njp-coder @njpCoder

Do you think the future of LLMs remains autoregressive, or will diffusion-based models eventually become the new standard?

Most LLMs write like this: Word 1 → Word 2 → Word 3 → Word 4 DiffusionGemma doesn't. Google DeepMind's new open-weights Diffusion LLM starts with an entire block of text and repeatedly refines it in parallel. Think: Traditional LLMs = Write DiffusionGemma = Draft and Edit Result? ⚡ Up to 4× faster generation ⚡ 1,000+ tokens/sec on H100 GPUs For years we've assumed AI must generate text one token at a time. DiffusionGemma challenges that assumption. #AI #LLM #DiffusionGemma

@AlfieJCarter Agents

Tip: in Claude Code, /goal: keep CI green sets the outcome once and lets it hold.

Claude Fable 5 is a shift in how we use AI: from executing tasks to managing responsibilities. It runs autonomously - reviewing, planning, self-verifying. You give it a goal, it works to maintain it. Tasks: "fix this bug." Responsibilities: "keep the tests green." A task ends. A responsibility runs.

@compileandpush Honestly, not the AI part. The real work was packaging, reproducibility, observability, retries, state management, cost controls, and failure recovery.

Claude Opus 4.8 just shipped Dynamic Workflows - up to 1,000 parallel agents on a single job. When to use it: → Chat: quick one-off → Skill: repeatable task → Subagent: messy side task → Agent Team: small crew → Workflow: massive parallel job (migration, audit, bulk code) Default answer is never a Workflow. Uber burned their entire 2026 AI budget in 4 months on Claude Code. Budget before you build.

Interesting idea from a recent article on visual AI: We're moving from AI that generates pixels to AI that generates the underlying system behind those pixels. The difference is subtle but important. Outputs are consumed. Representations can be edited, composed, versioned, and improved. That feels like the path from demos to real products. What do you think?

Cloud FinOps optimized compute utilization. AI FinOps will optimize intelligence utilization. The unit of optimization is changing: Server → Workflow The key question is no longer: "How much does an AI call cost?" It's: "Why is the system solving the same problem twice?" #AI #AgenticAI #Architecture

Built MYTHOS in 24 hours on @Replit. A system that lets you practice speaking, get real feedback, and improve over time - not just your words, but your clarity, confidence, and presence. cc @MannyBernabe @Franciscocrz @amasad @HayaOdeh Shoutout to the team for an incredible 10 years and the buildathon.

I thought I was bad at speaking. I wasn’t. I just never practiced it. We’re taught what to think. Not how to express it. So I built MYTHOS. An AI coach that analyzes how you speak: clarity, pacing, structure, presence and tells you exactly what to improve next. Built in 24 hours for @Replit Birthday Buildathon This might be one of the most important skills we ignore. cc @MannyBernabe @Franciscocrz @amasad @HayaOdeh

Started building something cool at the @replit 24 hour buildathon. MYTHOS. More soon 👀

nagajyothi just joined GX. SDE III who has a master's in aeronautical engineering. so she went from designing aircraft structures to building full-stack dashboards. mentors people on reddit on the side. the range is actually unhinged.

@Layton_Gott Data to insights in under a min

You're not failing because your product is bad. You're failing because you built a key… for a lock that doesn’t exist. We optimize code. We don’t validate problems. That’s the real bug.

@Jbm_dev Your data already has answers. We just help you see them — instantly. No learning curve required.

Hi guys 👋 I write code, break things, fix them, and repeat. Software Engineer here — always exploring new tech. If you’re building cool stuff too, let’s connect 🤝

Quick question for founders: when did you last look at your data - not the headline number, but actually look? Not the revenue total. Not the monthly growth %. I mean: sat down, opened your full dataset, and looked for something you didn't already know. If the answer is "not recently" - you're not alone. Almost nobody does. Here's what makes it hard: spreadsheets don't surface insights. They store data. Finding meaning requires asking the right questions and most of us don't know what questions to ask until after we needed to. So the insights sit there. Patient. Invisible. Here are 5 things almost certainly hiding in your spreadsheet right now: 1. Which 2–3 customers or products are driving the majority of your results (and which are quietly costing you) 2. A timing pattern - a day, week, or month where performance consistently shifts 3. Your best-performing segment - the one you'd double down on if you knew it existed 4. An outlier making your averages misleading 5. A trend that reversed 4–6 weeks ago that you haven't noticed yet None of this requires a data analyst. It requires looking at your data the right way. Drop a comment: when did you last actually dig into your numbers? #Dashlytics #Founders #DataAnalysis #StartupLife #BusinessIntelligence

@hirvesh @Habit_Pixel Awesome!

Most people building in tech right now are quietly working on something interesting… What are you building? Looking to connect with people working on: 🚀 Tech 🛠️ SaaS 💻 Full Stack 🧠 AI Tools 📱 Product Dev 📈 Marketing 🤖 AI Agents Drop it below — happy to share feedback or jam ideas 👇

@rcmisk Re-evaluate the ICP and see if we can get atleast 1000 true fans for the product