Grok 4.1 from Elon Musk hallucinates 3 times less than previous version
Grok 4.1 from Elon Musk is out – this is not just another update. The model was pumped up in emotional intelligence and significantly reduced the number of hallucinations. And it became much more empathetic and sensitive.
Shows even better results on EQ-Bench. This is a benchmark with tasks on various soft skills. Though there’s no comparison with the new version 5.1.
But the main result is perhaps different. The model hallucinates 3 times less than the previous version. This is really great. Because empathy is empathy, but accuracy is what determines whether you can trust the model in real tasks.
According to the company, Grok 4.1 significantly improves interaction quality through expanded creative, emotional and collaborative capabilities. The model became better at perceiving subtle user intentions, adheres to a more holistic communication style and preserves “personality”, while not losing accuracy.
To achieve results, xAI applied large-scale reinforcement learning infrastructure previously used for Grok 4. And optimized the style, character and usefulness of the new version. The company also developed methods that allow using advanced agent reasoning models as reward models.
In the benchmark for creative writing, the new model was inferior only to GPT-5.1 version.
It turns out Grok was taught to feel mood and write beautifully. But most importantly – the model hallucinates 3 times less.
Autor: AIvengo
For 5 years I have been working with machine learning and artificial intelligence. And this field never ceases to amaze, inspire and interest me.
Latest News
Hugging Face head predicted collapse of large language models bubbleClem Delangue from Hugging Face drew a red line in the discussion about the technology bubble. And this line doesn't run where everyone expects. The head of one of the largest AI platforms stated that there is a bubble, but it's not an AI bubble. It's a large language models bubble. And it could collapse as early as next year.
OpenAI released GPT-5.1-Codex-Max and surpassed Gemini 3 Pro in a dayOpenAI presented GPT-5.1-Codex-Max. This is a version of GPT-5.1 Thinking, specially tailored for programming tasks within the Codex coding agent. This is the first company model natively trained to work through multiple context windows using a process called compaction. The model is capable of working coherently with millions of tokens within one task.
Five IT founders earned over $200 billion from AI boomFive founders of IT companies can boast wealth of over 200 billion dollars each against the background of the AI boom. Just recently, as The Economic Times notes, having 100 billion dollars allowed access to the world elite club, but now the bar has doubled.
Japanese scientists created memory reading system via MRIA group of Japanese scientists from the NTT laboratory showed a system that generates text descriptions of what a person remembers, imagines or sees based on functional MRI data. Essentially, this is memory reading. And another big step toward mind reading.
