Slides: Run Frontier AI at Home — Alex Cheema, EXO Labs
Source Video
Run Frontier AI at Home — Alex Cheema, EXO Labs
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:
Why Most A Will Run Locally
ngineer
ELOPE
OCR text:
a) beOURT LES Cry a
Or ST EG h tare Cai ti irs (cell to aaa
ee ee ee renting our brains . re ae
ae oe ee ae ere Te) Tele TUsU LC Peale © el) Lg
brain
@tsarnick
i, )
of itas an exocortex,
Pra LG
OCR text:
The Hardware Lottery
Sara Hooker
Google Research, Brain Team
Abstract
Hardware, systems and algorithms research communities have historically
had different incentive structures and fluctuating motivation to engage with
each other explicitly. This historical treatment is odd given that hardware
and software have frequently determined which research ideas succeed (and
fail). This essay introduces the term hardware lottery to describe when a
research idea wins because it is suited to the available software and hard-
ware and not because the idea is superior to alternative research directions.
Examples from early computer science history illustrate how hardware lot-
teries can delay research progress by casting successful ideas as failures.
These lessons are particularly salicnt given the advent of domain spccial-
ized hardware which make it increasingly costly to stray off of the beaten
path of research ideas. This essay posits that the gains from progress in
computing are likely to become even more uneven, with certain research
directions moving into the fast-lane while progress on others is further ob-
structed.
OCR text:
Prefill stage Decode stage
KV Cache
Se Cone fteration 1 Iteration 2 Iteration 3 tteration 4 Iteration 5 Hteration 6
REreni)) (Decode) (Decode) (Decode) (Decode) (Decode) (Decode)
N 4
& What are LLMs?” <q ete 4 <a: I rea » os
OCR text:
/context
Context Usage
er claude -opus-4-5-20251101 - 63«/2e0k tokens (31?:)
, : 12 System prompt: 2.9k torxens (1.4%:)
a ») System tools: 14./K toxens (/.45)
bel Messages: 8 toxens (@.@%)
re Puen - Free space: 13/k (68.7%)
ae : i Autocorpact buffer: 45.8« toKens (22.575)
SlashCommand Tool - 4 com~™ands
L Total: 949 toxens
OCR text:
103
oO 126.3x
Yo improvement
10?
e
3°
£
:
1
= 10
a
Oo ee
=@— Local Accelerator Memory
2012 2015 2020 2025
Release Year
OCR text:
{Grok 4.20 Beta}
{4 Agents SCHED TE RAT}
alle el- ie {Lucas}
{Research and Fact Expert} {Creativity and Balance Expert}
{Grok}
{Captain / Coordinator}
Csr Uiiby {User input}
{Math/Code/Logic} i EAT SW eea:-3)
OCR text:
VedApr 8 2:35PM
EXC
Not Secure
hp/games.52415
exO
HOME
SOVOTNOO
INSTANCES
mike
442C237C
DELETE
alx-conity/G-5.1-74-8
3.E
WopgingFoce
js,s13,s14,i
READY
Lsly te chat!
s14
40x
LOAD HODEL
94nodels
34'C
GL-5.1-FP4-08
S
CH-5.1-094-0
abx-conity/6O-5.1-or.
378G8
s13
5x
TERSOR
RLX ADRA
30`c
16W
GLN-5.1-KXFP4-08
wtv
OCR text:
NotSecure
hp/ames52415
eXO
HOME
DOWNLOADS
14:48:27QUx
A3010401
4NO0ES
tell ne about ai.engineer
1X014:40:27
THTNKING
eAnalyze the Reeuesties
Dser Ii akingabeut "al.engineer
Constralnt:Repond irectly snd coocitely.
ldeotify
INSTANCES
442C237C
DELETE
alx-Cny/GLM-5.1-HFP4-08
TeSOrRUXROHANING
ing Foce
js,s13,s14,aie
SNINNTIN
ADOEL:
GLM-5.1-004-QI
TTFT1293.05TP516.9tk/s（59.1s/tok）
stv
OCR text:
@ 0 tetctn trong ok LS os tee = TS poe orer x Bis x Qe roger t .
° om
Q Kewes Googe 4 type aa £ Hi SAtMode
oO 8 oO a 8 Ss 8 oO ™ +
mee onto xO rhabpahge ixo EXO cae xO inte beyreds ier ease (487; AS orto
7 Ctoras Chrome
OCR text:
OOD reece Heownng hs oc TT 2108 1874 = 2} iso? x Boe x Seo a Bac x ¢ .
€ CO beanen 52415 et Gan;
a
Phae
Bal
sie dge tte
Pare re
rq au
re
bn ET
4 1
OCR text:
exO
HOMEDONLOADS
+NEW CNAT
OLOAD MOOEL
97sodels
g0t-055-280-0XFP4-08
）ADVANCED CPTIONS
CONVERSATIONS
gx10-de89
Ne valsd configurat lons for current
HiI
44`C
Testerdoy
28W
O5.5-358-A31-811
76.6to/s
4.7G8/119.608（4x）
TTFTS2545
Hi!
