Innovative Training Techniques

Thick thighs save lives? A landmark British Medical Journal study followed nearly 3,000 adults for over 12 years. People with smaller thighs had significantly higher rates of heart disease and premature death. That finding held up after controlling for belly fat, blood pressure, smoking, and every other risk factor measured. A follow up analysis of 72 studies covering more than 2.5 million people arrived at the same conclusion. But here is what most people get completely wrong about this research. It is not about fat distribution. It is about muscle. Your thighs are the largest muscle group in your body. When that muscle mass is high, your body clears blood sugar efficiently after meals, insulin resistance stays low, and your heart is protected from the inflammation that quietly builds over decades. A King's College London study of over 300 twin pairs found that the twin with greater leg strength had measurably better memory and faster cognitive function a decade later. Same genetics. Same upbringing. Different leg strength. Different brain outcomes. There is also a survival element most people never consider. When your body faces serious illness or major surgery it draws on muscle tissue for energy and recovery. People with more muscle going into those moments survive them at dramatically higher rates. Walking uphill. Climbing stairs. Squats. Resistance training. These are not vanity exercises. They are some of the most evidence-backed longevity tools available to you right now. If you are caring for an aging parent, watch their legs. Loss of lower body muscle is one of the earliest and most reversible signs of accelerating decline. ❤️

As LLMs become the core engine behind more and more AI products, customizing them with precision becomes critical. But the question I get most often is: “Should we use Supervised Fine-Tuning (SFT) or Reinforcement Fine-Tuning (RFT)?” Let’s break it down. 👀 𝗦𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗲𝗱 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴 (𝗦𝗙𝗧) 𝗨𝘀𝗲 𝘄𝗵𝗲𝗻: → You have a clean, labeled dataset (preferably >100k examples). → The task is verifiable and deterministic—think classification, factual QA, structured output. 𝗪𝗵𝘆 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀: → Efficient and reproducible. → Offline training with minimal infra orchestration. → Works well with modular fine-tuning (e.g., LoRA adapters). 𝗟𝗶𝗺𝗶𝘁𝗮𝘁𝗶𝗼𝗻𝘀: → Doesn't adapt well to subjective or multi-objective tasks. → Plateaus with small or noisy datasets. ♾️ 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴 (𝗥𝗙𝗧 / 𝗥𝗟𝗛𝗙) 𝗨𝘀𝗲 𝘄𝗵𝗲𝗻: → You’re optimizing for subjective quality, human preference, or multi-turn reasoning. → Labeled data is limited, but you can evaluate outputs programmatically (via reward models or heuristics). 𝗪𝗵𝘆 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀: → Incorporates task-specific feedback loops (e.g., correctness, engagement, success). → Allows for dynamic alignment with non-differentiable objectives. → Crucial for tasks like dialogue, summarization, tool use, and creativity. 𝗟𝗶𝗺𝗶𝘁𝗮𝘁𝗶𝗼𝗻𝘀: → More complex training pipelines. → Reward model design is critical and often brittle. Fireworks AI just dropped major updates to support both paradigms: 💡 𝗦𝗙𝗧 𝘃𝟮: https://lnkd.in/dZM8d54N → Optimized for speed, multi-token training, and massive context lengths → Supports modular LoRA, function-calling fine-tuning, and quantization-aware training 🚀 𝗥𝗙𝗧 (𝗕𝗲𝘁𝗮): https://lnkd.in/dWqB8WEh → Simplifies RLHF for open models (Llama, Qwen, Phi, DeepSeek) → Write your reward function, and Fireworks handles the rest Already showing performance on par with GPT-4o, at a fraction of the latency! If you’re building with open models and need production-grade tuning, these tools lower the barrier significantly. Start customizing your models on Fireworks AI: fireworks.ai/models

