Mime-Version: 1.0 Content-Type: multipart/alternative; boundary="9ca1af916137ee978587c0b748a5bc694230b7c9f260789bd4458503721a" Subject: =?UTF-8?q?=F0=9F=94=AE_Sunday_edition_#533:_AI_masters_math;_war_scenario?= =?UTF-8?q?s;_reinforcement_learning_critique;_Shanghai's_robot_movers++?= From: "Azeem Azhar, Exponential View" To: Hidden Recipient Date: Sun, 20 Jul 2025 03:29:17 +0000 X-Hiring: We are hiring, reach out at header-hacker@emailshot.io X-EmailShot-Signature: qtcP0S7vlD4EpxX-15znZU4EA_t_ZqDaOREtxzEanZdanr_ykQmYrG69SypJw5rhghixHUZ80wryEMOckuILDw== --9ca1af916137ee978587c0b748a5bc694230b7c9f260789bd4458503721a Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: quoted-printable View this post on the web at https://www.exponentialview.co/p/ev-533 Hi all, Welcome to our Sunday edition, when we take the time to go over the latest = developments, thinking and key questions shaping the exponential economy.= =20 Thanks for reading! Azeem For the most important themes of the week, check out our daily edition: Monday: Two Moonshots =E2=80=94 one hit, one miss [ https://substack.com/re= direct/b7d39f17-cd24-4304-af0e-bdfb07eb984e?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFO= bqArDfP822j-jnN48_jCfgM3m0rbAsF0l24U ] Tuesday: The Pentagon goes all-in on AI [ https://substack.com/redirect/726= 0cc82-a867-4b8f-9e5e-4afa62f10178?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP822= j-jnN48_jCfgM3m0rbAsF0l24U ] Wednesday: US doubles down on data and energy [ https://substack.com/redire= ct/3fa1fb07-6895-46af-92c7-617521d3bd01?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqAr= DfP822j-jnN48_jCfgM3m0rbAsF0l24U ] Thursday: AI=E2=80=99s inner monologue goes public [ https://substack.com/r= edirect/cb659fe9-c472-4e1b-b31f-3edb6b0f8254?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uF= ObqArDfP822j-jnN48_jCfgM3m0rbAsF0l24U ] Friday: OpenAI agents [ https://substack.com/redirect/3103ebf4-6d23-4234-b3= f4-a2645989df8f?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP822j-jnN48_jCfgM3m0rb= AsF0l24U ] If you=E2=80=99d rather stick to the weekly edition only, you can change yo= ur email preferences [ https://substack.com/redirect/596746d7-62ea-47b2-882= 6-bb91e4183abf?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP822j-jnN48_jCfgM3m0rbA= sF0l24U ] to opt-out of the daily cadence. The Tao of the Turing A new OpenAI model achieved gold-medal performance [ https://substack.com/r= edirect/d20ee0a4-6a9d-4a77-b951-4b76971ef776?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uF= ObqArDfP822j-jnN48_jCfgM3m0rbAsF0l24U ] in the International Math Olympiad = (IMO), the world=E2=80=99s most prestigious math competition. They used a = =E2=80=9Creasoning LLM that incorporates new experimental general-purpose t= echniques=E2=80=9D and the AI worked under the same time constraints as hum= ans with no access to tools.=20 The model thinks=E2=80=A6. for a long-time, for hours, in fact, according t= o Noam Brown, an OpenAI researcher. AI progress in math has been much faste= r than anyone expected, perhaps years faster than we might have estimated o= nly a few years ago. [ https://substack.com/redirect/4ca380eb-92bf-4cec-9d= 52-41f243f70716?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP822j-jnN48_jCfgM3m0rb= AsF0l24U ] This matters because the IMO tests creative reasoning beyond rote computati= on and requires detailed, logical proofs, demanding original arguments. Pro= blem designers intentionally seek [ https://substack.com/redirect/abbd0d36-= 7fbb-4934-96e8-747324872221?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP822j-jnN4= 8_jCfgM3m0rbAsF0l24U ] =E2=80=9Celegant, deceptively simple-looking problem= s which nevertheless require a great deal of ingenuity.