Slides: How LLMs work for Web Devs: GPT in 600 lines of Vanilla JS - Ishan Anand
Source Video
How LLMs work for Web Devs: GPT in 600 lines of Vanilla JS - Ishan Anand
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
OCR text:
INNOVATIONPARTNER
aws
PLATINUMSPONSORS
Graphite
WWindsurf
MongoDB
daily
augment code
WorkOs
OCR text:
“er e 1. Vist anteacishests.are-al-you-naed.i in Chrome
Worldl's Fair ere eet sei-2008 room
4. Download GPT2 model in the pinned message
HOW LLMS WORK FOR
WEB DEVS
GPT in 600 lines of Vanilla JavaScript
Ishan Anand
Spreadsheets-are-all-you-need.ai
f aws
“ 7
! |
OCR text:
Yaseen coisa)
This basic recipe, and the building blocks used in it, have not fundamentally changed since the
Transformer was introduced by Goog!e Brain tn 2017, and slightly tweaked to today's left-to-right
language made!s by OpenAl in GPT-1 and GPT-2.
For example, the neural network architecture of Meta‘s Llama 2 model series only adopts a few
changes that differentiate it from the original Transformer in 2017, or the Transformer as used in
GPT-2 or GPT-3. These are the following:
4)
RUC aes MCLs eleliCte
OCR text:
+ = Thread ao
fo) .
5 Rey 6 ° on
3 igsaad ete
a “I don't even see the R's. All! soe 15302, ‘618, 19777. 198, 3504, 1134,
‘ , 19772. 198, 1018.30, 198, 138322. 198, 1100, 302. 1658, 19772, 256-44,
1100, 3504, 1134, 19772, 13007
[eo] pi
< 7 s strawberry -
ai f Strawberry JY
atrenbderry,
a — Strenaber: .
. “strawbel
oe ; "Stramberry” BY
ma '
at . o
a
: % NDUR AW Gerla P0e4d WPM vss
[e} A
eR mal ate fs oS rere no ate a ea ea re
eres PU eo FE ee Xa)
Perec CU aie oc ORL eked
ia ‘
i a Mi ft emo
OCR text:
Laurning Troasberoble Views! Models From Notaral Language Seperinion Contrastive Language-Image Pretraining (CLIP)
he att” ing a ham hen iy Gane Rem Cobra Set tere
Crebtcay ‘(unas ste Pana teu’ ev There temas Lreagee suuarer
smerwt il in 8 nga id
wee ” acumen EE em oe
a Soot rete ete cae rope tose o
x Ss SSS Ss J aotrhanl ape mapt mercies Mid'maed XA
2 SS SS
= seer eee DLT in eaeranae
< LSrcar emer yp alechay chrapelienr ey
& SSI
—_ Gped ey thc veunetea cin 1)) Conarastre pre-purwg ‘UP Create autanel Cameo: Pen thet wes
- Resim.
: STS
% tet
2 Senne arr pera. wa a 2
fio eed ates raed ted en wD ole te | t=
ean eeoes =< > ——
5 encore
So meatertes emt owen Reply. .
a oe he eg ee [us * =
2 wooo
= Se eR ee
= Ty wots sat ws a en 1 Une ter pore chet prvecton tite ,
Sake teereepnere oak ote ss eae Om safe] [s
2 mx rs | Me
ry ee ae i
5 See | rem ee Be
- lS os eos 3
Ree ee
es hearer canted pe peed mee Sp © * .
ee Ge pee eR A ke et
1 batons yal Matibeaiing Push
Aeanuny wreve wach von ames ana eg] Fagare | Seemmmary of cut approach, Wile visndard image moet jeuatly timn an wnage featnoe extractor ad a haat cLansitier bo predict
Se | scene Label. CLIP penntly anne an wenage coooder and 1 tex encoter to predoc the carrect peunmge of 4 batch of (euage. tent) tateang
TUE TA] enamples Attest uae the learaed lent encoder symtheuacs 4 sero-shot lnarat <lassaiier by embedding the anaes of descreptions of the
eenn cen tet ecin flanged dataset’s clanes.
See:
a aT
OCR text:
SoYgaop ce
SOURNL I Tn a re eee eee cree So
ei
, v
. 8
' 6
eo . s
° @
°
‘¢ ATTENTION .
fl aWws
oe ~~ |
OCR text:
Multi-head Attention
Multilayer Perceptron
OCR text:
SimilarArchitecture,DifferentTraining
GPTAssistanttrainingpipeline
AIE Dataset Stage orgygeqy Rawinternet lettllons ofwrs Pretraining -10-10k(promgesponse) sq idel Astat eponses loqunthigqulty 100k-1Mcomperisons wrien bycontracor iowquanity Nigquality Reward Modeling Reinforcement Learning Prompts -- witenbycoracos
Algorithm Language modeling Languagemodeling ict the net token uopessepee predictreeards consisterntw preferences Reinforcement Leaming thereward
GPT InitfromST
Model Basemodel SFTmodel RMmodel RLmodel
Notes ap op e 1000sofGPUs 1.100GPUS deys of tvain can deploy thsmodel 1-100GPUS das cftaning 1-100GPUs apog loap ue) deys of vaning
https://www.youtube.com/watch?v=bZQun8Y4L2A
Microsoft smolo
OCR text:
Sy 2 ar EES =: See ERE OEP Sexmtmceee Me RAE GL ne ‘So
a :
a= ae chitecture, Different Training :
®
a
"Te GPT Assistant training pipeline a
: Crone tiate] Supervsed Finetuning Leer RY ett fea eae rant e
a a ee s bet gen a ee J *:
Les a mt : : we
. e
Se
cs) ease dey a eee co eT @
| ; © ©. © 0. ©) @O27°0 ~
Woes te ae ae Ok ChatGPT =
| a Microsoft ary?
OCR text:
Similar Architecture, Different Training
GPT Assistant training pipeline
Stage Pretraining Supervised Finetuning Reward Modeling Reinforcement Learning
Dataset
Raw internet
billions of words
low quality, large quantity
Demonstrations
Ideal Assistant responses
~10-100K prompt, response
written by contractors
low quantity, high quality
Comparisons
100K-1M comparisons
written by contractors
low quantity, high quality
Prompts
10K-100K prompts
written by contractors
low quantity, high quality
Algorithm
Language modeling
predict the next token
Language modeling
predict the next token
Binary classification
predict rewards consistent w preferences
Reinforcement Learning
generate tokens that maximize the reward
Model
GPT-3
Base model
SFT model
RM model
RL model
Notes
1000s of GPUs
months of training
e.g. GPT, LLaMA, PaLM
can deploy this model
1-100 GPUs
days of training
e.g. Vicuna-13B
can deploy this model
1-100 GPUs
days of training
e.g. ChatGPT, Claude
can deploy this model
OCR text:
Spread the word on spreadsheets-are-all-you-need
Visit Spreadsheets-are-all-you-need.ai sc peceenates one
3 « Mailing list & YouTube
SPREADSHEETS ARE ALL
* Members/Patrons YOU NEED.AI
* Discount on full class, office hours, ae
etc.
Available for Al consulting
AS SEEN ON
* Training, Strategy & Implementation Drie econ - ~enotormmertean
a a Microsoft = expat)
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.