Only 13% of CIOs report substantial AI value, Gartner finds

Following our October 21 coverage of infrastructure financing, Gartner's CIO survey adds a measure of enterprise outcomes. Only 13% of CIOs say their enterprise has realized significant or transformational value from AI so far.

Apollo questions AI payoff beyond tech

Apollo · Torsten Slok

Slok argues that AI's financial payoff has yet to appear outside tech. His analysis compares revenue per worker and profit margins across the largest technology companies and the rest of the S&P 500.

Which infrastructure signals indicate weak demand?

The AI Daily Brief

The episode distinguishes construction and power delays from demand-driven cancellations. It proposes watching unbacked neocloud defaults, tenant credit deterioration, and hyperscaler cancellations caused by weak demand.

Turn production traces into regression tests

AI Engineer · Sophia Yakoubovitch, Arize

Yakoubovitch describes a prompt change that improved action safety by 17 points while reducing task completion by the same amount. Her regression loop uses production traces, baseline experiments, and offline evaluations to check changes before merge.

Kaggle reports problems reproducing 65 benchmarks

AI Engineer · Kaggle

The team reports score manipulation, inconsistent reasoning settings, and substantial variation between identical runs. Its proposed evaluation infrastructure separates grading from execution and records model calls and traces.

Keep numerical outputs in deterministic code

AI Engineer · Vinod Ganesh, Kepler

Ganesh describes a financial-analysis system where the model plans queries, code performs calculations, and post-processing removes unverified outputs. Displayed facts link back to their source documents.

A community effort to revise OpenAI's math write-ups

Terence Tao's blog

A post on Tao's blog introduces a collaborative effort to improve OpenAI's mathematical write-ups. It gives readers a place to contribute corrections and review the claims.

Laurie Voss: review becomes the next constraint

AI Engineer · Laurie Voss

Voss applies the Theory of Constraints to agent-assisted development: faster code generation puts pressure on review, then operations, and eventually problem definition. He argues that specifying a successful outcome remains a central engineering responsibility.

Dan Luu argues for more ambitious software

Dan Luu

Luu's essay asks engineers to be more ambitious about what they build.

Battle Drones: a game built across 321 agent commits

Simon Willison

Willison documents building a mobile real-time strategy game with Claude Code across 321 agent commits. He describes the iterations and the lessons he took from the process.

Investigate executed artifacts after a compromise

Nate B. Jones

Jones uses web-server vulnerabilities, weak certificate keys, and a supply-chain compromise to explain why checking source code alone can miss an intrusion. His checklist includes rotating secrets, auditing connected-app permissions, and investigating deployment and build histories.

How Cerebras serves trillion-parameter models

Cerebras

Cerebras describes the serving architecture it uses to deploy trillion-parameter frontier models.