AI. As  it  turns  out,  there  are  lots  of similar  questions  that  have  been  asked in  both  fields  and  lots  of similar  techniques  that  we  can  use  to  study  them. And  imagine  this  powerful  device  that can  let  us  seamlessly  take  a  thought  in one  person's  mind  and transmit  it  into  another  person's  mind. Well,  we  have  this  device. It's  called  language. So  I  can  go  outside, see  the  beautiful  blue  sky  in  Atlanta  today, say,  the  sky  is  blue. And  you  will  have  a  picture of  a  blue  sky  in  Atlanta  in  your  mind. And  this  is  something  we  do  all  the  time. Of  course,  I'm  transmitting  information directly  into  your  brain  right  now, and  we  often  don't  appreciate  just  how  powerful  this  tool is  and  how  much  it  has done  for  human  culture  and  human  civilization. But  there  is  one  more  thing  that's happening  here  in  this  process. Not  only  does  the  listener, the  recipient  of  the  message, get  the  message,  but  they  are  now also  able  to  reason  about  the  speaker's  minds. You  can  ask,  well, I  guess,  Annie  believes  that  the  sky  is  blue. How  did  she  know?  Is  she  lying? What  was  her  source  of information  for  that?  What  else  does  she  know? Language  is  also  a  way  for  us  to probe  the  contents  of  other  people's  minds. And  in  fact,  that  same  idea  also  has  deep  root and  artificial  intelligence  with  the  tearing  test. The  tearing  test  from  1950  is a  setup  that  allows  a  human  observer  to interact  either  with  another  human  or with  a  computer  through  a  chat  interface. And,  you  know,  the  original  intent  is  to  figure out  whether  they're  speaking  to  a  human  or  a  computer. But  ultimately,  the  way  they're  doing  it  is probe  somebody  else's  cognition, knowledge,  reasoning  abilities  through  language,  right? So  we're  using  language  to interpret  the  underlying  thought. And  this  link  has  been  made  even  more prominent  with  the  development  of  large  language  models. So  these  models  are neural  networks  that  are  trained  on  large, large,  large  amounts  of  text, pretty  much  the  entire  Internet  these  days. And  they're  trained  with  this  very  basic  objective, the  next  word  prediction  task. So  the  fox  chase  the  blank. Maybe  a  model  would  predict  that  rabbit is  coming  next  as  a  plausible  continuation. And  then  there  is  additional  stuff  that  happens, a  fine  tune  that  allows  a  chatbot to  actually  answer  questions  and  provide  explanations. But  ultimately,  it's  a  big  prediction  engine. Train  on  language,  but  it  turns  out  that  these  models  now exhibit  a  lot  of  the  behaviors  that  in some  way  resemble  reasoning, and  they  exhibit  a  lot  of  knowledge  that  they've extracted  from  the  underlying  linguistic  training  data. And  so  as  a  result,  we  now  witness this  rapid  growth  of  a  new  discipline  that  you  might  call AI  psychology  where  cognitive  scientists  study artificial  models  as  if  they  were human  subjects  to  probe what  kind  of  cognitive  skills  and  abilities  they  have. So  you  might  have  studies  that  look  at  theory  of  mind. So  social  cognition  in  these  models, working  memory  capacity, personality  traits  and  political  biases,  and, you  know,  some  going  as  far  as  saying,  Well, maybe  we  can  actually  replace  human  participants  with these  language  models  because  they can  stimulate  human  responses  so  well. They  can't  quite,  but  people  have  pretty  far reaching  ideas  in  that  space.  And  so  that's  cool. But  there  is  a  danger  here,  right? So  there  is  potential  to start  conflating  language  and  cognition. So  just  because  these  capacities are  tightly  interconnected  and  we use  language  to  reason  about  cognition, doesn't  mean  that  they're  one  and  the  same. And  so  what  might  happen  is  you  might  fall to  a  fallacy  that someone  or  something  is  good  at  language. Therefore,  they're  good  at  thinking. And  in  fact,  it  happens  all  the  time  already. So  if  you  have  an  eloquent  speaker, maybe  a  politician  that  really draws  you  in  with  their  words,  speaks  very  persuasively, very  convincingly,  you  are more  likely  to  buy  their  argument, even  if  when  you look  at  the  content  more  closely,  it  kind  of  us. The  opposite  kind  of  fallacy  is  observing  someone  of something  being  bad  at thought  or  at  a  particular  cognitive  capacity, and  then  dismissing  all of  their  abilities,  including  language. And  so  that  has  happened  a  lot  with  language  models, especially  a  little  bit  earlier  on, people  say,  Oh,  look,  they  hallucinate, or  they  fail  at this  complicated  math  problem  that  I  gave  it. Therefore,  these  are  not  good  intelligence  models,  right? And  dismiss  all  the  things  that  they  can  do. And  in  particular,  they  are quite  remarkable  linguistic  abilities. And  so  the  main  claim  that  I'm  trying  to  make  here is  that  when  we  evaluate  cognitive  systems, we  really  should  try  and associate  how  good  they  are  at  language, and  there  are  other  traits  that  we  might  call  cognition, intelligence,  thought, and  whatever  else  you  want  to  put  under  that  umbrella. And  so  in  my  work, I  look  at  this  language  thought  relationship  in humans  and  human  brains  and  in  artificial  models. And  I  think  there  is  a  lot  of  potential  for these  two  areas  to  talk  to each  other  and  mutually  inform  one  another. In  the  talk  today,  I  will  first  give you  a  background  saying  why  we  think  we should  separate  language  from  other  kinds  of cognition  in  humans  by talking  to  you  about  the  language  network. Then  I  will  provide  a  framework  that  can  allow  us  to  talk about  the  language  cognition  relationship in  a  little  bit  more  depth, and  that's  the  notion  of  formal and  functional  linguistic  confidence. And  then  we're  going  to  look  at  three  case  studies, computer  code  comprehension,  legal  reasoning, and  generalized  world  knowledge to  see  how  some  of  these  ideas  play out  with  a  heavy  emphasis on  human  neuroscience  since  it's  a  neuron  ex  series, but  also  with  some  language  model  stuff toward  the  end  if  we  get  into  it. Okay,  so  let's  get  started. So  decades  of  research  have  shown  that language  processing  in  the  human  brain takes  place  within  its  own  dedicated  network. This  network  responds  to language  at  different  levels  of granularity,  words,  phrases,  sentences. It  responds  during  both  listening  and  reading, so  different  kinds  of language  comprehension  and  also during  speaking  and  writing, so  different  kinds  of  language  production. The  exact  location  of  that  network will  vary  from  person  to  person, so  this  is  just  a  rough  location  of  where  it  is. It's  typically  left  lateralized. And  in  addition  to  being responsive  to  all  different  kinds  of  language, this  network  is  selective  for  language. So  if  we  look  at  how  strongly  this  network  gets engaged  in  response  to  a  linguistic  stimulus, so  written  sentence,  a  spoken  sentence, you  will  see  high  activity,  high  engagement. But  if  we  measure  its  responses  in  response  to all  kinds  of  cognitive  tasks  that  are  not  linguistic, then  its  