Rogerio Ferraz @ferrazrog

@Suryanshti777 ROADMAP

Most people want to become AI Engineers Very few know what to learn next That's why I'm giving away one of the best Agentic AI Engineer Roadmaps for FREE to the first 4500 people only Inside you'll learn ✅ Python Fundamentals ✅ LLM Fundamentals ✅ LangChain LangGraph CrewAI AutoGen ✅ LCEL Runnables & Workflows ✅ Memory Systems ✅ Tool Integrations ✅ RAG Systems ✅ Multi Agent Systems ✅ Real World AI Projects ✅ Interview Questions & Answers 72 Hours Only How to get Follow me (so I can DM you) Like + RT Comment "ROADMAP" Once the limit is reached I'll stop sending it.

keyboard_arrow_left Previous keyboard_arrow_right Next

Confused by all the different types of AI Agents? Here’s a simple breakdown of the 5 core types and how they differ! ⬇️ 1. Simple Reflex Agents Work on if-then logic. No memory. Fast, but only suited for basic, predictable environments. 2. Model-Based Reflex Agents Add memory and a transition model. Can handle sequences and partial observability - ideal for robotics. 3. Goal-Based Agents Plan actions to reach a goal. Evaluate future states and paths. Used in search, planning, and games. 4. Utility-Based Agents Go beyond goals, optimize outcomes based on preferences, costs, and satisfaction. Think smarter agents. 5. Learning Agents Most advanced. Learn from experience and feedback. Adapt strategies, support all learning types, and evolve. 6. Common Building Blocks Across All Agents: Sensors, effectors, decision logic, performance modules, and interaction with the environment. This guide can help you understand how autonomous systems make decisions, adapt, and improve from simple bots to super-intelligent agents.

@NvidiaANZ GeForce Season

@NVIDIAGeForce GeForce Season

“The world is moving so fast that the person who says it can't be done is generally interrupted by someone doing it.” -- Elbert Hubbard #quote #goodmorning

Check out our new Data Science Agent 👇

Google AI Developers @googleaidevs

a year ago

Accelerate data science workflows with Google Colab's new Data Science Agent. This agent uses Gemini to act as your coding partner: upload your data, define your analysis goals, and watch a Colab notebook take shape. Now available for all Colab users → goo.gle/41DD7IF

20 218 1K 240K 659

@johnnyxbrown 2025

Você se lembra de quando ingressou no X? Eu me lembro! #MeuAniversárioNoX

Anthropic released a prompt improver in their developer console that automatically refines prompts It uses advanced techniques like chain-of-thought reasoning On tests it shows a 30% accuracy boost in classification and perfect word count adherence

A proibição de usar VPN deixa claro que o problema não é com o Twitter operando no Brasil, o negócio é proibir brasileiros de acessar o Twitter. O resto é lenda.

1/n Proxy-Tuning: A New Method in Customizing Large Language Models Harnessing the full potential of of large language models (LLMs) often requires fine-tuning them for specific tasks or domains. Traditional fine-tuning methods, while effective, present significant challenges. They are computationally expensive, demand access to the model's internal workings, and can lead to catastrophic forgetting, where the model loses previously learned knowledge. This is where proxy-tuning emerges as a groundbreaking solution. Proxy-tuning, as presented in the paper "Tuning Language Models by Proxy," introduces a novel paradigm for customizing LLMs without the need for direct access to their parameters or extensive computational resources. The core idea is remarkably simple yet ingenious: leverage the knowledge of a smaller, easily fine-tuned LLM to guide the larger model's behavior at the decoding stage. This is achieved by contrasting the predictions of a tuned "expert" model with its untuned counterpart, an "anti-expert," and applying the difference as an offset to the larger model's output probabilities. In essence, the expert acts as a tutor, subtly nudging the larger model towards the desired behavior without altering its underlying structure. The paper meticulously demonstrates the efficacy of proxy-tuning through a series of experiments spanning various tasks and scenarios. In instruction-tuning tasks, proxy-tuned models exhibited dramatic improvements over their base counterparts, nearly closing the performance gap with models fine-tuned through traditional methods. On knowledge-intensive tasks, they even surpassed the accuracy of directly fine-tuned models, suggesting that proxy-tuning can better preserve the factual knowledge embedded within the LLM. Moreover, the technique proved successful in adapting LLMs to the domain of code and achieving task-specific fine-tuning for question answering and math problems. One of the most compelling aspects of proxy-tuning is its ability to work with black-box models, where access to internal parameters is restricted. The paper showcases this capability by successfully applying proxy-tuning to GPT-3.5, a powerful but proprietary LLM, for temporal adaptation, enhancing its knowledge of recent events. This opens up exciting possibilities for customizing commercially available LLMs without compromising their proprietary nature. Proxy-tuning builds upon and distinguishes itself from previous approaches to LLM customization. While fine-tuning directly modifies model parameters, proxy-tuning operates at the decoding stage, offering greater efficiency and flexibility. Compared to prompt engineering, which relies on crafting specific input prompts, proxy-tuning provides a more systematic and robust way to guide LLM behavior. Additionally, it simplifies the concept of decoding-time experts by requiring only one expert and one anti-expert, both derived from the same small LLM.

