Open-source skills for controlling AI
https://github.com/mattpocock/skills
These Skills turn AI from an “improvising assistant” into a controlled engineering tool. They solve real pain points of working with AI: reduce chaos and hallucinations, save tokens, scale well, and increase predictability. LLMs quickly “get dumber” due to attention degradation in long sessions, so the approach deliberately breaks work into smaller pieces, preserves what is already understood, and passes context through documents instead of keeping everything in one long session.

Knowledge is transferred via CONTEXT.md and ADRs — the agent and the team speak the same language. Skills must be invoked manually (this is by design) and require some setup time. They work best with strong models + a solid agent harness (Claude Code, etc.). The biggest gains come on medium and large tasks where quality, reproducibility, and control over context matter.

Matt Pocock’s video is a practical tutorial on his popular open-source skills repository. He shows the full cycle of working with AI agents (Claude Code and others) for real engineering.

https://www.youtube.com/watch?v=M6mYodf0dJM

It covers installation and repository setup (/setup-matt-pocock-skills) — choosing an issue tracker (GitHub, Linear, local markdown files, etc.), triage labels, and domain documentation (CONTEXT.md + ADRs).

Core workflow:

• Ask Matt — a router that suggests where to start.
• Grill with docs — a deep agent interview that sharpens the idea, explores the code, and updates domain documentation.
• To-spec — compresses the conversation into a detailed spec (problem statement, user stories, implementation/testing decisions).
• To-tickets — breaks the spec into vertical “tracer-bullet” tickets (each fits inside the ~100–140k token “smart zone”).
• Implement + code-review — implementation with TDD elements and review.

Demonstrated via a refactoring / legacy-cleanup example: Grill with docs → the agent explores the namespace, asks 6–20 questions → shared plan → to-spec → to-tickets or implement (if it fits in the smart zone).

#bestpractices #agentskills

Comments

For several months now, new models have been able to operate autonomously for hours and generate rather complex projects.

AI is Burning Money
https://www.tomshardware.com/tech-industry/artificial-intelligence/amazon-accidentally-spent-usd1-8-million-using-claude-for-menial-coding-task-went-860-percent-over-budget-catastrophically-expensive-coding-blunders-discovered-in-internal-amazon-ai-usage-metrics
Amazon accidentally spent $1.8 million on Claude Sonnet for a simple task (matching authors with products). The budget was exceeded by 860%, the error went unnoticed for 5 months, and the project was never launched. A bug like this, which previously would have cost pennies, now burns hundreds of thousands of dollars due to token-based pricing. Naturally, no refunds are provided—the riches go to the providers, who tempt users in their advertisements to deploy more and more autonomous agents.

Software Continues to Degrade
https://ptrchm.com/posts/nothing-works-and-everyone-is-euphoric/
The author ironically describes the current wave of AI enthusiasm as a form of mass psychosis. We are constantly promised that artificial intelligence will radically boost productivity, automate programming, and deliver higher-quality software. But if code generation is practically a solved problem, why does the software around us continue to get worse?

Updates spark anxiety, user interfaces degrade, and bugs multiply—a banking app that constantly requires re-authorization; Slack stealing window focus; the LG website where the warranty claim form breaks at the very last step; or an in-car multimedia system that became even less reliable after an update, even compromising safe driving.

The problem is not a lack of tools. The teams building these major products almost certainly have access to the most advanced AI models. This text is not an anti-AI manifesto. On the contrary, the author believes AI grants developers extraordinary capabilities—allowing small teams to build high-quality, simple, and reliable products that were previously beyond their reach.

Discussion
https://news.ycombinator.com/item?id=49033004
A massive discussion on HN with nearly 700 comments. Users overwhelmingly confirm: software updates now bring anxiety rather than joy. AI has not made mainstream software better—it has simply accelerated the production of mediocre code and fostered irresponsibility ("the AI generated it, not us").

The Test/Reality Gap
https://www.devopsdigest.com/are-ai-coding-tools-hitting-a-ceiling
BlueOptima conducted the BARE (AI Refactoring Evaluation) study, testing 57 models on real-world tasks aimed at improving maintainability (refactoring production code, reducing complexity, and improving structure without changing behavior).

Conclusion: Models are increasingly passing syntactic code checks (> 80%), but they are almost incapable of making changes that actually improve code maintainability. There is also an interesting finding about a plateau: the latest releases of frontier models show almost no improvement on these tasks (stuck in the 17–23% range).

AI is for Prototypes
https://weeraman.com/the-prototype-isnt-the-product/
https://news.ycombinator.com/item?id=49132130
AI is excellent at rapid prototyping, but a prototype ≠ a product. The actual distance from "works on my laptop" to a production-ready system has barely shrunk: architecture, scaling, error handling, security, and maintenance still require the engineering judgment of responsible individuals. Without a deep understanding of the code, an AI-generated project cannot transition from the prototype stage to a stable product.

