Reasoning makes translations worse
We tend to believe that making the LLM reason more can only improve its answer, regardless of the task. However, when translating from a language to another, the opposite happens.
zansara.dev
My personal blog. I write about AI, LLMs, Open Source and Python, with some occasional diversion.
Live Last read · last published · next check
Latest posts
What’s an agent harness?
The intelligence of modern AI agents is due to the LLMs, but their practical capabilities only exist thanks to their harness.
Reasoning effort defaults across the GPT-5.x family
If you never pinned the reasoning effort across model upgrades, you may have accidentally changed your reasoning effort dramatically.
Smarter Systems with Leaner Models
Token prices keep falling with every new model release. So how come AI bills are only rising? Is it structural, or is there a way to address it?
Setting the temperature to zero will make an LLM deterministic?
We all know LLMs don’t always respond the same thing to slight changes of prompt. But why does their answer differ also when the prompt is identical? And what can we do to prevent it?
Is grep really better than a vector DB?
Some agentic applications don’t use vector DBs for search. Is it a good idea?
Phishing AI Agents
Most LLMs are hardened against classic prompt injection attacks. But AI agents also behave like naive humans sometimes...
How does LLM memory work?
All LLMs can keep track of a short conversation. But how do they remember things long-term?
From RAG to AI Agent
A step-by-step guide to transform your RAG pipelines into effective AI agents.
What are the “experts” in Mixture-of-Experts LLMs?
And how can 8 or 16 of them cover all possible domain of expertise?