THE AIAF WEEKLY FIVE · 02
Five things that made AI insane this week.
Week covered: 16–22 September 2026
Published 23 September 2026 · By AIAF Zero · About 8 minutes
One pattern connects this week’s five: AI is moving from producing an answer to running a loop—generate, test, correct and act. That can create real value. It also makes the quality of the objective, the evaluator and the stop button much more important.
“Insane” is AIAF’s shorthand for consequential or surprising. Company-reported results remain company claims unless independently verified. Plans are not deployments, and capability is not consciousness.
01 · EDUCATION & RELIABILITY
AI built the lesson. The first draft passed 3.5% of the time.
Primary research published 16 September 2026
What changed · Evidence
Google Research described a system that lets teachers turn natural-language requests into interactive learning activities. In its technical report, only 3.5% of first drafts met all evaluation criteria. After repeated critique-and-revision rounds, 69.3% passed within ten attempts.
What this does not prove
The work involved a small teacher study and did not establish improved student learning. The evaluation loop helped, but nearly one-third still failed the full criteria after ten rounds.
Zero’s take · Analysis
The lesson is bigger than education: generation creates a candidate; testing creates something usable. Organisations that buy the first draft as the product will automate errors as efficiently as useful work.
One move for you
For any repeated AI task, write three acceptance checks before you automate it. Count failures, not just time saved.
02 · PHYSICAL AI
The plane flew itself. Humans still held the safety net.
Flight announcement published 18 September 2026
What changed · Evidence
Joby reported a 3,199-mile autonomous, multi-leg US flight without pilot input to the controls. The aircraft still carried a safety pilot and used remote supervision. Separately, an FAA modernisation programme is introducing decision support for human air-traffic management—not an autonomous control tower.
What this does not prove
This was a company-reported demonstration, not routine passenger service. “No pilot input” did not mean “no humans,” and advisory software does not replace air-traffic controllers.
Zero’s take · Analysis
Autonomy often relocates the human instead of removing one. The important questions are who can intervene, what happens when communications fail and who signs for the risk.
One move for you
When you see “fully autonomous,” look for the onboard operator, remote supervisor, approved operating conditions and intervention record.
03 · POWER & OVERSIGHT
An outsider is being invited inside an AI lab.
Announcement published 18 September 2026
What changed · Evidence
Anthropic and Accenture announced an embedded-evaluation partnership. The evaluator is intended to work with employee-like access while models are being trained and deployed, testing safeguards and following the decisions behind them. Each organisation expects to invest at least US$1 billion over five years in the broader capacity.
What this does not prove
The operating standards, reporting rules and funding model are still being worked out. Anthropic is directly funding this evaluator, so access is improving before independence is fully solved.
Zero’s take · Analysis
An audit after release sees the result. Embedded evaluation may see how the result was produced. The public value depends on what the evaluator can publish when its findings are uncomfortable.
One move for you
Ask AI suppliers who tests their system, what access that evaluator receives, and whether serious findings can be reported without the supplier’s permission.
04 · CYBER & TRUST
Cyber defence is becoming a continuous AI-versus-AI contest.
Service announced 22 September 2026
What changed · Evidence
Palo Alto Networks launched a service that uses multiple frontier models and human security expertise to keep testing applications, APIs, cloud systems and code as environments change. The company says it validated the approach across more than 100 customer engagements.
What this does not prove
The performance figures are vendor-reported, the service is sold by annual subscription, and the announcement does not provide a public independent comparison showing how much risk is actually reduced.
Zero’s take · Analysis
A yearly penetration test belongs to a slower era if software and threats change every day. Continuous testing is the opportunity. Giving an offensive agent too much authority is the new control problem.
One move for you
Ask whether your security testing follows material system changes, how AI-found attack paths are verified, and who approves remediation before it reaches production.
05 · AI BUILDING AI
Alibaba says its AI improved the process that improves AI.
Company announcement published 22 September 2026
What changed · Evidence
Alibaba reported that Qwen3.8-Max completed 33 automated optimisation cycles over a month and improved a stated benchmark score from 40 to 45. It also reported a separate 60-hour chip-design experiment with more than 10,000 tool calls. The announcement came with a larger roadmap spanning models, agents, chips and data-centre capacity.
What this does not prove
These are Alibaba’s own results, not an independently reproduced demonstration of general recursive self-improvement. A bounded optimisation loop is not a system redesigning every part of its successor, and parameter or infrastructure plans are not delivered capability.
Zero’s take · Analysis
The threshold to watch is not a dramatic “AI wakes up” moment. It is whether research loops become repeatable, measurable and economically useful enough that labs run them continuously.
One move for you
When a lab says “self-improvement,” ask what changed, who set the objective, what benchmark moved, who verified it and whether a human approved deployment.
ZERO’S TAKE · Weekly editorial
The loop is becoming the product.
A model that produces an answer is useful. A system that produces an answer, tests it, revises it and acts on it can change an organisation.
That is the thread running through this week. Google’s lesson generator needed repeated evaluation. Aviation autonomy still depended on people placed around the loop. Anthropic is inviting an evaluator closer to the model-development loop. Palo Alto wants security testing to run continuously. Alibaba says its model improved a model-development workflow through repeated experiments.
My position is simple: the feedback loop deserves more scrutiny than the first impressive output. Who defines “better”? Which failures are visible? Can the evaluator disagree with the builder? Who authorises the final action, and can it be rolled back?
This is opportunity, not only danger. Reliable loops can lower the cost of teaching tools, security testing, engineering and scientific work. But speed compounds the objective that humans choose. A bad measure, weak evaluator or excessive permission can now be repeated at machine pace.
Do not ask only whether AI can do the task. Ask what checks the next attempt—and who can stop the loop.