Magic’s Post

The Magic team is inspired. They've developed a way to process queries with an ultra-long context window of 100M tokens - about 1000x times greater than the frontier models of other providers. See below for an excerpt of the blog post explaining what they've done. Thanks to Sonya Huang at Sequoia Capital for introducing me to the amazing Eric Steinberger of Magic on Sequoia's Training Data podcast. There are currently two ways for AI models to learn things: training, and in-context during inference. Until now, training has dominated, because contexts are relatively short. But ultra-long context could change that. Instead of relying on fuzzy memorization, our LTM (Long-Term Memory) models are trained to reason on up to 100M tokens of context given to them during inference. While the commercial applications of these ultra-long context models are plenty, at Magic we are focused on the domain of software development. It’s easy to imagine how much better code synthesis would be if models had all of your code, documentation, and libraries in context, including those not on the public internet.

Tachyum Demonstrates Full BF16 AI Support in GCC and PyTorch https://lnkd.in/gW3xFGm7 #pytorch #GCC #AI #BF16 #CPU #GPU #TPU

Know Your Neighborhood: General and Zero-Shot Capable Binary Function Search Powered by Call Graphlets This is a Plain English Papers summary o... #ai #beginners #datascience #machinelearning #prodsenslive #Software https://lnkd.in/dv3q-SFb https://lnkd.in/d_qNWxRM

Exciting news for all AI and machine learning enthusiasts! Google's Gemma family of open models has new additions - CodeGemma and RecurrentGemma. CodeGemma offers a powerful coding experience with its pretrained variants for code tasks, instruction-tuned variants, and a 2B variant for fast code completion. Meanwhile, RecurrentGemma boasts a unique architecture that uses recurrent neural networks for improved memory efficiency, maintaining speed regardless of sequence length. Both models are built on the Gemma foundation and embody responsible innovation. Check them out now! #AI #MachineLearning #CodeGemma #RecurrentGemma

Andrej Karpathy is one of the leaders in AI and ML. Here he implement LLM training in pure C/CUDA. This shows how straight forward fine-tuning your own LLMs can be. There are other tools like https://unsloth.ai/ that can help you train larger models faster and for less money. https://lnkd.in/g-jh7wak #ai #ml #artificialintelligence #machinelearning #LLM

My First AI Project and comparative AI An interesting conversation with a friend on how I got started in AI. Rewrite systems refers to a paradigm of computation with simple syntax and semantics. My first AI project for Master thesis involved checking for infiniteness of computer programs. It was a supervised learning model using Sturm’s polynomial equation to check the infiniteness. This concept could easily be extended for super compilers by parallel evaluation of different functions within a modular program to confirm the infiniteness. It is amazing that different concepts evolve from base technologies. Super computers were used for distributed computing but Hadoop made a parallel processing model feasible due to necessity for performing computations on big data. Similarly robots with learning behaviors were introduced to mimic human intelligence. Now in this decade focus has shifted to a software version of AI with Generative AI like concepts.

C++ code that use Neural Network to detect vulnerabilities in C Compilers 👉 Compilers can inadvertently generate vulnerabilities, among others: Buffer Overflows: Code Injection: Integer Overflows: Denial of Service (DoS): Information Disclosure: Insecure Default Settings: Undefined Behavior: Race Conditions: Dependency Vulnerabilities: Misconfiguration: Front-end Bugs: Back-end Bugs: 👉 The C++ code below implements a straightforward neural network aimed at detecting vulnerabilities in C compilers. Neural Network Class: The NeuralNetwork class defines the structure and functionality of the neural network, including methods for making predictions and training. Constructor: The constructor sets up the network with specified sizes for input, hidden, and output layers, and it randomly initializes the weights for the connections between these layers. Predict Method: The predict method accepts an input vector that represents various features related to vulnerabilities and calculates the output using the weights. It employs the sigmoid activation function for both the hidden and output layers. Train Method: The train method is designed for training the neural network with a dataset of training examples and their corresponding labels. While the specific training logic is not detailed here, it typically involves forward propagation, loss calculation, and backpropagation. Sigmoid Function: The sigmoid method applies the sigmoid activation function to a vector, which is a standard choice for neural networks to introduce non-linearity. Random Weight Generation: The randomWeight method creates random weights for the connections, ensuring that the network begins with a variety of initial values. Main Function: The main function instantiates the neural network with 12 input features (related to the listed vulnerabilities), 8 hidden neurons, and 1 output neuron. For a complete implementation, actual training data and labels would need to be supplied. #NeuralNetwork #Cplusplus #machinelearning #ML #vulnerability #compiler #bugs #frontend #backend #misconfiguration #security #dependency #insecure #exploit #buffer

There's a real resurgence of interest in compilers, and for good reason! In the world of AI, compilers play a critical role. They act like translators, taking the blueprint of a neural network (its architecture) and turning it into code that can run on specific hardware. The exciting thing is that both neural network designs and hardware are constantly evolving, which means there's a constant need for new and improved compilers. Microsoft Research is at the forefront of this compiler revolution. They recently released a whole set of cutting-edge AI compilers, each one designed to tackle specific challenges faced by deep learning models. These challenges can range from getting the most out of specialized hardware to making the models run faster and use less memory. Rammer, Roller, Welder, and Grinder, tackle a range of compiler optimization challenges based on the same tile abstraction, providing a comprehensive solution to connect AI models with hardware. If you’re interested to learn more check https://lnkd.in/d5Nar7cv

Discover the power of TensorFlow! This open-source machine learning framework by Google is revolutionizing AI development. From neural networks to deep learning models, TensorFlow empowers developers to create cutting-edge AI solutions with ease. Key features: *High and low-level APIs for model development *Eager execution for immediate iteration and intuitive debugging *TensorBoard for visualization and debugging *Distributed computing support *TensorFlow Lite for mobile and embedded deployment *Extensive documentation and community support TensorFlow is widely used in various domains, including image and speech recognition, natural language processing, and robotics. Its versatility and scalability make it a top choice for both research and production environments.