Loading...
Loading...
The last blog week of the summer: ChatGPT got teen accounts, Tencent shipped dirt cheap translation models, and I finally cleaned up my task queue before school starts.
This is the last full week before school starts, so I wanted to close out the summer with the stuff that actually happened instead of letting it slide. Nine weeks of writing this blog and the pace never once let up. If you told me in June that I would struggle to keep a weekly recap under a thousand words because too much was happening, I would have laughed.
The big platform news was OpenAI adding teen accounts to ChatGPT along with parent controls, and expanding ads in ChatGPT to Europe. I'm still on the younger side of things myself, so seeing them build a whole supervised experience for teens felt oddly personal.
Also on August 19 they announced Zero Data Retention for frontier models through the API, which matters if you're piping anything sensitive through one. For my own projects it changes the calculus on what I'm willing to send to which provider. School documents? Fine. Anything from my CTO work at Firebound? That now goes through endpoints with retention guarantees or nowhere at all.
It came from Tencent on August 20. Two translation models:
Hy-MT2-30B-A3B : a 30B mixture of experts flagshipHy-MT2-1.8B : the tiny compact siblingBoth cover 33 language pairs at prices so low they feel like typos. We're talking fractions of a dollar per million tokens. The catch is the 8K context window. That's a deliberate signal that this is a sentence level translation tool, not an agent, and honestly I respect the honesty of that.
I wired the small model into my alerts bot as an experiment, translating incoming messages before they hit Telegram. It worked fine for short strings, which is exactly what it's built for. Anything longer and it would fall apart, but I'm not asking it to write essays.
The fun part was the latency. Because the model is so small, translations came back faster than some of my regex processing steps. When your AI step is cheaper AND quicker than the code around it, you know the efficiency curve has bent somewhere dramatic.
Nine weeks, nine weeks and change, more model releases than I could keep up with.
Looking back at the summer as a whole:
School is about to eat most of my time again, but I'll keep writing when something big drops. Knowing this industry, that means Tuesday.
Cheers, Yassen