Yuzhao (Allen) Ni

Research Engineer at @GoogleDeepMind

[Yuzhao (Allen) Ni @YuzhaoNi]

Best has yet to come.

[Google DeepMind @GoogleDeepMind]

We’re rolling out three new models to make AI agents faster, smarter, and cheaper at scale:

🔵 Gemini 3.6 Flash: It uses fewer tokens than 3.5 Flash to deliver higher quality work at the exact same cost.

🔵 Gemini 3.5 Flash-Lite: A fast, cost-effective option for everyday tasks like processing documents and agentic search.

🔵 Gemini 3.5 Flash Cyber: A cybersecurity model built to find and patch critical software vulnerabilities.

[Yuzhao (Allen) Ni @YuzhaoNi]

@OpenAI I partially agreed. But 30% might be a bit exaggerated as other frontiers are able to score higher than 70%.

[Yuzhao (Allen) Ni @YuzhaoNi]

it is a good model

[Google DeepMind @GoogleDeepMind]

We’re shipping 2 major releases:

🔘 Nano Banana 2 Lite: our fastest and cheapest Gemini Image model
🔘 Gemini Omni Flash: now available via the Gemini API and in @GoogleAIStudio to help developers generate and edit high-quality videos.

[Yuzhao (Allen) Ni @YuzhaoNi]

5.Discipline matters: file must be https://t.co/hRMuMPiTKB (case-sensitive), kebab-case folders, no angle brackets in YAML, no https://t.co/JbMb85nkXC, keep under 5k words. And verify with scripts: “code is deterministic—language isn’t.”

Guide link: https://t.co/TyKqXt3i3S

[Yuzhao (Allen) Ni @YuzhaoNi]

Anthropic’s official Claude Skills build guide is 🔥
Skills = the best way to “freeze-dry” your expertise + workflow into a reusable instruction pack 🪄
Turn Claude from “general assistant” → “domain specialist.”

[Yuzhao (Allen) Ni @YuzhaoNi]

4.5 proven patterns: sequential workflows, multi-MCP coordination, iterative refinement w/ QC loops, context-aware tool choice, domain knowledge embedding.

[Yuzhao (Allen) Ni @YuzhaoNi]

3.Start with ONE hard task: iterate until it works, then distill into a Skill. Don’t boil the ocean. First usable Skill in 15–30 min.

[Yuzhao (Allen) Ni @YuzhaoNi]

2.Description field = life or death: must say what it does + when to use it + include real trigger phrases users will type. Too vague = never called. Too broad = chaos.

[Yuzhao (Allen) Ni @YuzhaoNi]

My 5 biggest takeaways 👇
https://t.co/roUQVulV4Y disclosure = the secret sauce: YAML frontmatter always-on → https://t.co/hRMuMPiTKB on-demand → references/ only when needed (token-efficient and deep).

[Yuzhao (Allen) Ni @YuzhaoNi]

Thoughts: AI benchmarks are saturating fast, showing incredible progress! Humanity's Last Exam (HLE) is the new frontier. #AI #Innovation

Prediction: By 2026, AI might surpass 60% on HLE, pushing us to redefine what we consider challenging for AI. Exciting times ahead! #AIFuture

[Jason Wei @_jasonwei]

Made this plot for an upcoming talk---crazy how quickly benchmarks get saturated these days. Looking forward to seeing how things play out for Humanity's Last Exam!

[Yuzhao (Allen) Ni @YuzhaoNi]

@i0ns0 @AnthonySusanto9 @CodeFryingPan But not many people will quit a 270000$ job and joins YC considering most of startup fails

[Yuzhao (Allen) Ni @YuzhaoNi]

@Dataminr do you have the original HD photos? @sakima_k