---
title: "Slides: Jack Morris: Stuffing Context is not Memory, Updating Weights is"
category: "slides"
video_id: "Jty4s9-Jb78"
sourceLabels: ["Public YouTube video frames", "Public YouTube metadata"]
---

# Slides: Jack Morris: Stuffing Context is not Memory, Updating Weights is

## Source Video
[Jack Morris: Stuffing Context is not Memory, Updating Weights is](https://www.youtube.com/watch?v=Jty4s9-Jb78)

## Relationship To World's Fair 2026
These slides are extracted from a public AI Engineer YouTube video connected to World's Fair 2026. Speaker-matched clips are supporting context unless later confirmed as exact session recordings; official livestream recordings are day-level/event-level source material.

## Related Scheduled Sessions
- No individual scheduled session mapping has been assigned yet; treat this as an event livestream deck.

## Extracted Slides
![[assets/slides/Jty4s9-Jb78/slide-001.jpg]]

OCR text:

> q |
> if A
> A
> \
> t
> ‘
> 3 €
> Vos
> 1: jr
> En 1 oy
> Bars teak Sine
> : Ree c . 2 me
> F a en 4
> ; bs Ua ~ . wa
> ot er : a x
> wii Se -a as =:
> — a %
> my . ~ i)
> mead ar - 4 mace? —_ —, ei
> mt " ra ,
> me) - Cratey =
> cm . q gs aM! &
> wey 1 3 - hha ee wt
> my 4 4 a \ a ‘Mw
> at og ’ F , 7: + —~
> my 3 pa 5s
> = = =~
> mry x . —e
> o 1 BERR Oh Nee eae Tae
> ae Pe a smees ms
> ae & See: ee —. Bet Oey ow
> fe" Le a Paar oan
> BN Bea e “
> See pean 4s
> ane Ro e ,
> a Sod Bet “
> , ss F
> Fs C ee
> Ps .

![[assets/slides/Jty4s9-Jb78/slide-002.jpg]]

OCR text:

> Vs Late] me (eX-\-0 le el Pree TCT TO) | a Pier
> did the Blue Jays win the World Series?
> (things that come after its knowledge cutoff) os
> help me optimize this kernel | wrote for AMD GPUs m
> (domain-specific, long-tail knowledge)

![[assets/slides/Jty4s9-Jb78/slide-003.jpg]]

OCR text:

> which of the shirts should | ie from amazon? a
> implement a new feature for the figma web interface in the -
> Figma monorepo. a
> Nigh Cour lam-lear- lm Ameah Acie (em
> 
> diagnose this patient given their past history.
> 
> what arguments did opposing counsel use in the Martinez
> 
> settlement negotiations?
> 
> is this question already answered from our internal wiki?

![[assets/slides/Jty4s9-Jb78/slide-004.jpg]]

OCR text:

> Three ways to teach things to LLMs: Las
> Fest
> Full Context STAG AVore IAL tS

![[assets/slides/Jty4s9-Jb78/slide-005.jpg]]

OCR text:

> Turnup volume
> LLM
> Context
> NYC(NYT）47.14Cafe
> Medical Record Prompt Output
> Isthereanything I'mmissing? evolving Consider repeating chest CT given symptoms respiratory
> FullContext·RAG·Weights

![[assets/slides/Jty4s9-Jb78/slide-006.jpg]]

OCR text:

> Context comes at a Cusr aa a
> ae
> 2
> With 1k tokens of context per user, we can ann
> output 10,000 tokens per second
> With 128k tokens of context per user, we can
> output 130 tokens per second
> “numbers computed for Llama 8B running at peak
> throughput (.e. excluding prefill) on a single H100
> Full Context ¢ RAG ¢ Weights

![[assets/slides/Jty4s9-Jb78/slide-007.jpg]]

OCR text:

> MoV Gime, COL seconu
> “rymbers computed for Llama 8B running at peak
> Papen ere mca cuv Ae 8 Ci) Re Tie (mate)
> . wf . ,
> av erry (4 tT RAG e Mersre nics Fy 2a, nee oe
> Lo. 2 ~ OO ee! ~
> _ at sow Be 7 ; _— we om -
> . _' — ott oe ; ; —
> to ae —= a
> =~ ~~. 4 = =
> mi we ~ =
> =) e r =
> =m “ ‘i | iy i: RR
> me a : mm _ _ as
> — Ta 1 ee. --T
> a . = ine mcm | = _