Step ups are an effective exercise for developing single leg strength and stability. They require you to apply force through one foot into the box, stand up under control, and maintain balance on that same leg. Depending on the variation, they also challenge different movement directions, joint angles, and coordination demands, which is why they transfer well to both athletic movement and everyday tasks. In this video, I’m demonstrating several step up variations I commonly rotate through training programs. On each clip, the knee drive and balance variation shown in the upper left corner adds an additional challenge by increasing the stability demand once the base movement is established. Step Up Variations Shown:
Static Step Up
Dynamic Step Up
Lateral Step Up
X-Over Step Up
Rotational Step Up
High Box Step Up Master the simpler variations first. Progress only when posture, balance, and position are consistent. Train smart. Let’s Work! 🔨 #fitness #performancetraining #strengthandconditioning

4 opensource LLM fine-tuning libraries you need to know about as an AI engineer. Fine-tuning used to require enterprise budgets and PhD-level expertise. Not anymore. Here are 4 libraries that makes LLM fine-tuning accessible for all: 1. Unsloth AI • 2x faster training with 80% VRAM reduction • Custom Triton kernels with manual backprop engine • Supports 4-bit to 16-bit quantization • Works on NVIDIA GPUs from V100 onwards (CUDA 7.0+) 2. HuggingFace TRL (Transformer Reinforcement Learning) • Specialized trainers: SFTTrainer, DPOTrainer, RewardTrainer • Full integration with Transformers and PEFT ecosystems • Scalable via Accelerate (single GPU to multi-node) • CLI interface for quick experimentation 3. Axolotl • Single YAML config for entire pipeline (preprocess → train → inference) • Advanced optimizations: Flash Attention, multipacking, sequence parallelism • Multi-GPU/multi-node support (FSDP, DeepSpeed, Ray) • Flexible data loading (local, HuggingFace, cloud storage) 4. LlamaFactory • Supports 100+ models including LLaMA, Mistral, Qwen, DeepSeek • Multiple training methods: full fine-tuning, LoRA, QLoRA (2/4/8-bit), DPO, PPO • Web UI and CLI interface - no coding required • Built-in experiment tracking (TensorBoard, Wandb, MLflow) These tools have removed the barriers. The question isn't whether you can fine-tune - it's which approach fits your use case. The best part? They're all 100% opensource. Which library are you planning to try first? Link to the GitHub Repos in the comments.

Single-leg Romanian deadlift & high pull: 1️⃣ Strengthens the entire posterior chain. This exercises requires significantly higher gluteal muscle forces than most other hamstring exercises, with the peak gluteal muscles forces being higher than reported during high-speed running (Prince 2014, Van Hooren 2022). 2️⃣ Improves eccentric hamstring strength at longer muscle lengths. The eccentric activity of the biceps femoris long head and semimembranosus is higher in the single-leg RDL compared to other hamstring exercises (Van Hooren 2022, McAllister 2014). 3️⃣ Fascicle lengthening during the single-leg RDL & high pull occurs with a high constant muscle activation. This makes it a very effective exercise to promote strength increases at longer fascicle lengths (Van Hooren 2022) 4️⃣ This exercise requires an explosive posterior pelvic tilt. The explosive pelvic tilt contributes to large joint forces at the hip during sprinting and facilitates an efficient transfer of power (Sado 2017, 2019). An explosive posterior tilt also enables soccer players to kick more powerful with less load on the adductors. 5️⃣ The single-leg stance requires greater stability and balance and makes the exercise more sport-specific. 6️⃣ As a progression the exercise can be performed with an elastic band or cable attached to the ankle of the swing leg. Resisting hip flexion facilitates the extension power of the contralateral leg through the ‘crossed extensor reflex’. By provoking this spinal reflex you can augment the extension force of the stance leg. #hamstrings #hamstring #glutes #singleleg #fasciclelength

