SafMed’s Post

QZ Platform revolutionizes your exams with ease... and much more

Berkeley Packet Filter (BPF) is an in-kernel execution engine that processes a virtual instruction set. eBPF is an extension of BPF to extend kernel functionality Netflix has released a command line tool named bpftop to perform performance optimizations and monitor eBPF programs. bpftop, written in Rust, enables eBPF runtime statistics using the BPF_ENABLE_STATS syscall. It collects data per program every second and presents it in tabular or time series format. Once terminated, it turns off statistics gathering. It employs libbpf-rs and ratatui crates for functionality. bpftop simplifies the performance optimization process for eBPF programs by enabling an efficient cycle of benchmarking, code refinement, and immediate feedback. #performance #performanceengineering

One thing I like the most about KodeKloud is the quick access to a K8s cluster to pratice with.

I have successfully passed ECES v3 course. With digitilization and crossing global communicative information, conducting robust encryption algorithms in parallel is non negotiable. Future looks bright. #ECES #encryption

I’m thrilled to share a new series of models I open sourced today called “Self-RAG Classifiers.” This work draws inspiration from the paper "Self-RAG" (as the name implies.) Self-RAG fine-tunes a single language model to generate reflection tokens. I've flipped the problem and developed a series of models that generate these tokens. The core LLM handles the most challenging task of generation. Separate classifiers handle reflection. This approach has the benefits of Self-RAG with more flexibility. Every RAG system will work best with different foundation models. This approach makes it easy to hot-swap the model responsible for generation and still generate reflections. This also stands for the individual classifiers. If you want to adapt one to your system you do not need to retrain the entire model, only the single classifier. I'm currently benchmarking the system performance for RAG. Stay tuned for updates! #largelanguagemodel #artificialintelligence #retrievalaugmentedgeneration Original Paper: https://lnkd.in/d5BGer4h Hugging Face collection: https://lnkd.in/djyxE3bj

Discover 5 techniques to optimize token usage without sacrificing accuracy #llms #mlops #chatgpt #rag #gemini #aiml #llmops https://lnkd.in/dCGYFEAh

👉 Swipe to discover: - What QEMU is - How QEMU helps you test smarter, not harder - What you can use it for (hint: QEMU can help accelerate your time to market) - Potential challenges of working with QEMU Learn more here: https://lnkd.in/eh5-AF9U #qemu #emulation #opensource #embedded #softwaretesting

It's always a pleasure talking at #DevConf2024, one of the best conferences out there! Until recently, #kubernetes did not support swap in a usable way. This was due to a discontinued development that kept the NodeSwap feature in an alpha state for an extended period. However, with collaborative efforts from multiple developers, I successfully continued the development and elevated swap support to full Beta in Kubernetes version 1.30. Currently swap on Kubernetes is fully supported bringing cgroup v2 support, newly introduced "swap behaviors" and a strong emphasis on system stability. In this talk I share the journey of bringing swap support to Kubernetes. This will include a technical overview, insights into the design choices we made and the challenges we encountered along the way alongside use-cases for using swap. In addition, we'll discuss our future plans and open questions that we still face. By the end of this talk, I hope you’ll have a deeper understanding of Kubernetes’ swap feature and feel equipped to contribute to its ongoing development! https://lnkd.in/dNmdUyNr

SW and even HW design and development in recent times can be summarized as tool-fetishism with almost an absence of decent mathematical abstraction behind it. On the other hand, in the realm of mathematics, abstract algebra, and powerful abstractions fetishism, we have the opposite, there is a lack of tooling to be applied to real problem concerns (languages like Haskell have popularized a small subset of that, still a small reflection of the powerful extent of abstract algebra). And because programming tooling lacks decent mathematical abstraction, all are ludicrously plagued with leaky weak abstractions that finally lead to the broken and overengineered systems that are the hallmark of the modern IT industry. Our Risk-Management Kernel design for our Cryptofisher System (nearly ready to be launched together with Roberto Crespo and Alfred Dietrich Steiof )has taken enormous lessons from this. With these lessons in our minds, we have faced the problem not by thinking to solve it with programming tooling, as is the norm in the IT industry, but by thinking with a domain-knowledge perspective and how it can be addressed mathematically both in terms of abstraction and numeric calculation by running controlling experiments over an exhaustive set of real market scenarios and Monte Carlo simulations. With this, we can come up with our first useful mathematical abstractions. Successive experiments led by these abstractions led us to discover a way to design our risk management kernel: with the powerful mathematical abstraction of a #Hamiltonian. A Hamiltonian is used in classical and quantum physics for computing the evolution of energy in a system, typically towards a minimum amount of energy. H = Potential + Kinetic energy. Our results led us to design a Hamiltonian that replaces the Energy with Signal and Noise Bayesian probabilities quantized with discrete variables. So then all that was left for us was to figure out how to program our Hamiltonian system and its evolution according to a simple addition of single Hamiltonians for every component: H = Hsignal + Hnoise + Hinteraction. What have we won with this? The very optimistic expected possibility to decompose the nightmare interaction between noise and signal (the Achilles' heel in most trading systems) into simple additive components Hinteraction = Hsignal + Hnoise In a composite manner, and the goal to reduce the noise-dependent components to a minimum.

Both the spot market and the futures market lack a sound and reliable mathematical model to evaluate pricing for given intervals in real-time markets. We do not have a mathematical and scientific method equivalent to the Black-Scholes equation , used to estimate the pricing of options (European options), with which we can assess the pricing in real-time markets of our assets and securities, especially in hyper-volatile novel markets like the cryptocurrency market. This absence of objective pricing in spot and futures markets is why, until now, we have not had a universal, scientifically-based risk management kernel to truly outperform the results of the spot and futures markets themselves by diminishing the risk associated with hyper-volatile real-time pricing fluctuations. Consequently, the best strategy, as we all know here, has been to hedge the volatile, unsecured assets or securities with derivatives and options, trading the risk asset by dealing with puts and calls of the option itself. This lack has hindered and impacted hyper-volatile markets with bubbles and manipulation of volumes in the renowned pump and dump strategy. Without a reliable, or at least approximate, method to estimate pricing—since there's no direct way to manage risk—the manipulation of these prices could be achieved through manipulation with relatively small volumes, creating bubble and burst cycles that hinder the crypto market itself in the middle and long term, leading to a continuous reset and new start with new investors: known among us as the crypto Groundhog Day. Over the last seven years, we have endeavored to develop our innovative risk-management kernel as part of our CryptoFisher platform (in the prelaunching phase with Roberto Crespo and Alfred Dietrich Steiof ) with which we could assess the pricing in the crypto real-time market through continuous pricing exploration with adaptive learning. This employs a new learning paradigm that, in contrast to deep learning and ML, doesn't require terabytes or even petabytes of data and long-term training. Instead, it continuously discovers new pricing conditions of the market through an uninterrupted novel learning process that significantly outperforms the market by enhancing the two types of bayesian success, achieved by first reducing the two types of associated bayesian errors. Our process programmatically uses a Hamiltonian pipeline in adaptive learning stages to increase the signal (successes) by first diminishing the noise (errors). The three sets of four Bayesian random outcomes are fed back (back- and forwards) into the system at every stage of the Hamiltonian pipeline with two discrete variables. The final outcome is very accurate instant pricing on the spot market (we have yet to test it in future markets) after removing most of the noise, quantified by two discrete variables, in every one of the three stages of the Hamiltonian pipeline.