Slides: Frontier results, on device - RL Nabors, Arize
Source Video
Frontier results, on device - RL Nabors, Arize
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:
a
oo)
nN
ew
ee
laa
a MDN web docs
a i ve
S
oy ~
=
i
Ss \w sd
142)
eI a
2
eo)
P
a a
s ,
“vA \
S
OCR text:
o | 4) 4=
You have agents. i
We can test them. | 3
OCR text:
; THE COST OF
eee gl ot
: in
OCR text:
8
=
5
S>
EI Tetra Lae RNA MU KML RAL CAS
i y D 99
S
: quality of experience...
3
3 — eee ene en ee ee ; 7 |
: -
OCR text:
Pe)
8
2. C
a
on
Do
Lap)
Pd
fe)
fs)
Ps
mag
=
a
S
a
D Couldn't connect to Claude
3
pees
a
9
= Refresh
a
ras ;
a i
© ei -
a i \
x
= 4
OCR text:
5
2 Chief Product Officers should not
low
Ey OTTER ACG CLOT OM MO CELL AY
S tokens with the democratization of
= SKU EG MR SOKO Ee
a — Gortee:, March 2026
S cf
y
OCR text:
2)
8 TASK-SPECIFIC MODELS CHEAT SHEET
S - Isacamera pointing at something?
a Use vision models like MobileNet, YOLO, MediaPipe
g
a
: ~
OCR text:
_ .
fs
2
S
ea =6(OLMs
¢
©
See eee) ee
Ss
i
1
8 a s
= -
OCR text:
5 *
&
5
S
Ea = Smaller) Language Models (SLMs)
=
A Smaller versions of LLMs containing several million
i to several billion parameters (LLMs may have
a hundreds of billions or even a trillion “6
OCR text:
a
fe}
fa
2
a .
fan)
fai)
of QI
S> ’
2) ‘i
re 4 a
ae teary ,
z a a n va
S i ia
P
i '
3
ee ie
Ss X
a
a
© s
= F
: >< B
: i |
a .
is)
3
—— i
OCR text:
cS)
ay ; :
o nergy consumption comparison
an
199)
D
a
3}
of
a LLMs
=
S
a SEES
199)
3
®
= Task-specific Models
a :
5 7 |
Ss 0 PE Ete ih) yf b |
a proportional energy consumed ) ~
OCR text:
7 ow oa a
xa ; ; Fi
Q x
aD o todo
a How hkel Successfully activated extension
om © Mow ae gis ttmaticoth yeas Padecho ccation?
oD o memory *
2 eT eer eee Successfully activated extension
= Poontig able cs hiehee for iy
Q
Ss
S)
72)
as
_
i
4
-
=>
ia)
Ss
a
Q
a
4
wn
my
FS 5
om 7 —
S
vn
3 Pe
= o
OCR text:
a ; ; a rn
im) HER SSUES cc isd Fea ce] OO Ce are peo CORGee Carers IE Cn POS a ieee aa a ae a a
hoe lie A Cor cea Soa CS CO
o TOG ferm ra rae ACen Md PO aed Caen Oe RT COCET ESET EG? FT EPO mann RreneL|
— ae eae Coe
on re eret a ec tana
ee Oe ee ee ee
a 3
® Pas] CAP OU Ss te) On COM Un DT ShSTEseaSan dd ROCESS DIT Ebel Cbraerteo hai NDS a eee a a
= COcnDER ES TOES LTO FIPS OP EU ES POMC OTE FTES ROEO LGC Os RIES CR ELENIR SECrea [RS ;
Pn ee ee fe eeu] BE geeg es ye gah or
S> fot ee CRC Ses ES BC Chan
° Evolutionary Context
“n No directesvidence Unike with ines urs va bere We oan
; ; ee eee ee ee ee ee oe
os Set oe rings fii stomach Gonnents cr tragkwsas s Phere s the . Lo .
a : ee eee eee
ne Debavonal tosssbevidienee fares tocaiter Ce ar yen wh noe tae ee
~~ cae . . Coa ee ee
aa serene SHOE SOTTED CSS Crane Um OY U2 C(010 GOD) SSR IG ED RS aT
Se a ee 4 eae Be oe
om LORS UAE BTS of Palco itelouty Sec UIs HSS Gr AE Ton OODLE,
A a Q ri Leer Tra eta
3 ee ee “a
3 SOSH eG ter) Deere tenet an SLC GTeRS TT CRs CORECOnNG (pO eR eh
a Me ldar at Et: a ce cy
as Pe ce Tc ee
D a oe
es F
a
aa
Q a
S ; no )
1) je H
~ A
n .