🌱 “𝐈 𝐝𝐨𝐧’𝐭 𝐟𝐨𝐫𝐜𝐞 𝐭𝐡𝐞𝐦 𝐭𝐨 𝐠𝐫𝐨𝐰. 𝐈 𝐫𝐞𝐦𝐨𝐯𝐞 𝐰𝐡𝐚𝐭 𝐬𝐭𝐨𝐩𝐬 𝐭𝐡𝐞𝐦.” This line hit me hard—because that’s what great teaching truly is. I once had a student who struggled not with ability, but with fear—fear of making mistakes, of raising their hand, of being wrong. Traditional instruction kept nudging them to “speak up more.” But what actually worked? Giving them a safe space to think quietly, letting them submit reflections anonymously, then slowly offering low-stakes speaking opportunities. They bloomed—on their own terms. 🔍 This is what barrier-free learning looks like. Not pushing students harder, but asking: What’s in their way—and how do I remove it? Some powerful methodologies that support this mindset: ✅ Inquiry-Based Learning – Let curiosity drive the lesson. ✅ Scaffolded Instruction – Support step-by-step until confidence builds. ✅ Metacognitive Reflection – Teach students to know how they learn. ✅ Growth-Oriented Assessment – Focus on progress, not just performance. 🌿 Students don’t need force. They need conditions to thrive. #LearnerCentered #Pedagogy #InquiryBasedLearning #GrowthMindset #TeachingStrategies #HolisticEducation #Scaffolding #ReflectivePractice #BarrierFreeLearning

How can smaller LLMs achieve strong reasoning? By combining data curation with supervised fine-tuning (SFT) and targeted reinforcement learning (RL). Microsoft released their first open reasoning/thinking models with Phi-4-reasoning distilled from OpenAI o3-mini. Implementation 0️⃣ Base model (Phi-4) and a diverse set of 'seed' prompts covering reasoning domains (STEM, coding, safety). 1️⃣ Filtered seeds to select challenging, "teachable" problems using LLM evaluators and difficulty heuristics. 2️⃣ Generated high-quality Chain-of-Thought reasoning traces with OpenAi o3-mini. 3️⃣ Perform Supervised Fine-Tuning (SFT) on the synthetic data. (Phi-4-reasoning). 4️⃣ Use Reinforcement Learning (RL) with GRPO with correctness and trace characteristics (e.g., length-aware accuracy). (Phi-4-reasoning-plus). Insights 💡 Used OpenAI o3 as “teacher” for high-quality synthetic reasoning traces. 📈 RL boosted performance, even on a small dataset (~6k math problems improved AIME >10%). 🧠 RL leads to higher accuracy, but generates ~50% more tokens. 🛠️ Uses dedicated reasoning tokens (<think>) and extended context length (32K via RoPE). 🔢 Best-of-N evaluation shows untapped potential and suggests room for better verification/selection methods. Paper: https://lnkd.in/egc4qYRH

Learning flourishes when students are exposed to a rich tapestry of strategies that activate different parts of the brain and heart. Beyond memorization and review, innovative approaches like peer teaching, role-playing, project-based learning, and multisensory exploration allow learners to engage deeply and authentically. For example, when students teach a concept to classmates, they strengthen their communication, metacognition, and confidence. Role-playing historical events or scientific processes builds empathy, critical thinking, and problem-solving. Project-based learning such as designing a community garden or creating a presentation fosters collaboration, creativity, and real-world application. Multisensory strategies like using manipulatives, visuals, movement, and sound especially benefit neurodiverse learners, enhancing retention, focus, and emotional connection to content. These methods don’t just improve academic outcomes they cultivate lifelong skills like adaptability, initiative, and resilience. When teachers intentionally layer strategies that match students’ strengths and needs, they create classrooms that are inclusive, dynamic, and deeply empowering. #LearningInEveryWay