New model announcements focus precisely on the output code becoming increasingly better and free of bugs—allowing models to generate more code for longer periods while passing all checks. However, does this translate to a better understanding of the actual work of a software engineer?

#agenticcoding #bestpractices

Comments

Google has strengthened its Flash lineup, DeepSeek has officially released V4 Flash with significantly improved agentic capabilities, Meta has entered the code-agent space with Muse, and Alibaba has updated Qwen Max.

Gemini 3.6 Flash and 3.5 Flash Cyber Updates
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/
Gemini 3.5 Flash Cyber is now available as a specialized model for cybersecurity (vulnerability search, validation, and patching), currently restricted to governments and trusted partners. The full Gemini 3.5 Pro is still undergoing closed testing.

Meanwhile, Gemini 3.6 Flash is now the main "workhorse" model. It shows improved code generation (sometimes outperforming Gemini 3.1 Pro) and handles multimodal tasks while consuming 17% fewer output tokens (according to Artificial Analysis). On certain benchmarks, like DeepSWE, token savings reach up to 65%.

DeepSeek V4 Flash 0731
https://api-docs.deepseek.com/updates/#date-2026-07-31
https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
https://arcprize.org/results/deepseek-v4-flash-0731
This is the official release (replacing the preview) of the model from DeepSeek, featuring 284–304B total parameters (13B active) and a 1M context window. It is currently in public beta via API, with open weights (MIT license). Similar to Google's updates, its agentic capabilities have been significantly enhanced; the Flash model now outperforms DeepSeek's own V4-Pro Preview on most agentic benchmarks, despite its smaller size and lower cost.

An update to DeepSeek-V4-Pro is also promised. The model is adapted for Codex and the Responses API, but currently, only deepseek-v4-flash is configurable, with support for deepseek-v4-pro expected soon. The company has announced an upcoming API price increase (as current rates are exceptionally low), but the model itself remains open for download.

Discussion
https://news.ycombinator.com/item?id=49214008
Discussions highlight how the model's low cost combined with high performance opens up new, previously cost-prohibitive use cases. Some users are already running it on rented GPUs (such as RTX 6000, MI300X, etc.). One user running 5–6 active sessions (effectively 12 threads) reports spending less than $5 per day.

DeepSeek-Centric AI Agent
https://reasonix.io/
https://github.com/esengine/deepseek-reasonix
Reasonix is a local agent (a single Go binary with no dependencies, under the MIT license) optimized specifically for DeepSeek with a cache-first loop. The project focuses on enabling Reasonix to handle long autonomous tasks without losing user control. It can be operated via the terminal, desktop application, browser, or directly in the editor via ACP, with all sessions shared.

Reasonix first outlines its work plan and then requests permission at each step for reading, writing, or executing commands. Each "move" is saved as a checkpoint outside of Git, meaning that even after several hours of operation, any change can be rolled back without affecting the commit history. The agent can be easily extended using MCP servers and custom skills.

Muse Spark 1.2 and Code
https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2
https://dev.meta.ai/
Meta is also entering the code assistant space. Muse Spark 1.2 is a model update fine-tuned specifically for coding and utilizing their proprietary Muse Code harness. It features improvements in code generation, debugging, understanding of large repositories, and long-horizon tasks.

Muse Code is currently in beta as a terminal-based code agent (supporting macOS/Linux, without Windows support). It features asynchronous background agents, planning modes (/plan, /grill, /goal), persistent sub-agents, and an event log for safe resumption of long sessions. The model is available via Muse Code and the Meta Model API (which includes a cheaper contributor tier that utilizes data sharing).

https://www.youtube.com/watch?v=c-V4MrY03Mc

The author of https://aicodingdaily.com/leaderboard tested Muse Spark 1.2 on 15 coding prompts (5 prompts × 3 projects: Flutter/Dart, mobile-to-API data synchronization, and CSV import in PHP). This test was conducted via Open Code and OpenRouter rather than Muse Code, measuring speed and real-world cost. The model proved to be quite slow and relatively expensive for the results achieved. The author notes that significantly better options are available at this price point, such as DeepSeek V4 Flash (even after the planned price increase) and Grok 4.5.

Qwen 3.8 Max
https://artificialanalysis.ai/models/qwen3-8-max
Alibaba's new flagship model (2.4T MoE). It initially secured top rankings (1st–2nd place) on the Artificial Analysis Intelligence Index, but dropped to 9th place, below Kimi K3, following a benchmark methodology update. The model tends to be highly verbose (generating many tokens). Open weights are expected to be released soon.