=E2=80=9D Is this a = system that can start to mimic or exceed expert human creativity in an impo= rtant domain? Mathematics is the universal language for describing the physical world wit= h applications across every domain from finance, the economy, climate, phys= ics, engineering, optimization, biology. And, of course, in improving AI sy= stems.=20 This could be quite the milestone=E2=80=A6=20 Or could it? Terence Tao, the =E2=80=9CMozart of Math=E2=80=9D famed for hi= s ability to excel across disciplines, and a measured optimist about the po= tentials of AI urges caution [ https://substack.com/redirect/87098702-1302-= 46c5-856f-e89f30bd82cf?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP822j-jnN48_jCf= gM3m0rbAsF0l24U ]: =E2=80=9Cin the absence of a controlled test methodology= that was not self=E2=80=91selected by the competing teams, one should be w= ary of making apples=E2=80=91to=E2=80=91apples comparisons =E2=80=A6 betwee= n such models and the human contestants.=E2=80=9D=20 Is it more a case of quod non erat demonstrandum? Tell me in the comments.= =20 Six paths to a war A new paper identifies six pathways through which advanced AI [ https://sub= stack.com/redirect/c4e9902a-e4d1-47da-ace8-bd59eb936628?j=3DeyJ1IjoiM2dmeXZ= tIn0.xu76uFObqArDfP822j-jnN48_jCfgM3m0rbAsF0l24U ] might increase the risk = of major war. One of the most dangerous pathways is purely human =E2=80=93 if national le= aders come to believe that losing the race to AGI would significantly weake= n their global standing, militarily or economically, they may take drastic = action. Suppose the US or China believes its rival is nearing a decisive breakthrou= gh; it may be tempted to take preventive action through sabotage, cyberatta= cks, or even military strikes to delay or derail the competitor=E2=80=99s p= rogress. One of the risks here is that we may not agree on what AGI is or w= hat it looks like; leaders might overreact to vague signs that a rival is c= lose to AGI. Their next step could lead to the point of no return. At the s= ame time, ambiguity about AGI=E2=80=99s exact implications could make them = hesitate. If this reminds you of the history of nuclear deterrence, you=E2=80=99re no= t wrong. See also: Nirit Weiss-Blatt highlights a UK AI Safety Institute paper critiquing cur= rent research on AI =E2=80=9Cscheming=E2=80=9D for significant methodologic= al flaws [ https://substack.com/redirect/f14b837d-9992-4fb8-a781-90e6babcb2= cf?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP822j-jnN48_jCfgM3m0rbAsF0l24U ]. Cosmos Institute argue this week that in building AI we must return to fun= damental questions [ https://substack.com/redirect/2f1cb74e-18dc-4ae7-acb0-= b09824ad1d86?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP822j-jnN48_jCfgM3m0rbAsF= 0l24U ] about human flourishing, not just engagement metrics =E2=80=93 they= call for a philosopher builder.=20 Harry Law urges that we need more productive critiques of AI from academia= than calling it a =E2=80=9Cbullshit generator=E2=80=9D [ https://substack.= com/redirect/94ecaa06-9ca0-4d80-89a4-c22d47677455?j=3DeyJ1IjoiM2dmeXZtIn0.x= u76uFObqArDfP822j-jnN48_jCfgM3m0rbAsF0l24U ]. a ReaLity check In the past two weeks, both ChatGPT Agent [ https://substack.com/redirect/f= 925e136-2a95-401d-a5ab-cd5e9c13a6dc?