GeekNews AI Roundup: 7 Highlights from Datasheet MCPs to Mathematics 2.0

The boundaries of AI tools and research are expanding rapidly — from precise querying of manufacturing datasheets to the evolution of mathematical proof paradigms. A recent batch of stories aggregated by GeekNews paints a pragmatic and clear-eyed picture of AI development. This article distills seven noteworthy items and examines the common thread running through them: AI is shifting from "impressive demos" toward "verifiable, deployable" applications.

Manufacturing and Open-Source Tools: Precision as the New Selling Point

PartsPlan is a highly industry-specific MCP (Model Context Protocol) server. It enables AI to look up component parameters from tables, charts, and drawings in manufacturer datasheets — and always cites the source document and page number. The key design choice: it will never speculate on values that don't exist in the datasheet. In a small-scale test shared by the developer, all 30 chart values were answered correctly — though as the documentation notes, this result cannot be generalized to overall accuracy across all components. This conservative "refuse to guess" approach is precisely the necessary threshold for AI tools entering professional engineering domains.

Hanok-LLM is an open-source LLM training project for the Korean language, covering the full pipeline from dataset preparation to deployment on Ollama and Hugging Face. Worth noting: the GeekNews detail page for this project was inaccessible at the time of writing, so this summary is based solely on the public RSS title and excerpt — training performance and implementation details cannot be verified. For developers looking to understand the end-to-end training workflow for low-resource language models, it still offers a useful reference path.

---

MCP (Model Context Protocol) is an open protocol standard led by Anthropic, designed to establish a structured interaction interface between AI models and external tools or data sources. Think of it as the "USB port" of the AI world — any tool that implements the MCP specification can be plug-and-play invoked by compatible AI clients (such as Claude Desktop). An MCP server is the execution endpoint of the protocol: it receives query requests from the model, accesses a specific data source, and returns structured results. As an MCP server, PartsPlan means engineers can query datasheets directly within an MCP-enabled AI chat environment — no need to manually upload PDFs or switch between multiple tools. This architecture shifts the responsibility for "precision" from the model itself to a dedicated server, thereby bypassing the hallucination problem that general-purpose LLMs are prone to in factual lookups.

Local-First Developer Tools: Privacy-Driven Pragmatism

BetterWispr is a free, open-source dictation app for macOS. Users can press a keyboard shortcut in any text input field to speak and have their words transcribed. It supports local models such as Whisper and Parakeet for offline transcription, with language support depending on the chosen model and hardware — and it will not silently fall back to cloud transcription. This "local by default, no sneaky cloud uploads" design directly addresses developers' strong demand for privacy and control, and represents the core advantage of open-source voice tools over commercial alternatives.

Vibe Coding: Capability and Risk

"Vibe Coding" combined with AI-assisted decompilation is enabling rapid browser ports of classic games. According to the original report, browser-playable versions of Halo, The Simpsons: Hit & Run, GTA, and other classics have appeared — some with a surprisingly high level of completeness in actual gameplay.

The real change isn't browser memory management — it's that AI has dramatically compressed the time required for manual code refactoring. But the flip side is that rapid re-creation makes it far harder for rights holders to respond with takedowns. The leap in technical efficiency is simultaneously amplifying the difficulty of copyright governance — an unavoidable gray area as AI-assisted development becomes mainstream.

---

"Vibe Coding" is a term coined by OpenAI co-founder Andrej Karpathy in early 2025 to describe a programming paradigm that relies heavily on AI code generation, where developers describe intent in natural language rather than writing code line by line. Developers simply describe "what kind of functionality I want," and AI handles the implementation details — even error messages get pasted directly back to the AI rather than being debugged manually. Decompilation, meanwhile, is the process of reverse-engineering a compiled executable back into human-readable source code. Combining the two means that even without the original source code, decompilation can provide a code base for AI to rewrite, which AI can then rapidly adapt to a new platform (such as a browser-based WebAssembly environment). This combination compresses what once took senior engineers months into an extremely short cycle — but in doing so, it lowers the barrier to copyright infringement to an unprecedented low.

