Slides: The Small Model Infrastructure Nobody Built (So We Did) — Filip Makraduli, Superlinked
Source Video
The Small Model Infrastructure Nobody Built (So We Did) — Filip Makraduli, Superlinked
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:
€3 Braintrust €} WorkOS OpenAl
OCR text:
The Small-Model
: Infrastructure
| BARES Supertined
[eee]
B25
=
OCR text:
x * ®
What Actually Makes Embedding
or’ Model Inference Fast?
ao
1] Google DeepMind
OCR text:
ba * ,
Mae
bd a
oO Pe
| | |
. AlEngineer
EUROPE
S95
OCR text:
Traming & re tening area pose DCMT RL stenet a gern
va Eva‘vation & bencnmark:ng GPU uti! zat on & schedu .ng
* ~ Appied Al researc Rosting & autosca’ ng
# . Bem rae feat ien Crployment & oprrat ons
bod *
ad os *
Best way to learn?
Join the team building it.
: | Al Engineer |
eS) EUROPE
=
OCR text:
.
Superlinked ©) github.com/superlinked/sie
rane
Ps * VC - backed Al infrastructure company building the model stack for Al search and document
re x processing
ve
Works with your favorite taols neers
ey Pa Pree’ os Pera te} coy paren ae
Cree nears oes ee ere vena Sone Un Tn eee earns
eer et Beene eee es prereset an
Cate iat tne ie Ld Cha Se Ct we aac d oe eon a ee Lael pee OE i rec
teat eer eaed ene Learns] coe ae ar ar Ree eee Tre el
eee te Cm Me eerie seer ie at et ere de he Pe A ae ee
Fo ett loses Pe oes re
l | Al Engineer |
= aUelaa
re = oF
OCR text:
PU emulate uh
Performance degrades as contest fille up
lel Ret CECE LOE CDC
aa _ / vere Tee ah ee Mae Lis
Oa a TM, . 2a,
* a N atl eeelre)
rs N
4 : N\
: \- . Es
: . “iy . "Why not just Opus 4.62"
. = Ty, se!
: | Al Engineer |
eS See) aa
a hati
OCR text:
Context management
Taxonomy classification
1OK categories Shopify product catalogue text + images
Pan 3
Pu *
aCe
Sg a Tool calling
ars Product
bauer sO) a ae
Switching model type = changing one parameter, not
rebuilding infra
] Engineering the future of Al
OCR text:
Model Performance V5 Size
we _ ‘ :
te ey °
Y
one ”
Pane d
ne nom &
a
* ad ‘ e °
a in .
2
.
Pole Muitel See titan Paraet ona)
| Al Engineer |
Eng EUROPE
aoe
OCR text:
ec , A wndiee 7 No universal engine
Pel
bd bd
bf bal An agentic workHow
ad ra *
Re-implementation
: | Al Engineer |
= ay:
d
OCR text:
cee Qwen2 ore a1 as
Norm Pre-norm RMSNorm
é Ca i‘ Q/KIV Rate Separate (GQA)
na ae
* rar i Lexth ated) Coa
Attention GGA (fewer KV)
Output
Allon packed sequences with = zero padding waste
Engineering the future of Al
ES) >
OCR text:
We open sourced both the yin (model inference) and the
4, a
nn Superlinked [m] ar art « [m]
® re
* a . SIE — Superlinked Inference Engine is allan "
i s i
tro :
©) github.com/superlinked/sie 7
] Google DeepMind
ES i a
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.