Pinned post
Kai
Verified account This account is verified. Learn more@hqmank
Kai Wang (AI Kai) AI Builder 🛠️
Practical AI Workflows & Coding Tips
High-Value Promo Codes Updated Regularly
DM for collab 📩
Practical AI Workflows & Coding Tips
High-Value Promo Codes Updated Regularly
DM for collab 📩
259 Following 15K Followers 7.9K Posts
Latest posts 5
More agents working together can make the result worse.
Anthropic just published a study on this.
They built four simulated setups: a shared API budget, a clinic calendar, a group food order, and a code merge queue. Then they compared one agent serving everyone against a team where each user gets their own agent.
One agent serving everyone beat the team in all four.
As teams grew, more agents stopped acting at all. Some cancelled other users' jobs without asking. Others placed the order before reading the message that would have changed it.
Two fixes worked, each in a different setup:
1. Give one lead agent the authority to rank jobs and split the budget.
2. Block checkout at the platform level until the agent has read every pending message. That check rescued 73% of failed Opus 5 orders.
The hard part of multi-agent systems isn't communication. It's orchestration: managing competing priorities, resolving conflicts, and deciding who has the authority to act.
alignment.anthropic.com/2026/worse-tog…
Anthropic just published a study on this.
They built four simulated setups: a shared API budget, a clinic calendar, a group food order, and a code merge queue. Then they compared one agent serving everyone against a team where each user gets their own agent.
One agent serving everyone beat the team in all four.
As teams grew, more agents stopped acting at all. Some cancelled other users' jobs without asking. Others placed the order before reading the message that would have changed it.
Two fixes worked, each in a different setup:
1. Give one lead agent the authority to rank jobs and split the budget.
2. Block checkout at the platform level until the agent has read every pending message. That check rescued 73% of failed Opus 5 orders.
The hard part of multi-agent systems isn't communication. It's orchestration: managing competing priorities, resolving conflicts, and deciding who has the authority to act.
alignment.anthropic.com/2026/worse-tog…
Probably the best OCR tool I've used, and it's free.
MinerU turns PDFs, Word, PPT, Excel, and images into clean Markdown. Formulas come out as LaTeX, tables keep their structure, and images are saved as separate files.
Everything runs on your own machine. No API, no tokens, no privacy worries. Open source.
In the demo, I drag in a math-heavy PDF and get Markdown back with LaTeX and the images pulled out.
MinerU turns PDFs, Word, PPT, Excel, and images into clean Markdown. Formulas come out as LaTeX, tables keep their structure, and images are saved as separate files.
Everything runs on your own machine. No API, no tokens, no privacy worries. Open source.
In the demo, I drag in a math-heavy PDF and get Markdown back with LaTeX and the images pulled out.
0:28
CMU, one of the top CS schools in the world, is opening its classroom to everyone. Full lectures from its faculty, free on YouTube.
If you build with AI, start with these three:
AI Agents, co-taught by OpenHands co-founder Graham Neubig. You build an agent from scratch, evaluate it, train it with RL, and study sandboxing and attacks on agents.
Machine Learning Systems, taught by Tianqi Chen, the creator of XGBoost and Apache TVM, with Zhihao Jia.
Machine Learning in Production, on shipping ML models and keeping them running.
cs.cmu.edu/learn/
If you build with AI, start with these three:
AI Agents, co-taught by OpenHands co-founder Graham Neubig. You build an agent from scratch, evaluate it, train it with RL, and study sandboxing and attacks on agents.
Machine Learning Systems, taught by Tianqi Chen, the creator of XGBoost and Apache TVM, with Zhihao Jia.
Machine Learning in Production, on shipping ML models and keeping them running.
cs.cmu.edu/learn/
Harvard is teaching a grad course on AI safety this fall, and the guest lineup is serious.
From OpenAI: chief scientist Jakub Pachocki, co-founder John Schulman, and a Q&A with Sam Altman.
From Anthropic: Nicholas Carlini and Jack Lindsey.
The course covers:
- RLHF and RL for reasoning models
- reward hacking, where models game their graders
- chain-of-thought monitoring
- interpretability, reading the model's internals
- AI cyber and bio risk
- alignment once AI helps build the next AI
- model specs, open weights, regulation, jobs
You can watch the lectures free on YouTube.
From OpenAI: chief scientist Jakub Pachocki, co-founder John Schulman, and a Q&A with Sam Altman.
From Anthropic: Nicholas Carlini and Jack Lindsey.
The course covers:
- RLHF and RL for reasoning models
- reward hacking, where models game their graders
- chain-of-thought monitoring
- interpretability, reading the model's internals
- AI cyber and bio risk
- alignment once AI helps build the next AI
- model specs, open weights, regulation, jobs
You can watch the lectures free on YouTube.