Daily Reading for 2026-09-09 - Learning in Public

Daily Reading for 2026-09-09

William Glaser
10 mins to read
daily-reading
Home / Daily Reading / Daily Reading for 2026-09-09

Daily Reading for 2026-09-09

AI and the Millennium Prize Problems

OpenAI says an unreleased internal model produced a resolution to the Navier–Stokes existence and smoothness problem, one of the seven $1M Millennium Prize Problems, after roughly 88 hours of agent compute (about 130 billion output tokens, with Lean verification via GPT-6 Astra). The result is shadowed by controversy: NYU’s Tristan Buckmaster and Anthropic’s Levent Alpöge allege OpenAI’s team only launched after hearing rumors of their near-year-long collaboration, and Buckmaster’s public questions about training-data influence go unanswered. Terence Tao separately warns that AI is mining “good, fruitful open problems” non-renewably — the mere rumor that someone is working a problem can now trigger a massive AI-powered effort to flatten it first, pushing researchers toward secrecy and reversing centuries of open-science tradition. Together these pieces frame an industrializing race where the frontier is no longer about solving problems but about who gets to discover them.

Watch for:

  • Whether OpenAI’s Navier–Stokes proof survives peer review and formal verification
  • Any disclosure about how model training data and user sessions influence results
  • Shifts in preprint-sharing norms as rumor-triggered scooping becomes routine

Wardley signals:

  • Components: OpenAI unreleased frontier model, GPT-6 Astra (Lean formalization), Claude and Codex (GPT-5.6 Sol), Anthropic, Lean proof verification, Millennium Prize Problems, open-science norms

  • Evolution stage: AI formal-mathematics agents at Genesis→Custom — novel, poorly understood, rapidly improving; Lean verification tooling approaching Product; the problems themselves are a fixed, scarce resource

  • Doctrine: “A bias towards open” directly violated by the reverse-open-science incentive Tao describes; “Exploit the landscape” and “There is no core” by labs racing to industrialize open problems

  • Climatic patterns: Red Queen (no choice over evolution — every lab is forced into the race); Creative Destruction as AI effort disrupts academic discovery; higher-order systems (agents) creating new sources of value

  • Diffusion markers: rumor-triggered effort mirrors security’s “just a rumor of a bug is enough” pattern; millions of dollars in LLM spend to get there first signals rapid, aggressive diffusion of the practice

  • Quoting Terence Tao

  • On the Navier–Stokes Millennium Prize Problem

wardley-beta
%% Q1 users: mathematicians/researchers and AI labs (OpenAI, Anthropic)
%% Q2 needs: solve/scout open problems, secure priority+attribution, preserve open science
%% Q3 capabilities: agentic proof search, formal verification, problem discovery, priority norms
%% Q4 components: agentic math agents, Lean verifier, open-problem corpus, training-data influence, priority/secrecy norms
%% Q5 maturity: agentic agents Genesis-Custom; Lean Product; open problems a scarce Product; data influence Genesis
%% Q6 managers: labs fund and operate agents+data; research community governs priority norms
anchor researcherNeed [0.95, 0.93]
anchor labNeed [0.95, 0.72]
component agenticMathAgents [0.38, 0.62]
component leanVerifier [0.68, 0.38]
component openProblemCorpus [0.66, 0.52]
component trainingDataInfluence [0.28, 0.35]
component priorityNorm [0.32, 0.22]
agenticMathAgents -> researcherNeed
agenticMathAgents -> labNeed
leanVerifier -> agenticMathAgents
openProblemCorpus -> agenticMathAgents
trainingDataInfluence -> agenticMathAgents
priorityNorm -> researcherNeed
%% note: signals named agents as Genesis-Custom; I hold agents at Custom given 88h solves, Lean at Product

ZAP Security Scanner Updates

ZAP’s August 2026 monthly update shows a mature web-application security scanner consolidating its core while adding an AI-assisted layer. The AJAX Spider and DOM XSS Active Scan Rule add-ons are being retired from the nightly and weekly releases because the Client Spider and OWASP PTK are now the more effective options — the old add-ons remain installable from the Marketplace but are no longer the default path. The headline addition is the opt-in LLM Support add-on, which connects ZAP to Claude, Gemini, Azure OpenAI, Ollama, or OpenAI-compatible endpoints, adds a tabbed chat panel, integrates with the MCP add-on so the LLM can call ZAP’s tools directly, and already powers AI-assisted features elsewhere: an LLM-driven OpenAPI importer and a “Review using LLM/AI” action on Alert Filters. Infrastructure work continues splitting authentication from verification, with configurable poll verification. A steady consolidation of commoditized crawling plus an emerging LLM-driven assistant.

