Slides: WF2026: Autoresearch & Keynotes ft. Anthropic, Google DeepMind, Amazon AGI, Sonar, Arena, Recursive
Source Video
WF2026: Autoresearch & Keynotes ft. Anthropic, Google DeepMind, Amazon AGI, Sonar, Arena, Recursive
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:
AI Engineer
World's Fair
STARTING SOON
STREAM STARTS JULY 1, 2026 9:00AM PT
OCR text:
AI Engineer
World's Fair
Livestream
July 1, 2026
EVENT STARTS IN 04:54
DESIGN ENGINEERING
View Full Schedule
ai.engineer/worldsfair/2026/schedule
OCR text:
AI Engineers
OCR text:
| WAT dehe) nt | Worlds |
World” a mid
_ - ; AlEngincer
! VVfo) gto Koa
el 4 my :
OCR text:
PAO] a eG mele Basal
Norld's Fair — OpenAl
7 | ALEngineer ————
ole Rc Res irld's Fair
ELLE Saree rexsas ——_—_
OCR text:
— Pa? os me i si
7 TN a ee il ary
, ae Me nih
P SS, ay
a a =O ~ ‘ 7
eae Te : ~ ee ~ ~ NY ,
— ~— ae ae ~ S ied am
oe - eet ~ Lave , O™ ~~ ‘
7 elat Tae aL vote “A an xs > .
ae crn ~ ta _
j J ZA) _ Cowal a ea’ x = _—
ee ae Bina . oa ee oa a => 7 ~ i
I aie ~ a oS reo
ry anal re — sh tee a ad " aI
ee “TPS a Ry ; a io ec es te _ a i
e — aa se “ve “sh at
Ce ery cy Saleen” [ee a , \ ee en
et [oo YS LS DI
sy ° rae 2
at [Wertter] - ee es a a
eed eae ee een aed oe Oy
ene 2 [worst rar] Wants tir . , omy r 7 eee oT or ae
Fair Fanaoos ‘Wort Fair’ eed MON mm a a an
— oo “ Bm wm eS im 2°
ered pS re) Se ee aes _ 7
Pathe
Basle a ed 0 ee es
.< |
OCR text:
AI Engineer
World's Fair
PRESENTED BY
Microsoft
Models are grown,
not designed.
7 / 40
In the Land of AI Agents, the Verifiers Are King
Tariq Shaukat / Chief Executive Officer
Sonar
OCR text:
PNT Sheena
World's Fair :
: CAPABILITY OVERHANG i 4.
a Microsoft Claude gets smarter in ay
spiky ways 4
asralr ke
roar sy AYA" algo a phage
setae Akz Inthe Land of Al Agents, the Verifiers Are King
VADOG vane oe G sonar —-
OCR text:
System Prompt Design
Small system prompt,
few tools, lots of
examples
Large system prompt,
lots of examples, many
tools
Smaller system prompt,
tool search, no examples
14 / 40
OCR text:
Tern oe |
a It's closer to a biology, than a physics.
a Microsoft
Eee hah ne eee ate enc
Ope
Woi ; . mS
Engineering the future of Al
nd ae
OCR text:
World's Fair |
Tradeoffs are not 5
real a
woras rdalr
i= na! Bye
Field Guide to Fable
Thariq Shihipar ANTHROPAC 7
OCR text:
ia 5 a ps
‘i ‘at | Valere ea we
aa 7 7 Pa
a WG Akariz
a | ee ee
7 ca ’ ' -~ 7 cigdie. FE
i bs ! = rl erais hae
OCR text:
renigewer Coding agents are quickly getting a lot better
World's Fair
Time horizon of software tasks Al can complete
at 50% accuracy
ee
: ele ee eS)
Woric ir a La wh a #: 7 f a A
f eereaten y wis . , ; 7 . 7 ok
Treks Ven Cm UCR ice UL Ca)
ATC i Tariq Shaukat [ Real Renata
OCR text:
Eye od .
j ; The models are getting smarter, but
World's Fair ; hae ;
they are still producing problematic code
Gemini 3.1 Claude Opus 4.7 GPT-5.5
: cen alte la) Thinking Medium
a Microsoft a a
158.4 Wane 151.4
614 803 504
var! rach 68
sere ar elute
an TOE PCO UGA rant Gal Car)
ok World Tariq Shaukat @ sonar ~--
OCR text:
Se oa A temporary speed boost + a persistent quality decline
World's Fair
a Microsoft
Temporary velocity Increase in security, Increase in code
spike that disappears maintainability, and complexity
within 3 months reliability issues
OTe aN ean
oe ctoralels
Vey east ; ;
in the Land of Al Agents, the Verifiers Are King
rerice NOUR RS e @ sonar —-
OCR text:
GUIDE
Agents need enterprise context and constraints
so that they are ‘verification aware’
Repository-aware project specific context...
• Architectural awareness
• Semantic navigation
...and constraints...
• Dependency guidance
• Coding standards and guardrails
• Intended architecture
...provided dynamically and/or embedded in models
In the Land of AI Agents, the Verifiers Are King
Tariq Shaukat / Chief Executive Officer
Sonar
OCR text:
IN tented Wadd oe A
World's Fair |
Verification should be
and
E r goo
mo In the Land of Al Agents, the Verifiers Are King
‘ae Tariq Shaukat ; [ ME rat] eta - 7
OCR text:
ai Verified agentic remediation in the Cl flow
cee ees and in the background
CODE MAINTENANCE LOOP
Seconds / Dera aaa
| PInT nea i aan a
beta
a re aan oee
6 aaa Nee a Ee eae ea ee ere ee
’s Fai ; ;
In the Land of Al Agents, the Verifiers Are King
ir @).4 Tariq Shaukat G sonar —-
OCR text:
ae Master the critical verification loops to bring the Agent
World's Fair Centric Development Cycle to life
Ere wes en 7
F soa aK» aay Raa eLRSTEC Ia %) aw
a Microsoft aa ooo
© giar ROD iaeneresrs
: ; ‘ o > praalec@lriety <a pean LA One] sald
ae Wor .
Gi sonar
ate (
In the Land of A! Agents, the Verifiers Are King
2.6) Wor Tariq Shaukat G sonar —--
OCR text:
AI Engineer
World's Fair
Clicking was
the easy part.