Testerdoy
n3.5-358-A31-8t
HI!
Testerdoy
03.5-358-A38-80（1
Chat Mar 31,09:48PM
0m3.5-358-A31-801t
Nor 31
Ryuichi'sMacBoo..
Sumnorise the first bit-
TEJ0
3.5-338-A38-8011
45'C
13W
What is this POF?
Mor31
Q03.5-358-A38-8t
TIFT1415s73.9tok/s
HiI
100.101.172.42tu7
Mor24
Qwen3.5-358-A38-81t
TTFT858s-81.5tok/s
Select a modetto get started.
Chat Mor 24,03:37 PM
Mor 24
gpt-0ss-20b-FP4-Q
sodel to.start chatting
43CONVERSATIONS
OCR text:
Real Local Al Benchmarks soonTM
; I 7 7
on ey 2 — 2
ore Ss
® i ae Fa | ran
on a marke re } | ; .
OCR text:
is ie
a = fp
Seep Bee li) 8 |
..
4 |
7
- ¢
a 4 z! Y Pt
4, si a _ y .*
BGS KS a P 4
on tee Ty i ;
OCR text:
© OF tee tromneg a x TZ 7t04 108 x bf vee oae2 x Shoo = Bio x + .
€ Cts arene orgipan1206 02782 er @an .;
See SEL Ey ca be - ar gan ae:
a Hecoroed peters aly appara dceyef ener
being eda.
g
8: cy =a, apes eee
i £ im dewt sent Goes | eee!
g " Garesniet
oe: ie glentitedets. 25
= _
& Abstecct
; 2 irteapeet copes mpterios iohts wt sl wiseasbo memes
- (Goraiped Copel erprabo an ties Oh erode Xd 8 ead
a peers anaes omen eres
g : mt) = i"
= iedierw tas Ci ae a5 OA er gee
. SS
Be a : peace
ig: foe DS apr ius Sy eam SSE minty same ae
Sere renee
uncer Es neh gry face eneeeecn as
i ee
Sao ee
een ne een Laine Pee SET ae
eA ‘Wat penarateadepe tan
OCR text:
exO
HOME
DONNLOADS
+NEW CHAT
15:09:38QUY
pasted-text.txt
pasted-text.txt
pasted-text.txt
Summarise these tve papers please
CONVERSATIONS
Chot Apr8,03:e9 PW
X015:09:38
-055-200-0074-Q5
Sumnorise these two pope.
02.42vt3
03:29PN
T56.3talsle
-00P4-QI
03:05PM
INSTANCES
pt-055-200-00F4-Q1
TIFT940983.1k/s
737529C8
AL313G
PipeLineRX Ring RWN
slx-community/opt-oss-2oe-RxrP4.
0:24PH
gpt-0ss-20o-FP4-Q
Mgging Fece
TTFT484990.4（ok/s
Ryvichi's Raclosk Prg
Ryaichi'sRecBek Prs→Rpichi's
Could you summorise thes.
RONNING
Kaclosk Prei 7-3})
pt-05-20-0FP4-08
03:01PM
Resdy te chot
hi!
03:01PM
DELETE
gt-0s1-200-0074-0
TT7T60881
93.1tok/s
PipeLine-LLN(CUDA)
opena1/gpt-oss-260
Huging Fece
Hi!
g10-de19
Testereoy
RUNNING
Q3.5-358-A38-811
TTFT 525a576.6tok/s
Hi!
Testerdoy
spt-0ss-200-KxFP4-08
DELETE ALL CHATS
Cence!
49CONVERSATIONS
OCR text:
Con
Stash
rae.py
app.svete.ts
Nstlee
defpipeline_avto.parattetc2s
on_Layer_Loaded（i.totat)
P
Changes
leoprefil-deco
layers[0]-PipeLineFirstLayer（Layers[e],sevice_rank,
tayers[-1]=Pipetinelastlayer(
edFles
tayers[-1],
device_rank,
world_size,
agroup,
Amens
1fisinstnce（Iner_odel_instace,Sptoss）:
inner_model_Instance.Layer_typesInner_model_instance.layer_types[
Addback6
Commitand Push...
Cor
If alayer type doesnt exist,e canset it too.
