Published on August 04, 2026 at 06:00 CEST (UTC+2)
LLMs reward expertise (576 points by MaxMussio)
LLMs reward expertise — Sean Gedecke argues that domain expertise, not generic prompt-crafting, is the real skill that separates users of large language models. He contrasts average users getting "sort-of-okay" output with Terence Tao's conversation with ChatGPT about the Jacobian Conjecture counterexample, where Tao's short, to-the-point messages and mathematical signaling elicit far more advanced responses from the same model. The piece concludes that LLMs have made everyone a generalist but have simultaneously raised the value of specialized knowledge, since experts can push models toward results that novices cannot reach even with unlimited tokens.
Amazonian civilization had estimated 3M people in 3% of forest area (69 points by marojejian)
Amazonian civilization had estimated 3M people in 3% of forest area — This Science article reports on lidar and remote-sensing surveys that have revealed dense earthworks, roads, and settlement structures hidden under the Amazon rainforest canopy. The findings suggest a sprawling pre-Columbian civilization with an estimated three million inhabitants concentrated in just 3% of the forest area. The discoveries overturn long-held assumptions that the Amazon was sparsely populated before European contact and highlight how aerial mapping technologies are transforming archaeology.
Ten advances in mathematics and theoretical computer science (472 points by milkshakes)
Ten advances in mathematics and theoretical computer science — OpenAI highlights ten research achievements where its models contributed to progress in pure mathematics and theoretical computer science. The post documents specific results and conjectures where AI-assisted reasoning, formal verification, or model-generated insights played a role alongside human researchers. It signals a shift toward treating frontier models as legitimate collaborators in rigorous mathematical discovery rather than just text generators.
That time when I failed the Microsoft interview (18 points by wofo)
That time when I failed the Microsoft interview — The author recounts applying for a Microsoft internship in 2015 as a CS undergraduate with strong open-source contributions, only to fail the phone interview. The post details the scheduling challenges of a 9-hour time difference, the pressure of classic "Cracking the Coding Interview" preparation, and the specific mistakes that cost him the position. It's a reflective, humbling retrospective about how confidence from real-world projects doesn't automatically translate to success in algorithmic whiteboard-style interviews.
Devtools must be open source (531 points by bryanmikaelian)
Devtools must be open source — The author reflects on how five years ago, few engineers wrote personal software tools; today, AI has made customizing and building personal developer tooling far more accessible. The post argues that developer tools fundamentally need to be open source so engineers can trust, inspect, modify, and extend the software that shapes their daily work. It positions open source not merely as an ideological choice but as a practical requirement for tooling that must adapt to individual workflows.
Show HN: Run an 80B Qwen in 4.3 GB of RAM on a Mac, and a 35B on an iPhone (28 points by leonickson)
Show HN: Run an 80B Qwen in 4.3 GB of RAM on a Mac, and a 35B on an iPhone — Swiftlet is a Swift + Metal runtime for Qwen3-Next and Qwen3.5/3.6 mixture-of-experts models that keeps only the small dense core resident in memory while streaming routed expert weights from disk on demand. This enables a Qwen3.6-35B-A3B model to run in 2.6 GB of RAM at 7–11 tok/s on an M5 Mac, and an 80B model in 4.3 GB at 4.5–5 tok/s, with the 35B also fitting on an iPhone 17. It's a striking demonstration that massive model architectures no longer require massive hardware when combining MoE sparsity with storage streaming.
Ask HN: Who is hiring? (August 2026) (126 points by whoishiring)
Ask HN: Who is hiring? (August 2026) — The monthly hiring thread features companies posting roles across the industry, with a notable concentration of AI/ML positions. The preview highlights Aptura AI, which builds evaluation datasets, calibrated rubrics, and reinforcement-learning environments for frontier AI labs targeting finance, healthcare, and legal applications. The thread reflects the current demand for applied AI engineers who can work on evaluation, RL environments, and domain-specific reliability.
Smaller, faster, safer: running Kimi and GLM at scale (167 points by ascorbic)
Smaller, faster, safer: running Kimi and GLM at scale — Cloudflare details how it serves Moonshot's Kimi K-series and Z.ai's GLM — large, long-context mixture-of-experts models — across its edge GPU infrastructure. The techniques include quantizing the KV cache, compressing model weights, and protecting shared caches when packing more requests onto hardware, all benchmarked with the open-source SGLang inference framework. The post shows how these optimizations reduce cost and improve throughput with no measurable loss in model accuracy.
Ask HN: Who wants to be hired? (August 2026) (77 points by whoishiring)
Ask HN: Who wants to be hired? (August 2026) — The companion job-seeker thread shows a workforce increasingly advertising AI-related skills. The preview highlights an AI full-stack engineer from Nepal with experience in TypeScript, LLM integration, RAG agents, MCP, and Cloudflare Workers, who currently leads engineering for an AI product. The thread illustrates that LLM application development is becoming a standard resume item rather than a specialist niche.