DeepSeek’s R1 learns to reason via pure RL with no / minimal SFT. Here’s how to understand the role of SFT and why it is used (or not) for reasoning models… TL;DR: R1 shows that we can learn to reason via pure RL, but this requires a lot of exploration/compute during the RL phase. SFT is still useful for reasoning models, but collecting data is harder due to the long CoT these model use. R1 generates a small amount of “cold start” SFT data synthetically, which removes model deficiencies and accelerates RL convergence. Model alignment. LLM post training usually has two parts: 1. Supervised finetuning (SFT): train the model on examples of “good” completions. 2. Reinforcement learning from human feedback (RLHF): train the LLM based on human preference pairs. SFT teaches the model correct style / formatting, while RLHF teaches the model to produce outputs that humans deem preferable. Why is SFT important? For a standard LLM, SFT provides a high-quality prior (or starting point) for RLHF. If we applied RLHF without SFT, the learning process would be less efficient. The LLM would have to do more “exploring” to find a good policy. By applying SFT, we get the model close to the solution we want, then RLHF applies the finishing touches. Where does SFT data come from? Usually, data for SFT is either synthetically generated or manually created by humans. Collecting data for SFT is expensive because we have to manually write a good response from scratch for the LLM! Reasoning models like R1 have a particular style of output. They begin by outputting a massive chain of thought encapsulated by a special "<think>" token. Then, the model generates an answer separately. SFT + Reasoning. How would we collect SFT data for a model like this? We can’t just ask humans to manually write out long CoT. This would be too time consuming and expensive! Our only option is to generate this data synthetically, but: - Generating (and verifying) this style of output is still difficult. - We don't want to hinder the model's ability to explore during the RL phase. R1-Zero - No SFT. Given the additional complexity of collecting SFT data for reasoning models, authors at DeepSeek first try to avoid SFT altogether! Interestingly, reasoning abilities are found to naturally emerge from pure RL. But, the resulting R1-Zero model has downsides; e.g., poor readability and language mixing. “Unlike DeepSeek-R1-Zero, to prevent the early unstable cold start phase of RL training from the base model, for DeepSeek-R1 we construct and collect a small amount of long CoT data to fine-tune the model as the initial RL actor.” - R1 paper R1 - Cold Start w/ SFT. The final R1 model creates a small dataset (few thousand examples) for SFT prior to RL. When we train over some SFT prior to RL (i.e., a “cold start”), we provide a better prior to RL, which eliminates instability during the initial phases of RL training, accelerates convergence, and avoids the previously-observed side effects.

Why they say your Calf Muscle are your Second Heart? If you care to know, I can quickly explain why. See, when you walk, climb stairs, or even stand and shift your weight, your calf muscles contract. Every contraction squeezes the deep veins in your lower leg. Inside those veins are one-way valves. When the calf contracts, blood is pushed upward toward the heart. When the muscle relaxes, one-way valves prevent it from falling back down. Miraculous, right? This is called the Calf Muscle Pump. Now, when it works well, you get good blood circulation, metabolism, and lymphatic flow. But when it doesn’t, some amount of blood pools in your legs. And you start seeing: heavy legs, ankle swelling, Varicose veins, reduced endurance, and cold feet. That’s why now researchers are calling your calf a Peripheral Heart, aka a Second Heart. Problem is still the same; modern life has quietly brought this system down. Too much sitting. Too little movement. Weak calf muscles. And that's why this post is important for you. Because now research is linking poor calf muscle function with: → Poor cardiovascular health → Poor metabolic regulation → Poor Lymphatic Drainage Health → Reduce longevity Even more interesting: Calf circumference is now being studied as a longevity biomarker. Why? Because, across various clinical and physiological studies, from metabolic experiments on the soleus muscle to vascular research on venous return, scientists are consistently finding that improving calf muscle function can significantly impact long-term health outcomes. So what should you actually do? Keep it simple. Focus on these 3: 1. Incline Walking on a treadmill or uphill (I repeat walking, not running) 20–30 mins at 12–15% incline, Slow pace (~4 km/h) → Activates both major calf muscles continuously, also supports the glute, thigh, and core muscles. 2. Standing Calf Raises on a platform Take a small platform (elevated edge) and stand on its edge, and take a support of a wall. Full stretch at the bottom, pause for 2-3 seconds. Full Heel raise, pause for 2–3 seconds. 3 sets × 10 reps → Builds strength in the gastrocnemius 3. Seated Calf Raises (Soleus Push-ups) Do them even while sitting on an office chair → Improves post-meal glucose metabolism Great for people who sit for long hours. Hope this helps, and you will take care of your calf muscles because now it's important. --- All research papers in the comment section. Repost if you believe small habits can impact long-term health. --- #HealthResearch #EvidenceBased #BusyProfessionals #MetabolicHealth #Longevity #FunctionalHealth