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP8= 22j-jnN48_jCfgM3m0rbAsF0l24U ] and Grok 4 [ https://substack.com/redirect/c= 93442d9-21c0-4181-8a98-a9036aa972b4?j=3DeyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP8= 22j-jnN48_jCfgM3m0rbAsF0l24U ] debuted with heavy use of reinforcement lear= ning to deliver major performance gains over their base models. But, as And= rej Karpathy put it, =E2=80=9C[i]t doesn=E2=80=99t feel like the full story= =2E [ https://substack.com/redirect/97= 79d566-08c9-4506-acb8-18dda81f39b7?j=3D= eyJ1IjoiM2dmeXZtIn0.xu76uFObqArDfP822j-jnN48_jCfgM3m0rbAsF0l24U ]=E2=80=9D = RL is powerful but it hits diminishing returns as tasks grow longer and mor= e complex. We likely need new learning paradigms to push the frontier. RL hasn=E2=80=99t yet had its big breakthrough moment, [ https://substack.= com/redirect/4e06105d-c2d6-4ddb-a246-01dd37a70597?j=3DeyJ1IjoiM2dmeXZtIn0.x= u76uFObqArDfP822j-jnN48_jCfgM3m0rbAsF0l24U ]where it suddenly scales up to = produce truly general-purpose, flexible agents. But even if RL does improve= , there=E2=80=99s a fundamental limitation: it only learns from outcomes (= =E2=80=9Cdid this work or not?=E2=80=9D) rather than from the process itsel= f... Unsubscribe https://substack.com/redirect/2/eyJlIjoiaHR0cHM6Ly93d3cuZXhwb25= lbnRpYWx2aWV3LmNvL2FjdGlvbi9kaXNhYmxlX2VtYWlsP3Rva2VuPWV5SjFjMlZ5WDJsa0lqb3= lNRGt3TVRjME1qWXNJbkJ2YzNSZmFXUWlPakUyT0RZME9UazBOU3dpYVdGMElqb3hOelV5T1Rre= k1UQXhMQ0psZUhBaU9qRTNPRFExTWpreE1ERXNJbWx6Y3lJNkluQjFZaTB5TWpVeUlpd2ljM1Zp= SWpvaVpHbHpZV0pzWlY5bGJXRnBiQ0o5LjEzSHU3UHJ6eXZEM0N5cS1ITzdDblA5TVBxWDIzT1p= WT3N6V2xHUXB3OVkiLCJwIjoxNjg2NDk5NDUsInMiOjIyNTIsImYiOnRydWUsInUiOjIwOTAxNz= QyNiwiaWF0IjoxNzUyOTkzMTAxLCJleHAiOjIwNjg1NjkxMDEsImlzcyI6InB1Yi0wIiwic3ViI= joibGluay1yZWRpcmVjdCJ9.KopWwEfrYfWvXZBLDCppadIUFDF0R9Ptno9us7uLeKw? --9ca1af916137ee978587c0b748a5bc694230b7c9f260789bd4458503721a Content-Type: text/html; charset="utf-8" Content-Transfer-Encoding: quoted-printable 🔮 Sunday edition= #533: AI masters math; war scenarios; reinforcement learning critique; Sha= nghai's robot movers++3D""
= An insider’s guide to AI and exponential technologies
͏   =   ­͏     ­͏     ­͏= ;     ­͏     ­͏     &#= 173;͏     ­͏     ­͏   &= #8199; ­͏     ­͏     ­͏=     ­͏     ­͏     = 73;͏     ­͏     ­͏   &#= 8199; ­͏     ­͏     ­͏ =     ­͏     ­͏     = 3;͏     ­͏     ­͏   = 199; ­͏     ­͏     ­͏ &= nbsp;   ­͏     ­͏     ­= ;͏     ­͏     ­͏   Q= 99; ­͏     ­͏     ­͏ &n= bsp;   ­͏     ­͏     ­= ͏     ­͏     ­͏   ̳= 9; ­͏     ­͏     ­͏ &nb= sp;   ­͏     ­͏     ­&= #847;     ­͏     ­͏    = ; ­͏     ­͏     ­͏ &nbs= p;   ­͏     ­͏     ­&#= 847;     ­͏     ­͏    = ­͏     ­͏     ­͏  = ;   ­͏     ­͏     ­= 47;     ­͏     ­͏     = ­͏     ­͏     ­͏  =   ­͏     ­͏     ­T= 7;     ­͏     ­͏     &= #173;͏     ­͏     ­͏   =   ­͏     ­͏     ­͏= ;     ­͏     ­͏     &#= 173;͏     ­͏     ­͏   &= #8199; ­͏     ­͏     ­͏=     ­͏     ­͏     = 73;͏     ­͏     ­͏   &#= 8199; ­͏     ­͏     ­͏ =     ­͏     ­͏     = 3;͏     ­͏     ­͏   = 199; ­͏     ­͏     ­͏ &= nbsp;   ­͏     ­͏     ­= ;͏     ­͏     ­͏   Q= 99; ­͏     ­͏     ­͏ &n= bsp;   ­͏     ­͏     ­= ͏     ­͏     ­͏   ̳= 9; ­͏     ­͏     ­͏ &nb= sp;   ­͏     ­͏     ­&= #847;     ­͏     ­͏    = ; ­͏     ­͏     ­͏ &nbs= p;   ­͏     ­͏     ­&#= 847;     ­͏     ­͏    = ­͏     ­͏     ­͏  = ;   ­͏     ­͏     ­= 47;     ­͏     ­͏     = ­͏     ­͏     ­͏  =   ­͏     ­͏     ­T= 7;     ­͏     ­͏     &= #173;͏     ­͏     ­͏   =   ­͏     ­͏     ­͏= ;     ­͏     ­͏     &#= 173;͏     ­͏     ­͏   &= #8199; ­͏     ­͏     ­͏=     ­͏     ­͏     = 73;͏     ­͏     ­͏   &#= 8199; ­͏     ­͏     ­͏ =     ­͏     ­͏     = 3;͏     ­͏     ­͏   = 199; ­͏     ­͏     ­͏ &= nbsp;   ­͏     ­͏     ­= ;͏     ­͏     ­͏   Q= 99; ­͏     ­͏     ­͏ &n= bsp;   ­͏     ­͏     ­= ͏     ­͏     ­͏   ̳= 9; ­͏     ­͏     ­͏ &nb= sp;   ­͏     ­͏     ­&= #847;     ­͏     ­͏    = ; ­͏     ­
=
Fo= rwarded this email? Subscribe here for more

