1. [AINews] Andrew Ng gets into AI Engineering — Substack

  • Why read: A useful overview of the skills that distinguish serious AI builders from casual tool users.
  • Summary: Andrew Ng’s framework covers building and deploying AI systems, software fundamentals, coding agents, and product judgment. Evaluation, error analysis, and governance are now core engineering work. Coding-agent fluency helps, but it builds on systems knowledge rather than replacing it. Develop people who can connect model behavior to production constraints and customer outcomes.
  • Read more

2. The Evolution of the Agent Harness — latent.space

  • Why read: Explains why agent reliability depends on both the model and its operating environment.
  • Summary: The harness provides the tools, memory, context, permissions, and feedback loops that let models work beyond a chat window. Benchmarks suggest the same model can perform very differently depending on the harness, making orchestration a product decision. As models absorb capabilities such as compaction and tool use, the harness shifts toward permissions, legibility, and review. Since human attention is scarce, agents need clear rules for when to act, interrupt, or ask for approval.
  • Read more

3. Designing Loops for Production-Grade Work — liquid.ai

  • Why read: A grounded example of what coding agents need to handle difficult production work.
  • Summary: Liquid AI tested agents on a tokenizer trainer requiring Rust systems work, tokenizer expertise, production-scale data, and outside verification. Both agents quickly produced plausible toy implementations, then failed on real data and external constraints. Progress came through repeated cycles of execution, failure analysis, repair, and reruns. Specify the outcome and constraints, then test against systems the agent cannot change.
  • Read more

4. AI Adoption Is an Architecture Problem — X (formerly Twitter)

  • Why read: Explains why buying AI seats rarely changes how work gets done.
  • Summary: Most organizations have a small group of enthusiastic users and a much larger group that does not use AI in daily work. Training alone is insufficient because useful deployment requires systems thinking, time, and process design. The proposed alternative is to consolidate operational data and embed AI into workflows that return structured outcomes. Measure adoption through reduced friction and completed work, not licenses or chat volume.
  • Read more

5. The Economics of the Intelligence Frontier — X (formerly Twitter)

  • Why read: Offers a useful way to think about where frontier AI remains valuable.
  • Summary: Every task has a minimum intelligence threshold and a point beyond which more intelligence is unnecessary. Above that point, buyers care more about cost, latency, reliability, and integration. Frontier labs can still create value by opening markets that previously fell below the capability threshold. Route models by task requirements, while treating routing, prompt adaptation, and long-running agent behavior as real engineering work.
  • Read more

6. How prompt caching works - Paged Attention and Automatic Prefix Caching plus practical tips — sankalp's blog

  • Why read: Practical advice for lowering cost and latency in context-heavy AI applications.
  • Summary: Prompt caching reuses computed KV states for stable prompt prefixes, avoiding repeated prefill work. The main practices are stable system prompts, append-only context, deterministic serialization, and avoiding changing tool definitions. Even equivalent JSON can miss the cache if its key order changes. Design caching at the application level, especially for agents with large recurring contexts.
  • Read more

7. Inside vLLM: Anatomy of a High-Throughput LLM Inference System — vllm.ai

  • Why read: Builds a practical mental model for how modern LLM serving achieves throughput.
  • Summary: vLLM combines scheduling, paged KV-cache allocation, continuous batching, prefix caching, and multi-GPU execution. Prefill is usually compute-bound, while decoding tokens is memory-bandwidth-bound, so the system schedules them differently. Block-based KV memory management keeps active requests efficiently batched. This helps explain tradeoffs among latency, throughput, context length, and GPU cost.
  • Read more

8. The AI Bullwhip — X (formerly Twitter)

  • Why read: Connects AI demand to the physical bottlenecks spreading through its supply chain.
  • Summary: GPU shortages pushed demand into memory, storage, CPUs, hard drives, power equipment, and construction capacity. Each bottleneck triggers investment that may arrive years later, raising the risk of overcapacity if software demand weakens. AI economics therefore depend on much more than model quality or chip supply. Stress-test infrastructure plans against lead times, substitution options, and a volatile capital-spending cycle.
  • Read more

