Invitation: WDCC Shanghai subconference on the forgetting-mechanism direction

Dear Dr. Wei, and colleagues at DeepSeek-AI,

I'm writing to invite someone from your DeepSeek-OCR team to join a subconference I'm chairing at WDCC 2026 in Shanghai this September, "Designing AI's Boundaries: Respecting Human Agency" — and to explain why your paper specifically, not AI compression research in general, is what prompted the invitation.

One of our working sessions examines six places where AI quietly stands in for a human capacity, each built around a real, well-built system rather than a hypothetical. Your "forgetting mechanism" proposal — rendering conversation history to images and progressively downsizing older ones, so token count and resolution decay together — is close to a perfect case for the session on representation. Figure 13 in your paper draws the parallel explicitly: memory clarity over time, visual clarity over distance, text clarity over resolution, presented as three instances of one curve. We think that parallel is genuinely productive and worth taking seriously exactly because it isn't naive — and also that it's worth sitting with a real tension inside it: human forgetting is driven by salience (the anomaly that nagged survives decades; the recent-but-boring vanishes in days), while the compression mechanism you propose is driven by recency (old fades regardless of what it was). Your own paper's quantitative section makes an even sharper point in passing — that a document's vision-token "decodability" holds fairly uniformly regardless of which specific content is in it, which is itself evidence that frequency-based salience and lived-experience salience are different things wearing the same name.

We don't run these sessions as prosecutions. Each one is built around participants whose work is directly at stake, sitting alongside practitioners of the very approach being examined, as co-explorers rather than defendants. Given that DeepSeek-OCR is the paper that made this argument concrete and quantitative — 97% precision at 10x compression, 60% at 20x — we think whoever on your team is thinking hardest about the forgetting-mechanism direction mentioned in your discussion section would find real interlocutors here, not just an audience.

The conference runs 26–29 September, hosted at SUES and Tongji University in Shanghai — practically convenient, we'd guess, relative to most invitations you receive. It's structured as sustained dialogue and AI-augmented working groups rather than paper presentations, there's no fee for participation, and remote participation is also available for anyone who can't travel. Details are at https://alwaysquestion.ai/.

We'd be glad to have anyone from your team who worked on the contexts-optical-compression direction — whether that's you directly, Yaofeng Sun, or Yukun Li — and equally glad to send more detail on the specific working session if that would help you decide.

With respect for the work,

--