🔮 Sunday edition #533: AI masters= math; war scenarios; reinforcement learning critique; Shanghai's robot mov= ers++

= Advanced AI capabilities, systemic risks, and infrastructure innovation
Azeem Azhar and Nathan Warren
<= /td>
∙
<= div class=3D"pencraft pc-reset color-paid-LmY0EP line-height-20-t4M0El font= -meta-MWBumP size-11-NuY2Zx weight-medium-fw81nC transform-uppercase-yKDgcq= reset-IxiVJZ meta-EgzBVA" translated=3D"" style=3D"list-style: none;font-s= ize: 11px;line-height: 20px;text-decoration: unset;color: rgb(94,73,217);ma= rgin: 0;font-family: 'SF Compact',-apple-system,system-ui,-apple-system,Bli= nkMacSystemFont,'Segoe UI',Roboto,Helvetica,Arial,sans-serif,'Apple Color E= moji','Segoe UI Emoji','Segoe UI Symbol';font-weight: 500;text-transform: u= ppercase;letter-spacing: .2px;">Preview

<= table class=3D"email-ufi-2-top" role=3D"presentation" width=3D"100%" border= =3D"0" cellspacing=3D"0" cellpadding=3D"0" style=3D"border-top: 1px solid r= gb(0,0,0,.1);border-bottom: 1px solid rgb(0,0,0,.1);min-width: 100%;"> 
<= /table>
3D""
= 3D""
3D""
<= /table>
<= img class=3D"icon" src=3D"https://substackcdn.com/image/fetch/$s_!5EGt!,w_3= 6,c_scale,f_png,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack.com= %2Ficon%2FNoteForwardIcon%3Fv%3D4%26height%3D36%26fill%3Dnone%26stroke%3D%2= 523808080%26strokeWidth%3D2" width=3D"18" height=3D"18" style=3D"border: no= ne;vertical-align: middle;max-width: 18px" alt=3D"">
<= tr>
=
 

Hi all,

Welcome to our Sunday edition, when we = take the time to go over the latest developments, thinking and key question= s shaping the exponential economy.

