Choose timezone
Your profile timezone:
Responsible AI-assisted coding and common training aspects
→
Europe/Zurich
61/1-009 - Room C (CERN)
Description
Contact
HSF (Caterina Doglioni. Claire Antel, Michel Hernandez Villanueva, Nicole Skidmore)
Supporting parties:
EVERSE + RSQKit
EuCAIF Fair & Sustainable WG
IRIS-HEP
IML
SMASH
COMCHA
Meeting content in 3 parts:
Part 1: Collect experiences/practices around responsible AI-assisted coding in HEP software towards a loose set of suggested guidelines - Short slides per topic + discussion (1h + 20 min)
Part 2: Training plans by HSF training: Slides by HSF training following community survey + discussion (30 min)
Part 3: Discuss any new conclusions, next steps, future (10 min)
Part 1 discussions purpose and structure:
- To gather community practices, experiences, challenges and needs around AI-assisted coding in HEP.
- Slides to be presented focussed on topics listed below (one topic at a time) and based on material uploaded by community (see “homework” below).
- Presentation format To Be Confirmed: Depending on amount of material collected, will collate slides into single deck or organisers will summarise material.
- May additionally host short “expert” presentations on a specific expert topic.
- Presentations to be followed by open discussion.
- For duration of meeting: People can contribute to this live doc at any point.
Before meeting:
- Many communities will be coming together, so we aim to have a semi-structured discussion starting from a solid base.
- Communities will have likely held internal community discussions already.
- To that end, we hope attendees can approach common discussion with a good idea of their own community challenges and needs already.
- To thus start on a solid basis, we invite community representatives and individuals to upload ~1-2 slides per topic to agenda - if any (slides can link to further material), ideally by Fri 24th.
- We’ll nevertheless aim to be open to new directions and conclusions!
- The live doc is available for any discussions beforehand.
After meeting/for those who cannot attend:
-
We continue discussions/comments on live doc, as well as document next steps where people can contribute further.
Common Software Forum Discussion
Zoom Meeting ID
62660214186
Host
Claire Antel
Passcode
92345060
Useful links
Join via phone
Zoom URL
-
-
How to upload material
*** For those with a CERN account ***
Log in and you should be able to upload material directly to the agenda.*** For those without ***
Use the public cernbox to drop material in there.-
- Glossary(ies) to be uploaded for reference.
- Spectrum of AI-assisted tools: From “no AI” to fully automated systems [presentation from Michael Sparks].
- Community context: Community opinions and impact.
[frame slide as ‘what are community opinions, how have or might LLMs impact the community?’]
Glossary
ref: hugging-faceLarge Language Models (LLMs)
Neural networks trained on vast text corpora that generate text by predicting sequences of tokens, basic units of text processing that typically represent words, parts of words, or individual characters [25]. For instance, the word “unhappily” might be tokenized as “un”, “##happi”, “##ly”, where ## marks tokens that are not the start of a word.
Agent
An autonomous AI system that can perceive its environment, make decisions, and take actions to achieve specific goals. Unlike traditional software that follows predetermined paths, agents can adapt and reason through complex scenarios.
Agentic AI
AI systems designed to act autonomously on behalf of users, making decisions and taking actions without constant human intervention. These systems can plan, execute tasks, and adapt to changing conditions.
Multi-Agent System (MAS)
A collection of multiple AI agents that work together, communicate, and coordinate to solve complex problems that would be difficult for a single agent to handle alone.
Agent Orchestration
The coordination and management of multiple agents working together, including task distribution, communication protocols, and result aggregation.
Retrieval-Augmented Generation (RAG)
A technique that combines information retrieval with text generation, allowing agents to access external knowledge sources to provide more accurate and up-to-date responses.Model Context Protocol (MCP)
An open standard for how AI systems integrate and share data with external tools, systems, and data sources.Token
The basic unit of text processing in AI models, which can represent words, parts of words, or characters.
Tokenization
The process of breaking down text into tokens that can be processed by AI models.
Context windows
Define the maximum number of tokens an LLM can consider when generating responses. State-of-the-art models typically handle hundreds of thousands to millions of tokens, constraining how much code and documentation they can simultaneously process. When context limits are exceeded, models lose track of earlier information. Even when information is contained within the context window, attention to mid-document details can degrade (“lost in the middle”), especially for models with very large context windows; this phenomenon is known as context rot.
In-context learning
Allows models to adapt their behavior based on examples and instructions provided within the current conversation, without permanent changes to the underlying model. This enables direction of model behavior through strategic provision of examples and formatting of instructions.
Prompting
Encompasses techniques for structuring inputs to elicit desired outputs, including clear requirement specification, strategic provision of examples, and structured formatting. Effective prompting can dramatically improve code quality and relevance.security
Principle of Least Privilege
A security concept where users and systems are given the minimum levels of access needed to perform their functions.
Telemetry
The automated collection and transmission of data from remote systems for monitoring and analysis.
Prompt Injection
A security vulnerability where malicious input is crafted to manipulate an AI model's behaviour or extract sensitive information.
Hallucination
When an AI model generates information that appears plausible but is factually incorrect or not based on its training data.ref: wikipedia, agentic-ai-terminology github, arxiv:2510.22254v2-"Ten Simple Rules for AI-Assisted Coding in Science"
Spectrum of AI-usage
From M. Sparks's page
-
No GenAI usage - We all start here, but there can be good reasons to choose this level. Examples: learning exercises, controlled data, secure environments, formal assurance, environmental or ethical concerns, or projects that have not agreed a policy.
-
Conversational interaction - Conversations asking for explanations, background, design trade-offs, unfamiliar errors and conceptual help. Examples: "Explain this compiler error", "What should I test?", "What assumptions does this numerical method make?", "What misunderstandings might the average developer have based on this code?" (for sanity checking API design)
-
Conversational artefact creation - Chat-based creation of small, reviewable outputs. Examples: a helper function, a test skeleton, a README paragraph, a regular expression, or a command-line usage example.
-
Conversational Creation with Zip Files (etc) - Providing selected files, snippets, logs or zip files so the tool can reason about the content. Examples: "Suggest tests for these files", "Review this small package structure", "Explain why this minimal example fails."
-
IDE autocomplete and inline suggestions - AI suggestions appear while editing. Examples: completing boilerplate, filling repetitive test cases, suggesting docstrings or small local code patterns.
-
Editor-integrated local assistant - The assistant can make bounded edits inside the editor. Examples: refactor a function, add type hints, generate tests for a module, update these docstrings.
-
Repository-aware assistance - The tool can search and reason across the repository. Examples: trace where a setting is used, identify tests for a feature, propose where to add a new option.
-
Constrained local tool-using agents - An agent can edit files, run commands and iterate locally under supervision. Examples: fix a failing test on a disposable checkout, run linting, prepare a patch without pushing.
-
Managed or cloud coding agents - A hosted or managed agent works in a configured environment, often asynchronously. Examples: ask a cloud coding agent to implement a small issue, run tests, and prepare a branch or PR for review.
-
CI/PR/repository agents - Agents are integrated into shared project workflows. Examples: review PRs, suggest CI fixes, triage issues, draft dependency updates, or maintain an AI staging repository.
-
Open-ended autonomous agents - Broad tool-using systems with persistence, scheduling or external integrations. Examples: agents that can act across repositories, services, email, cloud resources or multiple tools over time.
-
-