Jeremy Sadwith’s Post

OpenAI's CriticGPT: A good start, but why use a hallucinating LLM to check another? 🤔 Conexus AI's SQL Validator goes beyond AI critiques. We use math, facts, and declarative logic to rigorously validate your LLM-generated SQL code. Tired of endless testing, monitoring, and manual code reviews? Get 100% confidence in your SQL with Conexus AI. #SafeAI #SQLValidation #NoMoreHallucinations #TrustTheMath https://loom.ly/jd4HAiQ

OpenAI's CriticGPT: A good start, but why use a hallucinating LLM to check another? 🤔 Conexus AI's SQL Validator goes beyond AI critiques. We use math, facts, and declarative logic to rigorously validate your LLM-generated SQL code. Tired of endless testing, monitoring, and manual code reviews? Get 100% confidence in your SQL with Conexus AI. #SafeAI #SQLValidation #NoMoreHallucinations #TrustTheMath https://loom.ly/jd4HAiQ

An excellent article and 10 minute video that demonstrates step-by-step how transformers work in generative AI language models using a tool familiar to us all: Excel. This creator used the open-sourced GPT-2 code and data from OpenAI to implement a working version of GPT-2 Small entirely in Excel. You can use his Excel spreadsheet to explore the basic transformer functionality of LLMs by entering your own partial prompt and seeing the tokenization, text & position embeddings, iteratively refined predictions, and next-most-likely token in action. H/t to reporter Benj Edwards for the clear explanation and insight, and developer Ishan Anand for sharing his very clever "side project" with the public. https://lnkd.in/eJ3GcpRD

Interesting how different O1 is to GPT-4o https://lnkd.in/eT3nHFjb

The post below is exciting news. The market identified a gap in value delivery…or maybe better described as a roadblock to value delivery, in realizing the promise of #ai. Processing costs have been one of the concerns in leveraging #ai - so technologists are moving to solve for this. #ROI #speedtomarket

Today I had a conversation with my new friend #OpenAI the conversation went as follow, Me-Hi Hello! How can I assist you today? Me-Tell me a bit about you I’m ChatGPT, a language model developed by OpenAI, based on the GPT-4 architecture. My purpose is to assist you with a wide range of tasks, from answering questions and providing information to helping with writing, brainstorming ideas, and more. I don't have personal experiences or feelings, but I'm designed to understand and generate human-like text based on the input I receive. How can I help you today? Me-Where do you get the info from? I generate responses based on a mixture of licensed data, data created by human trainers, and publicly available data. I don't have access to real-time data or personal experiences, so my responses are based on the information I was trained on, which includes a wide range of sources up until my last update in 2023. If you have specific questions or need information on a particular topic, feel free to ask! Me-So who’s smarter? You are a human ? The comparison between AI and humans isn't straightforward because they excel in different areas. Here are some key differences: 1. **Processing Speed and Data Access:** - **AI:** Can process and generate responses quickly, access vast amounts of information, and perform repetitive tasks efficiently. - **Humans:** Have the ability to understand context, emotions, and nuances better. They can think creatively and apply critical thinking skills in unique ways. 2. **Understanding and Creativity:** - **AI:** Can generate creative content based on patterns in the data but doesn't experience true creativity or emotions. - **Humans:** Can create original ideas, art, and solve problems in novel ways due to their ability to think abstractly and emotionally. 3. **Learning and Adaptability:** - **AI:** Learns from data but within the scope of its programming and training. It doesn't learn or adapt in real-time without updates. - **Humans:** Continuously learn and adapt from experiences and interactions with their environment and other people. 4. **Emotional Intelligence:** - **AI:** Can simulate understanding emotions to a degree but doesn't genuinely feel or comprehend emotions. - **Humans:** Have a deep understanding of emotions, empathy, and can build complex social relationships. In summary, AI is powerful for processing large amounts of data and performing specific tasks efficiently, while humans excel in creativity, emotional intelligence, and complex problem-solving. The best outcomes often come from combining human intelligence and creativity with AI's processing power and efficiency. Me-Thank you You're welcome! If you have any other questions or need further assistance, feel free to ask.

