Dennis Zhuang

Migrating from Loki to GreptimeDB achieved: ✅ 30% reduction in TCO, ✅ 10x faster queries with no timeouts, ✅ and seamless cross-cloud deployment via unified APIs and a single control plane. #GreptimeDB #Loki #LogManagement #LogStorage

According to the TMA1 stats, Claude Code performance improved this week significantly compared to last week, rising from 20–30 tokens/s to 40–50 tokens/s.

@FordRacing @nuerburgring Congrats—absolute beast!

6:15.977 History made. The Ford GT Mk IV just clocked the fastest lap ever by an American OEM at the Nürburgring. Watch the full run on YouTube now: youtu.be/tTC10Kqb0XU @nuerburgring

Always thrilled by humanity's exploration of space!

Top 5 GreptimeDB blogs of 2025 — #4 Beyond ELK: using GreptimeDB as a log database. → 10GB of logs = 10GB+ of Elasticsearch storage. GreptimeDB: ~1/10th. → Compute-storage separation on object storage (half the cost of block storage on AWS) → Rust, no JVM, no more OOM kills → Pipeline DSL that parses raw log lines into typed, indexed columns — bloom filter, inverted index, fulltext, your call Full Vector + GreptimeDB walkthrough included, runs locally with Docker. greptime.com/blogs/2025-04-…

Dug into Claude Code's leaked client source. Their agent observability is seriously good — every detail I looked at had been thought through, from retry chains to privacy controls to how they layer and gate telemetry. Get observability right and you can do real product analytics. Get product analytics right and you tend to make better products.

GIF sucks! Here is the video.

Enjoy live and replay codex/cc sessions in TMA1! Install or upgrade: tma1.ai #AICoding #Agent

If all goes as planned, we'll release GreptimeDB v1.0 next week—exactly four years after starting the project.

gRPC based on Arrow flight is blazing fast.

@LotusDecoder 感谢使用，欢迎反馈👏🏻

这个实时本地记录 agent 用量的 插件，蛮好用。 一句话安装 “ Read tma1.ai/SKILL.md and follow the instructions to install and configure TMA1 for your AI agent ”

So to me, the key question is not only “How do we reduce hallucinations?” It is also:𝗛𝗼𝘄 𝗱𝗼 𝘄𝗲 𝗶𝗺𝗽𝗿𝗼𝘃𝗲 𝗵𝘂𝗺𝗮𝗻-𝗺𝗮𝗰𝗵𝗶𝗻𝗲 𝗶𝗻𝘁𝗲𝗿𝗽𝗿𝗲𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗲𝗻𝗼𝘂𝗴𝗵 𝘁𝗼 𝗺𝗮𝘁𝗰𝗵 𝘁𝗵𝗲 𝘀𝗽𝗲𝗲𝗱 𝗼𝗳 𝗺𝗮𝗰𝗵𝗶𝗻𝗲 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻? That may be the real software engineering problem of the LLM era.

LLM hallucinations matter, but they're not the whole problem. Humans make mistakes too — that's why bugs and software engineering practices exist in the first place. The bigger issue is that LLMs generate code faster than humans can review and understand it. What used to be written over months now appears in a day. The bottleneck isn't just model reliability — it's human-machine interpretability at that speed. We need a new AI communication method. #AICoding

Human-machine communication still happens at human speed. We still rely on reading, inspecting, testing, discussing, and reasoning through what we see. Our eyes and ears have not become 100x faster just because code generation has. So the real bottleneck is no longer just code production. It is interpretation. In that sense, the deeper challenge is not hallucination alone. It is that LLMs are so efficient that they weaken interpretability. They compress the time available for understanding, review, and assurance. If humans could instantly understand the code LLMs generate, hallucinations would be much less threatening. We already know how to deal with imperfect code — that is what engineering practice is for. The difficulty is that we now receive too much code, too quickly, for those practices to scale in the same way.

In the past, a meaningful amount of code might be produced over the course of three months, six months, or even a year. That gave teams time to review it carefully, write tests, discuss tradeoffs, build shared context, and involve multiple people in the validation process. Reliability did not come from perfect humans. It came from the fact that human systems had enough time to absorb and verify the work. Now, with LLMs, the same volume of code can appear in a single day. 🚅 That changes everything. 😅

Stateful LLM infra needs durable, queryable storage. Glad to see Greptime in the stack.

@MathewShen42 感谢推荐，欢迎反馈啊😀

推荐下 @killme20082 大佬的tma1，使用体验还是比较丝滑的

Repartition — one of the most important features in GreptimeDB v1.0-rc — lets you split and merge partitions online without physical data migration. Hot partition? Split it. Cold fragments piling up? Merge them back. We wrote a practical guide with step-by-step SQL examples covering both Mito Engine and Metric Engine tables, plus checklists for running this safely in production. Read it: greptime.com/blogs/2026-03-…

Of course, it's opensource: github.com/tma1-ai/tma1

I built TMA1.AI because I didn't want to sign up for yet another cloud account just to understand what my coding agents were doing locally — and I wanted more than cost tracking. More here: blog.fnil.net/local-observab…

Supported agents - Claude Code (metrics + logs) - Codex (logs + traces + metrics) - OpenClaw (traces + metrics) - Any OTel app using GenAI semantic conventions The dashboard auto-detects the data source and switches to the right view.