Yue Zhao

Assistant Professor @CSatUSC; | Anomaly & OOD Detection | AI Safety & Reliable LLMs | AI for Science | ML Systems | ⭐22k GitHub · Top 700 Worldwide

[Yue Zhao @yzhao010]

Do not forget to cancel your subscription to @TheEconomist and @WSJ :)

[Yue Zhao @yzhao010]

那学长有没有分享...没有这个 #PhD，哪个大公司给他机会去尝试的...

换句话说，没有读博的实习低成本进入企业，本科生、硕士有多少机会？

[Kimberly @king1818888]

今天跟刚博士毕业的学长讨论
一个暴论：这个时代博士学位没有卵用，paper没卵用，citation没卵用。

学术体系在AI的冲击下其实已经死了，只是还没埋。

早点去公司里，用公司的大集群，用公司报销额度充最好的AI PLAN，多多建立connections，攒足够多的钱，

然后在LLM kill the game之后，在世界毁灭之前，好好享受你的人生

[Yue Zhao @yzhao010]

This is extremely stupid and myopic to consider that they can re-(create) NVIDIA with mediocre management. This is the typical arrogance of the elite class, thinking everyone would agree.

Read the room!

[The Economist @TheEconomist]

Like many powerful leaders, Nvidia’s co-founder came to believe that he was too important for the normal rules to apply to him. But no boss is irreplaceable https://t.co/hm0diaD0Rl

Photo: Getty Images

[Yue Zhao @yzhao010]

现在很多#AI创业不是本身就是面向investor创业吗？

今年到明年肯定要死掉一大批AI创业公司 :)

[Akintola Steve @Akintola_steve]

What a joke of the century

[Yue Zhao @yzhao010]

ap loads 14 public data sources into one format, keeping messages, tool calls, results and labels together.

Use it to inspect traces, compare datasets or export data for training.

Code: https://t.co/vzVRDx0ULz

#AIAgents #OpenSource

[Yue Zhao @yzhao010]

I started working on agent auditing and kept getting stuck on the data: every dataset had a different trace format.

So we built AgentPile (ap) to handle that step.

Here is a one-minute demo with narration.

[Yue Zhao @yzhao010]

We welcome researchers with a PhD in AI, CS or a related field and relevant expertise. Help doctoral students sharpen their research plans with concise, constructive feedback.

Please share with colleagues! #PhD #MachineLearning

[Yue Zhao @yzhao010]

Help shape the next generation of AI faculty and researchers!

Join the AAAI-27 Doctoral Consortium PC. Rigorous but lightweight reviews: 3-5 applications, due Nov 7, 2026.

Sign up: https://t.co/VPs3vEwzSD

#AAAI2027 #AI #AcademicTwitter

[Yue Zhao @yzhao010]

This is exciting! Thank you @foresightinst !

[Foresight Institute @foresightinst]

https://t.co/5C91ao5INd

[Yue Zhao @yzhao010]

@adocomplete I think it could be a good program. I got in while the discount never applied, and none of the support responded in the last two months...

[Yue Zhao @yzhao010]

@ClaudeDevs I have been granted this 6-month max -- while still getting charged for $200/month. None of the supports works :(

[Yue Zhao @yzhao010]

30 days. Four support channels. Zero human replies. And on May 20 Anthropic charged me a second $200 for a benefit that was supposed to be free.

I maintain PyOD, the open-source anomaly detection library (9.8k stars, 42M+ downloads). In March 2026 I was accepted into the Claude for Open Source Program, for "Max 20x free for 6 months."

Instead, on both invoices Anthropic's billing applied a "6 Months of Pro (100% off)" promo to my Max 20x subscription. The promo discounts $0, so I was billed the full $200 each month:
- April 20, invoice -0011: $200
- May 20, invoice -0012: $200
Same incorrect promo line both times. $400 charged, and the benefit I was accepted for has never applied.

For a month, no human has replied. The Fin bot did not route my explicit human-escalation request. [email protected] and [email protected] have been silent since April 22. My GitHub issue was auto-labeled "invalid," then "stale."

I am not posting this out of frustration. I use Claude every day and want to keep doing so. Could @AnthropicAI or the Claude for Open Source team route ticket 215473995216928 to a human billing owner, reverse invoices -0011 and -0012, and apply the promised Max 20x OSS benefit? I can confirm the account details by DM.

#ClaudeCode #Anthropic #OpenSource #MachineLearning #AcademicTwitter

[Caiming Xiong @CaimingXiong]

In real healthcare operations, agents must do far more than answer medical questions. They need to read charts, interpret clinical and operational policies, verify coverage, route referrals, draft P2P scripts, and finalize care plans — where a single policy violation can mean a denied claim or missed patient outcome.

