🤖 Generative AI Lead @ AWS ☁️ (150k+) | Startup Advisor | Public Speaker | AI Outsider | Founder Thinkfluencer AI
🧠 🚀 𝗧𝗵𝗲 𝗘𝘃𝗼𝗹𝘃𝗶𝗻𝗴 𝗟𝗮𝗻𝗱𝘀𝗰𝗮𝗽𝗲 𝗼𝗳 𝗟𝗮𝗿𝗴𝗲 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀: 𝗪𝗵𝗮𝘁'𝘀 𝗰𝗼𝗺𝗶𝗻𝗴 𝗡𝗲𝘅𝘁? Previously (about a year ago), a single proprietary provider often led the race with others trailing. Today, the situation is different. Providers like Anthropic with the Claude family have emerged as frontier models in human intelligence. Google has closed the gap significantly with Gemini becoming a serious competitor. OpenAI continues to leverage the GPT-4 family despite high expectations. Here are some important insights: 𝟭. 𝗠𝘂𝗹𝘁𝗶-𝗠𝗼𝗱𝗲𝗹 𝗙𝘂𝘁𝘂𝗿𝗲: Customers need the flexibility to switch, combine, and mix different LLMs across various use cases. 𝟮. 𝗠𝘂𝗹𝘁𝗶-𝗠𝗼𝗱𝗮𝗹𝗶𝘁𝘆: Top proprietary models are now multi-modal, broadening the scope of potential use cases. 𝟯. 𝗢𝗽𝗲𝗻 𝗦𝗼𝘂𝗿𝗰𝗲 𝗖𝗼𝗺𝗽𝗲𝘁𝗶𝘁𝗶𝘃𝗲𝗻𝗲𝘀𝘀: Models like LLaMA3 and Mistral are proving to be competitive, especially in non-English languages. 𝟰. 𝗣𝗹𝗮𝘁𝗳𝗼𝗿𝗺 𝗘𝘅𝗽𝗮𝗻𝘀𝗶𝗼𝗻: Many platform providers, like Amazon Bedrock did 18 months ago, have expanded to support multiple LLM providers, showcasing a forward-thinking vision. 𝟱. 𝗞𝗲𝘆 𝗠𝗲𝘁𝗿𝗶𝗰𝘀: Latency, cost, and accuracy are crucial when choosing models. Proprietary providers often offer different model sizes to meet diverse customer needs. 𝟲. 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝗘𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻: Benchmarks like MMLU might need redefining to allow for further improvements, especially in reasoning capabilities. Yet, for 90% of the customer use cases, smaller and faster models like Haiku and Mistral AI small are more than enough. 𝟳. 𝗙𝘂𝗻𝗰𝘁𝗶𝗼𝗻 𝗖𝗮𝗹𝗹𝗶𝗻𝗴: Many top LLMs now support function calling, enabling LLMs to perform specific actions, crucial for agentic (well appointed by Andrew Ng) implementation. 𝟴. 𝗘𝗺𝗲𝗿𝗴𝗶𝗻𝗴 𝗧𝗿𝗲𝗻𝗱𝘀: Expect a trend towards smaller, fine-tuned models rather than new providers. IBM is already exploring this area. 𝟵. 𝗦𝗶𝗹𝗶𝗰𝗼𝗻 𝗖𝗼𝗺𝗽𝗲𝘁𝗶𝘁𝗶𝗼𝗻: Given the computational demands of training and running these models, we will soon see strong competitors to NVIDIA at the silicon level. Yesterday first example with Etched. 𝗢𝗻𝗲 𝟴𝘅𝗦𝗼𝗵𝘂 𝘀𝗲𝗿𝘃𝗲𝗿 𝗿𝗲𝗽𝗹𝗮𝗰𝗲𝘀 𝟭𝟲𝟬 𝗛𝟭𝟬𝟬 𝗚𝗣𝗨𝘀. 𝟭𝟬. 𝗡𝗲𝘄 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲𝘀: More capable and efficient architectures than transformers will emerge. Providers investing in this will gain a unique competitive advantage. AI21 Labs just released Jamba in Amazon Bedrock!!! What to Expect Next? Leave it in the comments 👇 #AI #LLM #MachineLearning #Innovation #TechTrends
It’s fascinating to see how the landscape of large language models is evolving. The rise of multi-modal and open-source models, alongside the expansion of platform providers, is certainly shaking things up. I’m particularly interested in autonomous agents based on LLMs and their abilities to handle increasingly complex tasks through function calling, enabling them to perform specific actions seamlessly. Which emerging trend do you think will have the most significant impact on the industry?
Very informative. Thanks for sharing Eduardo.
I am seeing latency and costs as major drivers for choice. Small Language Models are very appealing, especially at a fraction of the cost - when compared to the major LLMs. Thanks for sharing Eduardo Ordax. Very informative.
Reliable benchmarks are paramount, and the juncture at which the meta-model selects a specific model must be accorded the utmost priority. Excellent graph. In the context of the AI field, this statement makes sense. Reliable benchmarks are essential for assessing model performance accurately. The process where the meta-model (which oversees or selects among other models) chooses a specific model is critical, as it impacts the overall system's effectiveness and efficiency. Ergo, prioritizing this selection mechanism ensures optimal model deployment and performance.
Great pointers, any further details on IBM approach?
Loris Millet
🤖 agentic followed by AGI 🤔
This is good