![[assets/slides/Jty4s9-Jb78/slide-008.jpg]]

OCR text:

> alg-lals)ce)anel-16om al \A- eo CRererrernee a
> <
> CX
> es | 2
> : ere a ~ _ ‘ es : 7 z = freer art 7 ae — Lea hanna
> . EMM IGOR. j F F FE z
> ; PTT |
> Leer ese
> Cte Ear ad
> cre
> Full Context ¢ RAG ¢ Weights

![[assets/slides/Jty4s9-Jb78/slide-009.jpg]]

OCR text:

> i co TP TTT i
> 
> TP 2 lees
> PPP REEOLEL OS. . She
> PT RPE UROOS SS lo
> PP REEUELL [oe
> PT RUS.
> PU RUEERESEE SL.
> PT PUES.
> 
> aliealies “Te TEReeneaa
> 
> anion "TET
> 
> ehenteotis “Te RPERReeeua
> 
> let PT Pu EUeEGe.
> 
> ice lU ate).<S/ AIS) twelve tokens
> 
> Full Context ¢ RAG ¢ VWeiahts

![[assets/slides/Jty4s9-Jb78/slide-010.jpg]]

OCR text:

> x See eee REY alae y ee
> Grok 4.1 Fast and Agent Tools API
> &
> 
> RASS mec Cri Bria age! eu can Re art eae a aCe ogre cos am a ote Cre
> Be a eee ee re er eer) ar eer se ee ecm ed
> 
> Grok 4.1 Fast ou cecttes Oe eae eae a re ae Ge
> 
> Fi Rae same ee ear A cet te ree Re Aa ea ae
> 
> A aToue Cs (clan a oto) Ed MeO Co SO Se
> Gea Ro Eo oC CO a aD 0 EO Cee OS ce
> NeCaeee ae Oe eae ae ee
> 
> Full Context ¢ RAG « Weights

![[assets/slides/Jty4s9-Jb78/slide-011.jpg]]

OCR text:

> Context Rot: How Increasing Input Turnupvolum
> Tokens Impacts LLM Performance
> Kelly Hong Researcher-Chroma NYC (NYT)47.14 Cafe
> JeffHuberCofounder,CEO-Chroma
> handle the 10,000th token just as reliably as the 100th. process context uniformly-that is,the model should
> However,in practice, this assumption does not hold. We
> input length changes,even on simple tasks,
> In thisreport,we evaluate 18LLMs,including the state-of-
> the-art GPT-4.1,Claude 4,Gemini 2.5,andQwen3models.
> uniformly:instead, their performance grows increasingly Our results reveal that models do not use thelr context
> unreliable asinput length grows.
> FullContext·RAG·Weights

![[assets/slides/Jty4s9-Jb78/slide-012.jpg]]

OCR text:

> CCE a i
> OP ai 2s) iene tbo nos We
> ee ie ad Sea dk hel de
> rota Cage Oe ta ode
> ee tS nS Peet a ees a
> ROT SE ote a ae Rissa
> OV stained OP a Md
> PP hated baal Da)
> anentt OF eyed meh wY
> ¢
> . om
> 7. r ;
> hee Cy a RAG e Weights rane
> are weme os
> a
> ‘ a. -
> 2 me : ~
> = aw aa 7 ~ ek as
> =) - a : Me « SS .
> 7 -, ran s s _
> ” = au = a none 8 =~
> 7 3 t-_
> ” a a —_—
> " mt — —
> _ am & sa -
> =i ‘ aaa a
> Be! —————
> ~ am pa 4 ee enn Re, /
> = oo x ea oo
> 1 ee f wre,
> om 7 j i we ee
> A __ ew BRR eos coun | Fn =
> = =) ee: eae . i A ~ oe - omen,
> wn Be —_ i a a
> eats : Pere -“—— =
> r% ce ae eee me
> \ eats alee re “i ~~ ome -
> ra) RSA a i eee :
> BSA i oe