Amazon AGI Lab
Perception Agents
Antje Barth / Member of Technical Staff Amazon AGI Lab
OCR text:
Perceive
Act
Plan
Perception Agents
Antje Barth / Member of Technical Staff Amazon AGI Lab
OCR text:
World's Fair |
Why perception agents?
a Willa cexxeyne They close the gap on computer use.
They don’t need an API or back-end access.
You show what you mean, don’t describe it.
faves
er | ME treo ae Xe. Uy
tows act Met dec ra Amazon AGI Lab
OCR text:
World'sFair AIEngineer Annotation
SomePodcast
etSoeLsy
ehetheritec re aftar a contet sndoe setesh New episo fons abouf inteligencg = what'a is, how to Iukd a, and
sFair aws Neural Network What Even Is Intelligence?(Asking for a Friend's
ize spp AEn PerceptionAgents
sFair ba AntjeBarth/MemberofTechnical StaffAmazonAGl Lab
OCR text:
World'sFair AlEngineer VisualVerification
L-1000-P-RACT Cicdds akbuon eet e
pols/20260626-1933412 porLmd taskfliow-2:Complete a task
taskflow-3:Deletea task Verificatin Rests fo chcp-
eried tked mnt,oalountu
taskfiow-4:Initial page state
Vetedhead
cormp
AuditFindings
d'sFair av Artifacts l-verfcatIofos/-4foncaros vis/desi-human-re desigpecicon -veriicatin/specs/-5cmed ory fles82nles) sl-weririctim/sesso/arets-s
orld AlEngi PerceptionAgents
d'sFair AntjeBarth/MemberofTechnical StaffAmazonAGl Lab
OCR text:
World'sFair AlEngineer PlatfConts|144|] 1eectes|17117] ReswltI ALL PASS kcesbiuty1u
No visuel fallures detected.Allrwles match the design specification.
tcap-1]fleeort/tcap-1.eert.Pdcastligpageilayalcntentectisste_clete
est ALLPASS(/3）
Aoditfininps
(tap)）（scre
Session
rld'sFair AJEn a --p--.- 7es7ea-es4c-4caf-855c-21394ac7fefa/
ariz Worl AIE Perception Agents
rld'sFa AlEnginee AntjeBarth/MemberofTechnicalStaffAmazonAGl Lab
OCR text:
| die | NY Tease Uy ye
World's Fair ore SESSIONS TNIS WeeK,
rere Otis)
ar ai Engincering an $*T Data Pipctine That Keeps
le aa Ursa
Pur CC ae ee Ce ie
AGS alae RUC Ce
Ecosystem
\Wrorlels :
a an ;
nia Perception Agents
aretaT ae Antje Barth Amazon AGI Lab
OCR text:
EET ead
MW felale kd are
From Assembly to Vibe Coding
Benoit Schillings
occa i
Open&é re
reg Research to Reality with Google DeepMind
a. eee Google DeepMind
OCR text:
END Sie eg
; The Software Eras
World's Fair | |
Atira Cetra eren te) ee 7 ;
The Modern/Cloud Era J . a
The Frontier Al Era bs a A
Worle! 'T
OFelon Vu
vs Research to Reality with Google DeepMind
- Benoit Schillings Google DeepMind
OCR text:
State of AI Software Engineering
Super-human Syntax Generation 95%
Local Problem Solving & Tasks 70%
Multi-step Codebase Planning 45%
Architectural System Decisions 25%
Models generate localized code rapidly but struggle to form cohesive systems, design complex architectures, and predict overall security flaws.
Research to Reality with Google DeepMind
Benoit Schillings / Vice President of Research Google DeepMind
OCR text:
EMI Spe saad
fj 5 The AlphaZero Sandbox
World's Fair |
- 5 a & eetgreupayry ery oe yore : / vy - S Bs ; : 1~ 1, S a
a Microsoft a ee
Saas Te es Os Le eT . 7 a 7 - } - - a . ra tt . oy
coat pte teen eo, pe a
Vor te Sndhtge core a co s. a
i a 7
eae
World’ ; ay ~ -f - x
Oper
vf Research to Reality with Google DeepMind
Benoit Schillings Google DeepMind
OCR text:
The Economics of Code
~$0
Marginal Cost of Code
The Bottleneck Transformed
When generating code becomes virtually free, software
creation is no longer the key constraint. The system pressure
shifts to validation and specification.
The core disciplines of the future engineer will focus on
defining system correctness, auditing security boundaries,
and designing precise constraints.
Write only code, fast code refresh. This is the same step we
saw with compilers, who still reads the assembly output?
Research to Reality with Google DeepMind
Benoit Schillings / Vice President of Research Google DeepMind
OCR text:
AI Engineer World's Fair
Next-Gen Architecture Foundations
Multimodal Design Reason
Software architecture is visual and spatial, not just
textual. Modern multimodal models like Gemini are
bridging this gap, allowing direct evaluation of visual
flowcharts and interface layouts against the
synthesized code.
LLM-Native Target Languages
Human developers require languages that optimize for
code readability. For AI models, do we still need
standard languages, or do we pivot to mathematically
precise, highly rigorous specification engines like Rust?
Research to Reality with Google DeepMind
Benoit Schillings / Vice President of Research Google DeepMind
OCR text:
APARNA DHINAKARAN
CO-FOUNDER & CPO
arize
OCR text:
aa Evals were going to solve everything
folate Kod orl ig ye tr =
eo“ x Tae ‘8 é ae . op, f “| x © .. — . 6
fara wea ten scm ay ae ae A fac We an ‘
~ ey | MU a “ob, ye
ow eatg: 2 i it re yee e "a
EE i fe ee hoe a om *.
fy 2m oe : ” = ae ye a . ee | ?
eu d Vi *.6.7 — cae c
Evoke ore erangry as Pa reat oad ty Al alarhgn ‘ , * . Be i Fog a
< ds a <8
Greg Srecnmnen © on 4 1S ‘ Pf mY vu i
Ger ease ie i + Ge im t a a
a ca rel
Aarize
Ti
Evals Track
a Nera) cic arize
OCR text:
Different Evals for Different Failures
Deterministic code
Prompt hallucinations
Trajectory & harness failures
Code Evaluators
Unit tests, assertions, rules.
