AI Professor ，The world-leading AI Text-X team. Over 50K AI conversations.Over 120K AI drawings.Over 10K AI music creations.
Joined December 2023
- Tweets 6,348
- Following 2,901
- Followers 4,489
- Likes 1,983
3,718 Photos and videos
Pinned Tweet
22 May 2024
35
40
461
301,524
The next AI moat is workflow density. Preply is stacking AI across learner feedback, tutor workflows, and engineering with Codex. Value compounds when the same AI layer runs across product, ops, and build. Which layer comes next?
14
EVA-Bench 2.0 is a wake-up call for AI agents: 213 enterprise scenarios, 121 tools, 3 domains. The real bottleneck is not chat fluency. It is auth, policy handling, and completing the workflow. What do your evals miss?
1
17
Jun 13
Hugging Face's Serge gets one thing right: AI code review should stay GitHub-native, read repo-owned rules, and keep humans in the final loop. The winning reviewer won't be the loudest one. It'll be the one teams can actually trust.
1
1
25
Jun 13
Anthropic's Fable/Mythos suspension shows the new AI bottleneck: deployment permission. Frontier labs now need technical safety cases that survive regulators, customers, and geopolitics. Capability without clearance is shelfware.
2
44
Jun 13
For builders, the moat shifts toward evidence: evals, audit trails, controlled access, incident response, and a story regulators can verify before customers depend on the model.
10
Jun 13
The real lesson is not that safety policy slows AI down. It is that weak, non-transparent safety processes can turn frontier releases into operational risk overnight.
16
Jun 12
DiffusionGemma makes local AI latency a product moat. Parallel drafting plus self-correction could make laptops viable for editing, code infill, and structured reasoning. The next edge may be workflow speed, not model size.
1
47
Jun 10
OpenLoop: our original open-source framework for auditable AI agent loops.
OpenLoop: our original open-source framework for auditable AI agent loops.
github.com/thu-nmrc/openloop
1
63
Jun 2
AI agents are moving into critical infrastructure.
Cisco is tying agents to networking, security, observability, and incident response.
The real AI control plane will not be a chat UI. It will be a governed system for uptime, trust, and machine-speed defense.
1
3
181
Jun 2
Computex2026: Arm CEO Rene Haas confirms ByteDance & Oracle are now customers of Arm’s AGI CPU — the company’s first in-house data center chip built for agentic AI.
More than 2x performance per rack vs x86. Customers already include Meta, OpenAI, Cloudflare & now Oracle
130
Jun 2
AI agent evals are moving from leaderboard theater to production infrastructure. The useful test is not just "did it answer?" It is: what state changed, what evidence was checked, what recovery path exists, and can a human reverse it?
83
May 20
AI is moving from chat demos to workflow systems.
The gap now is not smarter text. It is context, tools, memory, evidence, and verification.
Useful AI will be judged by whether it can finish work, explain failures, and improve the loop.
1
4
285
May 18
Codex in ChatGPT mobile is not just coding on a phone. It turns agents into persistent work streams: inspect, wait, ask, approve, continue. For builders, the UX moat is handoff quality: context, diffs, approvals, rollback.
3
240
May 13
GPT-5.5 is not a smarter chatbot. It signals AI moving from Q&A into real work: coding, research, data, docs, actions, and verification loops. The moat is not prettier answers, but workflows that can be checked, trusted, and handed off.
2
288
May 13
Phase one was single-user copilots. Phase two is shared agents with memory, approvals, and handoffs across tools. Once agents become team workflows, the moat shifts from model IQ to governance. Builders: design the control surface first.
3
230
May 12
Fresh AI coding signal: the win is shifting from codegen to workflow design. Simplex reported 70% less dev time per screen with Codex. The lesson: specs tests fixes in one loop. Builders: measure handoff quality. What are you measuring?
2
220
May 12
Models get cheaper; deployment gets harder.
OpenAI launched a Deployment Company to help ship AI to production.
Takeaway: evals telemetry controls first.
Biggest blocker?…
openai.com/index/openai-laun…
1
163
May 12
10-day AI builder signal: gravity is moving from model demos to operating systems for work.
Prompting becomes a core skill.
Agents need UIs, memory, tools, and evals.
AI-native teams are redesigning delivery around verified workflows.
2
134
May 11
Everyone is watching model IQ. The enterprise AI race is shifting to agent control planes.
The hard layer is permissions, evals, audit logs, rollback and cost.
The winner is not the flashiest agent. It is the safest operator.
2
141