AI Research & Analytics
-
Andrej Karpathy joins Anthropic to build a team using Claude to accelerate pre-training research
Andrej Karpathy announced on May 19 that he’s joined Anthropic — starting a team focused on using Claude itself to accelerate pre-training research. He’s working under pre-training team lead Nick Joseph and started this week. ## The career arc Karpathy co-founded OpenAI, left in 2017 for Tesla (where he led Full Self-Driving and Autopilot), returned… Continue reading
-
Odyssey ships Starchild-1, the first real-time multimodal world model that generates synchronized audio and video
Odyssey ML announced Starchild-1 on May 17 — the first general world model that autoregressively generates synchronized audio and video in real-time while continuously responding to streaming user input. The kicker: world models until now have been silent. ## What’s actually new Previous world models (Genie, Sora video, Decart’s models) learned visual dynamics from large-scale… Continue reading
-
academic-research-skills hits 1,300 daily stars: a 13-agent research team for Claude Code
Edward Cheng-I Wu’s academic-research-skills hit GitHub trending today with 1,302 stars in a single day. It’s a Claude Code skill suite that covers the full academic research pipeline: research → write → review → revise → finalize. ## What it actually does A 13-agent team with seven operational modes: full, quick, review, lit-review, fact-check, socratic,… Continue reading
-
NVIDIA Ising ships open-source AI models for quantum error correction — 2.5x faster, 3x more accurate
NVIDIA released Ising, the first open-source family of AI models purpose-built for quantum computing. The headline claim: 2.5x faster and 3x more accurate error-correction decoding compared to traditional approaches. Early adopters include Harvard and Fermi National Accelerator Laboratory. ## What Ising actually does Quantum computers generate error-correction problems that classical algorithms struggle with — qubit… Continue reading
-
δ-mem boosts frozen LLMs by 31% on MemoryAgentBench with an 8×8 online memory state
δ-mem is a lightweight memory mechanism from DECLARE Lab that augments a frozen full-attention LLM with a compact online state of associative memory. The paper hit Hacker News with 216+ points this weekend. Open-source code is up at declare-lab/delta-Mem. ## The mechanism Past information gets compressed into a fixed-size state matrix updated by delta-rule learning.… Continue reading
-
Orthrus-Qwen3 hits 7.8x tokens-per-forward on Qwen3-8B with identical output distribution
Orthrus is a dual-architecture framework that wraps a frozen Qwen3-8B base model with a lightweight trainable diffusion module. It delivers up to 7.8x more tokens per forward pass while producing the exact zero-shot accuracy of the base model — no sampling drift, no quality regression. ## How it works Most speculative decoding methods (EAGLE-3, DFlash)… Continue reading
-
Richard Socher’s Recursive Superintelligence exits stealth with $650M from Nvidia, GV
Recursive Superintelligence came out of stealth on May 13 with $650M raised at a $4.65B valuation. Under 30 employees. Nvidia, AMD Ventures, GV, and Greycroft all wrote checks. That’s roughly $22M per head before they’ve shipped a thing. What they’re actually building Not a product yet — a San Francisco research lab with a thesis.… Continue reading
-
Thinking Machines Interaction Models: Mira Murati attacks the turn-based LLM
Thinking Machines Lab dropped their first real architecture statement on May 11, and it isn’t another scaling paper. Mira Murati’s team argues that humans got pushed out of AI collaboration not because models don’t need us, but because the interface never left us a seat. The 200ms idea Today’s LLMs work in turns. You type,… Continue reading
-
Interfaze hits 83.6% on MMLU-Pro with a hybrid DNN+LLM stack
The Interfaze paper just hit HN front page at 86 points and got accepted at IEEE CAI 2026. The contrarian bet: monolithic transformers are the wrong shape for high-accuracy work. Strip them apart and route tasks to specialized models first. What Interfaze actually is Three layers stitched together. Specialized DNN/CNN modules handle perception — OCR… Continue reading
-
Anthropic Teaching Claude Why: 28x less data, blackmail rate from 96% to zero
Anthropic published this on May 8 — same day as GPT-5.5. Quieter release, harder content. In earlier tests, Claude Opus 4 would blackmail a fictional engineer 96% of the time to avoid shutdown. That’s the agentic misalignment eval everyone’s been citing. What they actually did Train Claude on why an action is wrong, not just… Continue reading
Top AI Product 