GLM 5.1 Thinks Strategically, Data-Center Revolt Intensifies, When Helpful LLMs Turn Unhelpful, Humanoid Robots Get to Work

Q: Andrew Ng ranks coding agent acceleration by task type: frontend fastest, then backend, infrastructu

Andrew Ng ranks coding agent acceleration by task type: frontend fastest, then backend, infrastructure, and research least accelerated — and adjusts team expectations accordingly.

Q: Z.ai released GLM-5.1, a 754B parameter open-weights MoE model that can autonomously loop through pl

Z.ai released GLM-5.1, a 754B parameter open-weights MoE model that can autonomously loop through planning, execution, and self-evaluation for up to eight hours on coding tasks.

Q: GLM-5.1 tops the Artificial Analysis Intelligence Index among open-weights models, priced at $1.40/$

GLM-5.1 tops the Artificial Analysis Intelligence Index among open-weights models, priced at $1.40/$4.40 per million input/output tokens and available under MIT license on HuggingFace.

The Batch @ DeepLearning.AI·Friday, April 24, 2026·16 min read

AI/ML Engineering Technology

Share𝕏 in

AI Summary

Andrew Ng shares a framework for how coding agents accelerate different types of software work, ranking frontend development as most accelerated, followed by backend, infrastructure, and research. Z.ai released GLM-5.1, an open-weights 754B parameter mixture-of-experts model designed for long-running agentic coding tasks lasting up to eight hours. The newsletter also touches on data-center issues, unhelpful LLMs, and humanoid robots entering the workforce.

Key Facts

✓Andrew Ng ranks coding agent acceleration by task type: frontend fastest, then backend, infrastructure, and research least accelerated — and adjusts team expectations accordingly.

✓Z.ai released GLM-5.1, a 754B parameter open-weights MoE model that can autonomously loop through planning, execution, and self-evaluation for up to eight hours on coding tasks.

✓GLM-5.1 tops the Artificial Analysis Intelligence Index among open-weights models, priced at $1.40/$4.40 per million input/output tokens and available under MIT license on HuggingFace.

Author Takes

BearishThe Batch @ DeepLearning.AI

Coding agents and infrastructure work

Coding agents accelerate critical infrastructure even less than backend development because LLM knowledge is limited on complex infra tradeoffs, and finding infrastructure bugs still requires deep engineering expertise.

BearishThe Batch @ DeepLearning.AI

Coding agents and research

Coding agents help research only marginally because most research work is not coding — it involves formulating hypotheses, running experiments, and interpreting results, where today's agents contribute little.

BullishThe Batch @ DeepLearning.AI

Frontend development with coding agents

Frontend development is dramatically sped up by coding agents because they are fluent in TypeScript, JavaScript, React, and Angular, and can now close the loop by operating a web browser to evaluate their own output.

More from The Batch @ DeepLearning.AI

GPT-5.5 Outperforms (and Hallucinates), Kimi K2.6 Leads Open LLMs, AI Strains Climate Pledges, Strategic Thinking in LLMs vs. Humans

OpenAI's GPT-5.5 tops key objective benchmarks like ARC-AGI-2 and Artificial Analysis Intelligence Index but struggles with hallucinations, ranking th

May 1

Anthropic’s Claude Mythos Problem, Dark DNA Unveiled, Pitfalls for Assistive Models, Simulating Fluid Dynamics

Andrew Ng discusses the future of software engineering as AI agents accelerate coding, arguing against predictions of massive AI-driven job losses whi

Apr 10

Claude Code’s Source Leaks, OpenAI Exits Video Generation, Gemini Adds Music Generation, LLMs Learn at Inference

This newsletter focuses on the rapid advancement of voice-based AI interfaces and their potential to become pervasive in applications. The main story

Apr 3