response  is  low, and  that  applies  to  math and  logical  reasoning  and  problem  solving, different  kinds  of  conceptual  knowledge,  et  cetera. This  evidence  from  brain  imaging is  also  corroborated  from studies  of  patients  with  damage  to  the  language  network. So  people  who  have  really  extensive  damage  to their  left  hemisphere  essentially took  out  the  vast  majority  of  the  language  network. They,  of  course,  have  severe  problems with  all  different  kinds of  language  comprehension  and  production. This  phenomenon  is  known  as  global  aphasia. But  if  the  damage  is primarily  limited  to  the  language  areas, they  exhibit  unimpaired  performance in  all  these  other  different  kinds  of  cognition, such  as  arithmetic  and  problem  solving  and reasoning  about  cause  and  effects and  about  other  people's  state  of  mind. So  evidence  from  both  brain  imaging,  FMRI, and  from  global  aphasia  suggests  that the  language  network  is selective  for  linguistic  inputs  and  outputs. But  that's  not  the  end  of  the  story. Of  course,  the  language  network doesn't  operate  in  isolation. So  first  of  all, you  need  to  get  the  information into  the  language  network  somehow, right?  That's  kind  of  a  given. So  audition  for  speech, vision,  for  written  language,  or  sign  language. But  then  even  after  that, we  have  this  language  network  that's  responsible for  core  linguistic  knowledge. And  then  we  need  to  make  sense out  of  the  words  that  have  been  spoken  to  us. We  need  to  integrate  with  the  rest  of  the  information that  we  know  and  maybe  to  reason  over  it. So  then  a  host  of  other  networks  gets  involved, such  as  regions  responsible for  social  knowledge  and  reasoning, general  modeling  of  the  situation, keeping  track  of  different  agents  and  the  scene, a  whole  host  of  other  regions responsible  for  domain  specific  world  knowledge, including  knowledge  we  might  acquire  through  our  senses. A  network  responsible  for  general  cognitive  tasks, so  problem  solving  and  executive  control, and  putatively  some  regions  may  be responsible  for  semantic  tasks  specifically. What's  happening  is  that  there  is  a  constant  flow  of information  between  the  language  regions  and  all  of these  other  networks  that together  enable  us  to use  language  to  do  things  in  the  world. And  what  we  can  say  is  we  can  say,  Well,  look, we  know  these  are  all  different  capacities so  how  are  they  interrelated. What  are  the  mechanisms  supporting  them  in humans  and  in  artificial  systems  that  are  using  language? Are  they  using  language  in  a  human  way? And  so  what  we  can  do,  we  can  group these  capacities  into  two  categories. One  is  formal  linguistic  competence. So  these  are  language  specific  computations taking  place  in  the  language  network  in  humans that  allow  us  to interpret  language  according  to the  rules  of  its  grammar  and  its  vocabulary. But  in  addition,  we  also  have  functional  competence, which  is  a  whole  host  of  non  language  specific  abilities. These  systems  can  operate  over  non  linguistic  inputs. If  you're  showing  people  a  movie with  the  relevant  content, the  systems  will  come  online. But  in  order  to  use  language  to  do  things  in  the  world to  interpret  the  meaning  of  what's  being  said  to  you, you  need  all  these  other  systems  as  well. But  to  give  you  an  example, an  example  of  formal  competence  is  maybe a  sentence  like  the  keys  to  the  cabinet  R  on  the  table, knowing  that  the  verb  here  is  R  and  not  is, that's  an  example  of  formal  linguistic  competence  that requires  knowing  the  rules  of  English  syntax. And  surprisingly,  this  actually  has  been  quite  easy  for language  models  starting  pretty much  as  early  as  GPT  two  and  definitely  GPT  three. And  it's a  remarkable  scientific  and  engineering  breakthrough. People  these  days  are  so  focused  on  AGI that  the  linguistic  abilities  of these  models  actually  are  taken  for  granted. But  it's  really  non  trivial  and it's  not  something  that  linguists  are expecting  being  able  to  extract  the  richness  of grammatical  rules  for  English  and many  other  languages  just  from  pure  text  data, that  is  a  really  non  trivial  achievement that  tells  us  a  lot  about language,  the  cognitive  phenomenon. On  the  functional  competence  side, we  might  have  an  example  like  this. Six  birds  were  sitting  on  a  tree, three  flew  away,  but  then  one  came  back. There  are  now  four  birds. So  four  is  a  word. So  if  you're  predicting  this  paragraph  one  at  a  time, you  will  need  to  find  the  appropriate  word  form. But  in  addition  to  that,  you  also need  to  perform  some  basic  calculations. And  so  this  basic  math  is an  example  of  functional  linguistic  competence. And  this  is  where  it  can  be challenging  for  language  models. If  you  give  them  this  example, they  will  be  okay  in  part  because they've  seen  this  many  times. But  that's  some  of  the  functional  competence  domains  in mathematical  reasoning  and  social  cognition is  where  you  see  their  performance  start  to  break. And  even  when  they  do  succeed, it  might  be  for  the  wrong  reasons. So  their  success  might rely  on  memorizing  the  answer  or  figuring  out  how  to solve  this  narrow  class of  problems  without  generalizing  to a  broader  class  and  just  kind  of  figuring out  some  heuristics  from  their  training  data. And  so  one  thing  that the  EI  community  has realized  is  that  just  scaling  these  models up  and  just  leveraging  next  word  prediction  to its  core  is  no  longer enough  to  improve  functional  competence. And  so  these  models  are  now  being  fine  tuned  on specific  tasks  and  sometimes augmented  with  additional  modules  beyond  just  language, kind  of  like  what  we  have  in  the  human  brain. And  so  that  shift  is  already  apparent  in  AI, even  though  still  many  people  try to  have  a  model  that  kind  of  does  it  all  in  one  go. And  so  one  example  that  might  be  helpful  is, this  is  the  language  network. This  is  another  network  that  we call  the  multiple  demand  network  or  the  front  of ital  network  that  is  responsible for  general  reasoning  and  problem  solving. So  in  general,  the  harder the  problem  you're  trying  to  solve, the  more  your  brain  hurts, let's  see,  the  more  activity  we  will  see  in  that  network. And  so  in  response  to,  say, a  bird  problem,  what's  going  to happen  is  that  you're  going  to  read  the  description  here. It's  going  to  flow  through your  visual  corteses  into  the  language  network. The  language  network  will  extract the  meaning  behind  the  problem, but  it  cannot  do  math. And  so  what's  going  to  happen  is  that  it  might  pass the  relevant  information  to  the  multiple  demand  network, which  will  then  perform  the  calculation  and  return the  answer  for  which  you  can  then  verbalize. So  we  see  an  interaction  between these  networks  using  this  simple  example, and  we  think  that  it  happens  for the  vast  majority  of  linguistic  behaviors. Any  questions  at  this  point? Yeah.  