Chrome, Firefox, and Safari now support the CSS Custom Highlight API, letting you define reusable highlight styles via ::highlight(). The roundup also shows a paint‑order hack to correct text‑stroke bleed, and a new scroll‑axis‑lock property that finally enables true diagonal scrolling. These additions tighten the gap between design intent and native CSS capabilities.
Google just released Credentio, an open‑source C++ library that validates C2PA Content Credentials entirely on‑device. By keeping memory usage tiny, it can check gigabyte‑scale videos or high‑res images without sending data to the cloud, boosting privacy and latency for media apps. The repo is live, and Google invites community contributions.
DeepSeek Harness is in developer preview, letting developers build agent harnesses where every capability, models, tools, scheduling, UI, is a plugin that can be swapped or extended without touching core code. It records every interaction for full traceability and offers multiple runtime modes, from full toolsets to minimal benchmarking.
GitHub’s blog walks through a pull‑request‑centric workflow that uses four agent apps, Amplitude, Endor Labs, LaunchDarkly and PagerDuty, to answer scope, security, rollout and deployment questions without leaving GitHub. By bringing product analytics, vulnerability scanning, feature‑flag control and incident response into the PR, teams cut context‑switching and ship safer code faster.
Google unveiled HEIR, an open‑source compiler that turns standard AI models into encrypted‑inference versions. By letting servers compute on ciphertext, it lets sensitive sectors run AI without exposing raw data, shrinking the privacy‑cost trade‑off. The toolkit aims to make private AI a one‑click option for developers.
Qwen 3.8 27B FP8 is a 27‑billion‑parameter multimodal LLM released on Hugging Face with 8‑bit quantization, slashing memory and cost. It ships ready‑to‑run examples for Transformers, vLLM, SGLang and notebooks, letting developers deploy the model locally or in the cloud with minimal setup.
Simon Willison shows how to bypass a bloated tag list by prompting an LLM to invent descriptive labels for a post, then using vector embeddings to match those invented tags to the existing taxonomy. The trick automates tagging at scale and keeps blogs searchable without hand‑curated vocabularies.
The essay argues that AI interactions are too unpredictable for pure command‑style coding, demanding context, clarity, and feedback, hallmarks of leadership. By treating prompts as conversations and sharing intent, developers align AI output with goals, turning software work into guided collaboration.
Google’s new knowledge‑profiling framework shows that Gemini 3, GPT‑5 and similar frontier models store the vast majority of facts but fail to retrieve them, turning factual errors into ‘lost‑key’ problems. By classifying facts into encoding, recall and reasoning categories, the study points to post‑training retrieval tricks as the next lever for improving LLM reliability.
Google's new Gemini 3.7 Flash model boosts code generation accuracy by up to 30% and halves the price of its predecessor. It also outperforms 3.6 Flash in web‑dev layouts, finance and legal reasoning, and is now powering Gemini Spark for Google Workspace tasks.
In GPU Mode’s auto‑research contest, I used OpenAI Codex to generate a batched Householder QR factorization kernel that outperformed the baseline by 232×. The workflow combined automated code synthesis, iterative prompting, and targeted profiling, turning what was a research‑grade bottleneck into a production‑ready speedup.
A 1978 JIS encoding slip introduced several 'ghost' CJK characters that never existed in the real world. Decades later those phantom glyphs survived into Unicode, causing rendering glitches and confusing developers who encounter unmapped code points. Understanding their origin prevents mis‑interpretation and helps clean up text pipelines.
The author treats every post as a research experiment: if writing doesn’t teach him something new, he discards it. By insisting each article defends a controversial stance, he forces deep research and clearer thinking, turning blogging into a powerful learning tool for engineers and anyone who wants to sharpen their understanding.
You’re caught up.
That was all 13 stories in the 15 August issue.