ny \. »
e) iY
a} z
ae A
OCR text:
A . 7 eo: Ca
je) SIS aC a
-
o Se eet OD 0 OER FG COd ET CEO PE SSUES be tak Cred De LC
cr . . " Ong ey Ode
ae is moderately plausible but far from certain probabh. ar ah ececg
R TOTES Sn hes Gd CCT Cs CORT SL OD Cs FAST OPT SES CT MS OLTO RTE SLOT CO eee ee eer eee 2g
ma OL PrEbearGrce Danas rt
Q RP atl B-CLL } sie ec eS oe i Sst y
Lo Hore samy reasons ey a ee
cod _— . . Coat este ta 0R Fat dot Sm a
“ Evidence suggesting they mighthave had en
aa echolocation:
ray Concus:on
< ra Vici cospan cet] STAD EA 10 C se Miata he ee SSO ECTS ESE LSS
: ca ee en) a se Oe
[ee ete eed sured che cap structure: sinha fe
ia a a ET Parans J as are roe Le i eoN rome eer TTT Te ee
Sy bE EO KGS ATO SCOUTS De CA Tet a Sees OD Cet Ca ME ere arsa te teen 1c Ae ee ee ee ee ne ee ee ee
es Col bs REC Cre SOLED ITS TEC SS Eee] ALL CaS en rn ss
=
ae Ale ODOT (GS CMS STE ES OTST race ChE Pa TNT TE ae
oi 2s
= PaSCare ECOL ORE LOSS TSE DEOL MTG DI RETO SS (01D CR NECORSR SECTS
cag Sareea aot sr esc Ce erela ET RTO OTa a SOCOM SCRTTEO Ol Or OCICS Lee TE
= Mt To cestae tes it |
Ent CR eSOO LOS Leet Ghia tated PPR ame rea
Q 5
fo Sm Grelah siren Mca KO)EUL C01) MEO ALES RET te tates ms ROD er La SLAG =
: = F
= rd
a) = n
5 \
~
= ew | *
so A
3
OCR text:
Rachel-LeeNabors(they/them)nearestnabors.com
RIGHT-SIZINGAI
OCR text:
PS)
Q
im)
D
~~ Settings x
an
ta)
D
fe] —____
<4 he yt 2 ra ae _ 8s 7 eee a
oe) os
ot ~
i =
4 Current Cont-guration ce =
_ oy 7 a
S
a :
ry
Q
pias
tae
va)
=
Pa
I Golden Dataset |
a
= dé .
a) 1
Bea
OCR text:
Es)
a)
im)
a
on
a Age wm A Se
Golden dataset at a glance
Q
)
oe fr)
28 14 Po Wi 100%
S
oo
Po
fa)
Rp a: ee
a esky Hh moda
Q af rh cay na iy 7 fore]
D
A
;
S a.
: eal
Q g
=
OCR text:
p)
Q
8. Measures of Success
2
aa
fa)
a Dimension What it asks How it's measured
Q
Ss
S JSON validity Does the output parse? Try JSON . parse, count successes
—
Do £:N] tokens point to messages that
a Reference structural validity (re : ] pot g Regex extract refs, checkeachN € [1,
< actually exist? message_count)
PN
= Does the summary stay faithful to the thread, LLM-as-judge — Claude scores each
a Factual consistency es She Summary slay "at wucge — an
3 or invent claims? summary against its source
pee
= Word count vs context-specific limits (8-12
ig) Le compliance Does it stay in the target word band?
io ngth complian OES EE SEE NN MHTAEREL OKO for tist view, 19-46 for modal)
7
i p50 latency Typical TT summary Median across the eval set
s
S
o p95 latency Worst-case wait 95th percentile across the
=
2)
=] £ach evaluator returns a per-example score. Average across the 28-example golden dataset, with 3 repetitrons per example __
OCR text:
gQ Q serve Sib den @ 10k Tare Set O1 Ot Ct Ee Ga 2
Ps)
2 arize «+. ee
D
cm
ia) ,
oD
P=
9 ty
Sa
is} _
a Arize Phoenix =
=~ ea
= is
4 Trace the js
a =
Oo i i
3 Exponential :
= fe
a ~
g
5
= Get started Self-Host |
5 ea
: cy
a aad!
= X f)
=
OCR text:
" .
a
5
S
ea = Capability eval
a P J
A Asks, “What can this agent do well?” They should
i start at a low pass rate, targeting tasks the agent
a struggles with and giving teams a hill to clim
= —Anthropic, Demystifying Evals for Al Age
OCR text:
a as we ee 1 G7 cs
FS) ;
iS Cees ate e rs
a goider-summar.es ; , k=
a ;
® rere 2 q
P yeepgee om ata te a PA ; ;
9 ny e H ry FA
y , , A ,
fo}
a
a
a
ot
lo
; —_ — _ a
199)
3 ) Z : - |
os F : — — - : ,
a 7 — ae Pa 5 rn
any
n Os ant — ames ao ae cy a
=
a a
5 4
pod
wy 4
Cy
© ss
=} :
OCR text:
re
Q
im)
2
ro
p
D
rd
fo]
oa
Lo)
a
Ss “Small and Good Enough” Model
S$
d SAGE Model
3 ode
©
=
fe
2 o
in i
S
OCR text:
Rachel-LeeNabors(they/them)nearestnabors.com
p50latency(ms)-lowerisbetter
ClaudeSonnet(celling)
OQwen2.51.5B
Qwen31.7B
Gemma4E2B
Llama3.23B（recommended)
OCR text:
i be Peon ean ere ets
a eye rat Cana
roy f 5
Q aan SS Dare er eee eS
oO
—
food
) rer
ere Ste ete an ree Pe Sars PROT ECT Sal cece crs
fo) , , , s
S)
a) e .
Sono
nS ,
:
& an | | :
we '
al
ou ,
is
Sa
—
D
Q
e Dir etn ee a a Coram oe Le ey fe ang re nL aes ae Pad
“A CeCe ae on Pert ee ane Der en a eugee cone sere we
oS. ae a Ory fe Re ae ee ey 5 nA etee we aa! A
s) : . . &
a N
a fs
1S}
= f ee o
3 7 .
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.