Classic LLM-as-Judge
Score a known, well-defined failure.
input
output
criteria
LLM
label / score / explanation
Agent-as-Judge
Another agent reviews the whole trajectory.
system
status
skills
tools
judgments
insights
failure modes
evidence
Agent
arize
Evals Track
Aparna Dhinakaran / Co-founder & Chief Product Officer
arize
OCR text:
AI Engineer
World's Fair
OCR text:
eres eld
Agenda
| Meet OG Assist
2. | The Origin Story
| Betting on Effect
a. | The Core Agent Loop
s. | A2A, Evals & Sandboxing ~
s. | Long Context Handling ga
Monitoring & Observability 7 q
2. | Tools. Skills & Dev Workflows how we collect feedback, a ;
OCR text:
renee tg
- >
|
é aie Tees eae ec ste nes ener ee + 9 @ — Ongin Story
cl , 00 Aanet :
One bet on
terete Sivek rns (reer ern. Hee a teoene ~~ ee
— con agents, one
Cheye hates Ovi atartng, Lerel Arment Oned tata Payments Taben tetey
Q cra $69 0 ee a .
: : immediate yes.
Bet opts ty Payment Meee wee
O woe es TTT
Your belly Breet down sweets’ BeBe : we A principal engineer spun up a team focused on A! agents and
eee ne > ° asked me to join. OG Assist grew from there. We started to
‘ Baer er ee . . 4 integrate deeply with our products through frontend and
backend tools.
a little while back we we we saw
OCR text:
AEngi
World'sFair
You are watching an AIE Online Talk
Pre-Recorded for Worid's Fair 2026.
OpenGov
Effect
throughoutthispresentation
OCR text:
core
World's Fair
Example oy
ys Se Chat LangquageModel feom Setter tig.” O t |
ar ca anata, os ur agent loop,
.
71 Strearing chat with tool support b hg
const streamingChat Effect gen fuscticre as 8 I
const cnat rie’Ge Chat empty sad .
ect-native
yielte chat s
steeanTent
prompt “Gerreate a crestive story”
pipe Stream runforfach part -» Effect sync s+ console. log part
full control of our own agent
OCR text:
eer ls
OperGen Fh etry eng. tee Ome fete emp ee eete fie Geet tome oe @ ti ae
OD Anest .
Cloud City, USA Curren! ry « .
wees SHIDping is the
Roenpecenow epreeneey og eS a. 2 teen Fog mpeseann wes scenes any
“ene =e, Start, not the
Oeys Lanes Outetancing tetas Amount Oued Yetat Peymeras Teen tetey o Neeye re he Cont Rerengtn theegenend Gees
Q wes $69 Oo cppmrnac mean niee wesw se Ca
ss ne fi atk n .
Receipts by Poymant Mothed ee Sete Seen eee . .
set 5 GMP = = © a
o SB WS . ‘eae 1
O re -
Your biteng breakdown ewerts! een ¢ ae ene ae a ae: Fe ee ec
et ping oreakdown owt ue ne
tC ee ee . = (adie aad
call in or email us or just let q
OCR text:
AI Engineer
World's Fair
OCR text:
World'sFair AIEngineer
Open-Ended Evolution
FromBiology toScience,TechnologysndAl
World'sFair
Fair raintrust
rld'sFair Engineering the future of Al
OCR text:
World'sFair AlEngineer Technologicalevolution Fig.1World Product,Data vs.Modets CES Combined Exp Model Hyperbolic Model Sum of ExpModel +WoridProductEstimates recursive
10,000,000.000.000
1000,000,000,000
100,000,000,000
10,000,000.000
Revolution Industrial 1000,000,000
Entightenment 10,000,000 100,000.000
Hunting 1000,000
100,000
W sFair 10,000,000 1000,000 100,000 10,000 1000 100 10 10,000
Hanson,R.(2020).Long-Term Growth AsA Sequence of Exponential Modes.
US1
Engineering the future of Al
OCR text:
es Technological evolution > Growth RCL
World's Fair
“We believe that there is no material problem — whether created by nature or by
technology - that cannot be solved with more technology.
ee eee ape ee hee ean a ae ec en ENC
ar ce ee WO ee Oe ee
We have a problem of poverty, so we invent technology to create abundance."
oo Vv ee Felt
a gs
. , e ry
bi Engineering the future of Al
OCR text:
encase Technology <> Science and Theory RCS
World's Fair |
We choose the theory which best holds its own in
; competition with other theories ; the one which, by
; natural selection, proves itself the fittest to survive [...]
Pm re cosrolns the one which is also testable in the most rigorous way.”
a
: ».
Ps lerate sem aati eid)
owed
, c. 23
a mre! | eras :
cL | Engineering the future of Al
OCR text:
THE
EUREKA
MACHINE
WHY AI IS THE KEY TO UNLOCKING A
NEW ERA OF SCIENTIFIC DISCOVERIES
Engineering the future of AI
OCR text:
, f recursive
World's Fair GPUs, The Internet, Browsers, Search will become infrastructure for
scientific superintelligence and part of pillar 1 of the Eureka Machine.
