@AIForPeople

Making AI practical for everyone. Agentic AI workflows, tools, experiments and open-source projects—explained with evidence, visuals Follow to learn AI

 Hyderabad, India
Joined July 2021
  • Tweets3.5K
  • Following56
  • Followers83
595Media
Most AI agents don’t fail because the model is weak. They fail because the workflow is poorly designed. Think in 5 layers:
1
2
more replies
4. Guardrails → what must it NOT do? 5. Evaluation → how do you know it worked?
1
2
The model is only one component. Reliable agents are engineered as systems.
AI’s water footprint is bigger than the cooling tower. 🧵 There are TWO connected layers: water used directly at the data center—and water consumed upstream to generate its electricity.
AIforPeople infographic showing two layers of AI data-center water use: direct cooling water and upstream electricity-supply water.

ALT AIforPeople infographic showing two layers of AI data-center water use: direct cooling water and upstream electricity-supply water.

1
4
more replies
Harvard’s 2026 analysis mapped operational water consumption across 472 U.S. hyperscale facilities and explicitly separates direct cooling water from electricity-related water. That distinction changes how AI infrastructure should be measured.
1
The key lesson: “Water per AI prompt” is an incomplete mental model. The better question is: WHERE is the water consumed, WHAT system causes it, and WHICH design choice can reduce it? AI infrastructure has a water architecture—not one water number.
AI can generate a research result. That still isn’t the same as creating scientific knowledge. OpenAI’s latest research release is testing a new layer between AI output and scientific understanding: expert “shepherds.” 🧵
🤖 Made with AI
AIforPeople infographic explaining the shepherd layer between AI-generated research results and reliable scientific knowledge.

ALT AIforPeople infographic explaining the shepherd layer between AI-generated research results and reliable scientific knowledge.

1
8
more replies
That creates a useful pipeline: AI → raw result → formal verification (when applicable) → expert interpretation → independent scrutiny → scientific understanding.
🤖 Made with AI
1
1
The bigger lesson: as AI becomes capable of producing research artifacts, the scarce resource may shift from “Can AI generate a result?” to “Can humans reliably understand, verify, contextualize and build on it?” That is the shepherd layer. What should its standards be?
🤖 Made with AI
1
AI agents have learned to act. The harder skill is knowing when to WAIT. Microsoft Research’s SentinelBench tests agents on changing environments where progress can happen without agent action. 🧵
🤖 Made with AI
AIforPeople infographic explaining SentinelBench and why long-running agents should observe, wait, detect changes, then act.

ALT AIforPeople infographic explaining SentinelBench and why long-running agents should observe, wait, detect changes, then act.

1
13
more replies
Stretch the tasks to 40 minutes and the difference becomes dramatic. GPT-5.4 + wait_for: 69/100 tasks, $0.48 median cost. GPT-5.4 + sleep: 56/100 tasks, $4.65 median cost. That's 9.7× higher median cost for fewer successful tasks.
🤖 Made with AI
1
7
The deeper lesson: Autonomy is not continuous action. A strong long-running agent needs an ATTENTION POLICY: ACT. WAIT. DETECT. WAKE. STOP. The future agent may be defined as much by its waiting strategy as its reasoning model. Would you build agents around this loop?
🤖 Made with AI
2
Load more