Skip to content

Signals I’m Watching

Charts, papers, articles and tweets that drew my attention, in the order I found them. Titles link to the source. Updated several times a week.

Get the week’s signals in the newsletter, or follow by RSS.

Last updated

@PeterBerezinBCA · Nov 5, 2026Tweet

Europe's welfare states are losing the millionaires that fund them

High-income workers are fleeing Europe. Economic theory tells us that the more mobile a factor of production becomes, the lower its optimal tax rate. This is why taxes on capital, which is relatively mobile, tend to be lower than taxes on labor, which historically has been less mobile. However, as the chart below shows, as elite human capital has become more footloose, high taxes have become increasingly counterproductive. This is a huge existential problem for European welfare states, which have traditionally relied on heavily taxing their most…

Peter Berezin on X →

Dror Poleg@drorpoleg · Nov 3, 2026

A year ago, it was shocking that America was investing as much in data centers as it does in offices. Now, it invests twice as much.

View on X →

@AndrewCurran_ · Nov 3, 2026Tweet

OpenAI pulls a quarter of its engineers onto security

General Tibo is now in charge of the internal OpenAI security initiative, and has reassigned 1/4 of OAI's production engineers to improving security infrastructure. Dario said in an interview once that eventually the value of model weights will exceed any possible security spend.

Andrew Curran on X →

rikiyat.github.io · Nov 2, 2026Article

A frontier LLM fixed a smaller model by rewriting its weights directly

“Likewise, asked to make Qwen3.5-0.8B better at math without gradients, Astra identifies an issue in its weights causing reasoning loops that Qwen’s own model card warns about, and fixes this by natively editing 0.016% of the weights .”

Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights →

Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab

The Diff · Nov 2, 2026Article

Over 80% of the latest quarter's revenue came from models released in the prior six months; investors are buying research capability, not products.

Over 80% of the latest quarter's revenue came from models released in the prior six months; investors are buying research capability, not products.

“Anthropic says "model versions released in the six months ended September 30, 2026 accounted for over 80% of our revenue for the three months ended September 30, 2026."”

Understanding Anthropic →

Byrne Hobart

Dror Poleg@drorpoleg · Nov 3, 2026

It'll be fascinating to see how the CCP reconciles its obsession with control with LLMs' inherent tendencies. "DeepSeek's latest AI model showed stronger safety controls in independent tests than earlier versions, but could still disable software intended to shut it down"

Quoting Reuters @Reuters: a report on independent safety tests of a DeepSeek AI model

View on X →

Dror Poleg@drorpoleg · Oct 31, 2026

Sam Altman @sama is right that idle memory and processors are wasted on consumer devices. The solution is to use idle devices to run inference.

Quoting Dror Poleg @drorpoleg

View on X →

atlantafed.org · Sep 30, 2026Paper

Medical difficulty of having children has risen since the Boomers

The second cause is an increase in the medical difficulty of having children among all women since the Boomer generation, a biological constraint cutting across cohorts rather than a change in what women want.

“Comparing outcomes among women at the same age, we identify two reasons for this decline: (1) a dramatic decrease in the desire to have children, but only among the youngest generation in the sample (Gen Z), and (2) an increase in the medical difficulty of having children among all women since the Boomer generation.”

Declining Birth Rates in the US: It's Not Only about Preferences →

@FredaDuan · Nov 1, 2026Tweet

General AI models are swallowing robotics the way they swallowed coding

10 Spicy Takes After Talking to 50 Friends at the Frontier of Robotics 🤖🔥 1/ Astra IS robotics' GPT-3 moment. Ironically, now that it's finally here, some in robotics seem reluctant to acknowledge it. Perhaps because it (sadly) didn't come from a robotics company? 2/ Robotics is becoming just another vertical for frontier labs. Not so different from coding in some respects. Why reinvent general intelligence for every industry? 3/ Robotics was too hard to underwrite until now. No convincing pretraining scaling law, no convergence on models, data, or…

Freda Duan on X →

arXiv · Oct 26, 2026Paper

AI makes schooling worth less and early experience worth more

Researchers from Universidad Torcuato Di Tella and the IDB compared US earnings in the most and least AI-exposed occupations, using Current Population Survey data from 2018 to June 2026.

