Perplexity is moving into local AI, while agents are learning to experiment.

In today’s roundup: a new local computer from Perplexity, a threat forecast for enterprise software, and fresh approaches to training and evaluating AI agents.

🔥 Hot:

🔹 Perplexity unveiled the Portable Computer for local AI — The company is entering the device market, where AI services operate closer to users and their data. 🔹 Gartner estimates that $234 billion in enterprise software spending is at risk due to AI agents — This concerns spending on enterprise applications that autonomous agents could transform.

➡️ News:

🔹 Code Plus Equals AI launched an AI web application builder with token-based pricing — The service proxies OpenAI and Anthropic calls, while app creators receive 80% of the margin. 🔹 Frontend Design Pro added a skill set and machine-based quality checks for AI agents — The package helps agents create interfaces and automatically filter out low-quality results.

➡️ Useful materials:

🔹 Researchers proposed training code models using function-level feedback — The approach shifts process supervision from unclear intermediate steps to executable functions. 🔹 LLM agents learned to conduct controlled experiments in simulators — The authors examine whether an agent can not only reason and use tools, but also choose interventions in a system. 🔹 RENDER shows how memory format changes LLM evaluation results — The same dialogue can be represented as a memory record, a summary, or a raw fragment—and produce different responses.

➡️ Discussions and cases:

🔹 Qwen3.6-35B-A3B achieved 78,498 output tok/s on eight AMD MI350X GPUs — The author attributes the result to an open kernel and shows how the software ecosystem affects AMD’s practical speed. 🔹 Muse Glimmer outperformed Qwen3.8 in time and results in a user test — The comparison does not claim to be a rigorous benchmark, but it shows a noticeable difference between reasoning modes and models.

📝 If you would like to add other news and materials to the list, write in the comments.