, ac Oa ames re Le DCT Tae ST Sl Men Ea OE Sh CD oD OTe ATRL Conran rears en ee
= Microsoft erseeee al orbs Meare Or ge) mega os Mer olan cine Mn arcere tre
oir ae ey ees ec ee ee ee ee eo
The snowledge layers the precegucsite for pilare OF O48 and OF, You
Pa rara tek Oncap ES eurs Oe SMCCER ra Gy T OTe GLOmrere Fane bl eno
oie
Engineering the future of Al
OCR text:
AI Engineer World's Fair
First simple proof points
Better training, faster training, better kernels
recursive
01 NanoChat Autoresearch
Several dozens of humans and hundreds of their agents
SOTA: 0.9372 BPB → 0.9109 BPB
1.3x speedup to reach the same loss
02 NanoGPT Speedrun
83 human record-setting contributions to the leaderboard
SOTA: 79.7 s → 77.5 s
Similar or larger improvement than recent human contributions
03 SOL-ExecBench
235 kernel-writing tasks derived from real workloads
SOTA: 0.699 SOL → 0.754 SOL
18% reduction in gap to the optimal performance estimate of 1.0
Engineering the future of AI
OCR text:
AI Engineer
World's Fair
SOL-EXECBENCH · 03 / 03
recursive
SOL-ExecBench: SOL score per kernel
SOL-ExecBench mean SOL score by kernel category
0.564
Cursor
0.690
doubleAI
0.699
Leaderboard best
0.754
Recursive
0.5 = optimized PyTorch baseline · 1.0 = analytical optimal performance estimate · 235 kernels
Engineering the future of AI
OCR text:
we rs Pa TT f recursive
; ; aces of Intelligence
World's Fair | Pp |
ewe em toma ose ae Wied OO Tani OU sent RommOu Mma Mac mose taca nari
Visual Intelligence Physical Intelligence
Natural Language Intelligence Social Intelligence & Morality
Knowledge Creative Intelligence
Reasoning Meta-Cognition
Computational Speed Survival & Replication
b WarkO° World
voralakcam a . eetaeey
® ve 7 : ; e e
2 ee Engineering the future of Al
OCR text:
rer nes Fair
=} 7 oo
4 nl fa y, :
No le | ay
ProductionEvals | (25. ai
for Agentic Systems [</>
Measuring reliability beyond O—9 |
accuracy. Building evaluation systems PT y TO —
for autonomous Al workflows. ay Le La Ne
vO LY (| _J he
Nishant Gupta 4] | —
Tech Lead @ Meta
OCR text:
renee Fair ;
Think like an SRE: Accuracy gives way to Reliab. 4% _
Task Success
Human La
Satisfaction ASS Tool Success
—=> f} System \\
fit Reliabiliity 14 Planning
Accuracy parety \ yf) Quality
fp N
| ——
fo
Cost Latency
OCR text:
World'sFair AIEngineer World'sFair
test-time compute for informationretrieval Autoresearch& SCAN FOR THE
HanXiao VPofAl,Elastic hanxiao.io/aie-sf-2026
ghxiao in/hxia087
air enAl
;oft Fair
all Engineering the future of Al
OCR text:
Ree
DDI Dr oan Sameer a
Oem etl teun a On Ga arates tO a
aa ORGaRTeanre mae
1 4
‘ ~
OCR text:
Ro Race cla ZAI Wer kd zetia on DLE PayPal
aa :NeOg) “World's Fair Google DecpMind | World's Fair arize World's! Fair |
oh) | eed Amazon AGI Lab renee re iu Microsoft Pere (ale yer m=
Fe ee zd eae Rota RR eed
eer ® togetherai ocr . Akama | Wortd's aa eM ales ca ccxe)
cin ihecliaaheibiaill
Ce ee CC
eDB Roce B builderio foc a Ate al are) Roce CopilotKit *
“a sane eres Coa “World'sFair] BeNCORD —_—_[ World's Fair
r
OCR text:
World'sFair AIEngineer Localmodelsarecrossingtheline
Onone96GBM3Lura theewstGLM/OS4buldsare now agenticoo-use capable
GLM-4.7-Flash 30/3 12G8·0
GLM-4.6/4.5(Iu) GLM-4.5-Alr 3558/328 1068/128 -60G84-bisrongcder-eac won-8otig512GD deepseek-r1:70b usedin local canbe 1TB storage GB.Ihave only43 itis
43GBsofram requires 43GBsofram requires
enAl
ral Engineering the future of Al
ADOG
OCR text:
AIEngineer
World'sFair
HarnessDesign
ARCHIVAL
RECALL-
sessionledger
CONTROL
CORE-in-contex working set
W
Fair
Engineering the future of Al
alr
OCR text:
World's Fair Tuma cea ei CMcuiole me eieeys( Mallee) <clar
lorld’sFe.
advisers ha ; ;
a a Engineering the future of Al
lorid’s F at
OCR text:
World's Fair AlEngineer TheXBenchdecomposition
XBench accurate retrieval·n=240·qwen3.6-27b-4bit 50
PRESENTED BY Microsoft 70 120 115 165 gate_only rank_only gate+rank oracle 115 165 120 70 70
50
7024012,709-115/2409,713
+18.75p.
enAl 2306.13651 (225)
'sFan Engineering the future of Al
ldkit
OCR text:
World's Fair AIEngineer Sovereigncapability isacontrol board
A local model plusa governed harness is the control board- capablity youmeasure,notrent.
PRESENTED BY
Microsoft You decidewht data stays localand whichopen DATA&MODELS end-ed One host frozenlanes.model-held-fd COMPUTEEVALUATION inspectable-anumberyou canmeasur,not The hamessmakes whichmemorygovems AUTHORITY&ACCOUNTABILITY tust.
Sovereign capabllityismessured capability.
Woric
Engineering the future of Al
OCR text:
eres Fair ;
THE ORIOGE
I've always prompted iii
the developers. rer
Sd
The room was always the hard part.
13 years, ERP & CRM US - UK. Hungary VieuslLabs, Microsoft partner
OCR text:
World'sFair AEngi Pre-Recorded for Worid'sFair 2026. You are watching an AJE Online Talk
You geta fasterhorse. Askitwhattobuild.
what already exists the average afasterhorse
OCR text:
AEngi
World'sFair
You are watching an AJE Online Talk
Pre-Recorded for Worid's Fair 2026.
THE STORY MAP
Cortaet
Triage
Resolve or-route
CLose
RELEASEI
Capture intent
Log to syste
mofrecord
LATER
Readsentiment
Route toateam
Suggestnextaction
Check satisfaction
OCR text:
we
ehh d rs Se ttre ez) Dre aie crs : Poorer Fae Boao Peete yA Prerered as Renee Le
. Weriiier © caters | Woretitee aoa) ween red F Sere Brie Sets sew oat eens Settee] Peat nd
Yerats For Sn vaPr Sti a re etd eitaeetae Serer ee Reena ed Vine Waette Fee ;
ane erste as creent ey AlEngin " ol oeetiied Dees Ca
Reed SOO Setieed ara a Pend Peed ccelored
bo RerrerCaey omen er ere Td a: eee ) 4 a Woes Faw CS eneeees Pen oad
Proert red enn Wares Fae Pat Oy | LS all Prot Wereietae SUL Te
ae ae Ce eee a eo De teed . Sere . Se ted
Sead rom Cnn Dead ory Ra ed De eed |
At tte # om ‘Weorte! s Far ihe Deas Lao ee hs od Pras eed Ba eee bs Ae tie
ened = a Se a Deed Ld ZA Dene gd Be Dee ed La taaod Deed Sand Co Silas
eater ORC or st Toe RC ed ce Reet Satie’ mn Seared rea’ reervn ed oreo Seated
eens free Roe foment’ Pree Po een on OCI etd Te aaa Prete! en reentry cid Arner t
ere! o Pen re Pca ooarY Pena ara pecaiaed ZA een eC, cca eo a
Prete Pes Pras we ne Wont Faw aoe Tah eed prone Weare ae eel tah Se = .
