Reddit Plunges 23%: The Dynamic of AI Eating Into User Growth

Reddit's stock took a sharp dive, falling 23% in a single day. According to reports, the backdrop is a growing view that AI-powered search and summarization are putting pressure on user growth.

For a platform, having users satisfied by AI answers instead of opening the app directly could become a long-term drag on growth. While there's no way to confirm how much of this drop is actually AI-driven, it does look like another piece of evidence that AI is changing the very structure of traffic itself.

WSJ Looks Back at "The AI Supremacy OpenAI Lost"

The Wall Street Journal (WSJ) has published a long-form piece on how the dominance OpenAI once held is slipping, and its effort to reclaim the throne.

Release strategy, the rise of competitors, and the movement of talent and capital appear to be among the themes, but the article itself is behind a paywall, and the details couldn't be verified from the data gathered this time. We'll hold off on a definitive take, positioning it as a piece that conveys the big picture: the battle for industry leadership is becoming increasingly fluid.

Frontline of Local LLMs: Qwen 3.6 27B at 55 tps on 3× RTX 2080 Ti

A user in Reddit's LocalLLaMA community shared an example of running Qwen3.6-27B (Q5_K_S quantization) on three RTX 2080 Ti cards at 55 tokens per second (55 tps). The setup: a Threadripper 3970X, 128GB of RAM, and 3× 2080 Ti 11GB GPUs.

The poster draws out the speed by combining a tensor split, enabled flash attention, q8-quantized KV cache, and speculative decoding (draft-MTP). It's a knowledge-sharing post doubling as a request for advice on whether there's any more performance to squeeze out — a useful practical example showing that even used 2080 Ti cards can run mid-class models at high speed.

Will 1B-Class LLMs Disappear in 2026?

Also on LocalLLaMA, a discussion kicked off around the question of whether 1B-class LLMs will fade away in 2026.

The poster points out that Gemma 4 ships without a 1B model, Llama has seen no 1B update, and Qwen's recent releases haven't targeted the 1B range either. They rate Gemma 3 1B as one of the practical bests for now, while also sharing the impression that Bonsai's ternary LLMs hallucinate too much to be reliable.

A fascinating perspective, suggesting that the direction of small, on-device models may be shifting away from simple parameter reduction toward a different design approach.