Watch for:

  • Whether LLM-assisted alert review meaningfully reduces false positives
  • The LLM/MCP integration becoming the default way to drive ZAP programmatically
  • Full convergence on the Client Spider as the single supported crawler

Wardley signals:

  • Components: ZAP, Client Spider, AJAX Spider, DOM XSS rule, OWASP PTK, LLM Support add-on, MCP add-on, OpenAPI importer, Alert Filters, Authentication Helper

  • Evolution stage: web-application scanning core at Product→Commodity (consolidating duplicate crawlers, backwards compatibility emphasized); LLM-assisted scanning at Genesis→Custom (opt-in, forming, rapid iteration)

  • Doctrine: “Use appropriate tools” and “Use standards” (retiring the AJAX Spider and DOM XSS in favor of better replacements, keeping compatibility); “Bias towards open” via Marketplace and MCP integration

  • Climatic patterns: Commoditisation <> Centralisation as the crawler set collapses to one; components co-evolve as LLM and MCP tooling arrive around the existing core

  • Diffusion markers: retirement of legacy add-ons signals standardisation and accepted practice; multi-vendor LLM support and MCP integration show the AI-assisted layer spreading across add-ons

  • ZAP Updates - August 2026

wardley-beta
%% Q1 users: web application security testers
%% Q2 needs: find/verify vulnerabilities reliably and efficiently
%% Q3 capabilities: crawling, active scanning, AI-assisted alert review
%% Q4 components: ZAP core, Client Spider, OWASP PTK, LLM Support add-on, MCP integration, alert review
%% Q5 maturity: ZAP core Commodity; spiders/PTK Product; LLM support Genesis-Custom
%% Q6 managers: ZAP project maintains core; community contributes add-ons; LLM layer vendor-integrated
anchor testerNeed [0.95, 0.93]
component zapCore [0.78, 0.55]
component clientSpider [0.66, 0.5]
component owaspPtk [0.62, 0.42]
component llmAssistant [0.32, 0.58]
component mcpIntegration [0.48, 0.4]
component alertReview [0.42, 0.3]
zapCore -> testerNeed
clientSpider -> zapCore
owaspPtk -> zapCore
llmAssistant -> zapCore
mcpIntegration -> llmAssistant
alertReview -> llmAssistant
%% note: signals saw LLM scanning as Genesis; I place it at Custom (already shipping, multi-add-on)

The Socrates Agent

Daniel Miessler publishes the full source of “Socrates,” a Socratic-tutoring agent that never gives answers — it interrogates the learner with questions until they reason it out themselves. The roughly sixty-line Markdown file configures an agent (Claude Code, ChatGPT, Gemini, or any prompt-capable harness) with a strict persona: review the student’s existing work first, ask one small question at a time, shrink the problem when stuck, and withhold the answer no matter how it is phrased, with a single exception for safety. Miessler frames it as a tactical counter to AI eroding people’s ability to sit with a hard question — kids who outsource homework are outsourcing learning, and adults who skip “reps” by asking for summaries quietly lose a skill. The transcript is the deliverable: evidence that the student actually thought.

Watch for:

  • Whether the “one question at a time” pattern genuinely improves long-term retention
  • Adoption of no-answer tutoring agents in classrooms, curricula, and study tools
  • Jailbreak or product pressure to make such agents cave and supply answers

Wardley signals:

  • Components: Socrates agent file, Claude Code, agent harnesses (ChatGPT/Gemini), Markdown frontmatter agent config, the Socratic method itself

  • Evolution stage: Socratic/teaching agents at Genesis→Custom — novel practice, rapidly iterating, undefined market, limited evidence base

  • Doctrine: “Focus on user needs” and “Know your users” (the student’s learning matters more than the answer); “Use appropriate tools” (a single portable prompt file)

  • Climatic patterns: Red Queen in play as AI answer-generation commoditizes, forcing counter-patterns that preserve human skill

  • Diffusion markers: one portable file distributed via curl into an agent folder and pasteable into any harness — low-friction spread across Claude, ChatGPT, and Gemini ecosystems

  • The Socrates Agent

wardley-beta
%% Q1 users: students and lifelong learners (and parents/teachers observing)
%% Q2 needs: learn by doing, retain reasoning skills, avoid answer-outsourcing
%% Q3 capabilities: Socratic questioning, work review, problem-shrinking, transcripts
%% Q4 components: Socrates agent, agent harness, prompt file, general answer capability, transcript record
%% Q5 maturity: Socrates agent Genesis-Custom; harness Product; prompt file Custom; answer capability Commodity
%% Q6 managers: authors distribute the agent file; learners self-manage via harnesses
anchor learnerNeed [0.95, 0.93]
component socraticAgent [0.36, 0.62]
component agentHarness [0.7, 0.42]
component promptFile [0.55, 0.5]
component answerCapability [0.82, 0.66]
component transcriptRecord [0.3, 0.28]
socraticAgent -> learnerNeed
promptFile -> socraticAgent
agentHarness -> socraticAgent
answerCapability -> socraticAgent
transcriptRecord -> socraticAgent

ChatGPT Images 2.5

OpenAI’s ChatGPT Images 2.5 release improves instruction-following across multiple turns, responds faster, and better preserves subjects in reference photos. Two new API model IDs ship — gpt-image-2.5-sunburst, pitched for workflows where editing precision matters, and gpt-image-2.5-flare for fast, high-quality everyday generation. Simon Willison notes the models have already generated more than three billion images across ChatGPT Images and the GPT-Image API, and upgrades his openai_image.py CLI to accept one or more reference images, demonstrating a one-shot workflow (inject a reference chart, ask for a themed variant, get back an edited image). It reads as a reliability and control release on an already-commoditized capability rather than a leap in raw image quality — the competitive axis is now multi-turn editing and reference fidelity.

