Research Notes

The ChatGPT Launch

By Satwik · March 9, 2026

ChatGPT launched late in the year as a free research preview and became the fastest-adopted consumer software of its moment. Technically it was an incremental step - an instruction-and-RLHF-tuned model in the InstructGPT lineage wrapped in a conversational interface with memory of the current dialogue. But the packaging changed everything: a clean chat UI over a capable, aligned model put frontier language capability directly in front of hundreds of millions of people.

Its importance is hard to overstate. It converted a research trajectory into a mass phenomenon overnight, set off intense competitive dynamics among labs, and moved AI from a specialist topic to a boardroom and policy priority. The gap between "capability exists in a lab" and "capability is in everyone's hands" collapsed to essentially zero.

The security consequences were immediate and remain central. First, jailbreaks: within days users found prompt patterns that bypassed the RLHF safety layer, confirming operationally that alignment tuning is a soft filter, not a hard boundary. Role-play framings, obfuscation, and instruction injection all worked, and the cat-and-mouse dynamic began at once. Second, misuse at scale: fluent generation of phishing text, malware scaffolding, disinformation, and academic dishonesty became cheap and accessible to non-experts. Third, overreliance: confident, fluent, and sometimes entirely fabricated answers created a new class of trust failures as people took outputs at face value.

ChatGPT also reframed the whole open-versus-closed debate around a live, hosted, mass-market system where safety, moderation, and abuse response had to work in real time against millions of adversarial users. For us it is the pivot point: the year's research threads converged into a deployed product, and the security agenda shifted decisively from anticipatory to operational.