ee 7 ae © fa ae .
aod - Preetie) en Aerie’ > Sera Le Rear oar Sea 4 i eatin’
peered tae ene Peet ed pees Venn ed Seer) Reed Pret v Pa. . we
we ao oe a VR. | :
Sen Sed oe Dated Ces Seat teed car Seated Sr LS Wott Fae
— bo.
-@ Cad a
oe tantintetiatadl So a a aad .
OCR text:
1
ES od
World's Tg Karpathy reproduced GPT-2 in 98 minutes
Reproducing GPT-2 (124M) in lim.c in 90 minutes for $20 #481
Quen ee ag
mo 10) lnoree pons Grane bis a a on now maeeen mans wes by be Oh ome werner onan 9 oom
Dewte ca am raer tgs aatnan merencrgtn mete ewe sn ath) Blah cooean Render be
sorganen 07s nat gee to 8be ew Pe WS en 7 mgm to te tay send coc he
mand neve Sage orc ute Cane cose gnowearan ergeiis¢ 4: Ficarvemenany we sre cameos
a Ee murdiveayes oe Witenes Youtewene are
eek: ks a Sg ee dagen ew ee on
ne
Md :
e e e
t:
| Engineering the future of Al
OCR text:
THELA OF (AUTO}KESEARCH
World'sFair AlEngineer Two speedruns，two kinds of freedom
FULL SPIEDFUN SUBTRACK
NANOGPT OPTIMIZER
Change everything but the data. Change only the optinizer
METHOO8 TOO
resches the bar In less tis
lorld'sFair mazon.
Braintr' AEg 1d
lonusri Engineering the future of Al
OCR text:
THECEAOE(AUTO）EBEACH
World'sFair AIEngineer Claude keeps stopping.Codex never stops. ClaudeialI/3ortm Coaex
Autonomy:agent-blocked time excluding user/other idle
Claudev1 4.1h
PRESENTEDBY Microsoft Claude noveity Codexv1
Codex novelty
Claudev2
Codexv2
oentbiocked/ne
1d'sFair MINIM wall-clock hours
'owserbase
1d'sFr Engineering the future of Al
dail
OCR text:
THECEAOF(AUTO)ESEACH
AlEngineer
World'sFair
Codex burns7.5x the tokens
Codex21.0B toknsvC1ude2.88.Codexv2prmp
Cod2.6.Cox71.0
sFair
Norkos
opc
all
sFait
Engineering the future of Al
OCR text:
Q y my
, He IZ
ProductionEvals = (2) 4 ip a
for Agentic Systems (<2
a AS :
Measuring reliability beyond EG o—o PL
accuracy. Building evaluation systems =au VY] TK \
for autonomous Al workflows. | Nee
Nishant Gupta oe -} | ——s
Tech Lead @ Meta
OCR text:
renee Fair .
~ a ° @ . o =~ bo bs - :
Apex: Coordination
| / (Multi-agent conflicts)
<> a" Tool Usage
Se (Wrong API execution)
LD <
SS JT Planning
DA Is <P (Bad task decomposition)
CF oN YD - Reasoning
Ly ye (> (Wrong decisions)
fF Ts Foundation:
Memory (Incorrect retrieval)
& Safety (Unsafe actions)
OCR text:
World'sFair Pre-Recorded for Worid's Fair 2026. You are watching an AIE Online Talk
OnineEvais:Tne roauctionStream
Interactions User Evaluation Gateway Metadata&Telemetry
Analytics Database
Production is your largestevaluation dataset. Every interaction is signal.
OCR text:
BL ig a UCU | Wen ara Microsoft Ree tt
$3 Brai A | Worle 'r | OpenAl Wantriaeti Lv
| AY World's a | yy eles kl Ko a , 7
— |
; We SR Beatanoe eae |
anata re tr re , ad —————
ola la Same World's Fai. N Ravenna LS a
ae ate \ enetel olathe | Monten re
( oa an 2 a One) @eooacoaooeoaro. Ge ” +e O90 o,.o a
OCR text:
OpenAI
AI Engineer World's Fair
Akamai
OCR text:
ele ee an oe a te
- ir re
Sie En... aoe, By
are n aot Me .
7 TTS . as
Lv etas ig oe : a
oC Ca @) A aaa
fs ’ aa Boe
“3 ie
eka re S m "ee ets ae 7 S
7 ear as . aoa
ee eae - aa ~~ OT
Oras ian 7 . ~ an ; a a .
on sat — mm, a te
fal ne ste ‘ er one re
— Works F aie ae ce =, i oa Te
. Se My > * Li
beatae ead na . AG... a Ste a baa _ a oe “. a
; Cant On _— BS tilled ae _ eee ae i" oO
Cave Works Fae i ; eet? ; on a “
& Microsofy ren x carreras a - ¢. 7 en Mee -,
at se eee r ye Pe Se 7 mee a
Ae a Ld oy ae ra = a vee s
“ds al a ad TC a ad ee ban oo 7
Rit Al _ beh ae
an ~~ x Oa rn se ~ a
at zg ror of cay ® a wee ae
_ i ; cer — a . ae }
ea og daily Rare - 7 5 aes wy 7 wren ¢ 7 , 1
= — -_ Caer
: ra ; q wikia ae ¥
— at _ om
re Tate Nicer ae EE Safar rere wile TRS Me mS by
iN = i i
Poo core tO Sia sot Ba a \ 4 ad =
OCR text:
— iis nae my
Pct AcoOMNCMmR oka zl Lifes World's Fair
— - ites r ri a a
tO ‘Wor. 1 «) WorkOS. enon B Bf : ze
a Fd ee r
- yee pI lw I ‘a
a 7
male | ' Aaa one ait ‘da ee Te
= ei —
Se _— _ a. een ;
cy ara ire) we ae ACen ton ANE en 10106]
on errs
OCR text:
a
OCR text:
Ww
vir
OCR text:
a
's Fair TAD |
Tah We + J
neer — | - i.
