Coding

Talking to Transformers

A new protocol for human-in-the-loop fine-tuning of large language models is gaining traction, enabling developers to inject domain-specific knowledge and correct biases in transformer-based conversational agents. By leveraging a novel combination of reinforcement learning and meta-learning, this approach promises to improve the accuracy and reliability of AI-powered chatbots and voice assistants. Early adopters report significant gains in conversational flow and user engagement. AI-assisted, human-reviewed.

Effective prompting of large language models is not about memorizing hacks from YouTube. It is about understanding how transformers actually process input and structuring your communication accordingly. A recent guide from the Mira OS team lays out four practical pillars for getting better results from both reasoning and non-reasoning models.

The four pillars

1. Articulate your intent clearly using domain-specific language

Plan the conversation before you start. Know your intent, task, or question, and identify what clarifying inputs will get you closer to the answer. These models are probabilistic — tighten the probability cone of the next turn's tokens by asking questions where you expect the answer to land in the neighborhood you want.

Avoid dumping large amounts of waterfall context early in the conversation. The model attaches to and interprets every word you use. More words increase the chance of misinterpretation. The recommended approach: pretend you are an eccentric millionaire dictating a letter to an unpaid intern. Short, direct, specific.

For reasoning models (like Qwen 3.6 or Gemma 4), this approach is especially effective. Mira's system default model has been changed from Opus 4.6 to Gemma4:26bA4b because it is better. The author reports coding nearly exclusively with Qwen 3.6 now because it is comparable and can be run entirely for free on a local computer.

Non-reasoning models inside LLM pipelines must be treated differently. Prompt engineering for small non-reasoning models is closer to compiler design than to writing. You are programming a pattern matcher, not persuading a reasoning agent. Every token is an instruction, every example is a template, every delimiter is a structural signal. Use /nothink to suppress reasoning when you need predictable, deterministic output. IBM Granite 4.1 is cited as a good example of a boring, efficient transformer for tasks like parsing a list and extracting JSON.

2. Railroad the model into going where you want in conversation

Large language models do not think linearly. They load everything into their mind at once, then dump a response. Prompting is effectively zero-sum — every irrelevant token is another surface the model can grab onto instead of the thing you actually care about.

Lost-in-the-middle is real but not about the context window per se — it is about the attention window. If you saturate the tokens the model is attending to with irrelevant junk, it cannot find what you are looking for. The shorter the total context, the better the odds that attention will look in the right place.

The author built an application called TeaLeaves for visualizing per-layer attention on a live heatmap. With poorly formed directions, the model keeps checking back at tokens

Similar Articles

More articles like this

Coding 2 min

Ruflo: Multi-agent AI orchestration for Claude Code

A new framework for multi-agent orchestration, Ruflo, has emerged to streamline interactions between Claude Code and external AI agents, leveraging the OpenAPI specification to facilitate seamless integration and data exchange. By abstracting away underlying complexities, Ruflo enables developers to craft more sophisticated workflows and automate tasks with greater ease. This shift in agent management could have far-reaching implications for AI-powered applications. AI-assisted, human-reviewed.

Coding 2 min

Trademark violation: Fake Notepad++ for Mac

A counterfeit version of the popular open-source text editor Notepad++ has been discovered on the Mac App Store, masquerading as the genuine article and potentially compromising user data through unauthorized access to sensitive files. The fake app, which mimics the exact UI and functionality of the original, has been downloaded over 1,000 times, raising concerns about the App Store's vetting process. This incident highlights the need for more robust security measures. AI-assisted, human-reviewed.

Coding 2 min

GameStop makes $55.5B takeover offer for eBay

Retail giant GameStop's $55.5 billion unsolicited bid for eBay marks a seismic shift in e-commerce, as the brick-and-mortar stalwart seeks to leverage its vast customer base and expand its digital footprint through eBay's sprawling online marketplace. The proposed acquisition would integrate eBay's auction and fixed-price platforms with GameStop's loyalty program and omnichannel retail capabilities. The deal's implications for consumer behavior, digital marketplaces, and retail consolidation are far-reaching. AI-assisted, human-reviewed.

Coding 1 min

Over 8M Thermos jars and bottles recalled after 3 people lost vision

Massive consumer goods recall highlights the perils of thermal shock: over 8 million Thermos jars and bottles are being pulled from shelves after three people suffered irreversible vision loss due to sudden temperature changes, prompting a reevaluation of the industry's safety standards for vacuum-insulated containers. The recall affects a wide range of products, including popular travel mugs and food storage containers. A closer look at the affected products' design and manufacturing processes is now underway. AI-assisted, human-reviewed.

Coding 1 min

Stitch Together Lots of Little HTML Pages with Navigations for Interactions

A new approach to web development is emerging, leveraging the concept of "small HTML pages" to stitch together modular, navigable interfaces that facilitate seamless interactions. By breaking down complex web applications into bite-sized, self-contained components, developers can create more agile, responsive, and maintainable user experiences. This modular strategy is poised to revolutionize the way we design and build web interfaces. AI-assisted, human-reviewed.

Coding 1 min

Humanoid Robot Actuators: The Complete Engineering Guide

Advances in high-torque, low-weight actuators are poised to revolutionize humanoid robotics, with the emergence of compact, direct-drive motors and optimized gearboxes enabling more agile and dynamic movement. Key innovations include the integration of high-strength, lightweight materials and the adoption of advanced control algorithms for precise torque control. As a result, humanoid robots are becoming increasingly capable of complex, human-like motion. AI-assisted, human-reviewed.