OpenAI just launched a dedicated Superalignment team co-led by Ilya Sutskever and Jan Leike to solve the challenge of controlling future superintelligent AI. The company is dedicating 20% of its secured compute over the next four years to build a human-level automated alignment researcher. OpenAI warns superintelligence could arrive this decade, making existing human oversight methods obsolete.
Insilico Medicine just advanced an idiopathic pulmonary fibrosis drug candidate discovered using generative AI into Phase 2 clinical trials. Powered by its Pharma.AI platform and NVIDIA GPUs, the company reached Phase 1 in two and a half years at one-tenth the cost and one-third the time of traditional methods. Why it matters: It is real-world clinical proof that generative AI can design novel small molecules from scratch for human trials.
A new EPFL study found that between 33% and 46% of crowd workers on Amazon Mechanical Turk are using tools like ChatGPT to complete data annotation tasks. Training AI models on AI-generated responses risks compounding hallucinations and statistical errors over time. Basically, companies paying humans to clean up training data are getting synthetic text right back.
University of Kansas researchers just trained a machine learning classifier that accurately spots ChatGPT-generated science papers over 99% of the time, dropping to 92% at the paragraph level. The tool analyzes stylistic differences like vocabulary richness and punctuation habits that AI often misses. Why it matters: Academic publishers are desperate to catch AI slop, though small sample sizes and human light edits could still slip through.
NVIDIA Research just revealed Neuralangelo, an AI model that turns standard 2D video clips into high-fidelity 3D assets. Borrowing instant neural graphics primitives from Instant NeRF, it reconstructs tricky textures like roof tiles and marble with remarkable detail. The breakthrough makes building virtual environments far easier for game developers and robotics teams, and the code is already live on GitHub.
Researchers just used AI to discover abaucin, a powerful experimental antibiotic that kills a deadly hospital superbug. The model screened 6,680 untested compounds in an hour and a half, pinpointing a drug that targets Acinetobacter baumannii with minimal risk of resistance. Why it matters: AI is slashing candidate screening from years to hours, though clinical trials mean human use could take until 2030.
Researchers just introduced DragGAN, an interactive tool that allows users to manipulate images by simply clicking and dragging points across the frame. Instead of smearing existing pixels like Photoshop's warp tool, the AI re-generates the object in seconds, smoothly rotating 3D angles and hallucinating occluded features like teeth. Why it matters: It swaps out tedious prompt guessing for precise, flexible control.
UT Austin researchers just developed an AI language decoder that translates human thoughts into text using fMRI brain scans and GPT-1. Instead of requiring brain surgery, the system tracks blood flow patterns to reconstruct the gist of what someone hears or imagines. It offers real hope for paralyzed patients, though scientists warn strict mental privacy safeguards will be needed.
OpenAI researchers just proposed "consistency models," an experimental technique designed to generate AI images in just one or two steps instead of the hundreds needed by diffusion. While the early output quality isn't mind-blowing, it runs an order of magnitude faster and handles tasks like inpainting and upscaling. Why it matters: If image generators are ever going to run locally on phones without draining batteries, ultra-efficient architectures like this will be the key.
Researchers from Carnegie Mellon, Yale, and UC Berkeley found that OpenAI's GPT-4 outperformed skilled crowdworkers in data labeling tasks. Instead of spending $500,000 and 20,000 hours of human labor on 572,322 text scenarios, the team automated the process with GPT-4 for under $5,000. Why it matters: The results suggest AI can replace expensive human crowdwork at a fraction of the cost, threatening traditional data labeling services.
Stanford researchers just introduced Generative Agents, computer programs powered by ChatGPT that simulate believable human behavior through memory, reflection, and planning. In a two-day sandbox simulation akin to The Sims, 25 agents autonomously planned parties, ran for mayor, and were rated as more human-like than actual humans. Why it matters: It points to a future of emergent behaviors in gaming NPCs and virtual worlds, while raising new questions about parasocial attachment.
Apple researchers just introduced FaceLit, a framework that generates 3D faces directly from unannotated in-the-wild 2D photos. By integrating the Phong reflectance model into neural volume rendering, users can control lighting and angles with photorealistic consistency. It opens the door to creating studio-quality 3D digital avatars without specialized capture rigs.
Researchers led by John Hopkins University just laid out a roadmap for 'organoid intelligence'—a biocomputing paradigm combining 3D human brain cell cultures with brain-machine interfaces. By tapping into the brain's 100 billion interconnected neurons, scientists aim to bypass the physical limits of silicon chips, though practical deployment remains in its early stages.
Researchers from Google Brain, UC Berkeley, and Princeton just revealed that diffusion models like Stable Diffusion and Imagen memorize and spit out exact images from their training sets. While the duplication rate was only 0.03% across 300,000 prompts, larger models are even more likely to replicate originals. Why it matters: it breaks the claim that AI only creates novel work, creating massive copyright and privacy risks.