@actAVAai @iscreamnearby led and developed CHI-Bench (Clinical Healthcare In-situ Benchmark), the first long-horizon, policy-rich benchmark for AI agents operating across end-to-end U.S. healthcare workflows.

Key highlights:

▶️ High-fidelity simulators for Provider Prior Authorization, Payer Utilization Management, and Population Health Care Management, all exposed as MCP servers over patient, clinician, and insurer records.

🧪 Each trial runs 60–80 agent steps across 4–6 clinical stages, with access to 21 healthcare apps, 200+ MCP tools, and a 1,279-document operations handbook.

Leaderboard results across 30 frontier agents:
• Claude Code + Opus 4.6: 28% pass@1
• Codex + GPT-5.5: 21%
• Utilization review: 41%
• Care management: 32%
• Prior authorization: 29%
Reliability remains a major challenge: no agent exceeds 20% when the same case is repeated three times.

[Yue Zhao @yzhao010]

🎉 WSDM turns 20 in Hong Kong this February.

The 20th ACM Conference on Web Search and Data Mining is open for submissions: https://t.co/teEddxMziL

Topics in scope:
🤖 Foundation Models and Agentic Systems (LLMs, RAG, agent benchmarks)
🔎 Web Search (ranking, retrieval, query understanding)
💬 Intelligent Assistants (recommendation, conversational, personalization)
🕸️ Social Networks (graph mining, GNNs)
🛡️ Privacy, Fairness, Interpretability
✅ Information Integrity (misinformation, anomaly and fraud detection)
📊 Web Mining and Content Analysis
📡 Web of Things, Mobile Computing
👥 Crowdsourcing and Human Computation
🌟 Emerging and Creative Applications

Why submit:
✨ Findings track auto-considers main-track submissions for a second-shot venue
🎤 Keynotes from Bowen Zhou, Lars Backstrom, Jimmy Huang
📍 Cordis, Mong Kok, easy travel from Asia / NA / EU

🗓️ Key dates
✏️ Abstract: Aug 11, 2026
📝 Full paper: Aug 18, 2026
📬 Notifications: Oct 27, 2026
🎤 Conference: Feb 15-19, 2027

#WSDM2027 #AcademicTwitter #MachineLearning #LLMs #DataMining

[Yue Zhao @yzhao010]

@AnthropicAI Accepted into the Claude for Open Source Program in March 2026 (promised "Max 20x free for 6 months"). My April 20 receipt shows the promo applied to "Pro" at $0, not Max 20x, so I was still billed $200.

Ticket 215473995216928 has been open since April 20. Email follow-up sent April 27. In-app Fin chat silent for a week despite an explicit human-escalation request.

Could the billing team look at the SKU mismatch? The receipt itself shows the promotion went to the wrong product.

#ClaudeCode #OpenSource #AI

[Yue Zhao @yzhao010]

🚀 agent-style: 21 drop-in writing rules that shape how AI coding agents write like a tech pro.

Works with any AGENTS.md-compliant tool: Claude Code, Codex, Copilot, Cursor, Aider, and more. 12 rules from canonical writing authorities (Strunk & White, Orwell, Pinker, Gopen & Swan), 9 from field observation of LLM output.

v0.3.0 sanity bench, AI-tell violations on 10 fixed prose tasks:
• Claude Opus 4.7: 105 → 58 (-45%)
• GPT-5.4 via Codex: 51 → 28 (-45%)
• Gemini 3 Flash: 79 → 14 (-82%)

How it works:
• Injected into the agent's system prompt, so the first draft already reads clean
• Deterministic audit layer catches em-dashes, transition openers, cliché verbs, contractions, same-starts
• One `pip install agent-style` or a drop-in AGENTS.md block

213+ GitHub stars in 3 days.

Install: pip install agent-style
Repo: https://t.co/G9R2S3I0c7
Docs: https://t.co/T2PFrcSaXO

#AIAgents #LLMs #OpenSource #Python #PromptEngineering

[Yue Zhao @yzhao010]

🚀 PyOD 3 is out, and the big change is agentic.

PyOD has long been the Python default for anomaly detection (scikit-learn style). v3 adds a new layer on top:

• Describe your problem in natural language
• An agentic investigator picks detectors, tunes parameters, runs the analysis, and explains the result

Under the hood, ADEngine orchestrates 60+ detectors, now covering time series, graphs, text, and images, not just tabular data.

Still a one-liner when you want control:
from pyod.models.X import X; X().fit(data)

38M+ downloads, ~10k GitHub stars. Used across fraud, security, medical, and ML robustness.

Install: pip install pyod
Docs: https://t.co/ijn1oxuLVg
Repo: https://t.co/OAPQhbHFj8

#AIAgents #AnomalyDetection #MachineLearning #Python #OpenSource

[Yue Zhao @yzhao010]

This marks the restart of my Twitter/X experience. Will see what to post then!