The team wired LFortran to Enzyme’s LLVM‑level AD, extracting exact gradients from a 220‑line Fortran heat solver and exposing it as a JAX primitive. This lets engineers run gradient‑based optimization on decades‑old simulation code without rewriting, opening ML pipelines to high‑performance scientific models.
rjk::duck is a single‑header C++26 library that leverages static reflection to build type‑erased wrappers with far less boilerplate than Boost.TypeErasure or Folly.Poly. By annotating an interface once, you get owning and non‑owning semantics, vtable generation, and seamless runtime swapping of implementations. It works today on GCC with -std=c++26 -freflection, promising simpler, faster metaprogramming.
sx 2.0 adds a desktop app that uses a shared Dropbox (or other sync folder) as the backend for its AI‑skill package manager. Non‑technical users can drop markdown‑based skills into the folder and, with one click, have them auto‑installed across Claude, Cursor, Copilot, Gemini and more. Teams get versioned, searchable assets without git or terminals.
Spectral Compute’s SCALE compiler translates CUDA code directly to AMD GPUs, delivering up to 6× the performance of AMD’s HIP tools while preserving compatibility with existing CUDA codebases. The toolkit promises vendor‑agnostic deployment, letting developers ship the same binary to Nvidia and AMD accelerators without rewrites.
WorkOS has turned its documentation into a live AI interface using WebMCP, a new web‑standard that lets agents call structured tools directly on a page. By exposing navigation, search, and code samples as callable functions, AI can fetch exact answers without fragile HTML scraping, cutting hallucinations and sync headaches.
Claude's newest tokenizer turns the same TypeScript code into up to 73% more tokens than GPT, inflating the effective cost despite identical list prices. The analysis shows Opus 4.8 and Sonnet 5 charge the same per‑token rates but behave like $7.5/$37.5 for Opus and $4.5/$22.5 for Sonnet after accounting for token bloat. This reshapes budgeting for AI‑assisted development.
Salvatore Sanfilippo argues that with LLMs producing thousands of lines daily, reading every line is futile. Developers should steer design and architecture, using prompts to verify concepts, freeing time for strategy, new features, and quality assurance. Shifting focus boosts impact in an AI‑driven era.
By replacing most of a daily LinkedIn‑posting agent’s natural‑language SOP with a thin Python harness, the author cut token consumption by 94% and latency by 87% while keeping output quality. The trick: use LLMs only for candidate selection and draft generation, handling all deterministic steps in code.
An RL agent learns to write and launch training jobs for smaller models, then gets rewarded when those models improve. The whole pipeline runs on a $1.3k budget, showing meta‑RL can automate parts of the ML workflow without human hand‑crafting. It hints at a future where AI trains AI at scale.
Unmesh Joshi shows how domain‑specific languages act as a concrete contract for large language models, steering them toward precise, maintainable code. By iteratively building a DSL and using the LLM as a natural‑language interface, teams can keep design intent visible and avoid the drift of unchecked code generation.
Peter Gostev’s DOOMQL runs a terminal‑based shooter entirely inside SQLite, using GPT‑5.6 Sol to generate the SQL logic. Each frame is produced by a single query that outputs RGB pixels, proving a database can be a full game engine without graphics libraries. It’s a playful showcase of AI‑assisted, data‑driven game development.
LLM agents aren’t a single feedback cycle; they consist of three nested loops, inference, tool-use, and environment, that together drive autonomy. Understanding each loop clarifies why agents succeed or fail, and gives engineers a blueprint for building more reliable, self‑directed systems.
Using a DIY light‑sensor rig, the author measured end‑to‑end input latency across X11, Wayland, variable‑refresh‑rate (VRR) displays and the DXVK low‑latency fork. Wayland consistently outperformed X11, while VRR helped only when frame times fluctuated and the DXVK pacer showed modest gains in uncapped scenarios. The data debunks the myth that Wayland is slower for gamers.
Loop engineering is the practice of building autonomous AI-driven workflows that replace manual prompting. It originated from high‑profile AI teams and now powers tasks like automated PR triage and nightly test babysitting, but token costs and drift limit its adoption. Understanding its trade‑offs helps engineers decide if the hype translates into lasting productivity gains.
You’re caught up.
That was all 14 stories in the 14 July issue.