Does  it  matter  what  is  the  source  of  the  data? Because,  for  example,  a  video  is way  more  data  compared  to  a  text, for  example,  or  something  like  that. How  does  the  brain  deal  with  that? So  the  language  network  is selected  for  linguistic  information, and  we'll  see  going forward  how  much  we  can  push  that  selectivity, right,  whether  we  have  kind  of  exceptions  to  it. But  in  general,  information  from  videos, it  flows  primarily  through visual  cortes  right  from  low  level  to  higher  level, and,  you  know, if  you  think  that  it's  the  amount of  information  that  matters, then  you  would  expect  kind  of  the  opposite. You  would  expect  the  brain  region to  respond  more  to  videos than  to  language  because  language  is  more  bare  bones. But  in  the  language  network,  we  see a  flipped  result  where  it  prefers language  other  non  linguistic  kinds  of  inference. Yeah. I'm  not  so  familiar  with  this  because  you  mentioned  maybe the  upstream  logic  processing  area can  send  information  to the  language  processing  area  in  humans. Do  you  think  larger  language  model should  be  combined  with other  like  reasoning  model  to  work  together? No.  I.  Maybe  the  reasoning  model  produce some  input  and  inserts  the  answer  into  the  module. Yeah,  that's  great  because  I'm actually  not  coming  back  to  this  point, but  in  the  paper  where  we discuss  formal  and  functional  competence, we  actually  say,  Well,  look, if  the  human  brain  is  compartmentalized, where  you  have  language  processing  and then  cognitive  processing  in  other  parts  of  the  brain, then  if  we  want  a  model  that's  human  like, we  probably  want  that  same  kind  of  specialization. And  there  are  two  ways  of  doing  it. One  is  to  build  in  those  modules  in  advance. We  call  it  architectural  modularity. So  you  have  a  language  module, a  reasoning  module,  you  combine  them  together. Or  you  might  have  an  architecture  that's end  to  end  that  is  trained  holistically, but  something  in  that  architecture  permits  it  to develop  specialized  modules  within over  the  course  of  training. And  that's  a  hard  problem  of  how  do  we  do  that? What  are  the  kind  of  training  objectives  we  would  need. But  in  general,  I  think  it's possible  to  have  a  system  that  is  modularized in  the  end  but  doesn't  start  out  with  predefined  modules by  human  engineers.  Great. Okay,  so  with  that,  let's  dive  in. So  in  this  first  study,  we  asked, okay,  you  know,  how specialized  is  the  language  network  really,  right? Is  it  really  only  responsive  to language  or  maybe  it's  something about  language  that  is  shared  with  other  inputs. And  so  here  we  looked  at  computer  code  comprehension as  a  compositional  symbolic  system that  resembles  language  in  many  ways. And,  in  fact,  you  know,  it's they're  called  programming  languages. So  clearly,  there  is  a  parallel  there. And  so  if  you  think  about  what  happens  if you're  processing  a  little  problem  in  say, words  versus  code, you  have  to  start  with  interpreting individual  words  or  tokens. You  then  have  to  combine  them  into sentences  or  statements  using  syntactic  rules. You  then  have  to  link  individual  sentences  or statements  into  a  text  or  a  program, and  you  then  can  extract  the  underlying  meaning, and  they  perform  reasoning  over  it  to  get  at  the  answer. So  there  are  lots  of  similarities, and  so  we  also  know  that  computer  code  does  not  have an  evolutionary  specialized  module inside  the  human  brains  to  reason  definability. So  it  would  make  sense  for the  language  network  to  take  over. And  so  that  is  a  really  strong  test of  selectivity  for  language  versus  other  things. And  so  we  can  also apply  our  formal  functional  distinction  here. So  the  process  of  getting  to the  input  to  an  overall  meaning  of  a  program, we  can  call  formal  programming  competence. But  then  we  also  have  to  do  reasoning  over  it,  right? We  have  to  interpret  it,  predict  the  output, and  so  that  we  can  call functional  programming  competence. And  we  are  focusing  mainly  on formal  programming  competence  in  this  study. So  what  we  did  is  we  showed  people  little  snippets  of code  problems  in  Python  and matched  sentence  problems  and asked  them  to  predict  the  output. What  is  the  answer  going  to  be  here? We  also  had  a  whole  other  programming  language  we  looked at  that  was  as  different  as  Python  as  we could  think  of  and  it's  Scratch  Junior. It's  a  graphical  programming  language  for  kids. It  doesn't  have  any  text  in  it. So  it's  a  series  of  blocks  that  you combine  to  give  instructions  to  a  cartoon  character. And  then  this  was  also  matched with  corresponding  sentence  problems. So  we  look  first  at  responses within  the  language  network  to  Python  code. So  here  we  have  a  basic  sentence  reading  condition contrasted  with  the  non  word  reading  condition, so  nobody  could  have  predicted,  blah,  blah, blah,  versus  things  that look  like  language  but  aren't  linguistic. So  that's  our  general  localizer  that  we  use  to  identify the  language  regions  and  just make  sure  that  they  do  respond to  meaningful  sentences  more  than  this  control  condition. And  then  these  were  our  two  target  conditions, sentence  problems  and  code  problems. And  so  first,  we  see  our  expected  response that  this  network  responds  to  sentences, but  not  to  matched  controlled  non  words. It  also  responds  to  sentence  problems, which  makes  sense  because  again,  these  are  sentences. And  so  the  critical  condition  here, of  course,  is  our  code  problems. Is  it  going  to  be  as  strong  as  responses  to  sentences? And  it's  not.  It  is  a  little  bit more  than  non  word  reading, but  it  is  significantly  lower than  responses  to  our  sentence  problems. Suggesting  that  there  is  a  little  bit  of  a  response  in the  language  network  to  Python  code, but  it's  not  nearly  at the  level  of  English  sentences.  Yeah. How  is  the  language  delivered? Is  it  auditory  or  visual?  It's  visual  presentation. You.  But  we  see  the  same  thing  for  auditory  language, so  we  would  expect  to  sim. But  here  it's  matched,  so  it's visual  for  both  language  and  code.  Okay. Yeah,  well,  yeah,  so  visual  for  code seems  very  different  than  auditory  code,  at  least  for  me. Well,  yeah,  auditory  code  delivery  does  not  work. It's  like  a  known  problem. Which  is  very  interesting  because  there  are  a  lot  of insights  that  can  be  gleaned  from psycholinguistics  in  order  to  design maybe  programming  languages  that  will be  easier  to  comprehend  auditorily, but  that's  a  whole  other  tangent. Yeah.  I  never  trained  on  language  sentences,  type  code. Ah?  So  this  is  the  brain. This  is  the  brain. This  is  the  language  network. Yeah,  yeah.  So  most  of  the  stuff  I'm  talking  about for  the  next  20  minutes  is  brain  responses. Yeah,  yeah.  Yeah.  But  the  same  question. How  familiar  are  the  subjects  with  coding? Relative  to  how  familiar  are they  with  English  with  language? Yeah.  