Watch for:

  • The Sunburst-vs-Flare model split becoming a stable pattern for edit-precision vs speed
  • Reference-image editing fidelity emerging as the key differentiator between providers
  • Pricing, rate limits, and usage limits for the two new model IDs

Wardley signals:

  • Components: ChatGPT Images 2.5, gpt-image-2.5-sunburst, gpt-image-2.5-flare, openai_image.py CLI, reference-image workflows, GPT-Image API

  • Evolution stage: image generation at Product (rapidly increasing use, feature differentiation); the Sunburst/Flare split is a consolidation and differentiation move within that stage

  • Doctrine: “Use appropriate tools” (choose Sunburst vs Flare per workflow); “Set exceptional standards” on instruction-following and reference preservation

  • Climatic patterns: commoditisation of image generation driving differentiation on features and control; Jevon’s Paradox — faster and cheaper generation feeds the 3B+ image count

  • Diffusion markers: billion-scale usage, exposure as two named API models, and CLI tooling with reference-image support signal rapid, broad adoption

  • Introducing ChatGPT Images 2.5

wardley-beta
%% Q1 users: developers and content creators
%% Q2 needs: generate and precisely edit images at speed
%% Q3 capabilities: multi-turn generation, reference-faithful editing, CLI/API tooling
%% Q4 components: image gen model, Sunburst model, Flare model, reference-image flow, CLI tooling
%% Q5 maturity: image gen Product; Sunburst/Flare Product (differentiated); reference flow Custom; CLI Custom
%% Q6 managers: OpenAI runs models/API; users self-manage CLI and reference workflows
anchor creatorNeed [0.95, 0.93]
component imageGenModel [0.78, 0.6]
component sunburstModel [0.58, 0.52]
component flareModel [0.62, 0.46]
component referenceImageFlow [0.5, 0.55]
component cliTooling [0.52, 0.34]
imageGenModel -> creatorNeed
sunburstModel -> imageGenModel
flareModel -> imageGenModel
referenceImageFlow -> imageGenModel
cliTooling -> imageGenModel

AI-Assisted Data Breach Analysis

Niels Provos documents cutting the cost of a data-breach report nearly a hundredfold — from about $8 to roughly $0.11 — by optimizing his PlanAI framework and, decisively, switching from Anthropic’s API to OpenAI’s GPT-5.6 Luna. PlanAI assembles public breach reporting into technical analyses of whether security invariants would have prevented each breach. The initial Claude-Code/Fable 5.1 refresh carried whole documents through a branching graph, which drove up cost; refactoring to a sequential three-conversation flow, disabling reasoning on review workers, and caching notes in the system prompt helped, but the real win was model economics — Luna’s cached-input and output prices are roughly ten and eight times cheaper than Sonnet 5. The lesson is that token prices and cache-write costs are first-order concerns in choosing models, not afterthoughts.

Watch for:

  • Whether Anthropic ships a cheaper Haiku-class model to compete with Luna on report writing
  • Wider validation of the security-invariants thesis against real breach datasets
  • More cost-optimization write-ups as API pricing diverges further between labs

Wardley signals:

  • Components: PlanAI framework, security invariants, Claude Code with Fable 5.1, GPT-5.6 Luna, Haiku 4.5, Sonnet 5, prompt caching, data-breach analysis pipeline

  • Evolution stage: LLM report-writing at Product (rapidly improving, cost becoming a selection feature); prompt caching approaching Commodity

  • Doctrine: “Do better with less” and “Effectiveness over efficiency” (cost reduction as a first-class goal); “Know the details” of pricing and cache mechanics

  • Climatic patterns: commoditisation of model inference with divergent pricing between vendors; potential Jevon’s Paradox as cheaper reports enable more analyses

  • Diffusion markers: cache-write and cached-read price differences publicized and compared across vendors; model-selection guidance spreading as cost benchmarks

  • Reducing the Cost of Writing a Data Breach Report

wardley-beta
%% Q1 users: security analysts and breach-response teams
%% Q2 needs: low-cost, fast technical analysis of data breaches
%% Q3 capabilities: report assembly, invariant scoring, model/cost selection
%% Q4 components: LLM report writer, PlanAI framework, model provider, prompt caching, invariant analysis
%% Q5 maturity: report writer Product; PlanAI Custom; model provider Product; caching Commodity; invariants Custom
%% Q6 managers: analyst builds/runs PlanAI; model vendors price and cache; community validates invariants
anchor analystNeed [0.95, 0.93]
component reportWriter [0.6, 0.62]
component planaiFramework [0.5, 0.48]
component modelProvider [0.74, 0.55]
component promptCaching [0.82, 0.38]
component invariantAnalysis [0.44, 0.3]
reportWriter -> analystNeed
planaiFramework -> reportWriter
modelProvider -> reportWriter
promptCaching -> modelProvider
invariantAnalysis -> planaiFramework