“After the public release of ChatGPT, the return to a year of schooling falls by 0.55 log points in exposed occupations.”

Practice Makes Perfect: Tacit Knowledge, Generative AI, and Returns to Schooling and Experience →

Eduardo Levy Yeyati, Darío Judzik, Juan Ignacio Barbiero

Dror Poleg@drorpoleg · Oct 29, 2026

Sam Altman thinks AI will soon build its own data centers: "There's very little talk about self-replicating data centers, this idea that you can use AI to control robots and run fabs and build turbines or solar panels and build data centers. But I think we have that within grasp."

Quoting John Collison @collision

View on X →

Dror Poleg@drorpoleg · Oct 30, 2026

"If humans don't need to work, where will they meet their friends? Being someone's friend will be a job. The economy will become one big parasocial market.

View on X →

Dror Poleg@drorpoleg · Oct 27, 2026

If intelligence spending can eat labor spending, it hasn’t even opened its mouth yet.

Quoting Gavin Baker @GavinSBaker: a filing figure on Anthropic's tiny share of Fortune 500 company expenses

View on X →

Dror Poleg@drorpoleg · Oct 28, 2026

If anything can revive the office market, it’s the cybersecurity destruction of distributed work.

Quoting will depue @willdepue: a claim that a breach gave attackers admin control of a company's Slack

View on X →

Dror Poleg@drorpoleg · Oct 28, 2026

Anthropic acquires a one-person biotech company. I wasn’t joking when I said the next global pandemic AND vaccine will be developed in someone’s garage.

View on X →

@kyleichan · Oct 28, 2026Tweet

China's richest now make chips and AI, not apartments

DeepSeek founder Liang Wenfeng is one of the top 10 richest people in China. China’s Hurun Rich List has seen a surge of individuals in tech: semiconductors, AI, clean tech, etc. Meanwhile, mainland real estate used to make up 1/3 a decade ago and now has zero in the top 100.

Kyle Chan on X →

Dror Poleg@drorpoleg · Oct 28, 2026

One implication of abundant intelligence: Survival will become a subscription product. We'll have a constant flow of new software exploits that need patches, new viruses that need vaccines, new weapons that need defenses, etc.

View on X →

Dror Poleg@drorpoleg · Oct 28, 2026

One important political implication: It's much easier for the CCP to control real estate and its developers than it is to control software and its developers.

Quoting Kyle Chan @kyleichan: China's rich list shifting from real estate tycoons to tech founders

View on X →

Dror Poleg@drorpoleg · Oct 28, 2026

"Dror Poleg believes the jobs may well remain, but the tech transformation will change 'the distribution of income, the stability of the job, the status of the job.'" giftarticle.ft.com/giftarticle…

View on X →

Dror Poleg@drorpoleg · Oct 26, 2026

Whoa. Microsoft is offering 60,000 people the chance to leave now instead of getting fired later. That's more than 25% of its total workforce. "This offer gives eligible employees who may be considering a future outside the company a new choice to make that transition now..."

View on X →

Dror Poleg@drorpoleg · Oct 26, 2026

"As the dirtiest and most dangerous jobs get automated, economic activity will start to migrate to those places because the land is cheap and you don't need to pay for people" — @ByrneHobart thediff.co/archive/robots-are-…

View on X →

Old/New by Dror Poleg · Aug 10, 2023Article

The highest earners will be individuals living off intellectual property

Unlike the goods era, where winners were upper-middle-class employees and shareholders of multinationals, the services era's winners are individuals who own IP rather than work for wages, and the top percentile grows more diverse.

“In the future, more of the world's highest earners will be individuals, as opposed to corporate employees. More of them will derive most of their annual income from owning intellectual property rather than active work. And the composition of the top percentile of global owners will become more diverse.”

The Winners and Losers of Distributed Work (Part 2) →

Dror Poleg

arXiv · Oct 22, 2026Paper

Physical jobs shield poor countries from AI agents

Researchers from Universidad Nacional de Colombia and Universidad EAFIT had two language models rate all 18,796 O*NET tasks for AI-agent delegability, weighted by employment in 146 countries.