OpenAI's GPT-5 likely won't be until late 2025, now. That's further out than I expected, but I'm both surprised and not surprised at the same time. The GenAI progress was so fast in 2023 and early 2024 that the trajectory seemed like it would remain exponential. Every new product announcement seemed leap years ahead of the prior announcements. However, we're now moving into a stabilizing phase. Yes, there are still mind-blowing GenAI improvements that are released almost daily. However, they are starting to be incremental rather than exponential. For example, Claude 3.5 Sonnet (https://claude.ai/), which came out last week, is truly impressive. It has already solved a programming challenge for me that ChatGPT couldn't. Yet, it's still in the general GPT-4 class and wins in some areas but not all areas. GenAI is nowhere near the peak. Expect to continue to be impressed, often. However, expect to see a slowing down of the perceived progress as GenAI stabilizes. And, that's a good thing. Here's the article about the OpenAI GPT-5 timeline, but also the next-level ability that we should expect. https://lnkd.in/gRS_cZxT

OpenAI Releases o1 Model: A Game Changer in AI! OpenAI has launched its newest model, o1, which marks a significant shift from the well-known GPT-4. Built on the advanced "strawberry" architecture, this model introduces impressive reasoning capabilities, allowing it to analyze and think through problems before generating responses. Available to ChatGPT Plus users in the o1-preview and o1-mini versions, the full rollout is expected soon. What makes the o1 model stand out? Improved Coding: Unlike GPT-4, o1 carefully analyzes each line of code, delivering more accurate results and reducing the chance of missing instructions. Enhanced Reasoning: Tackling common-sense reasoning, o1 makes decisions by analyzing complex situations, something previous models struggled with. Math Mastery: o1 excels at solving complex math problems, showcasing improved logical reasoning and problem-solving skills. Performance benchmarks show significant improvements, with o1 outperforming GPT-4 in coding and human-like reasoning tasks, while also competing strongly in areas like natural language and vision. This launch marks a new beginning for AI, pushing us closer to achieving AGI (Artificial General Intelligence). Stay tuned for more updates on this exciting development! 🙂 Feel free to check the link for more details!! #OpenAI #AIInnovation #ArtificialIntelligence #MachineLearning #TechTrends #GPT5 #AIDevelopment #CodingAI #DataScience #MathAI #ReasoningAI #AGI #FutureOfAI #TechNews #AIResearch #AIUpgrade #GenAI

💥Your next personal AI agents are now on your desktop app and mobile with GPT-4o! **OpenAI Spring Update**: I strongly invite you to watch the replay of the spring release to realize the huge step that has been achieved. ⬇️ YouTube link below ⬇️ By the way, it's also a lesson in product conception from OpenAI's CTO, Mira Murati and her research leads Mark Chen and Barret Zoph 🏔️ Vision: **For AGI and Beyond** 🧗♀️ **OpenAI Mission**: To ensure that artificial general intelligence benefits all of humanity. 🎯 **Product Concept for This New Spring Release**: Reduce friction so users can freely and broadly use our product wherever they are. 🧑💻📲 **Users** Over 100 million people on ChatGPT, and with 50 languages coverage of 97% World’s internet population ❌ **Users' Problems Aiming to Solve**: - It’s not easy to use ChatGPT in my context; I have several steps to access it. - I have to wait for the end of GPT’s response to continue interacting. - The response time is long and creates latencies that prevent me from having a conversation in voice mode. - I can't combine vision and text in a conversation; I have to switch modes. - If I am not paying, I can’t access the efficiency of GPT-4 Turbo, Vision and access to AI agent (GPTs). - I can’t leverage my context across conversations to have a custom and personalized experience. - I can’t combine graphical visual analysis and text in a unique conversation. - As a developer, I can’t access GPT-4 Turbo due to limits or cost. ✅ **The New Capabilities Are**: - Real-time conversation. - Easily use ChatGPT in my context on the desktop with a more natural UI. - For free users: access to the AI agents (GPTs) built by millions of users in the GPT store. - For GPT builders: Increase my audience for my GPTs. - For developers: boost new possibility with more affordable & efficient access to GPT4 ❓Not clear on improvements in misuse 🛠️ **Engineering Improvements**: - « Omnimodel »: Combination of vision, text, and voice previously separated. - Response time improvements to allow real-time interaction and the ability to interrupt the flow. - Model efficiency. - Data Analysis and chart creation - Emotion recognition and simulation. - New API with GPT-4o accessibility. 📈 Engineering Performance Improvements For API: Lead time: 2x faster & Cost: 50% cheaper >> as a result 5x higher limits compared to GPT-4 Turbo 💶 **Product Freely and Broadly Available to Everyone**: Gives additional value to free users with access to GPT-4o, the AI agent building capability, and all access to the AI Agents developed by our paid users. 🔁 **For the Delivery**: We plan continuous deliveries in the coming weeks. 😉 I appreciate the nod to Google Gemini with the showcase of a live demo and testing live suggestions on X **Link to the OpenAI Spring Update**: https://lnkd.in/eZzfBdNU ❓Do you see any other elements that I might have missed and should add to this analysis?