Terminal
turboquant
Ueven sarding
bench
james mconvesionit
jamesmbx conversion 8bit
ogx
Local
James
s13
Loca(3)
Loca(6)
aait resp.start（con）
function CLlentResponse.start at xelaSfe7c
Centese(ntt://afe.c//dsa/gfs-2i//i）[N]
F11e/home/Larry/ex/.ven/Ltb/python3.13/site-packages/alohttp/cLient_reqre.o,Line 539,In staet
sessape,gayloadaait protocl.res(）type:Ignore[union-attr]
Lfunction DataQueue.read at dxelat4e193adb
F1le/home/Larry/exo/.ven/Lib/python3.13/site-packages/aiohttp/streans.py,Line 763,in read
alohttp.eLient_proto.Responseandler coject at xeBabedetb77o
aait slf.miter
LNone
alohttp.ellent_sroto.Responsehanfler coject at fxeBabedeto776
alohttp.cLiant_excegtLans.SocketTiaeoutError: Tineout en readingNata fron socket
INFO
dverLapping prefill:sent 120chunks,28342 takens（server_cached111,skip0)
INFO
INFO
KVcache added（torch):20342 tokens
02:05:27.4540PH
Server refix cache:saved 28342 tokens （entries:1+2)
02:09:39.0288P
INFO
INFO
]Prefill reqest:40570 takens,start_pos,overlappingaTre
Prefitl complete:28344 tokens 1n 8936as （2277 tok/s)
2826-64-08T14:69:48.194893+00
02:09:48.1416P
INFO
Sent coched:24K,arrays for positions0-28349
|patched|IF0-Preflx cache hit:28340 tokns,1281 blocks（2groups）
02:09:48.2043PN
INFO
INFO
OverLapp1ng prefitL:sent 120chunks,46568 tokens （serve_cached20340,skip0)
02:09:57.0796P
02:89:57.0793PH
INFO
INFO
KVcache adsed(tarch）:40568 tokens
Server prefix cache:saved 40568 tokens(entrles:2+3）
02:09:57.0885PN
INFO
UTF-
OCR text:
exO
HOMEDONLOADS
+NEW CHAT
15:10:48Q0
pasted-text.txt
marise this paper please!
CONVERSATIONS
Sumndrise this poper ple.
3:20PN
EX015:10:48
t-055-200-0094-05
TAINKINE,
Chat Aprs,03:e9 PM
03:09PM
Tronsformer-“Attention IsAl1 Youeed（Vaswnl et ol.,2817)
pt-05s-200-004-0
Reploce therecurrent or convolutionel 1oyers thot
Coel
Sumnorise these two pope.
ce-to-segence nodelswith onorchitecture
03:09PM
INSTANCES
t-055-280-00N-05
Sumnorise these two pope
ECS6DASB
DELETE
03:05PM
alx-comtunity/got-ass-20e-rFP4
g1-055-20-04-09
PipeLine-X RingNG
TFT9409ns-83.1 tok/s
Mggig Fece
Ryuichi'sRaclosk Pro
Summorise this poper
Ryach'Mecek PeRyichi's
03:04PN
RUNNING
Kadosk frei [3）
pt-015-20-074-Q8
TTFT4499590.4tok/s
y teChn!
Could you summorise thes.
03:01PM
gt-053-206-0074-01
hil
3:N
TTFT68s.93.1tok/s
Hi!
Testerdoy
(3.5-58-A34-801t
cpt-0ss-200-RFP4-08
TTFT 7281.2esTP59.7 tok/s(11.4ss/tek)
DELETE ALL CHATS
Conce!
S#CONVERSATIONS
OCR text:
exO
HOMEDONLOADS
+NEK CHAT
INSTANCES
ax-coamunity/gpt-gss-28o-RxFP4
Ryuichi's MocBoo..
PipeLine·EX Ring E
ging foce
Ryuichl's Racbosk Pro
CONVERSATIONS
Ralai'oceskreReihi's
S1'C
KacBoskPre17（7)
HiI
12W
READY
3:11PM
Btody ts Lhet!
1-055-280-00F4-05
TTFT6888.8to/s
38.2G8/128G8（30%）
22097AA8
DELETE
Sumarise this poper ple
03:11PM
PipeLIneVLUN (CUDA)
oenai/gpt-oss-28
Hgging Fece
LOABDIG
87.3.tok/s
gxle-0et9
Chat Apr8,03:e9 PM
LOADING
03:09PM
pt-055-200-0FP4-08
Summorise these two pape.
OLDAD MODEL
97eodels
03:09PM
gpt-053-20b-8074-0
gx10-de89
gpt-oss-20b
13G8
Sumngrise these two pope.
03:05PM
ADVANCED OPTIONS
pt-055-20-100P4-Q8
26.6G8/119.668
TTFT 94096583.1t0k/s
Sharsing Strategy:
100.201.172.42otun7
Pipelin
Sumorise this poper
100.83.156.7ta1lscale
192.168,100.2en
Iatercosnect:
pt-05-280-074-08
：192.168.100.1e7s7
172.17.8.16ocker
TCP/1P
(Fast)
ROMA
(CUDA)
VLLH
TT7T4849m90.4tok/s
inisun Devices:
03:81PM
gpt-055-20b-R0FP4-00
pt-055-2-FN4-Q5
hil
pasted-teat.txt
03:01M
13G8
DELETE ALL CHATS
Sumnarise this
SENO
PIPELINE
VLLS
51 CONVERSATIONS
DOWNLOAOING
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.