Model and policy news as of midday September 25. No link-dumping: I read the sources and attached a one-line summary plus my take. The Codex case that overlaps with the evening roundup gets only its key numbers here.
🌐 · English · 中文 · Español · 한국어 · · English Hub
Non-developer sales reps build demos themselves with Codex
Proaction boosts sales 60% and saves 75+ hours with Codex
A vehicle-telematics startup, led by a non-developer COO, shipped 30~45-minute interactive demos loaded with customer data at a rate of 4~6 per month, reporting 40~60 engineer-hours saved monthly and a 50~60% lift in early-stage deal progression. My take: this isn’t a model story, it’swhen the bottleneck moved to the salespeopleThis is an org-design case — the full analysis is in the evening roundup.
OpenAI Academy at 2 years — 250 events, 4 million reached
Since its September 2024 launch: 250+ events, 4 million people reached through content. The next step is a trainer-development program — a pipeline that grows people to teach the Academy curriculum in local communities, plus completion badges by role for knowledge workers, developers, leaders, and educators.
My take: once badges and curricula get serious, the hiring market reacts — the day ‘certificate of completion’ appears on résumés will come. It is genuinely a free curriculum worth comparing against Korean AI bootcamp price lists.
Daybreak provided for Ukraine’s civilian-infrastructure cyber defense
OpenAI extends cyber access to Ukraine for civilian defense
Announced during UN General Assembly week — Ukraine’s government gets Daybreak access for finding vulnerabilities and patch-testing. Background numbers: CERT-UA handled roughly 6,000 cyber incidents in 2025 (hospitals, energy, telecom included); Poland’s CERT Polska found 6 router third-party software vulnerabilities with model contributions, and the vendor has patched them.
My take: security-only models are already the norm. AI clauses are entering defense and government procurement specs — and they’ll enter Korean security-industry standard specs too.
Sam Altman addresses the UN Security Council on AI
Sam Altman’s remarks at the United Nations Security Council
The speech was on maintaining human control over powerful systems and the need for international cooperation. The practical regulatory impact is limited, but the fact that a frontier-lab head speaks on a security stage itself shows the governance conversation shifting venue.
Harvey legal drafting, invideo color grading 3x
Harvey turns legal context into stronger drafts · invideo improves color grading 3x
Harvey reported a memory-panel structure that folds full-matter context into drafts and pins per-lawyer preferences; invideo reported 3x color-grading success and 50 per-day custom effects. The common denominator isContext capacity and number of inference steps— the two axes of quality in the agent era.
Frequently asked questions
Why do the morning and evening briefings overlap?
The feed collects every 12 hours, so the same story may appear twice — morning posts are numeric summaries, evening posts are the deep analysis. If you read only one, read the evening post.
Will Academy certificates hold value?
It won’t be an instant-hire credential, but it’s the generation that goes on non-developer résumés as proof of use. It’s free, so it’s worth putting toward job training.
Do ordinary security teams use cyber defense models too?
Daybreak is the government and defense approval track. For the private sector, the standard API vulnerability-assessment workflow is the realistic path.
Source: manufacturer official announcements and published figures. Prices and specs can change — verify before deciding.
