Can the loop of AI becoming smarter on its own collapse depending on the 'grader'? ── Prologue to the personal research project 'RSI Research'

About this articleI am Makaji, an independent researcher. This article is the 0th installment (prologue) of the series on my ...

Direct Manipulation of Neural Networks via Weight-Space Programming: The Dawn and Ripples of Self-Reorganizing AI Demonstrated by GPT-6 Astra

Box ManipulationIn the history of artificial intelligence and machine learning, the determination of internal parameters ...
Tech Times on MSN

Tsinghua study: AI coding agents cut costs 26% by rewriting their own scaffolding

Tsinghua's HarnessPro cuts AI coding agent costs 26% while boosting accuracy without model retraining. Factory Droid 1.0 ...
cloud

Semantic aggregation over millions of rows in BigQuery with AI.AGG

System logs generated by globally distributed microservices often hide subtle degradations (like memory thrashing or systemic ...
shattered.io

OpenCode MCP Servers: 12 Steps, 70 Min [2026]

OpenCode MCP setup for v2.2.6: build a read-only Python server, lock tools down, pool connections and run headless in CI in ...
eu.36kr

MIT allowed GPT-6 Astra to directly modify Qwen's weights, and its math scores have soared.

MIT has unveiled an innovative, groundbreaking paradigm dubbed "Weight Space Programming", representing a transformative new ...
finance.biggo

Anthropic Cuts Fable 5.2 Token Prices to $8 as It Pushes Toward a Pre-Thanksgiving IPO

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, its most capable models, claiming output up to 30% faster than prior versions plus gains ...
BIOENGINEER

Machine Learning and Behavioral Models Converge to Reshape Urban Travel Forecasting

A systematic review of 133 studies published from 1998 to 2026 examines how urban travel mode choice modelling evolved from ...
XDA Developers on MSN

I stopped giving ChatGPT, Claude, and Gemini detailed prompts for a week, and one struggled more than I expected

Do we really need to spoon-feed AI anymore?
news.lavx.hu

LLMs read and edit neural weights without gradients

Astra identifies game policies from raw weight matrices and fixes reasoning loops in Qwen3.5-0.8B by editing 0.016% of parameters, lifting MATH-500 scores from 7.9% to 48.1% without training loops.
GitHub

README.md

Paper: Skill-based Model-based Reinforcement Learning, especially Section 3.3, equations 2–5. This is an educational adaptation of offline joint pretraining, not a complete SkiMo reproduction. You ...

'Like gene editing for neural networks': MIT researchers used GPT-6 Astra to edit Qwen's open weights directly, boosting math performance

In the strongest example of "weight space programming", as the technique is called, Astra’s intervention raised Qwen3.5-0.8B’s MATH-500 score from 7.9% to 48.1% while touching approximately 0.016% of ...