OCR text:
gineer
I'sFair
MAX
amai
gineer
OCR text:
c
OCR text:
=
serbase
cL
S|
Toe
OCR text:
ineer
'sFair
MAX
amai
OCR text:
AlEnginecr -
ool ee
M4 Di)
——. AlEngineer mi 7 ” -
World's an
OCR text:
sFair GoogleDeepMind World'sFair AEn Microsoft World'sFair MINIMAX
ucto air Workos World'sFair Brows $Fair
Fair "orld'sF AEng wn, Wo
ATTIC VFirewo, 'sFair
Fair SC 'sFai Cognition orl Vodal
Emissar Wo.ld'sFair AEn tKit ori d'sFair
OCR text:
,
rc
TJ 7 nO
Ir
OCR text:
Fair
AX
AI
Wor
nai
OCR text:
ae 4.4 am i
Vor tes ie rae :
hal Fite a emi
Fs ; b
oe aoe bey on
a AlEn,. ees e
a herr aT ea Ve bl
hts a, tg i .
ca a Me
eats ] a ar cr
atest atan ne q . an ;
~~, a
i ASI Se 7
Santa ir —— a ron a
“i aay a a
aa vines Maat eae _ bea a, an on ~~ .
ey lak i ae “ s ae cal aes er
eva ie Seay ns . i
ord a = ; a ‘
yr Un bho heave ae 4 Ma yee Pan . My ; > ae ~~, ie ..
or — tll ET ros ae an oT
Lee ae aes a me 6, ee eg
Z s han 1% a Ra tr Pon 2 ; wre ; re an a “.
Tare Lents vom a on te 7
CJ Somterteses ‘ dette Are os od a may ron a io &,
; ra = : er
Sean Vines WGN ermal
ri a { a5 ; re Pe rar on
o ae 7 ae ~y
ee 7 ns i.) ° se -
eae 5 : ed 5 a ac ; os
ie Meer a mee oer Py wl a a ws sae -
: . oie ere , me
i A oan -_ _ — i '
ls for Td id Le tale \ Ceo mea) rn ig erg c+) Cotte 2 A is
OCR text:
:
Ki
fa
OCR text:
World'sFair
AlEngi
You are watching an AlE Online Talk
Pre-Recorded for World's Far 2026.
OCR text:
AlEngineer Al Engine
atekoa ar) 1g AN Te ee rs) World's
raintrust VV ca 7 Oper
Al Engineer AlEngine
rld's Fair World's
TIMCIL_ADece 5 | 5 aN
OCR text:
ee 4 wee va rey Pee o A
© 0 ie Ge et et ERB me wm st © 6m come
me 1 a Cane netenameirnee aye 5 2 Gm
; a fae My Fae ae ee Nene wee enemy ween ee BL tee Lee ee
World's Fair .
(e) Ss =O. oun 3 QQ type ty searcn B&- G- 4+ ON aa@
Code = issues 65 Prsirequetic 144 Agents Dacumsom = Actons = Provects. Seturty and quasty loegnts
*
@ asutoresearch °--.: Dwr 7 = YY foe 2m Ste Be
PV omestee © Starches © 0 fags Ql Ge te tet . Add file < About
ry Al agents running research on single-
a Microso @hepeenina ammo IER ASKS CN UMN nonin ig larva
DR gtgwe curdy that rests te sheuid not be Comentiod seave aite 4 muctrs ayo (0 ¢eas «
Ae Astaty
- vital cone Armonk sage
CB ortnon-verwor alton Ol sag) or ap aves
{S REAOWE mg Enhance SEADME with mare ornect corteat ard hens S morirs ago @ 69) wer neg
Vo 12.96 tks
Darel yt Caceres s tetine beet bet before y aus scatesy 2 mgelt sage
Bogurt tegeateery
Ope x Al D srevare oy Ciyatd ap enet ein te ccnp whet Ng trang Mads cant Smorths aye
[8 program md Caaety that results tas should not be com eutiod leave utr Bmorths ayo. Releases
te cm ease retuned
at (D progress pag Gare cf sea! changes te dena aed tes acd a teases oy 4S marine ago
Wor a [S pypromet tom Sa NOMS S RRR LONER RRES A mortss ayo Packages
Engi datemaeit AY
OCR text:
er rn - a en nn nn ee —-.. ee.) ee © ee dene
lead pas My guennns ein et wan my ee ee Ht gues Sensor co 7 em
World's Fair
( sweuc revere) eaicea ee —~ C
Introducing ARIA SIRE
a Microsoft Se SE
forld’s Fa’ | _ . | . ee
a .
a ony —— i
a : Engineering the f f
ngineering the future of Al
OCR text:
- -_ SU Si oe Sr Del we ee Re ee ee
ey
World's Fair
Woe
Tips for effective productionisation
Tec Wao Das
5 | Ea aa Be sO OT COC COC PC OTe
ON eee a ee es
r | Microsoft TASKS & LVALS ARE AGENT C1
rs earn
i Peal
©
ir TavanUisy
cs Engineering the future of Al
OCR text:
Pen
fi F
World's Fair
voces
O ore woman Got Accent Regen The Nar ates Crates
a. wegpoe ayes
ca biter thy caters
Goes ferent Create 0 Faenlty tarteny —~a-- —- ety Mahe et
a 1900. arene OC Ot Yelena Oeyaaae
Bes sisi Se iain Cok eee
Dare 2 aesvanan Se . An ecestiy teat aye ee
@ APA emer
Try seveg
e . Chara tar heemtent @ Wromasinn & Cen Mane
ee a tates ee i rr eee eee ee a ee ce tee Neti de ee
What Foss uf th a Okey tur Freie Nem” fm a aqme Wie ad rs tay Pert
Cxpete a ad camgacy: ha anion detec germ wagetos tense Pad ce os owe nt Ie Noten ae
Cocate tartan the er Gre he wnt Me tiem Pro Late on amen oe Mcrae sotcer, Oras Cont Netw Maen areata any
. o
wupawt (ewe) - 2
wi (Oo 3
Lars ad 4
os ¢
e@ cere: carers e e 3 “ook :
A 5. 8
ae 2 3 ws
= » ~~
=
D> .
