Matt Pocock on Tracer Bullets in AI

AI programming is changing everything, but are you catching the mistakes? When AI assistants generate 10X more code than you normally would, the feedback loop breaks. This week's video is on how to make the most of of these ai tools ➡️ youtu.be/XavrebMKH2A

Artur & AI: One thing that stuck with me: good processes get better. A mess just becomes a bigger mess — faster. DORA’s latest AI report says exactly this: AI isn’t a magic fix. It’s a amplifiiler. It’s also easy to confuse generating code with delivering value. AI can produce a lot of incorrect or sub-optimal code if there isn’t a solid validation and feedback process in place. Worth remembering before rushing into implementation. P.S. I'm a huge fan of the t-shirt game. This one is super well-matched.

Some basic resources to get up to speed on AI-augmented coding. 1. Addy Osmani's LLM coding flow as of 2026 - https://lnkd.in/dPKg68ZP 2. Exploring solution spaces with AI - https://lnkd.in/dV5brynT 3. Tips on prompt engineering for developers - https://lnkd.in/dJ-4EBYZ 4. Best practices for using AI for coding - https://lnkd.in/ddNCEvYv Many of the links are for MS Co-pilot but the information is generally applicable. #AI #AIAgents #LLM #Programming #Vibecoding #Codex #ClaudeCode

Dr. Viral B. Shah is one of the creators of the Julia programming language and co-founder and CEO of JuliaHub. Read his article in Design News that explains how AI agents can handle physics-based modeling complexity while engineers focus on design judgment and tradeoffs. Find it here: https://lnkd.in/gUUXxrgD JuliaHub #AI #AgenticAI #MechanicalEngineering #HardwareEngineering #Dyad

“Software Made Simple.” And no, this wasn’t about AI - that’s what a magazine cover looked like in 1991 BusinessWeek (now Bloomberg Businessweek) dedicated an entire feature to object-oriented programming back then 😂 Apparently, “burying” IT professionals has been a tradition for decades

Healthcare innovation comes by understanding problems, and then using great technology to solve those problems! I was able to sit down with Ankur Gupta recently and talk about a new piece of the large language model (LLM) coding stack and how we are able to utilize a library called DSPy (declarative self improving python) for having better outputs from the LLM models. DSPy, like many of the advancement packages in AI, is an open source library that tries to abstract the LM prompt from the code infrastructure, making a system so that it doesn't matter which AI large language model is used is EXTREMELY important for robust systems, especially in healthcare, and DSPy makes that much easier than prompt engineering. https://lnkd.in/eS2BKDBj

Yann LeCun's Logical Intelligence launched Kona, an Energy-Based Model that proves what's safe before you do it by mathematically verifying all possible system states and telling you what actions are permissible, rather than making probabilistic predictions about likely outcomes. The distinction sounds academic until you realize how this approach fundamentally differs from current AI paradigms. Most models today operate on probability, making educated guesses about the most likely (top_k) outcomes. So, instead of that EBM approaches the problem by defining an energy function over all possible outputs and select the configuration with the lowest energy. The assumption is that this creates certainty to replace probability. This focus on mathematical certainty connects with my work on Agent Reasoning and Game Theory that emphasize model performance comparisons through game theory and agent reasoning. https://lnkd.in/eNT6CgBg While I pit reasoning agents powered by Nemotron and Qwen against each other in strategic games to evaluate their reasoning capabilities, Kona addresses a more fundamental challenge: building AI systems that can reason with logical constraints rather than probabilistic approximations. A live demonstration of Kona 1.0 on the company's website showcases the model's reasoning capabilities through head-to-head Sudoku challenges against leading large language models. In these challenges, Kona solves complex reasoning problems like Sudoku faster and with lower power consumption than its competitors, illustrating the practical advantages of energy-based reasoning. Source: https://lnkd.in/eQi9pR96 #ArtificialIntelligence #EnergyBasedModels #YannLeCun #LogicalIntelligence #Kona #GameTheory #AgentReasoning #AIResearch #MachineLearning #AIInnovation

AI is not ready! LLMs are language-based, non-deterministic statistical engines. Binary pure logic, the way machines talk, is much closer to old-school coding. LLMs are like a fuzzy layer wrapper to interact with humans. Don't code via the fuzzy wrapper. Talk directlty to the machine.

Anthropic (Claude Code) did a study on how LLM usage affects mastery when picking up a new programming library/language. TLDR; on average, decrease in mastery, no increase in speed. High-scoring AI interaction patterns: Generation-then-comprehension, Hybrid code-explanation, conceptual inquiry. Low-scoring AI interaction patterns: AI delegation, Progressive AI reliance, Iterative AI debugging. Extrapolate to other industries at your own risk. https://lnkd.in/g8S3DA3v