Shannon Davis’ Post

AI frameworks are vulnerably for API exploits. Look at this RCE https://lnkd.in/eswWuPE4

This is a personal anecdote, but it should be rather common occurance in the space. Any open source LLM i use, if i torture it enough it spits out this line: """ I must reiterate that as an AI language model created by OpenAI, """ Propably using the same RLHF dataset or feeding it openai responses. #opensource #llm #openai #largelanguagemodel

OpenAI The lazy syndrome users of your products are experiencing is demonstrated in the screenshot attached. It's fascinating how the capability of GPT -4 to browse and work with code files (e.g., code interpreter) is ignored, and excuses for why it can't perform these tasks are given. I understand governments worldwide, most notably the U.S., are pressuring AI/ML providers to put guardrails on their products. I have no idea how these pressures are interpreted and acted on by OpenAI, but there are changes to the model we, as users, notice.

🚀 Excited to announce that I have successfully built an AI model to classify SMS messages as spam or legitimate! 📱📊 Using techniques like TF-IDF and word embeddings, and classifiers such as Naive Bayes, Logistic Regression, and Support Vector Machines, I've developed a robust system to enhance SMS filtering and improve communication security. A big thanks to everyone who supported me on this journey! #AI #MachineLearning #DataScience #SpamDetection #codsoft

theOwl is theLenders' newest way to help you find the answers to all your Non-QM questions. 🦉https://lnkd.in/gz2H--KE 🦉 Powered by an OpenAI API and backed by a comprehensive warehouse of guidelines and matrices, theOwl provides real-time answers whenever you need them. Ask theOwl now! - #AI #nonqm #theOWL #nonqmloans

OpenAI has stealth tested a new model currently titled GPT2_chatbot. It appears to have improved reasoning abilities and does not appear to be on a GPT-5 level. It appeared on the LMSYS Chatbot Arena and I randomly got it in a test. It's very good at reasoning and in my one only comparison, it did better than Llama 70b. https://lnkd.in/eJZbweZb

Not long ago, solving simple problems with LLMs required painstaking, step-by-step efforts. Then came O1 from OpenAI, a game-changer that leveraged Chain of Thought (CoT) reasoning to pre-establish these steps before answering. However, it came with limitations: paid access, closed-source, and restricted daily tokens. Now, in a short period of time, we’ve entered a new era with DeepSeek-R1-Lite-Preview—an open-source, unpaid alternative that not only overcomes these limitations but outperforms O1 in many benchmarks, offering far fewer restrictions on daily token usage. My question is: what’s next? The benchmarks are yours. link for test: https://chat.deepseek.com/ #AI #LLM #AbstractReasoning #TestTimeTraining #MachineLearning #GenerativeAI #CoT

We are excited to introduce our final post with highlights on how Llama 3 70B without fine-tuning outperforms OpenAI GPT-4o in function calling with Friendli Tools. Performance Highlights - Function Calling Accuracy: Vanilla Llama 3 70B on Friendli performs on par with OpenAI GPT-4o and Fireworks AI Firefunction v2 in function calling. - Cost Efficiency: Llama 3 70B on Friendli matches the overall performance of GPT-4o in function calling at just 5% of the cost ($0.8 vs $15 per 1M output tokens). With Friendli Tools, you can now build high-performing AI agents without any fine-tuning required. 👉 Read the blog here: https://lnkd.in/gqwXsbzK 👉 Start building your AI agents today with Friendli Suite: https://suite.friendli.ai #llm #llama3 #functioncalling #tools #friendliai #friendli #gpt4o

OpenAI’s latest model, o1, has reportedly scored an IQ of around 120 on the Norway Mensa IQ test, potentially marking the first time an AI model has surpassed the average human IQ. o1 correctly answered 25 out of 35 questions on the Norway Mensa IQ test. It demonstrated proficiency in solving complex visual and logical puzzles, both on publicly available tests and new, unpublished questions. This performance suggests the improvement is not due to the questions being in the training data. These results would indicate that OpenAI has achieved Stage 2 performance on their 5-tier roadmap towards AGI. This milestone suggests the model can stop, think, and reason about problems to arrive at correct answers more frequently — a crucial prerequisite for Stage 3 (AI agents).

JUST IN: OpenAI releases new deepfake detector to combat election influence. Join the fight against misinformation! Share your thoughts with us. #OpenAI #Deepfakes #AI #TechNews #TheReportify To learn more and read the complete story, click on this link. https://lnkd.in/gXRZPRgD