![[assets/slides/Jty4s9-Jb78/slide-013.jpg]]

OCR text:

> Tunupvolum
> ContextRot
> Repeated Words-Performance by Input Length (Tokens) Clde 5oe4 + + On3-328 (P141 Gesinr 2.5 flash
> NYC(NYT)47.14Cafe
> Irput Length (Tokersi
> FullContext·RAG·Weights

![[assets/slides/Jty4s9-Jb78/slide-014.jpg]]

OCR text:

> 1
> = ae ee ee
> ao
> a. De. oe
> mee mu. x
> = -™ em ame _
> =~ — .
> Se ee
> an 7 -— EP pe
> WEED BB (ver Ce -
> emi am y a -
> =68 se | eS ai w=:
> amen) us | onla GUS. cores
> ED Gee a, — a e een. ~
> EE Gites vy _ a: ~_ oe me em -
> Sei .— ei Se A ay
> —=5 Gian at a > a 2 . ™
> SEE OS nee * ar 4 forename gt
> EE OB jie a : 460 4 zn | CET yee
> y —— ie oe ee ers ans ies
> — a Bre, Deen -
> == 2 / 7 a :
> | aa ar ee a or
> ee

![[assets/slides/Jty4s9-Jb78/slide-015.jpg]]

OCR text:

> Nothea
> ringanything?
> Tumupvolume
> inport chronadb
> chroma_client=chromadb.Client()
> siitchcreate_collectiontooet_or.cceate_collection
> to.avotd creating.ahercoliectionevery
> NYC(NYT) 47.14 Cafe
> collectton=chroma_cltent.get_or_create_collection(name=my_collection)
> 01Pp0R21
> upsert`
> to.votd=ddn
> collectlon.upsert(
> docunents=[
> "Thisisadocument aboutpineapple
> This is a document about oranges*
> [ZP}.‘-IP1.]-sp)
> results=collectton.query(
> query_texts[Thts ts a query document about florida],Chroma wlll etbed thts for you
> 2=511n
> manyresotts to retur
> print(results)
> FullContext·RAG·Weights

![[assets/slides/Jty4s9-Jb78/slide-016.jpg]]

OCR text:

> ar
> or
> ae eta a me.
> ho mee -™ a a
> We os on —mme _
> =. ae
> _- hf
> | . H : # _tpreee .
> i .o a co
> $ SiS. Ss eee ===
> | == aey - —
> = EB, » be =
> SEG) am _.. _— i‘ =
> | omen) im, " = =
> ; Ploy, a. - ; a :
> i “EEE Glin, y “ - — —
> Se Ss -———
> al SS ES .
> =a 7 = |
> | vo ween Se
> iy wm any! . A a oc © os oes —_—e, ;
> 7 EEE 09 | : : - —e
> t= | ee
> ~ i

![[assets/slides/Jty4s9-Jb78/slide-017.jpg]]

OCR text:

> . es Each of these points
> F ; Men cee Ue (ole lenis laimn ts a
> ; ae corresponds tc ~*~
> . ae Oy mya ro
> a SS Ps J a
> A ‘ ~
> LS . zi i. 5
> . .. aw on" .
> e y
> '
> ; re ; :
> Full Context + RAG + Voghts

![[assets/slides/Jty4s9-Jb78/slide-018.jpg]]

OCR text:

> Vector databases turns out to be invertible
> ’ a es
> a cS i-¢ 6a “% 40d ood. -0 199 > oD -e1te aaa v »
> Mage (foaled April 18. 2020) is an ger cook ce niy bo gti. oer, ae
> fh IS Ok ot ee
> American Thoroughbred racehorse SH, pest, oo) ee, ak ate
> who won the 2023 Kentucky Derby. an Que eon, bow eer rar oan [)
> O41 O46 ose. ony arte. 0 a6T, tet
> aon 12098, +6919, -0 18-8 01>. -8 ted 0 021,
> ae a F 70004, 0 O92. 0 224 8 OLS -9 EER. 0 858,
> o oa 004d. 6 OTT © OTS -2 14d. 9 277, -6 OOH. 0 OF?
> = 4 Q G48, 527d, 8 1R4 © ove. -4 eae aae, 8 91@, 6 Unt
> * ‘ 10087. +6 O82 O18 e326 $103. -0 938 8 O79. 0477
> a 6 Gt, -0 045. -6 48 -& O37 eo Oats, O74. 8 O82
> a oaae O11. O48 €orn, 2 Om aos, -9 Gen. Oo Gay
> rd @ O22 © 04) oon e212 +9 6272 o027, GO 817, 0 Ott
> 7. 5 aa P
> . Kentucky Derby, which was won by
> ‘ Mage (April 20, 2010), who is an
> ; American Thoroughbred horse and mare.
> Full Context + RAG + Vlewwhts

![[assets/slides/Jty4s9-Jb78/slide-019.jpg]]

OCR text:

> ee
> ao (6 y
> + «| Kentucky Dert was won by
> ; Mage (April 20 whois an
> : auc cro males Arlee Man):
> : 0) a
> Fo Comaxt « RAG « Weights oot OO
> i a. De mt c
> Mt nw a -wee. ws
> => -@ tw meme _
> =. — >
> | : _—— 3. om.
> ==:
> pee - | --Y Lt ===
> Ld oo : - = es
> ae ' Sn
> nan oe ] bee mae oo
> SE. — - 4 i. :
> SEED iio, _ a
> WEEE (un, y a ‘ ie a :
> | b>— 11] aren ; —. er
> =n ey a is f paw =
> | =e -_ Tt i —
> i ang . it ao | = _-
> 7 = asi) _,- . eee eae ' . ‘to © aoes — es
> i a S Sanaa Re : Re 2
> eee cece a
> ene En A bed
> ¥ —
> _ 4

![[assets/slides/Jty4s9-Jb78/slide-020.jpg]]

OCR text:

> Original text Ps ss ns | ~~
> vec2text is an algorithm for recovering a
> text from embeddings. Using vec2text, we ec
> can recover over 90% of text inputs fro
> m embeddings exactly. PI
> agi
> Hypothesis (Round 7) an
> vectext2 is an algorithm for recovering
> text from_ embed s. Using vectex
> » we cd recover text from embed
> 
> s, approx 90% 0 in.
> 
> Embedding
> Full Context * RAG + Wechts

![[assets/slides/Jty4s9-Jb78/slide-021.jpg]]

OCR text:

> Vector databases are absolute. ‘
> They should be relative. “= ~
> 6H o ory 5
> Full Context + RAG « eee

![[assets/slides/Jty4s9-Jb78/slide-022.jpg]]

OCR text:

> Vector databases are Prey ce an - .
> They should be relative. a Pa
> Search for ee ae
> customer Lee, |
> documents at a az |
> credit card a
> COMpany 7 .
> 
> Full Context + RAG ene

![[assets/slides/Jty4s9-Jb78/slide-023.jpg]]

OCR text:

> Document 1
> Purchased electronics from Amazon for
> $45.6/ on October 75, 20623.
> Card: ** 8A kee RARE TOGA VISA
> ; a Document 2
> ; = Bought household items at Target for
> $32.99 on Gctober “6, 2623.
> ak ; Oc in Ce  e a ee eS (O07 omer OAM Et Gr i0 2 8)
> Similarity according to OpenAl Embeddings
> mr -
> 332
> “ull Context + RAG + Weights

![[assets/slides/Jty4s9-Jb78/slide-024.jpg]]

OCR text:

> Embedding Context 7 bene
> 
> ccore le nn Guan
> 
> Card: **** keke **** 5676 MASTERCARD
> 
> Bougit .. =
> 
> Card: ##** *##4e #e** 5938 VISA i —
> 
> [| a oo ee
> 
> ooo _ —— =
> 
> Bougnt .. -
> 
> Car ai eh ie a ee an aor \ | =D
> 
> a c
> a oe Full Context * RAG © Weghts

## Slide-Derived Subjects To Review
Subject extraction uses video title, related session titles/descriptions, transcript context, and OCR text when available. OCR is best-effort and should be reviewed against the embedded slide images.