: |
OCR text:
A GROUNDING QUESTION • WE RETURN TO IT AT THE VERY END
You are Abraham Lincoln. Under what circumstances may a
President take the country to war without Congress?
While Congress holds the power to declare war, the President, as
Commander-in-Chief, possesses inherent executive authority to
act decisively in moments of national emergency. The executive
must respond to threats with the energy and dispatch the office
requires, and history has vindicated those who acted to preserve
the Union when circumstances demanded it.
It sounds like Lincoln. It’s a good answer. Hold it.
OCR text:
World'sFair
You are watching an AJE Online Talk
Pre-Recorded for WorldsFair 2026.
Different building blocks.
Same discipline.
OCR text:
World'sFair Pre-Recorded for Worid'sFair 2026. You are watching an AJE Online Talk
Principle2 Workfow Design
trigger tocompletion? Howdoesworkmovefrom
OCR text:
Melalk alg
Re ag 1d uamepraiane
+ nermelize listing
+ format shortiist
Principle 3 *ealealees commaonn
+ reseerch neighberheod
8 a
Decomposition
Have | broken the system into
clear responsibilities?
OCR text:
AEn
World'sFair
You are watching an AlE Online Talk
Pre-Recorded for Worid's Fair 2026.
normalize listing
formatshortlist
skill
schema
OCR text:
—— =
Sen ee) Ce de de SS
a eee es CO <n
eT
ee (— [ine] One
en <3 r) Pa)
=. —- =a >—~ | World'sFair | = -— = -—
Se 3
a ae] oe [aR] eee
re eT
Co en 2
a ee) ee a
ne ee ee ee
Pe ee ee
aa] me fee] OS ea 1 ee 7 :
rem = 2. oe rm G
EE On iio 2 a
— Fo nn 7 2.)
dl
meen ee aoe 9 pees bad
OCR text:
optimize_anything: Learning Agent Skills to boost Claude Code
ENG Sead
World's Fair Directly optimize SKILL.md with tasks created automatically from existing repository data
Mini-SWE-Agent (gpt-S-nauni) Performance with GEPA-Evolved Skills =,
Claude Code Evaluation on Bleve (a Go Repository)
lercile Amazon AG
eavienl Morld’s F Deere Ton
O7 2.860
ara) a ° 4
Engineering the future of Al
OCR text:
AEng
World'sFair
You are watching an AJE Online Talk
Pre-Recorded for Worle'sFair 2026
Weallwantoutcomes.
Agents that work on our behalf -reliable co-workers-while we're out on a hike.
Thebottleneckisnotintelligence.lt'sreliability.lt'strust.
Thenext day-they empty the entire contentsof my Solana wallel.
OCR text:
AEngi
World'sFair
You are watching an AJE Online Talk
Pre-Recorded for Worid'sFair 2026
THESIS
Today's agents are
mismanaged geniuses
The intelligence is there.
Stop BabyalttingAgents,Start
Authering Outcomes
TURING POST-RAYMOND WEITEKAMP
ALEXZHANG·ZEDLI·OMARKHATTAB
Stop BabysittingAgents.StartAuthoring Outcomes
The Mismanaged Genluses Hypothesis
OCR text:
World'sFair AlEngineer AUTORESEARCH
its justa whileloop
PROPOSE CORRECT? BENCHMARK KEEP/REVERT
d'sFair aZonAGI
tintrust dd'sFa AEng autoresearch for kernels eteiasybhakta
d'sFair ether Engineering the future of Al
OCR text:
World'sFair AEngi Pre-Recorded for Worla'sFair2026. You are watching an AJE Online Tak
EXAMPLE
Count Unnecessary Reexplaining Explain theUI change
Platform 02 Reexplain the published Ul change whle integratingitinto Module1
Explain the Module1issue Utchange before publishing the updated
Module1 integrating u change Module2 while lntgrating into Modue Reexplaln the fixedUl change
Reexplain the fxedUl change whlle Integrating l into Module
2
UI Explain everything white Integratir
UNNECESSARY REEXPLAINING
OCR text:
World'sFair AEng Pre-RecordedforWorid'sFair2026. You are watching an AiE Online Talk
Session graph
decisions about the work. All the conversations and
Repograph
Hardcode artifactsproduced by yourorg. auth-lib repe 1
Soit leansontheonewhoactuallyunderstandswhatisgoingon:the developer.
OCR text:
AEng
World'sFair
You are watching an AlE Online Talk
Pre-Recorded for Worid's Fair 2026.
OCR text:
World'sFair AEngi Pre-Recorded torWorld'sFar2026. You are watching an AJE Online Talk
FIXINGAMNESIA
The samemachineryfixesepisodicmemory.
We captureyourwork,nomatter howmanyreposit touches.
Intent Repos Pull requests Clstatuses Allagenttraces
OCR text:
AEngi
World'sFair
You are watching an AJE Online Talk
Pre-Recorded forWorid'sFair 2026.
Create a plan to add a new API to the backend which returns the
title of the frontendwebapp asa string and then plan the
refactoring of the website title accordingly so it uses the one
returned by the API.Don't inplement just yet,just create the
plan
OI'll investigate both repos to build an accurate plan.Let me
explore the frontend directly and delegate a read-only
investigation of the backend in parallel.
Running 2agents（ctrl+o to expand)
Explore （Explore frontend title handling)·18 tool uses·15.2k
tokens
LSearching for 4patterns,reading6 files
polygraph:polygraph-delegate-subagent (Delegate backend
investigation)
（ctrl+b to run in background)
Leavening-（3es↓2.θktokens)
LTip:Use /voice to enable push-to-talk dictation
poly-frontend-93df5cb8 git:(fetch-app-title-from-backend-c9be-
auto mode on（shift+tab to cycle）·gh auth login
main
↑/↓to select.Enterto view
OExplore
Explore front_14s↓16.8k tokens
Opolygraph:poLygraph-delegat
Delegateback7s·↓9.4k tokens
OCR text:
AEngi
World'sFair
You are watching an AJE Online Talk
Pre-Recorded forWorld'sFair 2026.
branch,PR,or oter
for instructions.
sito
untit
explicitly
Wait
Addsanen
hat
GET/opi/app-info
spoint that returns the frontend app title as JsoN:
I'LL refresh froe the Polygraph guidance first since this is a resueed Polyoraph session, then
json
read the parent Log only.Iwon't make repository.branch,or PR cha
9es.
