Medium

Salvatore Raieli

13.7K followers

Featured Book

Building AI Agents with LLMs, RAG, and Knowledge Graphs: A Practical Guide to Autonomous and Modern AI Agents

Building AI Agents with LLMs, RAG, and Knowledge Graphs: A Practical Guide to Autonomous and Modern AI Agents

Guide to intelligent AI agents using LLMs, RAG, knowledge graphs, and tools. Includes practical Python examples for deploying systems that re

2025

Gabriele Iuculano, et al.

Gabriele Iuculano, Salvatore Raieli

Stories

Pinned

In

TDS Archive

by

Salvatore Raieli

·

Oct 31, 2024

The Savant Syndrome: Is Pattern Recognition Equivalent to Intelligence?

Exploring the limits of artificial intelligence: why mastering patterns may not equal genuine reasoning

Are LLMs able to reasoning?

Pinned

In

TDS Archive

by

Salvatore Raieli

·

Dec 1, 2023

A Requiem for the Transformer?

Will be the transformer the model leading us to artificial general intelligence? Or will be replaced?

what will replace the transformer

Pinned

In

TDS Archive

by

Salvatore Raieli

·

Oct 22, 2022

How AI Could Help Preserve Art

Art masterpieces are a risk at any time; AI and new technologies can give a hand

how artificial intelligence is used in art

Pinned

In

TDS Archive

by

Salvatore Raieli

·

Oct 11, 2022

How artificial intelligence could save the Amazon rainforest

Amazonia is at risk and AI could help preserve it

How AI can save Amazon forest

Pinned

In

TDS Archive

by

Salvatore Raieli

·

Jun 30, 2022

Blending the power of AI with the delicacy of poetry

AI is now able to generate images from text, what if we furnish them with the words of great poets? A dreamy trip between poetry and AI

create images from text with DALL-E 2

In

Level Up Coding

by

Salvatore Raieli

·

Apr 1

Going Deep Requires Change: LLMs Have Been Using Residuals Wrong for 10 Years

Better scaling requires attention to depth

Residual connections in LLMs uniformly add all layers, causing signal dilution and unstable growth. Attention Residuals replace this with learned attention over layers, enabling selective use of past representations. Block AttnRes makes it scalable by grouping layers. It improves training stability, depth utilization, and performance, with minimal overhead, validated on large-scale models.

In

Level Up Coding

by

Salvatore Raieli

·

Mar 4

Teaching LLMs to Remember: Internalizing Documents Without Fine-Tuning

Never forget a document: Moving from fine-tuning to context internalization

Long contexts are crucial for LLM reasoning but make inference slow and memory-intensive due to quadratic attention. While context distillation embeds prompt knowledge into model weights, it is too costly per document. Doc-to-LoRA (D2L) addresses this by meta-learning a hypernetwork that generates LoRA adapters in a single forward pass, internalizing context without retraining. It reduces latency and memory, outperforms standard CD under constraints, and enables fast, scalable model adaptation.

In

Level Up Coding

by

Salvatore Raieli

·

Feb 23

The End of Prompt Engineering as We Know It (and the LLM Feels Fine)

How a new paper suggests we’ve been overcomplicating LLM prompting

A recent paper shows that simply repeating a prompt can significantly improve the performance of non-reasoning large language models. Across multiple models and benchmarks, prompt repetition boosts accuracy without increasing output length or requiring complex instructions. The findings challenge the belief that increasingly elaborate prompt engineering is necessary, suggesting that simpler strategies can often be just as effective.

In

Level Up Coding

by

Salvatore Raieli

·

Feb 20

Context Files Are Not What Your AI Agent Needs

Why more instructions and repository context don’t always improve agent performance (but surely increase costs)

The paper evaluates whether repository-level context files like AGENTS.md help coding agents complete tasks. Across multiple agents and benchmarks, they find that these files often reduce success rates and increase inference cost by over 20 % compared to no context. Agents tend to follow the instructions, but unnecessary requirements make tasks harder. Minimal, relevant context is recommended over large files.

In

Level Up Coding

by

Salvatore Raieli

·

Feb 16

Learning from the Crowd: Can One Agent Think Like Many?

How Collective Intelligence Can Be Compressed into a Single Model

Multi-agent LLM systems improve reasoning through debate but are costly and prone to error propagation. AgentArk distills multi-agent dynamics into a single model by training on reasoning trajectories, self-correction, and process-level signals. Shifting computation from inference to training, the approach preserves efficiency while retaining strong reasoning, robustness, and generalization across tasks.

Salvatore Raieli

Salvatore Raieli

13.7K followers

Book Author

Senior data scientist | about science, machine learning, and AI. Top writer in Artificial Intelligence

Following

Help

Status

About

Careers

Press

Blog

Store

Privacy

Rules

Terms

Text to speech