00:31:02 Harry Yu: Dear All, nice to meet you all. I am Assoc Prof. Hongqing Yu (Harry) from University of Derby invited by Prof. Lee. I am expert in GenAI for data analysis and reasoning. Really exciting to see any collabrations. 00:41:39 Andrzej: AccGPT on mattermost is not responding 00:42:13 Carla Marin Benito: Same here - does one need to sign up somewhere to start using it? 00:44:23 Uzziel Perez: Does the AccGPT have a Gitlab repo/documentation where people can contribute? 00:45:11 Harry Yu: Replying to "Does the AccGPT ha..." Same question 00:48:58 Luke Jason Van Leijenhorst: Replying to "AccGPT on mattermost..." The Mattermost bot should be accessible by everyone but indeed we have noticed that it is not working for everyone all the time, exact cause is unknown for now unfortunately. For some people it works if you create a direct chat with the bot (also named accgpt) 00:49:29 Albert: How much resources do you have available for the project, how much does the production deployment currently use and how is it funded? 00:50:05 Judita Mamuzic: So if it is not working for now for a user, it will never work until the problem is solved? 00:50:29 Michal Mazurek: I will try to ask this this question now 00:53:26 Andrzej: Can confirm adding /accgpt works 00:54:05 Michal Mazurek: Perfect! 00:55:57 Judita Mamuzic: How? Not for me 00:56:56 Andrzej: I opened a direct message to a AccGPT "user" and it didn't reply by itself to anything, but prepending /accgpt worked 00:57:20 Judita Mamuzic: Hmm tried the same, but no answer… 00:58:01 Luke Jason Van Leijenhorst: Replying to "How much resources d..." Currently resources are very limited. We have access to some GPUs to run LLMs locally but most of the heavy work is being done using Groq for which we received free access and usage to their developer tier for this project. 00:59:48 Luke Jason Van Leijenhorst: Indeed direct message with /accgpt followed by a question should work in the direct message chat with the AccGPT bot. Will take a few seconds after pressing enter before u see the answer. 00:59:58 Harry Yu: I have a question regarding how the different type of data are handled e.g. CSV data and PDF data. I am working in a research project related aircraft data analysis using LLM where we need to provide complex analysis and reasoing on the data to answer very specific question e.g. what is the best time to do service for this aircraft. Do you face the same problem? 01:04:20 Harry Yu: Happy for contribute and discusssion for ideas. I need to leave. Great talk, thank you. You can contact me: h.yu@derby.ac.uk 01:17:52 Pratik Jawahar: Inference is becoming low power as we speak. The hard part is training 01:25:20 Gabriele Benelli: Can people in the room say their name/institute? 02:19:20 Zach Marshall: Of course, ChATLAS has ATLAS internal information, so it should not be opened to other experiments. But the scripts are portable :) 02:21:54 Michal Mazurek: Replying to "Of course, ChATLAS h..." thanks! 02:26:09 Manuel Guijarro: Replying to "Of course, ChATLAS h..." For information that is accessible by anyone with a CERN account, we can just add it to the AccGPT knowledge base by addiing new entries in our Doc to GPT table: https://codimd.web.cern.ch/H_rvQFX9TryO3W-LTYO4zg ...and it will be added (if everything works well) next time we rebuild AccGPT vector DB. 02:28:12 Manuel Guijarro: Replying to "Of course, ChATLAS h..." Let us know about your use cases and we will add them to our list: http://cern.ch/llm-uc 02:28:55 Gordon T. Watts: Totally agree with the point about multiple models - a webapi allows for quick changes. 02:35:05 Maurizio De Giorgi: Maurizio (member of DBOD team): could it be of general interest sharing how to approach and tune pgvector index parameters and perhaps other aspects of using pgvector? There are also other potentially promising extensions to explore like for example https://github.com/timescale/pgvectorscale 02:35:47 Daniel Murnane: Replying to "Maurizio (member of ..." Please! Would be very nice 02:50:17 Maurizio De Giorgi: Replying to "Maurizio (member o..." Perhaps we can start by collecting what has been already done by who is using pgvector and check if we can contribute/improve it. We would also need to receive requests for motivating the exploration/evaluation of other extensions. Please come in touch in both cases. 02:57:44 Michael Sokoloff: Replying to "Maurizio (member of ..." @Maurizio De Giorgi I will ask Mohamed to contact you directly. I will also point him at the GitHub page you pointed at. 03:06:41 Mohsen Farid: Excellent question Zach 03:07:02 Jan Iven: Overall comment (not for the last talk): nice tech talks but missing the regulatory aspects? Were mentioned only indirectly so far ("phonebook", "ATLAS-only docs" examples), and perhaps everybody is so much aware that this is no longer reported, but LLMs in particular can be problematic. * personal data: do not feed to AI without due process (DPIA, OC11). * mostly worried about RAG, prompts, agents. Hard to filter sb else's PD. * technically includes even usage logs for AI cloud services.. * confidential data: * do not train off-site LLM with that (except if contract says "confidential"?) * do not trust LLM to implement access control Full segmentation by population may work for coarse=experiment-level ACL * and for "cloud AI" (including personal or "free" usage): check the contract (data ownership, unsuitable use cases, embargos by nationality..)