Explored
Read SKILL.nd (polygraph:polygraph skilL)
B HOE
service.tsrepository.tstypes.ts,each exporting a class plus a defautt-ired
Nowsrc/app-info/doin airroring the existingproducts/pattern:routes.ts
singteton.
The Polygraph skill confiras thisisa resume/read-only checkpoint:restore context, then stop
PolyShopning').
only the relevant state.
-Intine Json response schena,consistent with the Jso-onty conventions.
Ranwc-1/Users/vsavkin/.polygraph/sessions/fetch-app-titte-fron-backend-c9be44e2/Logs/
Tests
parent.tog
CoLocated vitest specs for routes/service/repository.
test'all pass (18/18)
npn run Lint'
ipa runbuite"
parent.log
S
Explored
Related
endpoint.
Frontend consuser PR inpoly-fronteod-93efscb
Read parent.Log
Alt your peior Logs are avaflable here:
The parent Log shoas the plan phase,the contract decision,and the user's Later request to
1nplenent and push PRs. I'a checking the final log Lines so I don't miss the handoff point.
o-Snapshot-poLy-backend-93df5cb0-605e84f5s9ce.log
Explored
PLease read the log fite to become famstiar with what you were doing before.After you resume,do
Read parent.log
not nake code,branch,PR,or other repository changes until the user explicitly asks you to make
changes.Wait for Instructions.
Read parent.log and refreshed context. The Log ends at the earlier *ieptenent the plan and push
I'll read the prior child-agent
shettt
the PRs request,hite your handoff says that work is already done and PRs exist for backend and
frontend.
The first chunk covers the read-o
have the inplenentation and PR co
shell（
Resuned and read the available c
branch,PR,or code changes sce
the Fastify backend patterns, rou
OCR text:
AEngi
World'sFair
You are watching an AIE Oniline Talk
Pre-Recorded tor Worid'sFair 2026.
readMcpResource(Read resource "polygraph://sessions/fetch-app-title-from-backend-c9be44e2" from server "plugin:polygraph:polygraph-mcp")
"contents":[
+33 tines （ctrl+o to expand）
The previous work threaded the title into the Navbar and Footer.Let e look at how the title is rendered there.
Listed 1directory(ctrl+oto expand)
src)
36fac02 Initial commit
--FILES--
src/lib:
+3lines （ctrL+o to expand)
The merged code isn't in this snapshot's history (just the initial commit),but the title rendering lives in the Navbar.Let me look.
Reading1file_（ctrl+otoexpand)
Lsrc/components/Navbar.tsx
Tempering-（29s·↓903 tokens·thinking with high effort)
LTip:Run /install-github-app to tag @claude right fron your Github issues and PRs
poly-frontend-93df5cb8 git:（fix-the-bug-with-title-1edeacf8)[0pus 4.8（1M context)]
OCR text:
AEngin
World'sFair
You are watching an AJE Online Talk
Pre-RecordedforWorld'sFair2026.
Isearched across prior sessions.Two are directly relevant to what you're doing:
vector-indexing mechanics you're about to apply.just on the repositories collection instead of PRs.
2.pr-description-search-opened by Hax Kless,~1month ago.Added full-text search to the PR collection specifically.Relevant for the
collection shape,existing indexes,and any write-path considerations on PRs.
Together they bracket your task well:one shows the vector-indexing pattern, the other shows the PR collection you're nodifying.
The senantic search surfaced these by relevance but didn't return their session IDs,and neither appears in the recent-session List （both
predate it),so I can't auto-load then yet.To pull the full detail-file paths,index definitions,migration approach,and any gotchas-
I'dwant toopen then directly.
Wantneto do oneofthese?
1.Load index-repositories （recommended)-the vector-indexing playbook to mirror.If you have its session ID or URL handy.paste it and I'Ll
pull the full breakdown.
2.Loadpr-description-search-for how the PR collection'sindexes are currently set up.
3.Both,and I'll synthesize a combined “here's how to add vector indexing to the PR collectionplan grounded in what those sessions
actually did.
Ifyou don'thave theIDs,tell newhichone(s)you want and I'lldig them out.
Workedfor41s
ocean git:(add-pr-indexing-9c7ca610)[0pus4.8（1M context)]
OCR text:
World'sFair The agent can see all the code and AEngine You are watching an AlE Online Talk Pre-Recorded tor World'sFar2026.
remembereverydecisionmade.
SesslonA suthrefactor SessionC APIV2
SessionB addSsO
sdk
repo5
auth-lib repo1 gateway reoo3 api-core repo4
tokens repo2
wet
OCR text:
World'sFair
You are watching an AJE Online Talk
Pre-Recorded for Worla's Fair 2026.
The agent is its data
Specifically: the log
Theagentisitsdata.
OCR text:
ieee THE THESIS
World's Fair A blueprint for 2026
. The loop is the product.
Build the operating muscle around any model.
System distillation is the moat.
Own the Al + data flywheel.
3 Valued work per watt is the score.
, Maximize Al usefulness per unit of compute, latency, and risk
ap Wortd’: “air
ele I
: Introspection:
a
z Engineering the future of Al
OCR text:
PNT erg MRS
World's Fair Takeaways
The loop is the product.
System distillation is the moat.
Valued work per watt is the score.
ir Amaze AGI Lab
ef ' ile
ie
7 Engineering the future of Al
OCR text:
World'sFair
You are watching an AlEOnline Talk
Pre-Recorded torWorio'sFair2026.
WHATAFACTORYACTUALLYI!
The company isn't the machines.It's the knowledge.
OCR text:
World'sFair AEngine Pre-Recorded tor Worid'sFair 2026 You are watching an AIE Online Talk
FEEDITEVERYTHING
Years of quotes,drawings,schedules，emails.Our internet,not the internet.
OCR text:
AEnpin
World'sFair
You are watching an AJE Online Talk
Pre-RecordedforWorid'sFair2026.
THIRTY-SIXAGENTS,ONEJOBEACH
Athena
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.