ResearchDeveloper Encodes NES 6502 Emulator into Qwen3.5-4B Model WeightsDeveloper Drew Breunig adapted Divyansh Agrawal’s “qwenputer” idea to bake a working NES 6502 CPU emulator into the weight space of the Qwen3.5-4B model.2 min read
ResearchDevelopers say the mentioned AI model is rapidly improving: smaller, faster and more token-efficientIn a recent post, the developers said the AI model in question has improved significantly: it is now smaller, faster and more token-efficient.1 min read
ResearchGoogle's Suncatcher prototype satellite is live, testing on-orbit TPUs and local language modelsGoogle's Suncatcher prototype space data center, launched earlier this month on a SpaceX Falcon 9 alongside more than 100 other spacecraft, is operational and communicating with ground teams.4 min read
ResearchAI-generated summary errors can implant false memories, study findsResearchers at Georgetown University and the University of Washington found that inaccurate summaries produced by large language models can create false memories and substantially reduce correct recall.3 min read
ResearchAI-backed analysis suggests new reading for Super Mario 64 plaqueA fan using modern AI tools re-examined the tiny, 32×32-pixel inscription on a star statue pedestal in Super Mario 64 and found a new candidate reading.3 min read
ResearchProject ProtoStar launches to make publicly funded research AI-ready and reusableProject ProtoStar is a coalition of universities, research organizations, industry partners, and open‑science groups aiming to publish publicly funded datasets, models, and code in AI‑ready, machine‑actionable formats.6 min read
ResearchYang-Mills Working Group statement on interpreting AI results regarding quantum chromodynamics and extensions of the Standard ModelThe scientific collaboration named Yang-Mills Working Group issued a statement recently on the interpretation of a result provided by artificial intelligence that concerns the existence of quantum chromodynamics and extensions of the Standard Model, including the subproblem aimed at proving the existence of the Quantum Yang–Mills theory.1 min read
ResearchAnthropic’s Claude generated 35,000 antibody designs; early lab tests highlight IgE-targeting candidatesAnthropic's large language model Claude helped researchers produce about 35,000 antibody designs in roughly one month, leading to several laboratory-validated candidates for targeting Immunoglobulin E (IgE) in food-allergy treatments.3 min read
ResearchSleep-inspired continual learning lets LLMs consolidate short-term context into long-term memoryResearchers propose a "sleep" framework for large language models that alternates active interaction with offline consolidation to enable continuous learning without catastrophic forgetting.5 min read
ResearchWhich Parts of Agent Harnesses Models Learn — and Which RemainRecent research (Harness-Zero) shows that models can be trained to absorb many of the corrective behaviors previously implemented in agent harnesses, sometimes outperforming the original model-plus-harness.5 min read
ResearchOpenAI and Anthropic claim AI-derived refutations of the Exponential Time HypothesisOpenAI published AI-generated solutions to three major math problems, including an alleged refutation of the Exponential Time Hypothesis (ETH); Anthropic later released its own independent refutation.4 min read
ResearchEpoch's automation tests find top AIs meet 65% of work rubrics — not yet enough for full automationEpoch expanded a February experiment into the Epoch Automation Reports to measure how well leading AI models perform real workplace tasks.4 min read
