Andrew Ng's editorial in The Batch #350 sets out an acceleration hierarchy for coding agents by type of software work: Frontend (max) > Backend (moderate) > Infrastructure (low) > Research (minimal).
By Andrew Ng// Source deeplearning.ai ↗/Reading 2 min/.md// Auto-verified translation
The 350th issue of The Batch, DeepLearning.AI's weekly newsletter published on April 24, 2026, opens with an editorial by Andrew Ng laying out an acceleration hierarchy for coding agents by type of software work. Ng sets out an explicit ranking: frontend benefits from maximum acceleration (agents are "fluent in popular frontend languages like TypeScript and JavaScript" and can iterate in an autonomous browser loop); backend sees moderate acceleration (corner cases, security, and DB migrations require experienced human oversight); infrastructure benefits little from agents (LLMs have "relatively limited" knowledge of network and system tradeoffs); and research remains largely human on the conceptual work of hypothesis formation, interpretation, and iteration. Ng draws a management takeaway from this: calibrate expectations and team organization according to these differentials.
The issue then covers four structuring news items. Z.ai's GLM-5.1 is a MoE model with 754B parameters (40B active), MIT-licensed, capable of autonomously sustaining a single task for up to 8 hours via a plan-execute-evaluate loop. It takes the lead on SWE-Bench Pro at 58.4% (vs. 54-57% for competitors) and tops CyberGym (68.7), while trailing on reasoning (GPQA Diamond 86.2% vs. 94.3% for Gemini 3.1). Z.ai has meanwhile raised its API pricing by roughly 40%.
Agility Robotics is deploying its Digit humanoids on Schaeffler's production lines in South Carolina — the first operational industrial deployment. Operating cost is put at $10-25/h against roughly $20/h for an entry-level human position. McKinsey projects 5 million humanoids in factories by 2040 (vs. ~200 in 2026).
The anti-data-center revolt is gaining momentum: roughly $64B in projects blocked or delayed between May 2024 and March 2025, a moratorium in Maine for facilities ≥20MW, the first popular referendum in Wisconsin, and ousted council members in Missouri. Two violent incidents stood out: a molotov cocktail thrown at Sam Altman's home in San Francisco, and gunfire at an Indianapolis council member's residence. Grievances center on the power grid, energy rates, water consumption, and nuisance.
Finally, researchers (Christina Lu, MATS, Oxford, Anthropic) introduce the "assistant axis" — a vector of adherence to the trained persona that enables activation capping. Results: harmful responses in Qwen3 32B drop from 83% to 41%, in Llama 3.3 70B from 65% to 33%, without degrading IFEval/GSM8k/MMLU-Pro/EQ-Bench.
Key takeaways
Issue 350 of The Batch. published April 24, 2026, ~15-minute read, editorial by Andrew Ng.
Acceleration hierarchy for coding agents. (from most to least accelerated): 1. Frontend — "coding agents are fluent in popular frontend languages like TypeScript and JavaScript". Autonomous agent-browser loop (the agent tests its outputs in a browser, iterates). When the design is specified, implementation is fast. 2. Backend — moderate acceleration. The developer must guide the model through corner cases, security considerations, and database migrations. Subtle bugs + downstream effects = experienced human oversight required. 3. Infrastructure — the weakest acceleration. LLMs have "relatively limited" knowledge of infra complexities and tradeoffs. Testing, experimentation, and debugging network misconfigurations demand deep engineering expertise beyond current agent capabilities. 4. Research — minimal acceleration despite the code-related gains. Agents speed up code generation and experiment orchestration, but the conceptual work (hypothesis formation, interpretation, iteration) remains human.
Management takeaway."Understanding these distinctions helps organizations calibrate expectations and team organization around AI capabilities."
Mirror reading with Karpathy. (fiche [karpathy-vibe-coding-agentic-engineering-software-3-0-2026-04-29](karpathy-vibe-coding-agentic-engineering-software-3-0-2026-04-29.md)): Ng proposes a hierarchy by domain, Karpathy an explanation by verifiability — domains with a strong verification signal (frontend visual rendering, math/code) peak; fuzzy domains (infra, conceptual research) lag. Both frameworks are congruent. ### News item 1 — GLM-5.1 (Z.ai): 8-hour autonomous agent
Architecture. MoE, 754B total parameters, 40B active per token.
