August 9, 2026
Silent thinking, loud backlash
Show HN: DeepSeek-V4 Latent Reasoning – moving "thinking" into latent space
AI says it can think quietly, but the comments are screaming receipts
TLDR: A developer showed off an AI model that supposedly keeps its reasoning hidden while still solving tricky logic tasks. Commenters weren't sold: they demanded a clear before-and-after comparison, mocked the writing, and questioned whether the whole thing was more hype than breakthrough.
A new Hacker News Show HN post promised something very sci-fi-sounding: an open model that does its "thinking" silently, keeping the step-by-step work hidden instead of printing it out. On paper, that sounds slick. The creator says it performs especially well on logic-style tests like tracking objects and spotting fallacies, while flopping badly on oddly specific stuff like bracket matching. In plain English: it looks smart in some situations, very confused in others, and the pitch is that the hidden-thinking trick is the reason.
But the real action was in the comments, where readers basically turned into forensic detectives. The loudest complaint? "Compared to what?" Multiple commenters were unimpressed that the post showed scores without a clean side-by-side against the original DeepSeek model underneath. That made some readers instantly suspicious, with one asking whether the numbers proved anything at all. Others went for the throat on style, calling the write-up "grating," "unreadable," and peak SEO slop. Ouch.
Then came the spiciest accusation: maybe not just the article, but the whole project felt AI-generated and self-congratulatory. That's where the mood shifted from nerdy skepticism to full-on comment-section side-eye. Even a simple question about a weird bad first response became part of the vibe: people weren't ready to celebrate a breakthrough until someone explained the glitches, the baseline, and whether a human had actually checked the work. The model may be hiding its thinking, but the community absolutely is not.
Key Points
- •The article introduces DeepSeek-V4-Flash-0731-Latent-Reasoning, an open-source model that keeps reasoning in latent space instead of emitting reasoning tokens in output text.
- •The model uses a frozen NVFP4-quantized DeepSeek-V4-Flash-0731 backbone, retains a 3-layer DSpark draft block, and adds a 35.7M-parameter latent reasoning head and decoder.
- •Initial BBH evaluation on 27 subtasks reports strong results on multi-step state-tracking tasks, including several 1.00 scores.
- •The article identifies dyck_languages at 0.26 as a genuine weak point and states the model performs less well on mechanical or syntax-heavy tasks.
- •The architecture compresses a 4096-d hidden state at layer 35 into a 1024-d variational latent and writes the decoded representation back into the residual stream at layer 42.