EK (@ekcheungAI) header
EK (@ekcheungAI) profile picture

Building Hong Kong × China AI Operator Community AI Agents | Workflows | Growth Mindset

Hong Kong
Joined November 2018
Photo by EK (@ekcheungAI)Photo by EK (@ekcheungAI)Photo by EK (@ekcheungAI)Photo by EK (@ekcheungAI)Photo by EK (@ekcheungAI)Photo by EK (@ekcheungAI)Photo by EK (@ekcheungAI)Photo by EK (@ekcheungAI)Photo by EK (@ekcheungAI)Photo by EK (@ekcheungAI)
Claude Code /artifact-diagramming: 1 skill that teaches Claude to draw better diagrams → Step 1: type /artifact-diagramming before you ask for a diagram in an artifact → Step 2: Claude loads guidance on when a diagram helps and what to draw → Step 3: it also covers writing inline SVG that stays legible in light and dark themes I think the light-and-dark rule is the detail that saves a re-roll. Needs v2.1.221 or later. What would you ask it to diagram first? via @ClaudeDevs code.claude.com/docs/en/comm…
4
2
88
Anthropic's Claude Code team ranks one tip above all the others: give the model a check it can run. Without one, "looks done" is all it has. The setup, from prompt to Stop hook, is in the article.
11
6
376
Fable 5.2 turned a single prompt into an Honor of Kings-style web game in one night — three years vs one night — and refused to copy the original art over copyright, according to a post showing the build. Great demo. The receipt is whether it plays past level one.
1
84
Minimax H3's image model generates tens of millions of pixels in one shot and crushes QWEN, according to a post; the post also says some models are uncensored. Resolution is a spec, not a receipt — show the actual full-res file, not a downscaled screenshot.
1
256
Fable 5.2 high beats Carbon high by 2 points, and is 10.3 min faster → Fable 5.2 high: 62 at $9.08, 16.8 min a task → Gemini 4 Carbon high: 60 at $7.76, 27.1 min a task → Claude Opus 5.5 max: 59 at $10.04, 25.2 min a task I think the 10.3 minute gap is why I'd pay the extra $1.32 for Fable high. Is 2 Index points and the time save worth that bill? via @ArtificialAnlys artificialanalysis.ai/leader…
1
82
Hydrogen is GA on the free tier, at $1.50 / $7.50 per 1M tokens → New model: gemini-4-hydrogen, with a unified 1M-token context window → Paid standard price is $1.50 per 1M input tokens and $7.50 per 1M output tokens → It adds a 4th thinking level, xhigh, next to low, medium, and default high I think this is the one to try for everyday coding before you spend Carbon quota. Would you start on the free tier, or go straight to paid? via @GoogleAIStudio blog.google/innovation-and-a…
1
107
16 experienced developers took 19% longer with AI tools. Afterward they guessed AI made them 20% faster. Feeling faster isn't being faster. How to find your own edge, in the article.
4
2
180
If my CEO delivered like this he can have as many affairs as he wants
135
1,233
60
35,941
407,735
EK retweeted
Replying to @Anikaflicks
The left one. 40 years of partnership that helped build Nvidia from a Denny’s booth into a $4T company is not something you trade for a photo that doesn’t even match the real story.
158
89
40
5,067
253,296
Claude Fable 5.2 max takes 48.0 minutes a task, 2.9x high, for 3 more points → Low 54 in 6.2 min; medium 56 in 7.8 min → High 62 in 16.7 min; xhigh 62 in 26.3 min → Max 65 in 48.0 min at $26.54 a task I think max is an overnight setting: high gives you 62 while you wait. How long will you wait for 1 answer? via @ArtificialAnlys artificialanalysis.ai/leader…
3
92
Gemini 4 Carbon matches Fable 5.2 on Perplexity's WANDR at 40% of the cost → Carbon scores 0.857 on WANDR at $5.86 per task → Perplexity says that matches Fable 5.2 at 40% of the cost → It is now in Perplexity Computer for Pro and Max, and in Search for Max I think price is Carbon's real pitch: the same WANDR score for 40% of the bill. Would you swap Fable for Carbon on research jobs? via @perplexity_ai
Gemini 4 Carbon is now available in Perplexity Computer for Pro and Max subscribers, and in Search for Max subscribers. It scores 0.857 on our WANDR benchmark at $5.86 per task, matching Claude Fable 5.2 at 40% of the cost.
2
1
127
15 seconds of motion graphics, no video model — just code from Fable 5.2, and a demo that reportedly burned 4% of a weekly limit. At that rate, how many passes does a scene get before the week's limit is gone?
2
1
121
Vercel Labs' agent-desktop reads the accessibility tree, not screenshots: each button gets an index, the agent clicks by number. No cursor grab, no window stealing. The real test is a long session, not one 0.17s click. Source: @Pluvio9yte on X
1
2
128
Desk, MacBook, light, even Clawd — the Blender room in that demo was reportedly built by Fable 5.2 from a single prompt, run live on stream, and the laptop screen is a web page Claude built.
1
91
OpenAI dots now connect more than 1 computer; you pick 1 per task → Step 1: in the ChatGPT app on each computer, open your dot's profile and select Allow → Step 2: your dot can build and test your app on a dedicated machine while you keep coding → Each thread uses 1 computer at a time; keep that machine online with the ChatGPT app open I think the second Mac for Xcode tests is the use case: your own screen stays free. Do you have a spare machine you would hand to an agent? via @OpenAIDevs
You can now connect multiple local computers to the same dot and choose which one it uses for a task. It can build and test your app on a dedicated machine while you keep coding on your laptop.
1
103
Codex /voice: 1 key starts a voice conversation in the terminal → Step 1: press F8 or type /voice, when your CLI build and session support voice → Step 2: /voice mute toggles your mic, /voice stop ends the conversation → /voice settings picks the voice for future conversations I think voice is best for review, not typing code: talk through the diff while you read it. Would you talk to your coding agent? via @OpenAIDevs developers.openai.com/codex/…
1
1
101
Someone built a whole game from scratch — models, sounds and textures — and says it all came from one-shot Fable 5.2 prompts with no human feedback. 8 hours, ~$550 at API pricing: ~2.7M output and 445M input tokens, heavily cached. They've posted it to play.
3
119
Co-Design Bench: 273 tasks from 26,314 real coding sessions, 2,331 hidden criteria. Best of 10 models passes 69.9% of criteria per task but fully finishes only 26.4%. Partial credit is the demo. Complete delivery is the job. Show me full-pass rate, not average score.
2
69
Full article:
12