This week brought the dramatic launch of GPT-5, which is running into its first technical problems, India continues to develop its own language models across 22 languages without a high budget, and AI-video is moving from viral hits to series on Netflix.

1. GPT-5: high expectations, but the new model is not running smoothly yet

OpenAI has officially introduced GPT-5 as the most advanced model for programming and agent tasks - in three versions (Standard, Mini, Nano), with intelligence that can be tuned across four modes from minimalist to highly complex. This model regularly outperforms o3 and other competitors in performance - e.g., in programming tests (SWE-bench 74.9 %), the AIME-2025 math competition (94.6 %), and creative writing on the EQ-Bench.

However, there was a problem during deployment - the router that is supposed to route queries to the appropriate version of the model failed. This led to a temporary restoration of access to previous versions of GPT-4 for paying users.

2. India bets on its own LLM: a linguistically diverse infrastructure under pressure of austerity

Facing the challenges of "120 languages and 19,500 dialects", India is investing in indigenous LLM solutions with limited computing budget. Projects like Sarvam AI (70 billion parameters with voice support), Soket AI, Gan.ai and Gnani.ai - all designed for multilingual use - are being launched.

The government's IndiaAI Mission initiative has allocated over 19,000 GPUs (including H100) for developing infrastructure capabilities and models, including Sarvam AI and others.

3. From viral ads to TV show scenes, AI video is taking over screens

Generative AI video is leaping from internet phenomena to the mainstream. Studio The Dor Brothers produced viral video clips with 16 million views, and studio Genre.ai created a game commercial for less than $2,000. Even Netflix is experimenting with AI-generated scenes, such as a slow-motion shot in a series The Eternaut.

4. Automated generation of synthetic data for code-based LLM

Researchers at Stanford, Princeton and Alibaba have designed a tool, SWE-smith, that generates realistic examples of bug fixes in code. For example, it uses automated unit tests and change reversion processes. The resulting data - dataset, model and code - are publicly available.

The Batch - DeepLearning.Ai by Andrew Ng / gnews.cz - GH