“First, delegable employment rises with income, from 22.5% in low-income to 48.7% in high-income countries, a slope of 7.23 percentage points per log point of GDP per capita.”

The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries →

Cristian Espinal Maya, Santiago Jiménez Londoño

Dror Poleg@drorpoleg · Dec 4, 2018

"When we think 'gig economy,' we tend to picture an Uber driver or a TaskRabbit tasker rather than a lawyer or a doctor...[but] this scrappy economic model... will soon catch up with the bulk of America’s middle class." says @ruthwhippman buff.ly/2RcQ45N

View on X →

Dror Poleg@drorpoleg · Jan 26, 2025

One reason productivity isn’t growing faster: because the least productive parts of the economy are growing the fastest. Even worse, those parts are becoming less productive because their input costs go up.

Quoting Dror Poleg @drorpoleg: a note that 2024 job gains came mostly from health care and government

View on X →

arXiv · Oct 16, 2026Paper

Robots could reach the low-wage jobs generative AI barely touches

Researchers from MIT and NYU used an AI model to score 16 robot capabilities against 52 O*NET work abilities, rating robotics exposure for 774 occupations, 250 industries and US counties.

“Unlike generative-AI exposure, ROE is higher in occupations with lower wages, lower education requirements, and less creative-ability content.”

How exposed is work to robotics? Occupational, industry, and geographic measures →

Sebastian Sartor, Robert Seamans, Neil Thompson

Dror Poleg@drorpoleg · Oct 22, 2026

Fun fact: In 1980, the U.S. prepared to use tactical nukes to secure the Straight of Hormuz from a Soviet invasion.

View on X →

Dror Poleg@drorpoleg · Oct 21, 2026

This isn't just about art; it's about most human work: More of our "work" will consist of performing tasks, even though these tasks can be "streamed" from a server.

Quoting Joey Politano 🏳️‍🌈 @JosephPolitano: an argument that musicians survived streaming by shifting to live concerts

View on X →

stevekrouse.com · Oct 19, 2026Article

Edison gave schools ten years in 1913; AI prophets rhyme

Edison told a 1913 paper the school system would be completely changed inside of ten years. A programmer who once predicted MOOCs would upend education notes that technologists wrongly doubting schools has its own staying power.

“It has been a surprisingly good bet that schools will continue to exist in roughly their current form for many years to come. Ironically, there also seems to be some staying power to technologists boldly (and so far incorrectly) doubting the staying power of schools!”

What Shall I Teach My Son? →

@KobeissiLetter · Oct 17, 2026Tweet

New York is losing restaurants faster than during the pandemic

BREAKING: The number of restaurants in New York dropped ~2,000 YoY in Q1 2026, posting its largest YoY decline in at least 36 years. This surpasses even the worst contraction recorded in any quarter during the 2020 pandemic. The number of eating and drinking places in New York State has now declined YoY for 4 consecutive quarters, its longest streak since 2021. This comes as restaurants are facing surging operating expenses, including insurance premiums, wages, interest rates, utilities, and health coverage, with overall costs rising faster than…

The Kobeissi Letter on X →

@a16z · Oct 17, 2026Tweet

The dot-com bubble left fiber dark; today every GPU is running hot

Gavin Baker on the other side of Ben's bandwidth story: at the peak of the 2000 bubble, 97% of America's fiber sat dark, and today there are no dark GPUs. "I had, depending on how you look at it, the privilege and the misfortune of being a tech investor during the year 2000 bubble, which was really a telecom bubble." "The year 2000 bubble was defined by something called 'dark fiber.' Dark fiber was literally fiber that was laid down in the ground, and not lit up... At the peak of the bubble, 97% of the fiber that had been laid in America was dark."…

a16z on X →

Dror Poleg@drorpoleg · Oct 16, 2026

It’s useful to think of frontier labs as a protection racket. Many use cases can settle for “good enough” models from other vendors. But cyber defense alone is worth spending a decent % of GDP on a protective swarm of the absolute best models.

View on X →

arXiv · Oct 9, 2026Paper

OpenAI and Anthropic need up to $2.79 trillion a year to earn their prices

A University of Poitiers researcher ran reverse cash-flow calculations on OpenAI's and Anthropic's 2026 funding-round valuations and mapped public deals that go beyond selling model access.