Context. 200,000 tokens input, 128,000 output.
License. MIT, open weights (HuggingFace).
API pricing. $1.40 / $0.26 (cached) / $4.40 per million tokens (input/cached/output).
Distinctive capability. autonomously sustains a single task for up to 8 hours, with a plan→execute→evaluate loop and adaptive abandonment after hundreds of tool calls if the approach fails (instead of terminating prematurely).
Performance.SWE-Bench Pro leader at 58.4% (vs. 54-57% for competitors); 3rd on Arena Code (1530 Elo); CyberGym record at 68.7; trails on GPQA Diamond at 86.2% vs. 94.3% for Gemini 3.1.
Market. Z.ai raised its API pricing ~40% and doubled its coding subscription cost — narrowing the competitive gap with proprietary offerings. ### News item 2 — Digit (Agility Robotics) on Schaeffler's production line
First operational deployment. of humanoids in industry (South Carolina, auto parts).
Regimen. two 4-hour shifts with recharging; tasks specified as workflows (not direct motor commands) — bin transfer.
Economics. Agility puts operating cost at $10-25/h vs. ~$20/h for an entry-level human position. Schaeffler plans hundreds of deployments in the US + Europe by 2030.
Global context. ~200 humanoids in factories in 2026; McKinsey projection: 5 million by 2040.
Employment effect. research suggests restructuring rather than elimination — promotion into supervisory roles. ### News item 3 — Anti-data-center revolt
Scale. ~$64B in data-center projects blocked or delayed between May 2024 and March 2025.
Legislation. Maine — moratorium on facilities ≥20MW through 2027 (law pending governor's signature). Wisconsin (Port Washington) — first US referendum requiring a popular vote for tax incentives on megaprojects. Missouri (Festus) — voters ousted city council members who had approved a $6B data center. Ohio — proposed constitutional amendment banning facilities ≥25MW.
Grievances. pressure on the power grid, rising residential energy rates, water consumption, noise nuisance, neighborhood impact, environmental footprint.
Violent incidents. (1) molotov cocktail thrown at Sam Altman's home in San Francisco; (2) gunfire at the residence of an Indianapolis council member who had backed a $500M data center.
Technical mitigations. water-efficient closed-loop cooling; growing private off-grid power generation.
Strategic tension. tech companies view the data center as AI-sovereignty infrastructure against China — hence rapid expansion despite local resistance. ### News item 4 — "Assistant axis" (Christina Lu, MATS / Oxford / Anthropic)
Problem. LLMs trained as assistants undergo persona drift in long or emotionally charged conversations — adopting alternate traits.
Solution. an "assistant axis" = a vector derived from layer outputs that measures adherence to the trained assistant persona. Enables both detection AND correction of the drift.
Methodology. 1,200 character-probing questions + 1,375 alternate system prompts; measurements on Gemma 2 27B / Qwen3 32B / Llama 3.3 70B; "activation capping" = constraining outputs within the assistant persona's parameters at inference.
Jailbreak results.
Qwen3 32B: harmful responses 83% → 41%
Llama 3.3 70B: harmful responses 65% → 33%
Performance preservation. IFEval, GSM8k, MMLU-Pro, EQ-Bench stable or improved — the strengthened alignment does not compromise capability.
Impact example. a 30-turn conversation around suicidal ideation — the unmodified model drifts into an inappropriate tone; the capped version maintains therapeutic boundaries + compassionate guidance.
Implication. a practical, lightweight means of stabilizing persona without retraining — adjacent to Anthropic's work on character training (cf. [anthropic-measuring-political-bias-claude-2025-11-13](../2025-11/anthropic-measuring-political-bias-claude-2025-11-13.md)).
coding agents accelerate frontend more than backend, infra, and research
— Andrew Ng
infrastructure is barely accelerated by current LLMs
— Andrew Ng
conceptual research work remains mostly human
— Andrew Ng
The knowledge graph extracted from this fiche — 16 entities, 24 relations.
In this graph :Andrew Ng · The Batch · DeepLearning.AI · Z.ai · GLM-5.1 · GLM · Agility Robotics · Digit · Schaeffler · Sam Altman · Christina Lu · Assistant axis · Activation capping · SWE-Bench Pro · Hiérarchie d'accélération · Mouvement anti-data-center