There's  a  diversity, even  among  native  speakers  of  how. So  this  was  one  of the  first  studies  on  computer  code  in  the  brain, and  so  we  picked  a  pretty  lenient  threshold. The  people  needed  to  be  a  proficient  with  Python, but  we  didn't  impose  any  strict  criteria  in  terms  of how  long  they  needed  to  know  it, I  think  our  lower  bar  is  one  year. And  so  that's  the  minimum. But  of  course,  you  know,  it's  usually  more. We've  collected  that  information. We've  tried  to  look  at  individual  differences, but  the  sample  size  is  simply  too  small, so  I  wouldn't  make  anything  out  of  it. Um,  I  will  say  that  for  foreign  languages, you  would  see  engagement  of  the  language  network. Even  if  you  are  learning  a  new  natural  language, you  would  see  much  more  activity than  you  would  see  to  Python. It's  true  that  the  level  of  experience  is  different, but  oftentimes  it  might  actually  go  the other  way  where  if  you're  proficient  and  fluent, you  activate  the  language  regions  less. It's  easy  for  you  essentially. Okay.  So  now,  in  contrast  with  Oh, and  so  now  we  can  also  look  at Scratch  juniors  or  second  programming  language. So  again,  responses  to  sentence  is  high, responses  to  non  words  are  low, and  we're  looking  at  our  critical  code  condition. And  here  we  see  very  low  responses. So  if  the  programming  language doesn't  look  anything  like  English, like  Python  does,  then the  language  network  really  doesn't  care. And  so  we  can  conclude  that  it responds  a  little  bit  to  Python and  not  at  all  the  Scratche. And  then  we  can  also  look  at  the  multiple  demand  network, which  we  know  is  responsible  for  general  problem  solving. And  so  in  contrast, it  has  really  strong  responses  to  code  problems. So  first  of  all,  it responds  to  sentence  problems  as  well, which  makes  sense  because  they  are  performing  reasoning. It's  math  problems  in  half  of  them, string  manipulation  in  the  other  half. But  then  we  see  this  additional response  and  code  problems, suggesting  that  code  interpretation,  right, is  a  form  of  programming  competence, also  loads  on  the  multiple  demand  network. And  so  we  see  this  for  both  Python and  Scratch  junior,  right, where  in  Scratch  Junior, the  Gulf  here  is  even  larger  numerically. Yeah.  C  activation  of  those  two  numbers? Not  in  this  study.  So  the  question  of how  we  look  at  the  co  activation. Let  me  wrap  up  this  part, and  then  we'll  come  back  to  this. So  to  sum  up  these  results, we  see  a  little  bit  of  response  in  the  language  network to  fight  in  code  and  not  at  all  for  Scratch  junior. In  contrast,  the  multiple  demand  network  exhibits robust  responses  for  code  comprehension specifically  to  both  of  these  languages, suggesting  that  that's  the  main  network  responsible for  extracting  meaning  from  computer  code. And  then  in  some  follow  up  work, we  actually  did  some  decoding  using  code  embeddings  from a  code  neural  network  model  showing  that  both of  these  brain  networks  have  information about  the  identity  of the  computer  program  that's  being  read. Information  is  there  to  some  extent. But  then  another  research  group did  some  follow  up  analysis, suggesting  that  probably  what's  happening  is  this  kind  of false  recognition  situation for  the  language  network  where we  see  an  initial  activation  and  then  a  drop. So  it's  like,  Oh,  is  it  language?  Oh,  no,  it's  not. So  that's  our  best  hypothesis  of  what's  happening. And  then,  of  course,  with  Scratch  junior, you  don't  see  this  false  activation because  it  looks  nothing  like  language. So  the  false  alarm  didn't  happen  to  begin  with. This  doesn't  fully  answer  your  question, but  it  kind  of  goes  in the  same  direction  of  what  is  actually  going  out  there. Alright,  so  we  can  conclude  the  form  of programming  competence  relies  primarily on  the  multiple  demand  network. So  we  still  have maintained  this  selectivity  profile for  the  language  network. Okay,  the  next  study  is a  more  fun  collaboration  that  we had  with  a  legal  reasoning  researcher, where  we  can  see  some  of these  ideas  essentially  play  out  in a  more  complex  setting  for a  pretty  complex  cognitive  ability. So  in  general,  we  live  in  a  society that  requires  a  lot  of governance  of  social  relations  between  individuals. So  laws  exist  in  order  to  give  people  an  idea of  how  they  should  behave  in  this  complex  environment. And  so  the  key  principle  here  is  taking a  law  and  applying  it  to  the  world around  you  to  know  what  you  should  be  doing, what  you  should  not  be  doing. And  the  process  of  interpreting the  law  in  relation  to your  environment  is  called  legal  reasoning. So  that's  the  subject  of  our  study  here. And  the  lead  author  on  this  study  is  Eric  Martinez, who  has  a  law  degree  which  wasn't  enough  for  him, so  then  he  got  a  PhD  in  cognitive  science, and  he's  really  interested  in  legal  reasoning and  interpreting  legal  texts from  a  cognitive  science  perspective. The  legal  literature  makes  a  lot  of  claims about  how  things  should  be  or  how  things  are, but  we  are  still  in  the  earlier  stages  of  having empirical  evidence  for  what  is  actually going  on  in  the  human  mind as  we  are  interpreting  legal  documents. And  so  we  can  use the  formal  functional  framework  here  as  well. So  we  can  talk  about  formal  legal  competence, and  that's  the  process  of  extracting meaning  from  legal  texts, which  are  often  written  in a  very  particular  way  we  might  call  legalese, and  so  you  might  require specialized  training  in  order to  actually  understand  these  documents. But  then  there  is  the  process of  functional  legal  competence, and  that's  the  legal  reasoning  itself. Applying  legal  knowledge  to  specific  cases  to  figure out  how  a  law  or a  contract  might  bear  on  the  situation  at  hand. And  on  the  functional  competence  side, there  have  been  a  lot  of  debates  in the  legal  literature  about  how things  are,  how  things  should  be. And  one  side  of  that  debate  called  formalism. Applying  legal  rules  is  a  formological  process, and  so  actually  one  inspiration  for  the  study was  the  feedback  we  got  on our  computer  code  study  from  people  on  the  Internet, Oh,  well,  it  looks  like computer  code  is  just  like  reading  a  contract,  right? You're  doing  this  very  formal, structured  process  of  interpreting very  formally  written  language. And  so  I  bet  the  responses to  contracts  in  the  brain would  be  kind  of  the  same  as  the  code. So  many  people  have  that  kind  of  intuition. But  on  the  other  side  is  the  realism  School,  which  said, Well,  look,  applying  legal  rules is  not  just  about  formal  reasoning. In  fact,  it  depends  on the  context  and  the  social  factors  around  it. What  were  the  consequences?  