haiyan zhang

Head of Gen AI, Gaming @ Netflix, Prev: Gaming AI & CoS @Xbox, #AI @MSFTResearch, Innovation @IDEO, Co-Founder @OpenIDEO, Inventor & TV @BBC, @BAFTA, D&AD.

[haiyan zhang @haiyan]

On #WorldParkinsonsDay really proud and honoured to know @ems_lawton who champions for people with Parkinson’s Disease and shows the world every day how remarkable she is ❤️❤️❤️💪🏻

[BBC Stories @bbcstories]

#WorldParkinsonsDay Big Life Fix: Emma loves drawing but her battle with Parkinson's stops her from even writing her own name. In 2016, Innovator and Inventor Dr @Haiyan Zhang attempts to make her dreams possible again. 💪👩‍🔧

[haiyan zhang @haiyan]

When did sunsets become beautiful?

Less than 300 years ago, people did not look up at the sky and admire the beauty of the sunset. It took one painter to change that. And within one human generation, the whole world saw differently.

I've been thinking about what that means for AI, and whether we are building systems that help us see new things, or just better mirrors of what we already know. Full piece👇

https://t.co/1MVEGXk22i

[haiyan zhang @haiyan]

Agent skills! It’s the new frontier!

[Shubham Saboo @Saboo_Shubham_]

https://t.co/ka5oPvyyKW

[haiyan zhang @haiyan]

@Colteastwood @MrboomstickXL Happy Bday!

[haiyan zhang @haiyan]

Also my Claw drew its own self portrait, that she wanted me to share with you 👀

[haiyan zhang @haiyan]

I've been experimenting with @openclaw for the past two months. Some moments have been genuinely great, and some have been deeply frustrating, especially around deterministic tool use for custom CLIs and calls that were never in any model's training data. I flipped around hooking up Claw to the frontier models, while keeping an eye on the $$$ meter. So after a while, I asked Claude Code to analyze my Claw chat logs (ie. arguments) and distill a benchmark dataset based on exactly where models had failed me (!)

The result was a set of bespoke requests, 28 prompts across two sets: assistant-specific tasks (correct CLI syntax, memory hygiene, calendar edge cases, following instructions without mid-run re-planning) and general reasoning. This becomes a personal benchmark dataset I can use to evaluate new models. I then ran every model I’m considering and ranked them by what actually matters for a daily driver: latency (70%), benchmark accuracy (25%), and cost per call (5%). What I really want (!) set a default model, let it run, not worry about how much the tokens are costing, not be frustrated that my Claw is just … dumb 😂 because of the model choice.

Check out my Personal Claw Capability Leaderboard!

The standout models at the top were fast and accurate, surprisingly Ling 2.6, Gemma-4 26B, and GPT-5.4 all scored within a point of each other while staying under 2.5 seconds. Bigger, slower models with higher benchmark scores ranked lower because wall-clock responsiveness matters more than raw capability when you're waiting on your Claw to get back to you on Telegram.

Really impressed how well Ling-2.6-1t and Gemma-4-26B ranked! GPT-5.4 and GPT-5.4-nano also held up, without breaking the piggy bank 🐽💸

Encourage folks to generate your personal capability benchmark- and hold the models to account!

[Romain Huet @romainhuet]

We’ve seen Claude Code users bring in Codex for code review and use GPT-5.4 for more complex tasks, so we thought: why not make that easier?

Today we’re open sourcing a plugin for it! You can call Codex from Claude Code with your ChatGPT subscription.

We love an open ecosystem!

[haiyan zhang @haiyan]

@PeterDawoud @OpenAI @AnthropicAI

[haiyan zhang @haiyan]

Heat is on! @OpenAI vs @AnthropicAI 🔥🔥☃️

[haiyan zhang @haiyan]

@steinekin They learned from the best.

[haiyan zhang @haiyan]

Heat is on! @OpenAI vs @AnthropicAI 🔥🔥☃️

[haiyan zhang @haiyan]

Haven’t seen that! Will take a look

For openclaw I got a token sub from
https://t.co/OGrJX4DQ0E

Which is like $20/mo but they give you bonus $10 in tokens. And you can switch between all the models.

They also offer MiniMax2.5 for free - hoping they’ll upgrade to 2.7 free soon

Minimax is decent for just claw conversation, summaries. So you basically get that free and can switch to paid tokens for heavier tasks.

[haiyan zhang @haiyan]

People often ask me what AI I am using these days, and the product performance on these is never static: Noticed recently with @AnthropicAI update to increase to 1M context window for Claude Opus 4.6 - the model seems to have got really lazy at web search, almost never doing it on a query and then only when I specifically ask for it to web search…. Because it’s hallucinating or using outdated training data.

So I guess be more pushy.

I still find @ChatGPTapp to give the best overall accurate answers - just is way too verbose all the time to be super helpful.

[haiyan zhang @haiyan]

@perclae @AnthropicAI I’ve been really impressed with GitHub Copilot CLI. Of course when I worked at Microsoft we had unlimited tokens, so that helped 🥲

[haiyan zhang @haiyan]

@rdaureli @AnthropicAI Web search FTW!🤘

[haiyan zhang @haiyan]

@matvelloso it’s a lot of compute for what might, 90% of the time, be static or the same for your audience

[haiyan zhang @haiyan]

@matvelloso it’s like a world model, but for websites!

[haiyan zhang @haiyan]

@Glassdoorgamer @AmericanTimdog that was never my call, but I’m rooting for you guys

[haiyan zhang @haiyan]

🫡❤️ Thank you!

[Pure Xbox @purexbox]

Two Xbox Executives Announce Their Departures From Microsoft. https://t.co/b1jFVlfIDU