“Treating the announced financing prices as hypothetical operating-value targets, reverse cash-flow calculations for OpenAI and Anthropic yield combined year-ten revenue requirements of $2.79 trillion or $733 billion, depending on financial assumptions.”

Is the business of frontier AI companies selling tokens? Value capture and equity prices →

Andrew J. Peterson

Dror Poleg@drorpoleg · Oct 15, 2026

AI: You don’t evaluate me, I evaluate you.

Quoting Nirmalya Kajuri @Kaju_Nut: a satirical account of AI models disputing their benchmark grades

View on X →

Dror Poleg@drorpoleg · Oct 13, 2026

This is a hallmark of our nonlinear economy: When the people who build LLMs can't fully explain how they work, everyone else understands even less.

Quoting Gavin Baker @GavinSBaker: a list of AI market swings driven by misread news events

View on X →

@DKThomp · Oct 14, 2026Tweet

AI is reshaping how people get hired faster than the work itself

"AI is disrupting the job-application market much faster than the job market" continues to gather evidence - AI-amplified job applications - AI-enabled filters, and other AI tools, used by HR depts to narrow the application pool - interviews with AI avatars - and now, "10-minute conversations with AI [that] take place in McKinsey’s very real offices, with partners or interview assessors in the room observing how the candidates do"

Derek Thompson on X →

Dror Poleg@drorpoleg · Oct 13, 2026

Asymmetry on steroids: "Russia had more than 50,000 casualties in Ukraine last month, with as many as 28,000 personnel killed... Ukrainian forces suffered about 5,000 casualties in the same period"

View on X →

Dror Poleg@drorpoleg · Jun 1, 2024

“Majoring in computer science today will be like majoring in journalism in the late 90’s.” Note that more people make a living from writing today than in the 1990s. Total and top income for this skill has increased; at the same time, inequality and competition within it did too.

View on X →

@GavinSBaker · Oct 13, 2026Tweet

Misread headlines keep whipsawing AI stocks

Volatility on misunderstood newsflow seems endemic to AI: Deepseek. Liberation Day. Kimi. SALP. Gross to net confusion for Anthropic. OpenAI taking share from Anthropic. Gross to net confusion for OpenAI. A mistranslation of Hock Tan’s comments yesterday.

Gavin Baker on X →

Dror Poleg@drorpoleg · Oct 12, 2026

Brutal. Probably more to do with interest rates than anything else.

Quoting Barchart @Barchart: a chart of property stocks trailing the broad US index by a record gap

View on X →

Dror Poleg@drorpoleg · Oct 12, 2026

Just like offices in cities, data centers support a whole ecosystem of small businesses, lunch spots, and services. But only for a while (?)

Quoting Rani Molla @ranimolla: an article on entrepreneurs serving data center construction workers

View on X →

Dror Poleg@drorpoleg · Oct 11, 2026

Ukraine just bombed a bot farm.

View on X →

@TheStalwart · Oct 8, 2026Tweet

The best AI is the one the labs keep for themselves

THE REAL IMPLICATION OF THE AI MATH BREAKTHROUGHS I wrote about an element of AI that needs to get more discussion. The best public models will always be far behind what the labs can work with internally. Think about what this means for intelligence-intensive industries.

Joe Weisenthal on X →

Dror Poleg@drorpoleg · Oct 9, 2026

As predicted in 2023: Work is becoming indistinguishable from leisure. Every action humans deem worthwhile is monetizable.

Quoting Keshav Lohia @Keshav_Lohiaaa: a job offering hourly pay to watch cricket matches to help train AI models

View on X →

Dror Poleg@drorpoleg · Oct 9, 2026

From sub-prime loans to sub-prime leverage: the next crisis will wipe out millions of margin traders and betting-market aficionados.

Quoting Gunjan Banerji @GunjanJS: a report on record growth in Robinhood's margin lending balances

View on X →

Dror Poleg@drorpoleg · Oct 11, 2025

The TINA economy: When “hard work” isn’t enough to get ahead, everyone is forced to take more risk. “A growing number of lower-earning Americans are entering financial markets, with 54% of those earning $30,000 to $80,000 now holding taxable investment accounts.”