What  was  the  intent? What  were  the  harms  involved? And  so  it's  not  just  about  pure  logic, it's  about  contextualized  reasoning about  the  social  implications  of  the  rule  violation. And  so,  can  take this  framework  and  actually start  making  neural  predictions  about  it. On  the  formal  legal  competence  side, we  have  our  two  potential  players of  the  language  network  and  the  multiple  demand  network. And  the  question  is,  which  of  these  networks  takes this  extra  toll  from  interpreting  legal  documents? On  the  functional  competence  side,  under  formalism, we  would  expect  the  multiple  demand  network  to  be involved  in  actually  reasoning  about  the  text. And  under  the  realism, we  would  also  expect  involvement from  social  cognition  regions, such  as  the  theory  of  mind  networks. So  we  have  some  interesting  brain  predictions to  work  with  here  under both  the  formal  and  functional  side. And  so  we  have  people  read  two  versions  of  contracts, the  Legal  Ese  versus  the  plain. The  content  is  exactly  the  same, but  LegalEs  has  center  embedded  clauses, passive  voice  sentences,  stop  in  all  caps, which  is  supposed  to  make  things  easier  to  read, but  actually  kind  of  doesn't  help. And  a  lot  of  low  frequency  words and  phrases  from  Latin,  et  cetera. And  then  the  plain  but  they  can  rewrite  all  of that  in  plain  English  while  preserving  the  meaning. And  so  we  looked  at  two  groups  of  participants. We  had  lawyers  and  non  lawyers  with a  comparable  degree  of  graduate  education, and  we  looked  at  responses  in the  multiple  demand  network  to  legal  As  and  plane. And  we  saw  that  there  is  indeed  increased  activation  in both  groups  to  legalese  versus  plane, suggesting  that  indeed,  it  takes an  extra  cognitive  toll  to process  legalese  contracts  relative  to  plane, even  if  you  have  a  lot  of  experience. So  even  lawyers  show  this second  and  then  we can  also  look  at  responses  in  the  language  network. And  here  we  actually  see a  difference  in  the  lawyer  group, but  not  in  the  non  lawyer  group. And  essentially  suggesting  that  specialization for  reading  legal  text  that  does  emerge manifests  in  the  language  network. So  we  see  more  involvement  here. So  we  see  the  multiple  demand  network engaged  across  both  groups  of  people, and  the  language  network  engage  for  lawyers specifically  after  they  read  the  contract, we  have  people  read  scenarios  about a  real  situation  where  that  contract  was  in  play. And  so  say  there  are two  roommates  with  a  rental  agreement, and  now  there  is  a  conflict, and  our  participants  had  to  answer different  questions  about  the  resulting  situation and  the  applicability  of  the  contract. And  so  we  have  them  grouped  into  a  logic  category, a  moral  category,  and  different  kinds  of  legal  questions. And  so  we  first  look  at responses  in  the  multiple  demand  networks. So  first  of  all,  we  see strong  responses  to  a  logic  question, which  was  just  like,  a  math  related  problem and  nothing  in  response  to  a  moral  question. So  that's  good  news  for  us.  It  means our  positive  control  and  negative  control are  working  as  expected. And  then  the  responses  to  the  legal  questions were  something  somewhere  in  between. So  first  of  all,  we  see  that  there  is heterogeneity  in between  different  kinds  of  legal  questions, suggesting  as  maybe  is  expected, that  legal  reasoning  is  not  this  one  monolithic  thing. There  are  many  sub  processes  going  on. And  we  see  some  engagement, but  it's  not  nearly  as  strong  as  to  large  questions. And  it's  matched  with  difficulties, so  difficulty  is  not  the  explaining  factor  here. And  then  we  see  the  same  pattern  of responses  in  our  non  lawyer  group. In  the  theory  of  mind  network, we  see  responses  which  the  zero  basin  here  is  shifted, but  essentially  we  see  much  more  engagement in  response  to  moral  questions  and legal  questions  relative  to the  logic  questions  in  both  groups, in  non  lawyers,  we  see a  little  bit  more  for  response  to  moral  questions, which  some  of  our  lawyers  have reported  treating  moral  questions as  legal,  so  maybe  that's  why. But  essentially,  we  see  that  in  neither  case, do  we  see  this  clear  patterning of  legal  reasoning  questions  with  pure  logic? So  we  see  that  both  networks  are involved  during  legal  reasoning for  lawyers  and  non  lawyers, which  suggests  that  formal  legal  competence recruits  the  multiple  demand  network. In  lawyers,  there's  also  additional  recruitment of  the  language  network. And  functional  legal  competence  recruits both  the  multiple  demand  network  and the  theory  of  mind  network  in  both  groups, which  essentially  provides  support  for  social  realism, suggesting  that  it's  not  just  pure  logic  that  there is  social  reasoning  involved  here, and  we  get  similar  results  from  looking  at the  correlational  patterns  in  the  brain  across these  different  kinds  of  questions  and  tasks. I  will  say,  however,  that  this  is  work  in  preparation. So  take  this  with  a  little  bit  of  a  grain  of  salt. We'll  be  finalizing  these  results  in the  coming  weeks.  Alright. And  so  now  we  are moving  on  to  generalized  world  knowledge, which  is  perhaps  the  domain  I'm  most  excited  about. It's  also  the  messiest  domain where  we  really  start  to  see difficulties  trying  to  separate out  language  from  knowledge  and  language  from  reasoning, just  because  they're  intertwined  so  much. And  so  I  think  this  is a  very  exciting  area  to  be  working  towards. So  we'll  talk  about  whatever  we  have  time  for  now. But  then,  you  know,  I  hope  that  a  lot  of my  future  work  will  be  in  this  domain  as  well. So  language  has  a  lot  of  information  about  the  world. So  models  trained  on  language  will  learn  a  lot  of  stuff just  by  memorizing  different  words or  tracking  occurrence  patterns. So  factual  information  like Paris  is  the  capital  of  France, so  birds  lay  eggs. That  can  be  learned  directly  from  the  text. But  what  is  more  is  that  by  tracking co  occurrences  between  words, we  can  learn  a  lot  of  the  distributional  information. So  I'm  a  model  or a  human  would  encounter sentences  like  the  sky  is  blue  today, the  sky  was  pitch  black,  the  sky  is  pink, and  keeping  track  of  that  information provides  a  distribution  over  possible  colors  of  the  sky, what  colors  are  possible  for  the  sky, which  colors  are  more  frequent  versus  less  frequent. And  it's  mainly  this  distributional  knowledge that  we're  interested  in  here  today. And  so  one  specific  kind of  this  knowledge  is  generalized  event  knowledge. So  it's  storage  of  common  templates of  templates  of  common  events  observed  in  the  world. So  the  fox  chase  the  rabbit  is  a  good  plausible  event. The  rabbit  chase  the  fox,  less  plausible. The  fox  cheese  the  teacher. Also  not  very  plausible,  but  possible,  right? So  these  don't  violate any  fundamental  rules  of  the  world  around  us. And  so  because  all  of  this  information  is  in  principle, could  be  available  in  your  linguistic  input, and  lots  of  information  like  that  we learn  from  language  instead  of  direct  observation, we  ask,  does  generalized  event  knowledge rely  on  the  language  network? And  so  here,  we're  first  referring to  the  language  network  in  the  brain. And  so  we  show  people  sentences like  the  Fox  is  chasing  the  rabbit versus  the  Rabbit  is  chasing  the  fox. We  also  show  them  line  drawings  of  the  same  kind  of events  and  had  them  do  one  of the  two  tasks  in  different  blocks. So  one  is  a  semantic  task. So  is  the  sentence  or  picture  plausible  or  implausible, and  the  other  one  is  a  perceptual  control  task. So  the  stims  is  just  moving  on  the  screen, left  or  right,  very  slowly. You  have  to  report  which  way  it's  moving. And  so  our  hypothesis  were,  well, the  language  network  is not  engaged  in  pictorial  even  semantics, it  cares  about  language,  not  meaning. So  it  wouldn't  care  about  pictures that  would  only  care  about  sentences. Alternatively,  the  language  network cares  about  meaning,  and  therefore, it  would  be  engaged  in  pictorial  even  semantics, maybe  just  as  much  as  in  sentence  semantics. And  so  we'll  look  at  responses  in  the  language  network. We  see  high  responses  to sentences  during  the  semantic  task. That  makes  sense.  