Seth Mottaghi

💡 Struggling to keep up with AI and Machine Learning? Here are 12 blogs that share real, practical insights — not just research buzzwords. 1)Andrej Karpathy Explains neural networks and LLMs from scratch with simple, practical insights that make complex AI topics easy to follow. 🔗 https://karpathy.ai/ 2)Sebastian Raschka, PhD Clear and practical posts on training, fine-tuning, and evaluating LLMs with hands-on Python code examples. 🔗 https://lnkd.in/dScSjap8 3) Interconnects by Nathan Lambert Covers AI alignment, open-source model trends, and ecosystem updates from a researcher deeply involved in AI policy. 🔗 https://lnkd.in/dtpfKgQJ 4)Lil’Log by Lilian Weng Breaks down AI safety, alignment, and reinforcement learning based on real work experience at OpenAI. 🔗 https://lnkd.in/diTr58jG 5) Chip Huyen Writes about MLOps, model deployment, and designing ML systems that actually scale in production. 🔗 https://huyenchip.com/ 6) Eugene Yan Shares applied ML and data science lessons from building recommender systems and AI products used by millions. 🔗 https://lnkd.in/deYQp7V8 7) Philipp Schmid Hands-on tutorials for training and deploying large language models on AWS with Hugging Face tools. 🔗 https://www.philschmid.de/ 8) Jason Liu Writes about LLMs, data workflows, and building real-world AI solutions using open-source technologies. 🔗 https://jxnl.co/writing/ 9)Hamel H. Covers MLOps, fine-tuning, and product strategy with practical advice for ML engineers and startup teams. 🔗 https://hamel.dev/ 10) Berkeley AI Research (BAIR) Research blog featuring the latest work in computer vision, NLP, and robotics from Berkeley’s AI lab. 🔗 https://lnkd.in/db9G6MNT 11) Hugging Face Blog Updates, tutorials, and community highlights from one of the biggest open-source AI platforms. 🔗 https://lnkd.in/dV7QTyau 12) Google DeepMind In-depth posts about AI breakthroughs, safety, and research directly from Google’s top AI team. 🔗 https://lnkd.in/dQFczkhg If you found this list useful — 👉 Follow Shubham Vora for more AI resources and tool breakdowns 💾 Save this post to explore later 🔁 Repost to help others stay updated on AI

121

42 Comments

From Claude Code to Cowork: The Changing Economics of Labor driven by LLM Agents Blog: https://lnkd.in/dUHcitB2 In 2025 coding agents crossed a critical capability threshold, leading to significant productivity gains and widespread adoption. 2026 is shaping up to be the year agentic AI transforms knowledge work. While Dario Amodei warns about massive white-collar layoffs, Garry Tan and Aaron Levie point to the Jevons Paradox, arguing that falling costs will unleash enough new demand to offset any job losses. Who will prevail in this debate will depend on the productivity multiplier and the elasticity of demand.

9

Researchers at Haize Labs are tackling the "Last Mile Problem" in AI by using fuzzing to evaluate GenAI systems. They found that traditional evaluation methods are inadequate for the brittleness of GenAI applications. The most important finding is their approach to scaling evaluation: "Scaling Evaluation with Agents as Judges" "Scaling Evaluation with RL-Tuned Judges" They also highlight the distinction between fuzzing and adversarial testing in AI. Case studies include: "Haizing a Major European Bank's AI App" "Haizing a F500 Bank's Voice Agents" "Scaling Voice Agent Evals with Verdict" These data points demonstrate their practical application of fuzzing to real-world GenAI systems. by AI Engineer https://lnkd.in/ev53hPEv

1

Here's an example of something that I think is at the core of the difficulty of creating reliable complex agentic workflows: I was having a chat with Perplexity (a tool that generate answers using top models like GPT-5.2, Claude 4.6, and Gemini 3.1 Pro) and it made a very basic mistake in its response. It told me that if a>b, I should choose option 1, while in reality the exact opposite was true and it acknowledged its mistake when I pointed it out. My point is that these are the kind of basic mistakes that can happen at the core of any agentic process and of course guardrails need to be in place to catch and correct them, otherwise, they will cascade and derail the whole process. Of course, humans make similar mistakes too and processes are needed to be in place to catch them and prevent them from cascading down the workflow. However, creating fully autonomous agentic systems while the core model is capable of making such basic mistakes is a real challenge and every decision and architecture needs to be made with these types of challenges in mind. I know this is not news to a lot of people who read this post, but I thought I'd just share my personal experience :)

5

Rethinking how information flows in Large Language Models We’ve scaled LLMs by adding more parameters, data, and compute—but connectivity itself is becoming the next bottleneck. In my latest blog, I dive into Manifold-Constrained Hyper-Connections (mHC) by DeepSeek, a new architectural idea that: 1. Moves beyond single-stream residual connections 2. Enables stable multi-stream information flow 3. Uses mathematical constraints (via Sinkhorn normalization) to make expressive models trainable at scale. The key takeaway: constraints don’t limit intelligence—they enable it. mHC shows how principled architecture design can unlock deeper, more expressive, and more stable AI systems. If you’re interested in LLM architecture, scaling challenges, or the future of model design, you might enjoy this read 👇 🔗 https://lnkd.in/gS9fiMZ3 #AI #LLMs #DeepLearning #MachineLearning #AIArchitecture #GenerativeAI

7