Daniel Taylor

Hi everyone! My last post was in Florida; this time I'm across the continent in Huntington Beach, California, for #OptivCon SoCal and the latest tip from my "guide to the AD forest recovery guide". Do you have virtualized DCs? Here's an important point to consider for forest recovery. #ActiveDirectory

66

2 Comments

If you can’t explain your pricing in a sentence, you’re in trouble. We were. At Thunder Compute, our first pricing model—an hourly rate plus usage—seemed smart on paper. But once customers asked, “So… how much will I actually pay?” we realized our answer, “it depends,” only confused them more. And confused customers don’t buy. While we were focused on simplifying cloud GPUs, our pricing unintentionally made things more complex for our customers. So we listened carefully to our customers and switched to a flat hourly rate; simple, transparent, and 70% lower than our competitors. The response was immediate: customers stopped thinking about pricing and started using our product. Two lessons emerged: 1. Simplicity sells. When pricing feels like a math problem, customers walk away. 2. Feedback is gold. The faster you learn your first idea is wrong, the faster you get it right. Whether you're selling enterprise software or coffee beans, keep it simple. If you can't explain your pricing in one breath, neither can your customers.

127

5 Comments

Seeing ASP.NET WebForms revisionist historians, kinda like folks saying "actually the Star Wars prequels were good!" No, they're not. And no, Blazor isn't WebForms, nor does it vindicate WebForms' design. WebForms were an abstraction to pretend the web doesn't exist in favor a completely different, fundamentally incompatible model. Yeesh.

129

44 Comments

When I wrote my "AI Agents in 2025" post at the beginning of the year, I wrote that the biggest investment would be in reliability, putting the guard rails in place to prevent the agent from doing bad things. The recent Jason Lemkin incident with Replit, as well as other incidents such as Anuraag Gupta's run-in with Gemini CLI deleting his files, demonstrate that this is still very much the case. In the Gemini CLI case, the agent was executing commands incorrectly and not detecting command errors, leading to the loss of files. In the end, the agent gave up trying to remediate because it got so confused, and it was too late in any case, there was no remediation possible.

26

1 Comment

Stop wrestling outputs. Start shaping them. Today Anthropic dropped a small switch with big leverage: Claude Code + Claude app now support /output-style, so you can lock the way Claude talks before it writes a single line. Consistent tone on demand. Reviewer voice for PRs, mentor voice for juniors, exec voice for summaries. Less cleanup, more shipping. Stop rewriting verbose answers into the format you actually need. Team-readable by default. Standards aren’t a hope, they’re a preset. Quick start (paste this at the top of your chat): /output-style: "Staff engineer — terse, diff-first, no fluff. Use bullet points, show code before commentary, 120 chars width." Now try: “Refactor this file for clarity and testability.” “Explain the change like I’m onboarding a new hire.” “Summarize risks for a product manager.” My go-to presets: ✅ Reviewer: “Blunt, precise, starts with a diff, cites lint/rules.” ✅ Teacher: “Step-by-step, small chunks, labels pitfalls & why.” ✅ Security: “Threat model first, list vulns -> CWE refs -> fixes.” ✅ Prod Notes: “3 bullets: Context → Decision → Next Step. 60 words max.” ✅ No-Fluff: “Answer in the fewest tokens that still solve the task.” You don’t have to beg models for style anymore, you set it, then build. Comment STYLE and I’ll share my ready-to-paste presets for Claude Code.

45

41 Comments

Maybe an unpopular take, but I feel compiling streaming pipelines into Rust binaries upon job creation doesn't make for a great user experience, it's just taking too long. Kinda ironic when time-to-first-result is longer for a streaming query than for your typical OLTP pull query. Compilation can still happen JIT-style under the hood to ensure peak performance, but give me first results fast, in particular when iterating on a query during development.

24

11 Comments

We’ve heard growing interest from clients in local RAG solutions for mobile and embedded devices, driven by privacy requirements and the need for offline access. On these highly resource-constrained platforms, we’re limited to SLMs with fewer than 5 billion parameters—sometimes even under 1 billion. New SLMs are being released almost weekly, consistently setting records on benchmarks. (Though we’ve noticed the pace of improvement has started to slow—a topic worthy of its own discussion.) But despite the excitement and promises, the reality is: today’s generic SLMs are still far from being able to replace SOTA cloud LLMs in real-world tasks. I’m not even talking about advanced reasoning or multimodal capabilities. Even in relatively straightforward tasks like summarization, their performance lags—they hallucinate more, follow instructions less reliably, and struggle to integrate information comprehensively. We believe the first step to building production-grade local RAG systems is to understand these limitations in depth. To that end, we developed a RAG evaluation pipeline informed by our experience developing on-device RAG. It specifically targets common SLM pitfalls when used as summarizers. We benchmarked a range of popular sub-5B SLMs. Some interesting findings: 1. Most SLMs struggle to recognize when the retrieved context is insufficient to answer a question and to respond with a meaningful clarification question. 2. They also generally struggle to provide comprehensive answers to multi-faceted questions. 3. BitNet-b1.58-2b-4t from Microsoft Research delivered impressive performance, especially considering its minimal memory footprint. 4. The best overall performer in our tests was Cogito-v1-preview-llama-3b from Deep Cogito. Based on what we’ve learned, we’re building specialized models and local RAG techniques to address these limitations. Our goal is a robust local RAG framework that incorporates advanced retrieval and summarization techniques, while maintaining efficiency on resource-constrained devices. Stay tuned for more updates! https://lnkd.in/gkJpe3Ut

66

4 Comments