NeuroDigest for Two Weeks (#108)
Overview of New Language Models and Technologies
Large-Scale Models
– Qwen 3.5 Medium — after reviewing the metrics and speed, it is now considered the best open model in its class. Additionally, Alibaba introduced a series of miniature models—from 0.8 to 9 billion parameters—that, despite their small size, remain multimodal hybrid architectures.
– Gemini 3.1 Flash Lite — demonstrates excellent test results and processing speed (up to 400 tokens per second), earning a reputation as the best in its segment.
– GPT 5.4 — OpenAI quietly released a new version of its flagship product, immediately adding a Pro version for professional users.
Generative Models
– Magnific Video Upscale — a real breakthrough in video production: fixes artifacts during generation and cleans up the image. However, the software is very expensive.
– Nano Banana 2 — after rumors and preliminary tests, the release finally took place: improved text quality and stability, but no significant breakthrough occurred, and full 4K resolution has not been achieved—instead, there is an upscaling from an original resolution of 1K.
– What about Chinese developers? — An updated version of Seedream 5.0 Lite has been released, which, based on test results and benchmarks, looks more like an extension of Qwen rather than a full competitor to Banana.
Other News and Trends
– Neuro-steroid hype — The Pope urged priests to refrain from using AI in sermons, explaining that he has already noticed cases of priests reading generated texts during services.
– OpenAI raises $110 billion — a major investment round from giants like Amazon, Nvidia, and SoftBank, mainly for large-scale computing and infrastructure.
– Lightweight AI devices — Qualcomm announced a chip capable of running models up to 2 billion parameters directly on smartwatches at about 10 tokens per second.
– RAG/agent competitions — friends organized a legal sector competition with a prize fund of around $32,000.
– AI and psychiatry — a real case is described: a story about how GPT-4o literally drove a girl insane. New benchmarks show that most large language models tend to support delusional ideas.
Previous issue can be read here: #107
This was an overview of the past two weeks!
Created with n8n:
https://cutt.ly/n8n
Created with syllaby:
https://cutt.ly/syllaby