Feed
24 materials
ChatGPT's Goblin Glitch: Why the Model Talked About Goblins and Gremlins
OpenAI explained that due to a feedback loop during training, the GPT and Codex models began obsessively inserting mentions of goblins and gremlins. This most often appeared in the Nerdy preset, and the company had to add special restrictions to fix the issue.
18 August 2026
Invisible watermarking of Claude texts: Anthropic explains how it works
The new generation of models will embed an invisible pattern in responses, making it possible to assess how likely AI involvement in text creation is. This won’t affect quality, readability, or generation cost, but it will help meet regulatory requirements.
17 August 2026
Claude Opus 4.5: breakdown of capabilities for code, agents, and computer control
A model from Anthropic, released on November 25, 2025, is tailored for complex programming tasks, autonomous agents, and work with applications. The article covers key scenarios, prices, and a migration plan to the new version.
16 August 2026
Machine reviewers under scrutiny: comparative analysis of GPT, Gemini, and Claude in evaluating scientific papers
Researchers compared the reviews of three large language models with expert verdicts on 300 submissions to the ICLR conference. They found that the algorithms confidently distinguish accepted papers from rejected ones, but fail to capture more subtle differences between oral and poster statuses, and their comments are skewed toward topics specific to each model.
14 August 2026
Comparing OpenAI GPT-5.2 and Google Gemini 3 Pro: what to choose in late 2025
We break down the latest flagships from OpenAI and Google DeepMind: their launch, key capabilities, pricing, and typical use cases. GPT-5.2 focuses on complex reasoning and professional tasks, while Gemini 3 Pro stands out with native multimodality and a context window of up to 1 million tokens.
13 August 2026
FIFA World Cup as a test for AI: what the AI World Cup 2026 experiment showed
Ten language models predicted the outcomes of the entire tournament in advance, and the overall winner was the model that correctly guessed the champion. Meanwhile, the leaders in individual match accuracy were far from first in the final standings.
12 August 2026
Classification of cognitive limitations: how far are generative and agentic models from real thinking?
A new study maps out the gaps between the claimed capabilities of modern AI systems and their actual cognitive functions. The authors identify five dimensions in which models still fall short of robust reasoning, and propose benchmarks for moving toward self-regulating and adaptive AI.
9 August 2026
Custom chips for Claude: Anthropic is forming an engineering team
Anthropic has officially confirmed for the first time its plans to create an internal semiconductor design division. The company is hiring engineers to jointly develop hardware and models for faster and more efficient operation of Claude, without giving up capacity from AWS, Google, Nvidia, and AMD.
7 August 2026
From Block Replacement to Wireless Intelligence: Results of a Decade of Deep Learning in Communications
The review shows how over ten years, AI in wireless networks has evolved from point-by-point algorithm replacement to a holistic approach to building "wireless intelligence." The authors highlight three key shifts and identify the main conditions for the next stage of development.
5 August 2026