Build Fast with AI

I had a dream today, so I recreated it using AI : The backrooms as a waterslide🌊: Imagine going down a waterslide that takes you straight into the Backrooms. Created with Flux 3 by @bfl_ai.

Kimi 3.1 Pro (High Reasoning) has me debugging the AI more than my Unity game. I've been using Kimi 3.1 Pro with High Reasoning to improve an existing 2.5D Android mining game in Unity. Even with visual references and gameplay footage, I’m getting black screens, floating drills, and messy UI. I keep pointing out the same problems, but they’re still there after multiple attempts. At this point, I’m spending more time correcting it than building the game. Here are the references I gave it alongside what it actually produced. How is your experience with this model?

Tested OpenAI’s newly released GPT-6.1 Terra in Codex CLI Ultra mode against DeepSeek V4.1 Flash at max effort, using the exact same prompt. The task: Build a realistic, interactive European family MPV/crossover in Three.js and WebGL, featuring a detailed interior, opening doors and trunk, steering, working lights, and an orbit camera. No prebuilt complete car models. The setups: GPT-6.1 Terra: Ultra mode with automatic delegation, Codex CLI 0.164.0 DeepSeek V4.1 Flash: Max effort, AgenC Core 0.19.19 This is a comparison of the two coding setups, including their harness behavior, not a controlled single-agent comparison. Both completed the task autonomously.

My ChatGPT window now is an aquarium with fish swimming across my screen and a Pac-Man game playing itself while I wait for the agent to finish. Introducing Luma: a Windows app that adds 9 animated wallpapers to ChatGPT without blocking your chat. 🐠 Aquarium 👾 Pac-Man arcade 🪼 Jellyfish 🌀 Neon vortex and more! Link in replies👀

Kling 4.0 vs Seedance 2.5, same prompt at 720p. Kling delivers richer colors, deeper blacks, and stronger contrast. Seedance nails the transitions, motion accuracy, and prompt adherence. Better visuals look great, but in real-world AI video workflows, following the prompt matters just as much.

Most observability tools charge by what they index, so teams quietly sample logs, cap retention at a week, and drop fields like user_id. Then the one trace you need at 3am is already gone. @Coralogix takes a different route. Nothing gets indexed. All telemetry lands in your own cloud storage, in open formats, kept as long as you want. One query language reads logs, metrics, traces and security data together. And you can just ask it what broke: in chat, from Slack or your IDE via MCP, or from a CI script via the CLI. What's the field your team stopped logging because it got too expensive?

What the hell is going on with Claude? I'm paying $100/month for the Max plan and using Opus 5.5, yet my 5-hour usage limits are getting wiped in just 10 minutes of completely normal Claude Code sessions. Weekly limits are draining like crazy too. How are we supposed to get any actual work done when usage limits are this ridiculous, even on a $100 plan? Anyone else experiencing this sudden drop in usage?

Opus 5.5 is insanely good at creating motion graphics videos, but the sound effects are still its weak point. That’s why I use Wav for all the SFX and voiceovers in my Claude videos. It makes such a huge difference to the final result. Truly a game changer @musicgpt_com !