A Sober Research Perspective: What AI Still Can't Do

Amid widespread optimism, several research-oriented items offer a necessary reality check.

AI Still Struggles with Algorithmic Innovation

Epoch AI's InnovationEval evaluates whether models can independently produce new algorithms — from ideation through implementation to experimentation — that are comparable to human-developed techniques that have stood the test of follow-on research. Using the improvement of GRPO over SDPO as a calibration baseline, Claude Fable 5 scored only 2%, and GPT-5.6 Sol reached just 15%. The original paper also notes that "selection of best-run results" and similarity to existing research both constrain how results should be interpreted. In other words, current models still have a significant gap to close before achieving genuinely original algorithmic breakthroughs.

---

GRPO (Group Relative Policy Optimization) and SDPO are both policy optimization algorithms used in reinforcement learning to train large language models — specific methods within the RLHF (Reinforcement Learning from Human Feedback) technical family. GRPO was proposed by the DeepSeek^Ad team as an improvement over earlier algorithms like PPO, offering gains in computational efficiency and training stability, and was used in training models such as DeepSeek-R1. InnovationEval uses "the improvement of GRPO over SDPO" as a calibration baseline to set a quantifiable reference for "what counts as a meaningful algorithmic innovation" — the new algorithm proposed by the model must achieve a substantively similar magnitude of improvement to be counted as a success. By this standard, both Claude and GPT-5 series scores fall far below expectations, indicating that current models still have a fundamental gap compared to top human researchers when it comes to "proposing genuinely novel and effective optimization ideas."

The Trap of Single-Benchmark Numbers

An article titled Sorry, You Still Have to Think for Yourself cuts to the heart of the matter: even if an AI agent can rewrite code, the trade-offs between compatibility, latency, throughput, and memory limits still require human judgment. A Campfire Once comparison of rewrites in Rust and Elixir showed that the two implementations had inconsistent constraints and load conditions; when the Rust event buffer was increased, delivery rate improved — but maximum latency also shot up significantly. This case vividly illustrates the limits of single-benchmark numbers: performance has never been something a single figure can capture.

Mathematics 2.0: From Proof Scarcity to Proof Abundance

Fields Medal laureate Terence Tao has proposed an imaginative thesis: thanks to AI and mechanized verification, mathematics may be moving from an era of "proof scarcity" to one of "proof abundance." At the same time, he remains clear-eyed — many problems at the frontier remain unsolved. In his vision of "Math 2.0," the goal should not be to optimize for sheer quantity of proofs, but to broaden the range of meaningful research directions. This perspective echoes the research findings above: AI is a tool that amplifies human research capacity, not a black box that replaces human judgment.

---

"Mechanized verification" refers to using formal proof assistants such as Lean, Coq, or Isabelle to translate mathematical proofs into a formal language that computers can automatically check — thereby eliminating logical gaps that human reviewers might miss. Traditional mathematical proof publication and verification relies heavily on peer review, which is slow and prone to accumulating hidden errors. In recent years, the growing adoption of tools like Lean 4 in the mathematics community, along with advances in AI models (such as AlphaProof) on automated formal proofs, is changing this landscape. Terence Tao himself has actively promoted integrating formal verification into everyday mathematical research workflows, publicly stating that AI-assisted formal verification allows him to handle complex, detail-heavy proofs with greater confidence. A "proof-abundant" era means that more intermediate results and local lemmas can be rapidly produced and reliably verified — building a denser knowledge network for mathematical research rather than relying solely on a handful of landmark breakthroughs.

Conclusion: Pragmatism and Clarity in Tandem

Looking across these seven items, a clear throughline emerges: the AI field is undergoing a shift from "can it be done?" to "can it be trusted? can it be controlled?" PartsPlan's no-speculation principle, BetterWispr's local-first approach, InnovationEval's rigorous calibration, and Tao's emphasis on research direction — all point toward the same consensus: truly valuable AI applications are built on verifiability and human judgment. For developers, this is both an opportunity and a reminder: the more powerful the tools become, the more important it is that we preserve the capacity for independent thinking.