Today, with @Tim_Dettmers, @huggingface, & @mobius_labs, we're releasing FSDP/QLoRA, a new project that lets you efficiently train very large (70b) models on a home computer with consumer gaming GPUs. 1/🧵 answer.ai/posts/2024-03-…

I spent all day learning Metahuman Animator and managed to create a digital version of myself. It's using one of our new Scan Store Metahuman identities and textures (my face), captured using an iPhone. Can't wait to finish these and get them out there.

People are just starting to realize the power of Local LLMs Especially with new Apple chips. It's a game-changer Let me show you: Falcon is 180B LLM. It is the size of GPT3. How fast would it run on your Mac? Apple M3 Max: 3.5 tokens/sec We can compare it to 2 x A100 (2 x 30k$) 2x A100 80GB: 7 tokens/sec Now add to this the recent speed of @NousResearch & @teknium with models like openHermes And the fact that soon they will add Vision functionality to send images x.com/Teknium1/statu… and we get real competition to commercial models from OpenAI and Anthropic More of independent companies start to publish their model weights Take 01.ai team from China. It is currently leading the OSS leaderboard in quality & context size 01 team has 200k context window model. It's the 7 copies of Shakespeare's Hamlet which can fit onto a single prompt GPT4 turbo has only 128K 7B models like Mistral are already running at 40+ tok/s! This is insanely fast! I have run a test on my old M1 a gif below to give you a feeling of the speed of it. Absolutely stunning work is done by @ggerganov. Georgi & llama.cpp community have literally shaped the progress of open source models with llama.cpp project. It allows you to run any LLM on your computer. You can even run llama on your ios with it I am incredibly bullish on Apple metal & local computations. There are many interesting projects rising around it, one that came to my mind is @opentensor – decentralized commodity of computing resources like selling your GPU to someone And I haven't even started on the power of webGPU & Web LLM What other interesting stuff have you seen? share in comments

keyboard_arrow_left Previous keyboard_arrow_right Next

Teknium 🪽 @Teknium

3 years ago

I should have multimodal vision Hermes next week if all goes well

42 24 528 109K 37

FREE Machine Learning Book! If you want to get started with Machine Learning and Deep learning, check out this Deep Learning with Python Book! This book covers: - Neural Networks - Convolutional Neural Networks - Recurrent Neural Networks - LSTM - Generative Adversarial Networks Here is the book link: tanthiamhuat.files.wordpress.com/2018/03/deeple…

"A falta de amor é a maior de todas as pobrezas." -- Madre Teresa de Calcutá #frase #boanoite

All the hype about machine learning for big applications like self-driving cars, playing Go, passing programming tests. I'm more exited by the "small" ML successes that solve concrete problems. Hidden in organizations, no PR. Just a data scientist (team) doing the real work.

The 3-point seat belt was invented by Nils Bohlin (Volvo) in 1959. Modern seat belts have saved millions of lives. The reason the three-point seatbelt is so widely adopted is actually because Volvo opened up the patent so that any car manufacturer could use it in their design.