AI Engineer hosted Sarah Chieng, head of developer experience at Cerebras, and Cova Lee of OpenAI's developer experience team, to unpack the AI content system they say drove 200 million views. They did not pitch a single tool. They mapped a pipeline.
The engine is speed. Chieng and Lee point to GPT 5.6 Soul Ultra Fast, a model inside ChatGPT powered by Cerebras that they said runs at 750 tokens per second. That throughput matters when you are chewing through long scripts and large files for video, they said. Cerebras, which also powers models with Mistral and Cognition, positions that model as one of the fastest inside ChatGPT.
The front end is talking. Lee scripts with WhisperFlow by speaking concepts aloud, then refines in ChatGPT Voice Mode. Chieng prefers Notion with a Notion connector for swipe files and scripts, while Lee has moved to ChatGPT Space, an integrated workspace the team frames as a Notion alternative. For demos they lean on Coast, which records computer activity locally so you can find and re-record a moment, and CleanShot for annotation, blurs and titles.
Editing is where they let AI be hands on. HyperFrames by HeyGen stitches angles, adds text, renders and cuts long form into shorts via API, plus spins up AI avatars for quick fixes. Chieng was blunt about the limit: avatars are not good enough to do the whole thing, but they are good for patching a wrong word or date. Astra handles computer-use control of Premiere for granular cuts, and ChatGPT can drive CapCut directly. Chieng called CapCut, the desktop version of TikTok's editor, phenomenal. The workflow even fixes the obvious failure: AI generated captions are always wrong, Chieng said, so ChatGPT rewrites them, adds sound effects and animates emphasized phrases.
Packaging decides performance. Chieng said 50% of a YouTube video's performance comes from title and thumbnail, and the pair said spending half your time there can be the difference between 500 views and 5 million. Visuals carry memory, they noted, citing 65% recall after three days with a visual aid versus 10% without. Their research stack hunts outliers with a YouTube tool that slides from 1x to 20x average views to find breakout formats. Dots monitors performance and trends proactively, Buffer schedules across platforms and helps with captions and timing, and the Twitter API pulls data to refine what works.
That last part explains a warning Lee gave: if you only study best-performing content, you are always slightly behind. The outlier lens is meant to catch deviation early. Other pieces fill the system too, from Figma for brand and ChatGPT Image Gen for backgrounds and animated titles to After Effects and Blender for 3D, plus gear notes that keep costs concrete: Lee's former Canon EOS R50 at $300 versus her current Sony FX3 at $2,500 for body plus $2,500 for lens, a $60 Rode mic mini, and a $500 Sony ZV-1 as a starter pick.
There is also a tell about how OpenAI is thinking about audience. An unnamed speaker said the company has a very big interest in developers not just being that audience but expanding to basically everyone, and that it is rethinking what it means to be a developer. Lee tied success to point of view cultivated over real work, not prompts alone. The system works when engineering comes first and content second, a reaping and sowing loop they credited to devrel peers. Relinquish a little control, as one OpenAI creator did by editing a 15-minute video in one take by talking his instructions into the footage, and you get speed. Keep the POV, and you keep the reason people watch.