Sentences  extractive  meaning. We  see  low  responses  to both  stimulus  types  for  perceptual  tasks. It's  a  hard  task.  They're  tracking  it  on  the  screen. They're  probably  not  even  thinking  about  the  meaning, so  we  see  low  responses. And  so  the  critical  condition  here is  pictures  during  the  semantic  task. Is  it  going  to  be  high  as  for  sentences  semantic  or  low? And  we  see  this  result  with  something  in  between. And  so  we  didn't  get a  conclusive  resolution  to  our  hypothesis  here. So  we  have  to  say,  like, the  language  network  is  somewhat engaged  in  pictorial  and  semantics. There  is  some  response, but  it's  not  as  strong  as  four  sentences. And  we  replicated this  result  across  multiple  experiments. We  see  a  similar  picture  emerge in  naturalistic  stimulus  comprehension. So  listening  to  stories  versus  watching  movies, we  see  a  similar  kind  of  drop  off, but  no  engagement  for  the  music  stimul. We  don't  see  the  same  kind  of  response  for single  objects  categorization,  though. So  asking,  Oh,  does  this  animal  live  in water  or  can  this  object  be  found  in  the  kitchen? The  response  in  green  here  is  very  low. So  maybe  it's  something  special  about events  that  is  causing  this  response  and  the  language. But  also  we  have  this  really  valuable  piece of  causal  evidence  from  two  participants  with global  aphasia  who  have really  severe  language  deficits alongside  12  age  match  controls  that  we  had. So  for  sentence  picture  matching  task, asking  which  of  the  line  drawings  goes  with  a  sentence, they  perform  at  chance  or  near  chance. So  these  are  two  participants  with  aphasia,  sa  controls. So  that  makes  sense. They  cannot  process  language properly.  They're  bad  at  this. In  contrast,  for  picture  plausibility  judgments, they  perform  at  the  level  or close  to  the  level  of  healthy  controls, suggesting  a  dissociation  between their  linguistic  skills  and  their  event  knowledge. And  so  this  causal  evidence helps  us  draw  a  novell  conclusion  that the  language  network  is  recruited  but not  required  for  pictorial  Evan  semantics. But  that  doesn't  answer  the  question. Why  is  it  recruited? But,  uh,  actually,  before  I  get  to  that. So,  one  other  question  you  might  ask. Okay,  the  language  network  does not  does  have  a  preference, but  other  brain  regions  that  respond  equally strongly  to  verbal  and  pictorial  semantics. And  the  answer  is  yes, there  is  a  set  of  brain  regions  that  we're  finding that  actually  seem  to  care  about  semantic  tasks, regardless  of  whether  it  comes in  form  of  sentences  or  pictures, and  they  are  separate  from both  the  language  regions and  the  multiple  demand  regions. And  so  happy  to  talk  about  that  more. This  is  a  work  that  we  are  also preparing  to  put  out  currently. So  stay  tuned  for  semantic  processing  in  the  brain. Okay,  and  the  second  question  that lingers  from  that  conclusion  is  okay. You  know,  why  do  we  see this  response  in  the  language  network  to  pictures? What  is  it  doing  here? And  one  explanation,  well, it's  just  task  irrelevant  activation. You  see  this,  you  know,  a  picture,  Fox  chasing  rabbit. You  have  activity  for  Fox,  for  rabbit, for  chasing,  but  it's  not  helping  you  do  the  task. An  alternative  explanation  is  that,  well, the  information  that  you've  extracted  from language  actually  is  helping you  do  the  task  by  tracking those  distributional  properties  of  the  input, you  can  reason  about  event  plausibility. It's  not  the  only  route  which  is why  people  with  aphasia  can  still  do  it, but  it's  a  possible  route  for  solving  the  task. And  so  we  don't  have  definitive a  definitive  answer  as to  which  of  these  explanation  is  correct. But  what  we  can  do  is  we  can  start following  up  on  the  second  explanation,  asking, Okay,  what  about  language  models trained  on  purely  distributional  properties  of  the  input? Can  they  do  this?  And  so  that  will be  an  existence  proof  that  it  is possible  to  solve  this  task with  distribution  of  language  information  alone. And  so  we  looked  at  language  models along  with  my  colleague  here,  Karina  Cow. What  we  did  is  we  had  minimal  sentence  pairs, so  we  gave  models, sentences  like  the  Fox  chase the  rabbit  versus  the  rabbit  chase,  the  fox. And  we  looked  at how  likely  a  model  judged  a  particular  sentence  to  be. So  you  can  see  just  what  is the  probability  of  a  model outputting  sentence  one  versus  sentence  two. We  expect  that  probability  to  be  higher  for the  plausible  sentence  versus  the  implausible  sentence. And  if  a  model  got  a  particular  pair,  correct, we  give  it  an  accuracy  score  of  one, if  not  zero  and  then  average over  a  bunch  of  minimal  pairs. And  so  we  did  it  for  two  kinds  of  sentences. One  is  animate  inanimate  interactions. So  the  teacher  bought  the  laptop versus  the  laptop  bought  the  teacher, which  essentially  means  that the  second  sentence  is  impossible. It  violates  the  animacy  restriction  on  the  verb. And  so  here  we  see  that  the  language  models in  green  here  actually perform  close  to  the  ceiling  and  close to  humans  who  are  at  ceiling  here. And  even  very  basic our  control  language  models  before  neural  network  era, they  perform  fairly  well, as  well,  although  not  as  well. We  also  looked  at  animate  animate  interactions, so  the  fox  chase  the  rabbit versus  the  Rabbit  chase  the  fox, and  there  we  actually  see  a  performance  gap  where language  models  of  the  recent  generation are  not  as  good  as  humans. And  then  our  control  language  models, which  are  much  more  basic, they  actually  are  a  chance or  only  a  little  bit  about  chance. So  there  is  a  meaningful  difference here  that's  being  captured. And  we  call  that  the  gap between  the  impossible  and  the  unlikely. There  might  be  many  explanations  for  this  gap. If  we  go  back  to  the  formal  functional  distinction, we  might  say,  well,  maybe the  selection  restrictions  on  the  verb, knowing  that  bot  requires  an  animate  subject, that  actually  might  belong to  formal  linguistic  competence, and  that's  why  models  learn it  