Windows XP 2002 for the Itanium: Unbridled rage (68 points by jandeboevrie)
Windows XP 2002 for the Itanium: Unbridled rage — A retrocomputing post detailing the author's painful journey to get Windows XP ia64 build 2600 running under a fork of QEMU with working Itanium Merced emulation. The write-up covers building cross-compilers, firmware complications, and the general misery of emulating an obscure architecture. It's part technical documentation, part war story, celebrating the persistence required to resurrect historical software on incompatible hardware.
Domain expertise is the new "prompt engineering" — Article 1's Terence Tao example demonstrates that the gap between expert and novice LLM output is not about prompt syntax but about the user's ability to evaluate, guide, and steer the conversation. As models become increasingly capable, the competitive advantage shifts from "knowing how to ask" to "knowing what a good answer looks like." Why it matters: Prompt-craft as a standalone skill is being commoditized; the durable skill is deep domain knowledge paired with AI collaboration. Implication: Organizations should invest in upskilling domain experts to work directly with models (in mathematics, law, medicine, engineering) rather than building layers of prompt intermediation.
AI is becoming a research-grade mathematical collaborator, not just a text tool — OpenAI's "Ten advances in mathematics and theoretical computer science" (article 3) combined with the Tao conversation (article 1) signals that frontier models are producing original, verified contributions to rigorous fields. These are not demos but potentially publishable research outcomes involving formal reasoning and counterexample discovery. Why it matters: Verification, not generation, is now the bottleneck — even expert collaborations require careful checking of model outputs. Implication: Expect new research workflows where human mathematicians validate AI-proposed conjectures, and expect "AI co-authored" papers in pure math to become normal within a couple of years.
MoE + aggressive memory management is the dominant path to edge AI — Article 6 (Swiftlet running an 80B Qwen in 4.3 GB RAM) and article 8 (Cloudflare serving Kimi/GLM with KV-cache quantization and weight compression) both converge on the same architecture pattern: mixture-of-experts models where only the dense core stays resident and experts are streamed or compressed. This makes frontier-class model capabilities feasible on phones, laptops, and edge datacenters. Why it matters: The defining constraint for AI deployment is no longer total parameter count but memory bandwidth and cache management — which means inference engineering (not model architecture alone) determines what's deployable where. Implication: Products should be designed around MoE streaming runtimes; expect a wave of "small-device, large-model" consumer apps that were impossible 12 months ago.
The AI job market is bifurcating into "LLM app builders" and "AI reliability engineers" — The hiring thread (article 7) shows demand for people building evaluation datasets and RL environments where "mistakes are expensive," while the candidate thread (article 9) is full of engineers fluent in RAG, MCP, and LLM integration. The former is a high-value, specialized niche; the latter is rapidly becoming table-stakes. Why it matters: Model capability is outpacing trust infrastructure — finance, healthcare, and legal adoption will be gated by who can build calibrated rubrics and safety environments, not by who can bolt RAG onto a chatbot. Implication: Engineers who move from "wiring LLMs into apps" to "building the evaluation and RL infrastructure that makes LLMs trustworthy" will capture disproportionate career leverage.
Inference efficiency has become a competitive moat — Cloudflare's post (article 8) is explicit: quantizing KV caches and compressing weights "enable us to support more customers at lower costs, with no change in model accuracy." This is the same dynamic that made companies like NVIDIA and OpenAI giants — whoever serves tokens cheapest, wins. The fact that these optimizations are being upstreamed to SGLang suggests a commoditization race where efficiency gains rapidly spread. Why it matters: As open-weight models converge in quality, serving cost becomes the primary product differentiator. Implication: Teams should benchmark models on delivered tokens per dollar on their exact hardware, not just on public leaderboards; edge providers offering "smaller, faster, safer" models will undercut centralized API pricing.
The push-pull between local/private inference and shared-edge safety — Article 6 puts an 80B model in a personal device (extreme privacy, no telemetry), while article 8 positions Cloudflare as the safe shared infrastructure layer ("protecting the cache those requests share") — and titles itself "smaller, faster, safer." These are two competing visions for the future of AI computation: fully on-device vs. trusted-edge-with-shared-infrastructure. Why it matters: Security models and compliance requirements will differ radically depending on which wins in a given vertical (healthcare, finance, defense may prefer local; consumer apps may prefer edge). Implication: Builders should architect apps to be deployment-agnostic — able to run the same model on-device or at the edge depending on regulatory, privacy, and latency constraints. The MoE-streaming pattern in articles 6 and 8 is the enabling technology for both.
Open-source infrastructure is consolidating around a few serving/runtime stacks — The ecosystem is quietly consolidating: SGLang (article 8) handles serving, Swiftlet-style runtimes handle local MoE execution, and the Cloudflare team explicitly upstreams their patches to keep SGLang the common substrate. Meanwhile, article 5 argues devtools must be open source because engineering workflows depend on inspectability and customizability. Why it matters: Open-source inference infrastructure will likely become the default substrate for both cloud and local AI — meaning competitive advantage shifts upward to data, evaluation loops, and vertical applications, not the serving layer. Implication: Startups should treat inference as commodity infrastructure and invest in proprietary evaluation data, RL environments, and expert-in-the-loop workflows — which are precisely the roles the hiring thread (article 7) is racing to fill.
Analysis generated by deepseek-reasoner