ChatGPT Adoption Widens Its Reach — Faster Growth Among Users 35+ and a Better Gender Balance

According to OpenAI's officially released Q1 2026 adoption trends report, the pace of ChatGPT's user growth is accelerating. Particularly notable: growth is fastest among users aged 35 and over, and the male-female ratio is starting to become more balanced than before.

At the end of 2024, the user base skewed heavily toward "young men interested in tech," but entering 2026, it is increasingly reaching mainstream business users and seniors. This trend suggests that AI is shifting from "a tool for engineers only" to "an everyday tool used broadly."

That said, a growing user base does not necessarily mean deeper understanding — prompt quality and depth of use likely still vary widely.

From E-Waste to AI Cluster — Repurposing Old Smartphones for LLM Inference

A project on Reddit's LocalLLaMA community is drawing attention: building a load-distributed LLM cluster out of smartphones destined for disposal.

The setup uses the SoCs of unwanted smartphones as compute nodes, distributing inference tasks across multiple devices. Cluster builds with single-board computers have been around for a while, but smartphone SoCs often pack relatively capable GPUs, which could give them an edge on cost-performance.

Actual inference speeds are likely limited, but it's an intriguing effort for edge inference experiments and educational use. It also offers one answer to the e-waste problem, which deserves credit.

Applying Prompt Caching to RL Training Delivers a 7.5x Speedup

A report on Reddit's MachineLearning community is sparking discussion: applying prompt caching to reinforcement learning (RL) training for long-prompt / short-response patterns achieved a 7.5x speedup.

RL training commonly generates multiple responses to the same prompt and compares and evaluates them; this approach reuses the KV cache for the prompt portion, skipping prompt processing on each pass. It is especially effective for RL tasks with long contexts and could translate into practical reductions in training costs.

Prompt caching is already offered in the OpenAI and Anthropic APIs, but applying it to RL training pipelines is a new angle — one that should prove a useful insight for teams training their own models.

Preparatory Fixes for MTP/MMProj Progress in llama.cpp

Word in the LocalLLaMA community: preparatory fixes for MTP (Multi-Token Prediction) and MMProj (multimodal projection) are progressing on the B9109 branch of llama.cpp.

MTP predicts multiple tokens at once to improve inference speed, and adoption has been spreading in recent open-source models. MMProj is a projection layer for adding multimodal capabilities, such as image understanding, to text models.

These fixes front-run support for new model architectures to be released down the road — a fresh reminder of just how much llama.cpp's response speed matters to the local LLM community.

Hokuriku Bank Tackles Overloaded Call Centers with Voice AI

According to ITmedia, Hokuriku Bank has begun a trial deployment of voice AI. To address call centers pushed to the breaking point, the bank is introducing a system in which voice AI automatically handles "overflow calls" (incoming calls that fail to reach an operator).

Staffing shortages and customer-service bottlenecks have long been challenges for Japan's regional banks, and automated voice response is expected to deliver quick wins. Depending on the results of the trial, adoption could spread to other financial institutions.

"Meta-Meta-Prompting": A Technique for Making AI Agents Work

An article titled "Meta-Meta-Prompting: The Secret to Making AI Agents Work" is generating discussion on Hacker News.

Insufficient prompt design is cited as a major reason AI agents fail to work as expected. The author argues that "meta"-level instruction design — having the AI evaluate and improve the prompt itself — is what matters. Concretely, the article proposes an approach that designs not just the task instructions but also "how those instructions should be interpreted."

Practical techniques include building in loops where the agent self-evaluates and enumerating failure patterns up front — making it a worthwhile read for developers working on agents.