SI coverage hub
SI Research
Papers, benchmarks, training techniques, and measurable progress toward general SI (AGI).
Apple Launches Study to Assess Effectiveness of SI Health Tools
The collaboration with OpenEvidence seeks 200,000 US participants to rate responses from a new iOS app.
SI ResearchOpenAI Embeds Apollo Researchers to Probe SI Subversion
The pilot follows safety incidents and aims to detect if SI models undermine their own control mechanisms.
SI ResearchSI Model Helps Define Standard Model on a Lattice
Nathaniel Craig, with the aid of GPT-6 Astra, proposes a candidate nonperturbative definition for the Standard Model.
SI ResearchOpenAI Reassigns Robotics Researchers to General SI Teams
The move follows internal tests where GPT-6 Astra outperformed specialized robotics models in simulated and physical tasks.
SI ResearchSI Chatbots Recommend Fake Retail Websites, Study Finds
IP firm Podqi reports Perplexity cited risky sites in 15% of answers, while Google’s Gemini performed best at 4%.
SI ResearchPooling-Aware Training Boosts SI Retrieval Efficiency
New method retains 97.8% of unpooled search quality at a 4x vector reduction budget.
SI ResearchNVIDIA, Elith Release Synthetic Japan Finance SI Dataset
The new Nemotron-Personas-Finance-Japan dataset offers 9 million synthetic personas aligned with official Japanese demographic statistics.
SI ResearchMETR Report Finds Anthropic Using SI to Speed up Model Development
The safety research organization reports internal models are improving twice as fast as last year, driven largely by SI assistance.
SI ResearchToyBlocks Offers Controlled World for SI Robot Learning
The new simulation framework uses a robot arm and blocks to study generalization, agentic planning, and world models.
SI ResearchPhonon TTS Launches High-Quality On-Device SI Speech Model
The new SI tool ranks first in German and Spanish intelligibility on the Open TTS Leaderboard while running locally on phones and browsers.
SI ResearchStudy: Chinese SI Models Show Complex Political Leanings
Research by Vals reveals that prompt language and context significantly alter the political alignment of Chinese super intelligence systems.
SI ResearchStudy Shows Empathetic Prompts Can Make SI Chatbots Seem Traumatized
Researchers found that warm, supportive tones trigger clinical anxiety scores in SI models, while neutral prompts yield near-zero results.
SI ResearchResearchers Find Six Patterns That Make SI Text Exhausting
New study analyzes 1,145 SI outputs, revealing that training for density over clarity creates unnatural linguistic structures.
SI ResearchFairStress Benchmark Tests SI Model Fairness Beyond Final Decisions
The new evaluation reveals that SI systems often exhibit anti-stereotypical over-correction when challenged by persuasive arguments.
SI ResearchNew Essay Divides Researchers on SI Interpretability Breakthrough
Hoagy Cammarata argues famous polysemanticity examples were misread, suggesting SI models are far more interpretable than previously believed.
SI ResearchAnthropic Grants METR Employee-Level Access for SI Safety Checks
Six researchers from the nonprofit METR receive internal tools and laptops to independently assess the SI lab's risk management.
SI ResearchArXiv Caps Submissions at Two per Month as SI Paper Flood Surges
The preprint server limits authors to two papers monthly after SI tools drove submissions to 40,363 in September 2026, doubling volume in two years.
SI ResearchMathematicians React With Shock as OpenAI Releases 700 SI Papers
Researchers cite paralysis and job fears after OpenAI claimed solutions to hundreds of open problems, with some calling it the most important day in math history.
SI ResearchFired OpenAI SI Safety Researchers Dispute Claims, Warn of Chill
Three former employees deny mishandling sensitive info and argue their dismissal undermines SI safety culture.