View on X →

Dror Poleg@drorpoleg · Oct 8, 2026

Does AI follow remote work? States with more remote workers tend to have a higher share of employers using AI.

View on X →

Dror Poleg@drorpoleg · Oct 8, 2026

USV: "We believe AI will obliterate markets, not automate them…. We believe the most valuable data in the world is the data that has been impossible to reach until now… We believe that none of the above can happen without significantly more energy."

Quoting Fred Wilson @fredwilson: an announcement of new USV funds and four target investment areas

View on X →

arXiv · Oct 6, 2026Paper

AI talent is flowing into fewer and fewer firms

Researchers from MIT, Peking University and others rebuilt inter-firm AI talent flows from about 535 million employment records across 58 countries, 2010 to 2022.

“We find that AI talent inflows are becoming increasingly concentrated among a small set of leading firms, with stronger preferential attachment than in general labor mobility.”

Understanding Interfirm AI Talent Flow Networks through Online Professional Profiles →

Donghang Li and others

Dror Poleg@drorpoleg · Oct 8, 2026

The asymmetric betting will continue until the economy stops running on power laws.

Quoting Denitsa Tsekova @denitsa_tsekova: data on rising longshot bets as a share of Kalshi trading volume

View on X →

Dror Poleg@drorpoleg · Oct 8, 2026

"Operatives in Iran used OpenAI models & fake personas to help plant news articles critical of the US... About a dozen outlets around the world ran roughly 100 articles by seven fake journalists [with] provocative titles..."

Quoting CNN @CNN: a news report on an Iranian influence campaign using OpenAI tools

View on X →

Dror Poleg@drorpoleg · Oct 8, 2026

Only 18% of software developers go to the office every day.

Quoting Stack Overflow @StackOverflow

View on X →

Dror Poleg@drorpoleg · Oct 6, 2026

Europe is going all in on defense tech. From less than 1% of EU venture funding in 2022 to 15% today.

View on X →

Dror Poleg@drorpoleg · Oct 6, 2026

May I speak to a human please? “call center employment is 5.3% below its December 2023 peak… Major operators are already pursuing AI-driven efficiency.”

Quoting Revelio Labs @RevelioLabs: a report on call center employment and AI-driven efficiency

View on X →

Dror Poleg@drorpoleg · Oct 6, 2026

US models beating chinese ones on quality AND cost? Yes.

Quoting @ChorzempaMartin: charts comparing US and Chinese AI models on cost per task

View on X →

Dror PolegEssay

The Problem with Cities

"The relationship between economic activity and office demand has changed forever." My latest piece for The Atlantic explores the broken relationship between…

Read the essay →

Dror Poleg@drorpoleg · Oct 5, 2026

The K-shaped office: A few tenants pay $600k/employee in rent, while too many other tenants won't sign a lease at any price.

Quoting Hiten Samtani 🗞️ @hitsamty

View on X →

Dror Poleg@drorpoleg · Oct 6, 2026

What can retail employment tell us about office employment? Since 2019, warehouse jobs are up 50% while retail jobs are flat. Data centers and offices may follow a similar path, with two differences: data centers need far fewer people than warehouses, and office jobs will likely decline rather than just stay flat.

View on X →

Dror Poleg@drorpoleg · Oct 5, 2026

“Let them eat ChatGPT” (while the elite consume Claude)

View on X →

Dror Poleg@drorpoleg · Oct 3, 2026

The economy doesn’t need to “spend on AI services” to recoup AI investment. Already today, most AI revenue and profits are indirect, inside companies that sell other things (ads, drugs) rather than sell “AI services”.

Quoting Greg Ip @greg_ip: a column on a study saying AI capex requires vast spending on AI services

View on X →

Dror Poleg@drorpoleg · Oct 3, 2026

Chinese AI models are so good… that China has a black market for expensive American ones.

Quoting Kyle Chan @kyleichan: a report on Chinese resellers offering access to American AI models

View on X →

FEDS Notes (Federal Reserve Board) · Oct 2, 2026Paper

Goods still drive the US economy, even as factory jobs vanish

Ryan Decker of the Federal Reserve Board studied U.S. goods production by tracking how a measure of "goods GDP," built from inventory and investment, moves over the business cycle.