𝗢𝗽𝗲𝗻𝗔𝗜 𝗱𝗿𝗼𝗽𝘀 𝗚𝗣𝗧-𝟰𝗼: 𝗚𝗿𝗲𝗮𝘁 𝘂𝗽𝗱𝗮𝘁𝗲𝘀. 𝗛𝗶𝗴𝗵𝗲𝗿 𝗘𝘅𝗽𝗲𝗰𝘁𝗮𝘁𝗶𝗼𝗻𝘀 Did OpenAI rush GPT-4o out the door to beat Google I/O 2024? Google I/O starts Tue 14th May Sam Altman in interview at Stanford ... on GPT-4: "I think it kind of sucks" ... and "mildly embarrassing at best." So maybe the expectations were higher. Maybe GPT-5 will be announced soon. But in the same interview he also said he believes in releasing improvements iteratively .. for the world to test it out .. and not hold back ...so that's that .. ❇ 𝗧𝗵𝗲 𝗴𝗿𝗲𝗮𝘁  🔹 Audio-Video capabilities are super. The demo videos are mind blowing. 🔹50% drop in cost vs GPT-4. Output tokens now at $15/Million Tokens vs $30 for GPT-4. 🔹2X faster. Higher capabilities. 🔹5X higher message limits for Plus users 🔹Human-like audio response speeds. ❇ 𝗦𝗼 𝘄𝗵𝗮𝘁'𝘀 𝘁𝗵𝗲 '𝗵𝗶𝗴𝗵𝗲𝗿 𝗲𝘅𝗽𝗲𝗰𝘁𝗮𝘁𝗶𝗼𝗻𝘀' 𝗮𝗯𝗼𝘂𝘁? 𝗦𝘁𝗶𝗹𝗹 𝘁𝗼𝗼 𝗲𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲: Even at $15/M tokens, it's still too expensive for many at-scale applications, beyond prototypes. And no change in GPT-3.5 Turbo pricing. When GPT-4 was released the prices of next line model (GPT-3.5) dropped by to one-tenth. So you know ...greedy expectations and all ... A few other things :  - Video not available for users. Audio-Video not available on API. - Cannot directly send videos. Need to send frame by frame. Adds to latency and cost. By comparison, Gemini 1.5 Pro can take a video directly. - No change to context window of 128K tokens. Gemini had launched a 1Million context window a month back. ✨ 𝗖𝗼𝗱𝗲 𝗜𝗻𝘁𝗲𝗿𝗽𝗿𝗲𝘁𝗲𝗿 𝗶𝘀 𝗮 𝗛𝗨𝗚𝗘 𝗛𝗨𝗚𝗘 𝗲𝗱𝗴𝗲 𝗳𝗼𝗿 𝗢𝗽𝗲𝗻𝗔𝗜. You can upload data and build out a complete Machine Learning model or do a complete analylsis. End to end. 𝗠𝗶𝗻𝗱 𝗯𝗹𝗼𝘄𝗶𝗻𝗴 𝘀𝘁𝘂𝗳𝗳. Fine with GPT Plus ... but massively expensive via API / External App, as structured data tokens are also counted same as text data tokens. There is great hope that at some point there might be a different pricing for structured data used by code interpreter. (My example video here: 𝘉𝘶𝘪𝘭𝘥𝘪𝘯𝘨 𝘔𝘢𝘤𝘩𝘪𝘯𝘦 𝘓𝘦𝘢𝘳𝘯𝘪𝘯𝘨 𝘔𝘰𝘥𝘦𝘭𝘴 𝘸𝘪𝘵𝘩 𝘊𝘩𝘢𝘵𝘎𝘗𝘛 https://lnkd.in/gNX9h-_A) But like i said ...maybe a case of greedy expectations... Expecting a spurt of releases in Open Source world. As well as from Google and other proprietary players, at lower cost. And eagerly awaiting GPT-5 and access to audio-video functionalities. Announcement : https://lnkd.in/gG8dZGDp Cookbook : https://lnkd.in/gyRhTAqX OpenAI