alongside  with  general  grammatical  rules. Versus  knowing  that  the  rabbit  chased  the  fox  here, it  doesn't  violate  any  linguistic  rules  per  se. It  is  more  about the  graded  event  knowledge  that  we  have  about  the  world, and  so  that  maybe  we  can  put  that into  functional  linguistic  competence. So  we  can  start  using  this  framework  to tease  apart  different  processes, different  aspects  of  general  event  knowledge. And  I  should  start  wrapping  up  here. But  to  finish  up  the  event  section, does  generalized  event  knowledge rely  on  the  language  network  and  the  brain? The  language  network  is  recruited  but  not  required. And  does  generalized  event  knowledge naturally  arise  in  large  language  models? Yes,  for  impossible  events  relying  on selection  and  restrictions  and less  so  for  graded  event  knowledge. That's  definitely  above  chance,  but  it's  not  as  good. Then  we  have  recently  expanded our  knowledge  framework  to a  whole  host  of  different  domains. We  call  it  elements  of  world  knowledge  or  EOC. We've  looked  at  a  lot  of  them. Specifically,  we'll  leverage  here  contextual  knowledge. So  the  Fox  cheese,  the  rabbit. There  you  can  solve  that  problem  just  by  memorizing. If  you  just  have  those  sentences  in  your  training  corpus, the  model  will  be  able  to  tease  them  apart. But  here,  we  have  sentences  like  Ben  and Alok  are  friends  versus  Ben  and  Alok  are  enemies, and  um  the  sentence  Ben  and  Alok of  friends  is  plausible  in  the  context  of  Ben  likes  Ag, but  implausible  in  the  context  of  Ben  Hits  Log. And  then  the  enemies  one  is  the  flip, so  the  context  determines the  plausibility  of  the  target  sentence. And  so  we  tested  a  lot of  different  language  models  from  Hugging  face, the  open  language  model  repository on  a  lot  of  these  domains, and  we  see  that um  our  best  performing  domain  social  interactions, the  later  generation  language  models actually  achieve  good  accuracy. They  are  close  to  one.  Was  on  spatial  relations, we  see  models  performing  much  worse  overall. This  bar  here  is  the  human  ceiling, so  we  see  that  humans  here  are  not quite  perfect  for  some  interesting  reasons, actually,  but  there  is a  huge  gap  between  humans  and  models  here, which  is  the  main  thing  we  care  about. And  we  can  do  the  same  thing  for  a  bunch  of  different. The  main  takeaway,  really,  is  that social  interactions  and  social  properties are  easiest  for  models, and  knowledge  about the  physical  worlds  of  physical  relations  and spatial  relations  are  domains which  models  struggle  with  the  most. And  overall,  it's  interesting  that  you  see  in  general, all  models  kind  of  cluster  together. Of  course,  there  are  differences  between  them, but  overall,  domains  that  are  hard, are  hard  for  all  models, domins  that  are  easier  are  easier  for  all  models. And  so  we  show  here  that basic  world  knowledge  in  language  models  varies drastically  by  domain  with  social  knowledge  being easier  than  physical  and  spatial  knowledge. And  the  goal  here  is  really  to  create  a  framework. So  we  have  ways  of  generating new  sentences  with  those  same  templates. There  are  ways  to  expand  it  to  potentially  new  domains. So  if  there's  any  interest  in this  kind  of  world  knowledge  framework, we're  happy  to  talk  about  expanding  and  building  upon  it. But  to  conclude  the  talk,  formal  linguistic  competence is  knowledge  of  linguistic  rules  and  patterns. Functional  linguistic  competence  are non  language  specific  skills required  for  real  life  language  use. This  distinction  is  grounded  in  human  neuroscience. It  helps  clarify  the  discourse around  language  and  thought, and  it  can  be  applied  to  domains  beyond language  like  programming  and  legal  reasoning. And  finally,  it's  applicable  to  both  humans  and large  language  models  and  AI  systems  in  general. With  that,  there  are  two review  papers  that  I  want  to  highlight. One,  about  the  language  network  with my  PhD  advisor  at  Fedorenko  and  Tamaegev and  our  paper  about Associate  language  and  thought  in large  language  models  with  my  colleague, KamaHowd,  I  was  a  senior  author, and  a  few  other  amazing  authors  as  well. With  that,  I'd  like  to  thank my  amazing  collaborators  and  all  of  you for  your  attention. Yeah. For  the  wild  knowledge  question, I  wonder  if  you're  able  to differentiate  between  sentences  that  are plausible  and  likely  to  be  expressed  versus sentences  that  are  plausible that  you  would  never  see  them  on  the  Internet. Yeah,  and  there  is  a  little  bit  of  work  on  that. So  the  question  was  sentences  that  are impossible  versus  possible  versus,  like,  common,  right? Frequent  or  describe  events  that  are  impossible, possible,  but  rare,  possible  by  frequent. And  so  language  models, at  least,  you  know,  up  to,  like,  a  year  ago, but  probably  still  are  good  at  distinguishing common  and  uncommon  events  and  bad distinguishing  uncommon  versus  impossible  events, which  is  kind  of  what  you  would  expect  because  there  are statistical  models  that  extract regularities  from  the  input. There  is  another  paper  came  out  recently that  tries  to  bring that  distinction  one  step  further  and  say, there  are  also  impossible  events and  inconceivable  events. So  impossible  events  is  like  I  don't  know, like  an  elephant  flying  up  in the  sky  versus  inconceivable  is  like, colorless  green  ideas  furiously. So  you  can't  even  imagine  what  it  would  look  like. And  so  that  gradient  might actually  be  even  more  fine  grain  that  refused  originally. But  it  is  a  very  useful  distinction because  that's  exactly  where  we  can  break past  the  statistical  regularities that  govern  language  models  and, you  know,  to  some  extent  govern  humans  too. Yeah.  I'm  really  interested  in  the  fact  that the  large  language  models were  most  deficient  in  physical  and  spatial. Least  efficient.  Mr.  Amo. Yeah,  yeah.  Yes.  Have  you  looked  at activation  in  the  pin  language  network  in humans  in  terms  of physical  physicality  of  the  sentences or  its  relation  to  movement? Yeah,  so  kind  of  in,  fitting with  the  overall  formal  functional  competence  framework, when  we  look  at  how  the  brain responds  to  say  complex  narratives, we  actually  see  that  words  loading  heavily  on,  say, social  content  engage  social  cognition  regions, words  like  family,  relationship, love  engage  so  social  regions, whereas  words  that  have to  do  with  colors  or  different  textures  or  weight, they  engage  a  different  set  of  regions  which,  you  know, we  would  expect  is responsible  for  processing physical  knowledge  about  the  world. So  it's  not  the  language  network where  we  see  those  differences,  it's  outside. So,  you  know,  these  words  they  seem  to recruit  relevant  domain  specific  regions. Yeah.  I'm  going  to  go  back  to  the  Evok  slides  with  Theon. With  the  results?  