Notably, running such a massive open model locally is impractical for average users, which is why many are currently relying on the smaller Qwen 3.5 and Qwen 3.6 models.

#newllmmodel #deepseek #metamuse #qwen

Comments

A week of minor updates: Grok 4.6, Gemini 3.7 Flash, DeepSeek V4-Pro 0813, GLM-5.3, Qwen 3.8 27B, MAI-Code 1.1 Flash, and Muse Glimmer—all have improved code generation capabilities without changes to their underlying architecture.

Grok 4.6
https://x.ai/news/grok-4-6
On August 12, xAI released a post-training-only update based on the same ≈1.5T base as Grok 4.5. Regarding code generation, it achieved the 1 public result on FrontierCode v1.1 Extended (61.3%), APEX-SWE (56.4%), and saw CursorBench v3.2 rise from 66.7% to 69.9%. The key improvement is that the model now autonomously tests and verifies its code during long agentic sessions, requiring significantly fewer redundant steps (≈53 steps compared to ≈103 for Claude Opus 5).

The weak point remains terminal performance (Terminal-Bench 3.0 score is only 26%). Elon Musk mentioned that Grok 4.7 is just a few weeks away. These updates are steadily pushing the model higher in the Artificial Analysis rankings, bringing it closer to top-tier models.

DeepSeek V4-Pro 0813
https://api-docs.deepseek.com/updates/
On August 13, the Pro tier received the same agentic post-training update previously introduced in V4-Flash 0731. This is a full production release rather than a preview. Operationally, they added support for the OpenAI Responses API, a one-click Codex setup, and reasoning level controls (low/high/max). The weights for the 0813 version have not been published.

Gemini 3.7 Flash
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/
Released by Google on August 13, just three weeks after 3.6 Flash. This update represents the largest leap for the Flash lineup in coding tasks: DeepSWE v1.1 improved from 49.0% to 65.3% (according to benchmarks), and FrontierCode 1.1 Main rose from 34.4% to 43.6%. In Google's comparisons, it outperforms Claude Sonnet 5 and GPT-5.6 Terra. Additionally, it is nearly twice as cheap at launch ($0.75/$3.75 per million tokens until the end of 2026) and features a 1M context window.

GLM-5.3
https://z.ai/blog/glm-5.3
Released by Z.ai on August 14. Built on the same 743B base as GLM-5.2, all improvements stem from post-training (IndexShare + SAO). Coding performance: +50% on the internal Z.ai Code Bench, Terminal-Bench 3.0 at 28.3% (the best score among open-weight models, compared to GLM-5.2's 4.6%), and DeepSWE 1.1 at 66.9%. Reports indicate that its cybersecurity capabilities (CyberGym 84.5%, ExploitBench up from 24.4% to 54.4%) increased so significantly that the weights release was delayed by two weeks for safety reviews. Currently, it is only available via GLM Coding Plan and ZCode.

Qwen 3.8 27B
https://huggingface.co/Qwen/Qwen3.8-27B
Released by Alibaba on August 14, this is a dense multimodal 27B model under the Apache 2.0 license—distilled from the flagship Qwen3.8 Max for local deployment. All benchmarks show a noticeable improvement over Qwen3.6-27B, as expected. It features a native context window of 262K (expandable to 1M via YaRN) and runs on standard high-end GPUs and Ryzen AI Max.

MAI-Code 1.1 Flash
https://microsoft.ai/models/mai-code-1-flash/
https://github.blog/changelog/2026-08-11-mai-code-1-1-flash-available-in-github-copilot/
Microsoft updated its small coding model (an MoE with ≈5B active parameters) on August 11 and launched it in GitHub Copilot. Developed in response to developer feedback, the update delivers a +22% improvement on Terminal-Bench 2.1, +15% on .NET, and adds native vision capabilities (supporting screenshots and diagrams). For routine light tasks, other efficiency metrics are even more notable: it is 25% more token-efficient and four times cheaper (73% lower list price than version 1.0), trained from scratch on a clean enterprise dataset without distillation.

Muse Glimmer
https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
On August 10, Meta released its first open-weights model in 16 months (since Llama 4)—a 30B model under the Apache 2.0 license. This is a distillation of the proprietary Muse Spark, designed for local deployment on a single consumer GPU (4-bit quantization requires ≈17GB VRAM). In coding, it leads its class in agentic tasks but falls short of Qwen3.6-27B (which has already been succeeded by 3.8). Notably, Mark Zuckerberg also promised to eventually open-source the weights for the more powerful Muse Spark 1.2.

#newllmmodel #grok #gemini #deepseek #glm #qwen #microsoft #meta

Comments