Akinwale AkinolaHead, Project Management| JNC International LtdSurulere, Lagos, Nigeria
Quick simple analysis to have data for decision-making. For example, how many additional batteries to add to an existing system to increase the backup time by a certain amount of time? This is not my subject matter area and I would have depended on an expert to get it done. I was interested in the financial implications but AI did the analysis part for me pronto. Saving Changes...
Katty BibicProject Manager| TeocoHuntersville, Nc, United States
Hello Claudia, the company I work for hasn't explored GenAI yet, but I do see the benefits that this could bring to the company in multiple projects and line of business within the company. There is no doubt that human intervention will still be required, as GenAI is not perfect, but definitely will give PMs more time to focus on more strategic activities. Thanks. Saving Changes...
DIVINE KAYIIProject Manager ( Principal Project Engineer)| Shell Development CompanyPort Harcourt, Ri, Nigeria
I have use Chat GPT to generate project reports and enhance communication to project stakeholders. Saving Changes...
I have been using ChatGPT/Gemini/Bing/Mistral (self hosted) in the following activities:
1. white page, the start of activity where I get blocked because of not knowing how to start. ChatGPT is a great tool to give a set of bullets of what you want to describe, and it provides a nice start easy to continue (thus reducing the overall time of writing)
2. meeting minutes summarization when meeting online (when I am the host; remember to inform the meeting attendees that you are using a recording tool and the purpose).
3. web search, when the search objective is not easily described in simple words (Gemini / Bing).
4. Basic coding tasks or QA-related activities (creating scripts, tests, and other types of simple code). It works pretty well, although it may require different schemes to work properly (sequences, self-review agents, etc.)
5. Image content extraction and translation. Saving Changes...
Emmanuel UdoOther| Inflatus Consultants and Partners Port Harcourt, Nigeria
Jan 04, 2024 4:56 AM
Replying to Sergio Luis Conte
...
Challenges remain the same from long time ago (Gen AI is outside there from long time ago): AI entities have a grade of confidence associated to them that must be published with the results. So, the final decision still remains in human being hands. The same for the outcomes: helps me to search patterns in a hugh amount of data to create the information I need adding that today I can find data public and outside my organization.
We are still in the very early stages of learning about AI. Saving Changes...
Emmanuel UdoOther| Inflatus Consultants and Partners Port Harcourt, Nigeria
We are still the very early stages of learning about AI. Saving Changes...
Farhad AbdollahyanManaging Director| Cyrus Associados Apoio em ProjetosSao Paulo, Sp, Brazil
I'm using ChatGPT as an assistant and a secretary in writing draft e-mails and reports that I review before sending or sharing to avoid hallucinations (which I found pretty recurrent).
I use it intensively to translate from and into French, English, and Portuguese with excellent results. Saving Changes...
Sure, I can share an experience related to an AI project where language models (LLMs) were used for classification, and the challenges faced due to biases and inconsistencies in the model's output.
In this particular project, the objective was to develop an AI-powered content moderation system that could automatically classify user-generated content into different categories based on their nature and appropriateness. We leveraged a large language model trained on a vast corpus of text data to perform this classification task.
Initially, the LLM showed promising results, accurately classifying a wide range of content during our testing phase. However, as we started deploying the system in production and exposed it to a more diverse set of real-world data, we encountered several challenges.
First, we noticed that the LLM exhibited certain biases, likely stemming from the training data it was exposed to. For instance, it tended to misclassify content related to marginalized communities or underrepresented groups, often labeling them as inappropriate or offensive, even when they were not. This raised concerns about the fairness and inclusivity of our system.
Another issue we faced was the LLM's tendency to hallucinate or generate inconsistent outputs when presented with certain types of content or edge cases. Sometimes, it would confidently classify a piece of content into a particular category, only to contradict itself when the same content was presented slightly differently.
We realized that these inconsistencies and biases were likely due to the limited diversity and potential skew in the training data used by the LLM. While we had provided a large amount of data for training, it might not have been representative enough of the real-world scenarios our system would encounter.
To mitigate these issues, we took several steps:
1. **Data Augmentation**: We actively worked on expanding and diversifying our training data by collecting and annotating a wider range of content from various sources, ensuring better representation of different perspectives, communities, and contexts.
2. **Human-in-the-Loop**: We introduced a human-in-the-loop process, where a team of human reviewers would provide feedback and corrections on the LLM's classifications, allowing for continuous learning and improvement.
3. **Ensemble Approach**: We experimented with ensembling multiple LLMs trained on different data subsets, leveraging their collective strengths and mitigating individual biases.
4. **Explainable AI**: We integrated explainable AI techniques to gain insights into the LLM's decision-making process, helping us identify and address potential biases or inconsistencies more effectively.
While the journey was challenging, these measures helped us improve the reliability, fairness, and consistency of our content moderation system. We learned that relying solely on a single LLM trained on a limited dataset can lead to biases and inconsistencies, and a more holistic approach combining diverse data, human oversight, ensemble methods, and explainable AI techniques is crucial for building trustworthy and robust AI systems. Saving Changes...
David BrezlerOwner| Brezler, LLCWhite Plains, Ny, United States
Similar to Pascal Brunet in the Data Landscape of Gen AI training, I found myself presented with a problem related to Python coding in order to extract data from a large set of documents. The critical issue - and the reason I leveraged the AI tool - was that the Python coder I had on the team was unavailable at precisely the time I needed to engage, and we were running up against a very tight deadline. I was experience a hiccup in the code, and the data extraction wasn't happening. So I asked the GPT tool to analyze the code for issues. It was able to designate not only the problem area, but sample code to patch the problem. Within a few minutes, I had corrected the code, resolved the issue, and completed the data extraction, presenting the completed project to the client on time. Saving Changes...
Anonymous
I've used it to quickly iterate on prototypes, content (slide presentations for topics on AI usage for jobseekers), and social media content refinement. Saving Changes...
PMI Team Member