Yeah,  the  neural  network  performances. What  are  material  dynamics? Because  that  one,  it  seems  to  do  pretty  well. Right  next  to  the  social  property. Yeah,  material  dynamics  is  like  I  have  some  cloth. I  watched  it  rip  versus  I  have  I  don't  know, milk,  I  watched  it  rip. And  so  these  are  the  kind of  properties  we're  looking  at  here. And  with  this  example,  you,  of  course, can  see  that  a  lot  of  that  information can  be  recovered  from  simple  word  co  occurrences. So  just  looking  at  cloth  and rip  together  versus  looking  at,  you  know, milk  and  rip  together, that  can  give  away  a  lot  of  that  information. And  so  that's  probably  why  it's  easier. The  sentences  in  this  domain  are  very  short,  right? So  here,  there  are  a  bunch  of  factors  other  than the  domain  itself  that  will  influence  LLM  performance. We  have  looked  at  them. I  used  to  have  a  little  disclaimer  down  here. So  these  are  things  that  you  might, you  know,  some  of  that  have  to do  with  formal  confidence  maybe. They  explain  some  of  the  variants, but  not  all  of  the  variants. The  domain  still  contributes  on  top  of those  differences  in  sentence  length and  basic  word  coccurrences. Yeah.  So  Python  experts might  have  activated  that  network  less. And  then  when  looking  at  the  legal  data, Lawyers  increase  the  activation versus  non  legal  grad  students? Within  legal,  do  you  think that  there  was  a  scaling  based  on  level  of expertise  that  non  expert  to  expert  lawyers, but  then  within  lawyers,  there's  variability  based  on how  proficient  they  are  with  the  gel  system,  et  cetera. Mm  hmm  leading  to  less  activation  from  hiring  experts, similar  to  Python  experts. Mm  hmm.  Let  me  give  you an  example  from  a different  domain  that  I  think  will  help. So,  work  from  Evslab by  Sana  Malek  Marda who  is  defending  in  a  week  or  something. She's  done  a  lot  of  work on  bilingualism  and  multilingualism. And  so  there,  what  she's  finding  is that  you  of  polyglots, people  who  speak  five  or  more  languages. They  have  activation  in the  language  network  that  essentially  goes  like  this. The  better  you  know  a  language, the  more  activity  you  see  in  the  language  network, except  when  it's  your  native  language  when  it  drops. So  there  is  a  meaningful  difference between  a  native  language  and  other  ones, but  kind  of  the  better  you  understand  the  language, the  more  activity  you  will  see  in  the  language  network. And  the  multiple  demand  network,  essentially, the  better  you  know  the  language, the  less  you  need  to  activate the  multiple  demand  network. And,  if  you  experience  speaking  a  foreign  language, and  if  you  don't  know  it  very  well, your  brain  gets  tired. That's  the  multiple  demand  network  getting  engaged. But  if  you're  fluent  in  a  language, you  will  not  experience  that, and,  you  know,  multiple demand  network  is  no  longer  active. So  we  see  these  kinds  of  trade  offs. They're  not  necessarily  linear,  right? And  so  from  the  lawyer  data  alone, it's  a  little  bit  hard  to  say what  exactly  is  going  on  here. But  we  might  expect some  potentially  nonlinear  curve  where we  see  first  language  network  getting  specialized. So  we  do  see  this  increased  activity,  but  maybe  later  on, it  actually  drops  again  because  it  becomes easy  and  so  you  don't  need  to  engage  it  as  much. So  hard  to  say,  but I  think  something  like  that  is  possible. Yeah.  For  the  first  example, does  it  matter  for  the  language  used? Is  it  more  interpretable  or  not? For  example,  Python  is  relatively  readable  compared  to C  or  something  that  is more  machine  native  compared  to  human  native. Is  there  any  work  that  you  know  along? Yeah,  so  the  work  I  was  referring  to is  from  Marina  BednGroup  Johns  Hopkins. And  so  they  are  the  ones  who put  forth  this  hypothesis  that the  language  network  has this  false  alarm  response  to  Python, which  leads  to  activate  a  little  bit  of  not  at  all. So  if  that's  true,  then  we  would  expect  some  kind of  gradient  where  the  more  English  like  the  language  is, the  more  activity  we  will  see  in  the  language  network. And  so,  you  know,  how  do  you  quantify English  likeness  that's  where, you  know,  it  might  get  a  little  bit tricky,  but  that's  roughly. So  I  would  expect  essentially all  programming  languages  to fall  somewhere  in  between  Python and  Scratch  junior,  right? The  most  language  English,  and  the  least  English,  like. Yeah.  So  regarding  the  Fox  Chase  the  rabbit  example, so  what  kind  of  prompts  are  you  using  Because  I feel  like  the  right  kind  of  prompt, you  can  get  the  right  kind  of  answers. It's  the  prompt  cate  account  for  this  kind of  So  for  those  for  that  initial  study, we  are  not  using prompting  these  are  pre  trained  models before  they  ever  fine  tuned. And  what  we're  looking  is  we're  looking  at the  log  probabilities  of  the  output,  right. So  literally  just  using this  model  for  its  original  purpose, which  is  text  generation  and  looking  what  is the  probability  score  for generating  one  sentence  versus  the  other. That  tends  to  give  you  more  information  than prompting  because  essentially  in  prompting, you  don't  know  if  a  model  is failing  because  it  didn't  understand the  prompt  versus  it  doesn't  have  the  relevant  knowledge. And  so  there  are  studies  showing that  for  log  probabilities, a  model  might  succeed  even  when  a  prompt  would  fail. We  actually  just  had  a  follow  up  paper  accepted  to a  workshop  like  a  day  or  two  ago  where we  did  log  probabilities  versus prompting  on  fine  tune  and  non  fine  tuned  models, and  we  show  that  log  probabilities  are  still  better. So  instruction  tuning  doesn't necessarily  help  with  prompt  based  responses. So  this  is  actually potentially  a  more  sensitive  way at  getting  at  this  underlying  knowledge. It's  12  18. Oh,  so  maybe  that  should  be  the  last  question. Question  from  Benjamin  Fanelli. I  just  wanted  to  ask,  how  much  of human  intelligence  coded  language? Can  you  think  of  any  other  media  that might  have  capturable  intelligence? Question  is  how  much  intelligence is  captured  in  language. And  so,  as  I  said, lots  of  information  is  captured  in  language. The  vast  majority  of  the  information  we  have about  the  world  is  learned  indirectly  through  language, through  other  people,  right? All  of  the  I  don't  know, physics  and  chemistry  and  astronomy,  right? It's  not  us  observing  the  phenomena. It's  us  learning  about  them  from textbooks  and  from  one  another, same  with  social  knowledge,  right, gossiping  about  one  another  to  watch  out  for. That's  different  from  the  core  question I'm  trying  to  target,  which  is, are  the  mechanisms  responsible  for language  processing  the  same  as the  mechanisms  in  the  brain or  in  an  artificial  system  used  for reasoning  or  is  general  world  knowledge stored  in  the  same  circuits  as  linguistic  knowledge. And  so  that's  where  we  see  a  lot  of  dissociation. So  even  if  information  comes  to us  linguistically,  which  a  lot  of  it  does, it  might  then  proceed  to  other  brain  networks  and systems  for  this  more targeted  domain  specific  processing. And  a  good  last  question,  please. Relax. I