Thanks for reading!

=

Azeem

Up= grade to paid

For the most important themes of t= he week, check out our daily edition:

If you’d rather stick to the weekly ed= ition only, you can change your email preferences to opt-out of the daily caden= ce.

The Tao of the Turing

A new OpenAI model achieved gold-medal performance in the Internat= ional Math Olympiad (IMO), the world’s most prestigious math competit= ion. They used a “reasoning LLM that incorporates new experimental g= eneral-purpose techniques” and the AI worked under the same time cons= traints as humans with no access to tools.

Th= e model thinks…. for a long-time, for hours, in fact, according to No= am Brown, an OpenAI researcher. AI progress in math has been much faster th= an anyone expected, perhaps years faster than we might have estimated only a few years ag= o.

This matters because the IMO tests creative r= easoning beyond rote computation and requires detailed, logical proofs, dem= anding original arguments. Problem designers intentionally seek “elegant,= deceptively simple-looking problems which nevertheless require a great dea= l of ingenuity.” Is this a system that can start to mimic or exceed e= xpert human creativity in an important domain?

Mathe= matics is the universal language for describing the physical world with app= lications across every domain from finance, the economy, climate, physics, = engineering, optimization, biology. And, of course, in improving AI systems= =2E

This could be quite the milestone…

Is it more a case of quod non<= span> erat demonstrandum? Tell me in the comments.

Six paths to a war

A new paper identifies six pathways through which advanced AI might increas= e the risk of major war.

One of the most dangerous p= athways is purely human – if national leaders come to believe that lo= sing the race to AGI would significantly weaken their global standing, mili= tarily or economically, they may take drastic action.

Suppo= se the US or China believes its rival is nearing a decisive breakthrough; i= t may be tempted to take preventive action through sabotage, cyberattacks, = or even military strikes to delay or derail the competitor’s progress= =2E One of the risks here is that we ma= y not agree on what AGI is or what it=20= looks like; leaders might overreact to vague signs that a rival is close to= AGI. Their next step could lead to the point of no return. At the same tim= e, ambiguity about AGI’s exact implications could make them hesitate.=

If this reminds you of the history of nuclear deterrence, = you’re not wrong.

See also:

a ReaLity check=

In the past two weeks, both ChatGPT Agent and Grok 4 debu= ted with heavy use of reinforcement learning to deliver major performance g= ains over their base models. But, as Andrej Karpathy put it, “[i]t doesn’t fe= el like the full story.” RL is powerful but it hits diminis= hing returns as tasks grow longer and more complex. We likely need new lear= ning paradigms to push the frontier.

= RL hasn’t yet had its big breakthrough moment, where it suddenly scales u= p to produce truly general-purpose, flexible agents. But even if RL does im= prove, there’s a fundamental limitation: it only learns from outcomes= (“did this work or not?”) rather than from the process itself.= =2E.

Keep reading with = a 7-day free trial

Subs= cribe to Exponential Vie= w to keep reading this post and get 7 days of free access to the ful= l post archives.

Start trial

A subscription gets you:

<= td style=3D"font-weight: light;">Daily and weekly analysis on what matters = in AI & technology
3D""
3D""Access to $2,000+ in AI tools (Annual only)
3D""Networking and c= ommunity space
 
3D""3D"" --9ca1af916137ee978587c0b748a5bc694230b7c9f260789bd4458503721a--
=
3D""Like
<= /td>
3D""Comment
3D""Restack
 

&#= 169; 2025 EPIIPLUS1 Ltd
Copyright © 2023 EPIIPLUS1 Ltd= , All rights reserved
Unsubscribe

3D"Get3D"Start=