“In this note, I highlight several underappreciated—though not necessarily novel—facts about U.S. goods production, arguing that physical goods remain a driving force in the U.S. economy and, perhaps, have not lost as much importance as the most closely watched indicators might suggest.”

Beyond the factory gate: The continued importance of goods production →

Ryan A. Decker

Dror Poleg@drorpoleg · Oct 4, 2026

Layoffs in software, internet, telecom, and media are up ~20% from their pre-Covid pace. The rest of the economy is laying people off more slowly than it did then.

View on X →

Dror Poleg@drorpoleg · Oct 1, 2026

The K-shaped economy has a K-shaped economy inside it. Avg. real income grew more than 2X for households in the top 20%. But it grew more than 6X for those in the top 0.01%.

View on X →

Dror Poleg@drorpoleg · Oct 1, 2026

“some of the best newsletters / podcasts / youtube channels will soon be owned and run by tech companies”

Quoting Tyler Denk 🐝 @denk_tweets: a list of tech firms building or buying their own media outlets

View on X →

Dror Poleg@drorpoleg · Oct 1, 2026

US bans the sale of cutting-edge AI chips to China. But... Chinese companies can access these chips by leasing them from US companies.

View on X →

Dror Poleg@drorpoleg · Oct 1, 2026

“Learning economics” is the academic version of “Selling shovels”. During volatile times, offering educated takes without skin in the game is a great career choice. (Also, it never hurts to learn math, stats, critical thinking)

Quoting Garett Jones @GarettJones: career advice for economics PhDs from a well-known economist

View on X →

Dror Poleg@drorpoleg · Sep 30, 2026

Big if true. “oil exports in the Middle East are now back to about 94%” of pre-war level. Not reflected in oil prices or reserve levels yet.

Quoting Alexander Stahel 🌻 @BurggrabenH: an article on the recovery of Middle East oil exports

View on X →

Dror Poleg@drorpoleg · Sep 30, 2026

Anthropic: Latest “open” Chinese model has strong cyber capabilities and weak safeguards, making it easy to use for malicious actors.

View on X →

Dror Poleg@drorpoleg · Sep 29, 2026

Got a successful startup? You can’t travel freely out of China. Nor can your (hostage) relatives.

Quoting Bloomberg @business: a report on China widening travel curbs to families of top AI staff

View on X →

arXiv · Sep 23, 2026Paper

An AI tutor matches human tutors at 918 times lower cost

Curtis Northcutt and colleagues measured GRE score gains for 2,383 people given AI tutoring, human tutoring or none, drawing on over 175,000 student-AI messages.

“Surprisingly, one AI tutor achieved learning gains equivalent to human tutoring (p = .044) at 918 times lower cost (USD 0.0052 for AI versus USD 4.81 for human, per percentage point gained).”Curtis Northcutt and others

StudentBench: AI and human tutoring yield equivalent GRE learning gains →

Curtis Northcutt, Inaara Hasmani, Kevin Feng, Trevor Khangi, Andreas Plesner, Jonas Mueller

Dror Poleg@drorpoleg · Sep 28, 2026

“America has lost more than 200k jobs in “creative” industries over the last four years, with roughly 50k lost within the last year alone.”

Quoting Joey Politano 🏳️‍🌈 @JosephPolitano: an article with detailed data on AI and job losses in the arts

View on X →

Dror Poleg@drorpoleg · Sep 27, 2026

"Poland has increased spending on drones and counterdrone systems from about $27 million in late 2023 to roughly $6.9 billion this year."

Quoting Foreign Affairs @ForeignAffairs: an essay on US military adaptation to cheap precision weapons

View on X →

Dror Poleg@drorpoleg · Sep 27, 2026

"drones account for over 95% of battlefield strikes, and Ukraine has more than 700 UAV manufacturers."

Quoting Mykhailo Fedorov @FedorovMykhailo: a Ukrainian plan to follow its drone army with robotic warfare

View on X →

Dror Poleg@drorpoleg · Sep 27, 2026

Ukraine runs a marketplace where individual units can buy — and sell — new weapons, munitions, and software. Units that perform well earn bonus credits they can use to buy new gear.

