OpenAI is bringing advertisers into ChatGPT’s image tool, while AI buyers are asking how much of the underlying system they should control.
Today’s issue covers visual ads, the ownership trade-off, and Formycon’s AI work inside regulated drug development.
OpenAI is testing visual ads inside ChatGPT’s image generator, putting paid products beside a task where people are already deciding what something should look like.
US trials start later this month with an initial group of advertisers.
The format can show products in use, but OpenAI says the ad stays labelled, separate from the generated image and unable to change ChatGPT’s answer.
Placement also depends on a new kind of brand-safety check. Before serving an ad, the system assesses whether the conversation is emotionally vulnerable or sensitive.
The rollout is further along in Europe:
Free and Go users in 31 countries have seen ads for six weeks.
Plus, Pro and Enterprise accounts remain ad-free.
EU rules require a public record of ChatGPT ads by January.
OpenAI says ChatGPT reaches 1.2 billion people a week and that its ads business passed a $1 billion annual run rate in under 200 days. Those are company figures, but they explain the speed.
Companies are starting to split AI work into three buckets: use closed-source models like ChatGPT or Claude, connect those models to company data, or run open-source models locally.
The AI Daily Brief’s 5th October episode focused on companies that want more control over their AI systems. That usually does not mean building a model from scratch.
Closed-source models are still the easiest route for everyday drafting, research, support summaries and admin. They are fast to start using, and most teams already know the interfaces.
Open-source or locally hosted models become more relevant when the work touches sensitive company data, token costs are getting too high, or the team needs the model to keep running even if an outside provider changes terms.
The trade-offs are practical:
Closed-source models usually win on ease of use and capability.
They can create software, token and running costs that grow with usage.
Open-source or locally hosted models can reduce some running costs, but add maintenance, security and testing work.
Sensitive customer, financial or legal data may change the answer.
The wider team still needs a tool they can actually use.
Founders context: Closed-source versus open-source AI is a real operating decision, not just a technical preference. The main factors are:
software, token and running costs for closed-source models;
maintenance costs for open-source or locally hosted models;
access to sensitive company data;
ease of use for the wider team;
model capability.
AI IMPLEMENTATION
This is how Formycon is using AI to reduce the administrative load behind biosimilar drug development, where a trial can leave several thousand pages of regulated records.
The 200-person German company is moving current and future clinical trial files into Phlexglobal’s electronic trial master file. Existing records will come across from older systems rather than starting with an empty cabinet.
The system uses AI to manage documents while preserving the controls needed for data integrity, quality checks and regulatory compliance.
Formycon expects three practical changes:
One central record for trial design, conduct and monitoring.
Faster document handling without dropping required review steps.
Less manual work when current files move out of legacy systems.
Its chief scientific officer says the new system could improve efficiency by about 20%. That is Formycon’s internal assessment, not a measured result from a completed rollout.
The company is also using separate AI software to analyse manufacturing-process data. The aim is to speed work with contract manufacturers and predict structures that help teams assess product quality.
The pattern is useful beyond pharmaceuticals. Formycon has put AI around dense records and repeated checks, while people remain responsible for the scientific and compliance decisions.
📡 Further Context
🗣 Sam Altman says society should accept some AI harm. OpenAI’s chief drew a line between tolerable misuse such as hacks and scams and extreme loss-of-control risks, while keeping his disagreement with Anthropic over tighter rules.
🕵 Meta’s Muse can build pages about people around its user. The personal agent stores relationship context and trains on interactions by default; Meta disputes one tester’s account of access without the expected permissions.
🚓 A US judge calls one Flock search unconstitutional mass surveillance. The ruling concerns a warrantless search across a connected licence-plate camera network and does not bind other courts.
🤝 Huawei and Qualcomm cross-license 5G, compute and AI patents. The companies did not disclose financial terms; Qualcomm will also buy some Huawei US patents after regulatory approval.
🧱 China reportedly stockpiled 343 immersion lithography tools. The report says older DUV scanners can still produce advanced chips through extra manufacturing steps, complicating export-control assumptions.
🤖 AI Tool Context
Bentomux: Runs several coding agents side by side and surfaces permission prompts before they stall.
DailyHelm: Reviews store, advertising, SEO and operations data overnight, then flags what needs attention.
Spira Maxima: Turns scripts, photos or recordings into social videos with presenters, B-roll, captions and music.
FastRouter.ai: Routes requests across more than 200 models with cost, latency, failover and usage controls.
