Schema-Guided Reasoning (SGR) is a structured prompting technique that uses predefined schemas (SO) to guide large language models through explicit reasoning workflows, encoding expert cognitive processes directly into inference.

RT @LangChainAI: 🏆 Winning RAG Solution. A powerful RAG implementation that analyzes company annual reports through LangChain's framework.….

The real gem is not DeepSeek-r1. Even though it has MoE architecture, you still need 8xH100/H200 to run it efficiently. No, the real gem is deepseek-r1-llama-70B. This distillate works with reasoning so well (when given a chance via CoT/SO), so it beats all GPT-4o flavours (2/2).

Early preview of v2 of my LLM Benchmark. It tests various LLMs in different business workloads, but now makes use of reasoning capabilities of each model (places Chain-of-Thought slots in response schema and arranges the entire response layout via response schema/SO) (1/2)

I used to make jokes about Grok models of First version was . not good. Grok-2 made it into 12th place of our LLM Product Benchmark. This was not expected at all.

Mistral AI is good. But it is a common misconception that “it’s all open source”. Their latest and best models offered via the platform are closed.

Still, Google being Google, working with Gemini Pro 1.0 can be a pain, especially in EU. I recommend to skip it and use other similar models. This is a sneak preview of @trustbittech LLM Benchmark for March. Full report will be published on our website soon-ish. (3/3)

Claude 3 Haiku is a low-cost alternative with big context window 200k. Is a bit similar to Gemini Pro 1.0 in cost and performance, although Gemini has a smaller context. Both models provide an inexpensive alternative to running a good fine-tune of Mistral 7B locally. (2/3).

Claude 3 Opus catches up with Chat GPT-4 on product oriented LLM benchmarks. It also scores perfect 100 score in document comprehension. Kudos to @AnthropicAI for listening to the needs of business and enterprise customers! (1/3)

If we prove that this ChatGPT thing actually works, how can we quickly catch up with it? . How can we employ this new flavour of AI across the company in a pragmatic way?. More on that is in the latest release of ML Under the Hood.

In less than an hour, I'll be talking on the webinar about LLMs in the business. Webinar is free. Language - English. Begins at 10:00 CET. Link: Join @vermorel and me!

RT @rushing_andrei: While the AI tech is very novel, powerful and exciting, if it doesn't solve any real customers problems -- it's complet….

A simple ChatGPT example I always present at the start of my AI Coding Workshop. Draw a screen mockup, copy to ChatGPT and ask to generate the layout. Result is usually usable right away for simple cases.

Microsoft now has power to distribute the most precious resource (GPUs) between two top players at will.

In my opinion, this is the perfect scenario. Sam won't have to start from scratch, and OpenAI can begin to decelerate peacefully. The only downside now is that we'll have to switch all endpoints to Azure API 😂 (3/3).

The new CEO of OpenAI is the former head of Twitch, who sold it to Amazon for $1 billion and was sufficiently bored to accept such an offer. He had previously mentioned that it was time to slow down. (2/3).

Sam Altman is moving to Microsoft, where he will lead a new AI research division, as just stated by Satya Nadella. His colleagues are joining him in this move. Microsoft promises to provide resources and move quickly. (1/3).

ChatGPT, draw me a mermaid diagram! $MERMAID_CODE. Now you know how good ChatGPT is at this 😆

@eMundoGmbH You can read more about "AI in Software Engineering" event on LinkedIn: I'll be sharing a distilled report of findings and take-aways in my next newsletter at (2/2)

You need just 2 hours to become a lot more productive in coding. Here is what an experienced software craftsman (Lead IT consultant at @emundogmbh) wrote about a workshop on "AI in Software Engineering" (1/2)