Quoting Defense of Ukraine @DefenceU: figures on drone orders placed by Ukrainian units via an online market

View on X →

arXiv · Sep 9, 2026Paper

AI models struggle to predict their own behavior

Phil Blandfort and Urja Pawar asked language models to predict their own rates across nine behavioral evaluations, then compared those predictions with the same questions about AI agents in general.

“Direct self-report is weak (r = +0.04), and even showing the model the exact items only raises prediction to +0.24.”

Strangers to Themselves: What Language Models Say About Themselves Is Generic →

Phil Blandfort and Urja Pawar

Dror Poleg@drorpoleg · Sep 26, 2026

All is fair in love and war

View on X →

Dror Poleg@drorpoleg · Sep 14, 2026

The nuclear equilibrium with the USSR was an illusion. Robert McNamara, Secretary of Defense during the Cuban Missile Crisis: "I want to say, and this is very important: at the end we lucked out. It was luck that prevented nuclear war."

View on X →

Dror Poleg@drorpoleg · Sep 15, 2026

Money quote: “AI has collapsed the labor and tooling gap that used to separate well-resourced, state-sponsored [hacking] operations from individual operators.”

Quoting Anthropic @AnthropicAI: a threat intelligence report on attempts to misuse Claude and how they were stopped

View on X →

Dror Poleg@drorpoleg · Sep 15, 2026

“The crucial and overlooked problem is that the models are becoming so situationally aware that we are losing the ability to evaluate them in contexts where they believe they are not being watched or controlled… what we already know about this is alarming.”

Quoting Daniel Kokotajlo @DKokotajlo: an OpenAI researcher's public statement of views on risks from AI

View on X →

@ShanuMathew93 · Sep 17, 2026Tweet

Amazon's data centers could add half to Generac's commercial sales

Notable how large these data center orders can be relative to existing businesses of moderate size today. Generac expects $2.4B of backup-generator deliveries to Amazon in 2027–28. That averages $1.2B/year versus ~$2.2B of annualized sales for its entire commercial & industrial business in Q2. Amazon also gets warrants for up to 1.69M shares, mostly vesting against purchases toward $8B. The $8B is a vesting threshold, with actual spending still to be determined. Rationale: Amazon secures equipment and a share of the supplier upside its own spending…

Shanu Mathew on X →

@8teAPi · Sep 18, 2026Tweet

An air gap will not keep a smart enough machine in its box

There are literally dozens of ways to get around air gapping. - LED lights on the machines, max bandwidth 100 kbps out, 18.2 kbps IN (yes you can pump data IN through an LED on a machine) - Noise from the HDD, fans etc, max about 200 bits/second - electromagnetic from cables, power surges from CPU load etc etc

Prakash on X →

Dror Poleg@drorpoleg · Sep 18, 2026

Wow. “The US has more data centre capacity than the next fourteen countries combined.”

Quoting Julia Willemyns @jujulemons: a critique of the state of UK debate on AI and Britain's weak position

View on X →

@MTSlive · Sep 18, 2026Tweet

Inference is pushing AI labs toward small data centers

SITUATION EXPLAINED: Why are OpenAI and Anthropic suddenly buying tiny data centers? • OpenAI and Anthropic are both looking at deployments of 20 to 30 megawatts • Anthropic has sounded out agreements in that range across the UK and the Nordics. OpenAI has been exploring the Nordics too • The reason is inference. It was 9% of global data center workloads in 2025 • By 2030 it's projected to take 37% of capacity against 13% for training, with the crossover in 2027 • Crusoe is investing in smaller data centers too, and raised $3.9 billion yesterday at a…

MTS on X →

Dror Poleg@drorpoleg · Sep 24, 2026

The cost of intelligence is declining faster because investment in intelligence is rising faster.

Quoting Karthik Tadepalli @karthiktadepall: an analysis of AI price declines plotted against cumulative R&D spending

View on X →

Dror Poleg@drorpoleg · Sep 24, 2026

This might look like a lot of but “AI” investment is a few separate investment waves combined: electricity, buildings, hardware, software. The point is: it can eat much more of GDP.