9. Deep Dive: The Next Trillion-Dollar Futures Market — X (formerly Twitter)

  • Why read: Explains what would be needed for compute to become a tradable financial asset.
  • Summary: Compute futures could help AI buyers, data-center builders, and hardware financiers manage volatile GPU rental prices and hardware obsolescence. The main obstacle is interchangeability: identical GPUs can perform differently across providers and workloads. Durable contracts may need performance grades, delivery standards, and credible benchmarks instead of generic GPU-hours. If the market develops, its forward curve could inform AI infrastructure investment.
  • Read more

10. The Case for Software Craftsmanship in the Era of Vibes — zed.dev

  • Why read: Makes the case for raising the quality bar as code generation becomes abundant.
  • Summary: AI makes code easier to produce, but growing complexity makes humans and agents less effective over time. Judge the work by reliability, understandability, changeability, and user experience, not lines produced. AI can help engineers learn unfamiliar domains and pursue larger product goals. People still own system design and cannot use urgency to justify accumulating debt.
  • Read more

11. Fast and Hard Code — pocoo.org

  • Why read: Shows how coding agents may make efficient languages and systems techniques more accessible.
  • Summary: As agents reduce the cost of working in unfamiliar languages, developers may use Rust, Zig, eBPF, low-level networking, and similar tools more often. The result could be faster, smaller systems rather than simply more code. Domain knowledge still matters because agents need clear goals, constraints, and validation. Use AI to broaden technical ambition while keeping performance and safety checks strict.
  • Read more

12. The habits of AI writing, and what to do about them — X (formerly Twitter)

  • Why read: A practical editing guide for making AI-assisted writing clearer and more useful.
  • Summary: The test is whether writing is clear, specific, trustworthy, and useful, not whether it contains supposed AI tells. Common problems include empty profundity, generic language, excessive structure, and repetitive punctuation. Use models to spot vague phrasing, find sources, and test readability, then make the editorial calls yourself. Keep the writer’s voice and concrete meaning intact.
  • Read more

13. How I write with AI — X (formerly Twitter)

  • Why read: A practical example of using AI for research while keeping authorship of the prose.
  • Summary: The author uses AI to find sourced examples, search notes, challenge claims, and identify gaps in drafts. He avoids generated final prose, outlines, and summary-first workflows because they often preserve generic structure and erase distinctive ideas. Better results come from supplying domain context and asking for links, quotes, and evidence. Use AI as a research assistant and skeptical reader, not as the writer.
  • Read more

14. Please Don't Smooth the Metrics — Kellblog

  • Why read: Warns that investor-friendly reporting can hide the signals operators need.
  • Summary: Rolling averages clarify long-term trends but can hide sudden churn, missed targets, or weaker customer quality. In the example, smoothing delayed recognition of an ICP and renewal problem by several quarters. Keep raw quarterly movement, dollar amounts, and plan comparisons alongside smoothed views. When a metric looks noisy, investigate the noise before removing it.
  • Read more

15. Autonomy and Innovation — Stratechery by Ben Thompson

  • Why read: Argues that AI-driven cybersecurity will need autonomous defense as well as autonomous offense.
  • Summary: Cybersecurity is dual-use: the same capability can find vulnerabilities, exploit them, or patch them. Offensive agents can improve quickly because an exploit only needs to work once; defenders must preserve production systems safely. Partial automation simply moves the bottleneck from discovery to remediation. Build closed defensive loops that identify, patch, deploy, monitor, and roll back changes safely.
  • Read more

Themes from yesterday

  • Agent effectiveness increasingly depends on production loops, external verification, and harness design rather than model choice alone.
  • AI adoption is a workflow and data-architecture problem, not a training or seat-purchasing problem.
  • The infrastructure economy is expanding from GPUs into memory, storage, power, construction, and financial markets.
  • Human advantage remains concentrated in taste, specification, judgment, attention, and accountability.