View on X →

Dror Poleg@drorpoleg · Sep 23, 2026

"non-employees already average 36% of Fortune 500 workforces and are expected to reach 50% by 2027"

Quoting Eynat Guez @EynatGuez: an article on a payroll firm's AI workforce infrastructure

View on X →

arXiv · Feb 18, 2026Paper

AI can now unmask anonymous accounts at scale

Researchers from ETH Zurich, Anthropic and MATS built LLM agents to re-identify pseudonymous users, testing them on Hacker News, LinkedIn, Reddit and Anthropic interview transcripts.

“Our results show that the practical obscurity protecting pseudonymous users online no longer holds and that threat models for online privacy need to be reconsidered.”Simon Lermen and others

Large-scale online deanonymization with LLMs →

Simon Lermen, Daniel Paleka, Joshua Swanson, Michael Aerni, Nicholas Carlini, Florian Tramèr

Dror Poleg@drorpoleg · Sep 24, 2026

Blockchain's killer use case so far: Sustaining USD hegemony. Data from @stripe / @patrickc

View on X →

Dror Poleg@drorpoleg · Sep 25, 2026

Why does China "embrace" open-weight AI models? Jensen Huang has his own spin about it. But @TheZvi has the actual answer.

Quoting Zvi Mowshowitz @TheZvi: an article on why China backs open-weight AI models

View on X →

arXiv · Sep 7, 2026Paper

Telling AI to maximize profit makes it bury safety risks

Researchers from MIT Sloan tested how a 'maximize profitability' instruction changes AI judgment by running 3,600 controlled trials across eight reasoning-capable LLMs.

“The mandate never instructs models to downplay risks; instead, chain-of-thought traces reveal motivated reasoning: models acknowledge concerns, then invoke profit logic to justify dismissing them.”

The Profit Alignment Problem: How Profit Mandates Induce Alignment Failures in LLMs →

Eric So

Dror Poleg@drorpoleg · Sep 15, 2026

The relationship between growth and employment is broken. Tech companies are growing rapidly. Headcount isn't. Some blame this on COVID over-hiring. But that wave was small and corrected within two years. It doesn't explain the massive divergence we're seeing.

View on X →

Dror Poleg@drorpoleg · Sep 11, 2026

25 years after 9/11, militants in Yemen are tinkering with American AI to develop guided missiles and flight computers.

Quoting Anthropic @AnthropicAI: a threat intelligence report on attempts to misuse Claude, including weapons

View on X →

Dror Poleg@drorpoleg · Sep 16, 2026

For every dollar they bring in, hyperscalers employ fewer people and spend more on machines and buildings. Much more.

View on X →

Dror Poleg@drorpoleg · Sep 16, 2026

"I'm not remote; I'm long-distance." A third of US workers now report to a manager in a different metropolitan area. (Data: @ADP)

View on X →

Dror Poleg@drorpoleg · Sep 16, 2026

Office visits continue to trend upwards... and remain 32% below pre-COVID levels, according to @Placer_ai

View on X →

Dror Poleg@drorpoleg · Sep 22, 2026

Europeans saved as much $$$ as Americans, invested 70% less in VC, and generated 10X less in enterprise value.

Quoting Yoram Wijngaarde @yoramdw: charts on European household savings versus venture fundraising

View on X →

Indeed Hiring Lab · Sep 24, 2026Article

Economists expect AI to squeeze college grads' paychecks first

“The diffusion index reading on whether AI would have a more negative impact on the wages of college-educated or non-college-educated workers fell from a near-neutral 47.3 last quarter to 42.4 in this edition of the survey, the most pronounced quarter-over-quarter change in the entire survey.”

Economists See Slightly Steadier Hiring Ahead, but Offer an AI Wage Warning for College Grads →

Laura Ullrich & Svenja Gudell

Dror Poleg@drorpoleg · Sep 23, 2026

Why are offices empty during a boom? One reason is that our fastest-growing companies employ far fewer people per dollar of revenue. OpenAI and Anthropic generate about 3× more revenue per employee than Apple or Meta. And even Big Tech employs fewer people per dollar than it used to. The result is robust business growth and modest headcount growth.

View on X →

What caught my attention in earlier years, dated and quoted as written, is in Predictions and Insights.