Anthropic's FlashProve verifies AI outputs without revealing model weights

GPT-6 Astra edits other AI models’ weights directly in Weight-Space Programming

Researchers demonstrated Weight-Space Programming, in which GPT-6 Astra directly edits the numerical weights of other AI models instead of relying on step-by-step optimization or retraining. Astra identified which of four games a network was built to play and matched training languages of small models without running them. It generated weights for simple games from rules alone, averaging 517 points in Breakout versus 25 for a trained baseline, and wrote a 9x9 Go network that beat a comparably sized model trained on roughly 20 million games. On Qwen3.5-0.8B, changing just 0.016 percent of weights lifted MATH-500 accuracy from 7.9 to 48.1 percent. Jürgen Schmidhuber called the approach old hat; researchers point to security audits and targeted behavior changes.

GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights →

Topics: GPT-6 Astra

Anthropic’s FlashProve proves which model wrote a response without exposing weights

Anthropic released FlashProve, an open-source research prototype using zkSNARK cryptography to verify whether an AI response came from the claimed model without revealing its weights. The provider fingerprints the model’s weights and delivers a compact proof with each response. FlashProve generates proofs at roughly 110 tokens per second on a single H100 GPU for Meta’s Llama 4 Maverick, a 400-billion-parameter model; each proof is about 19 megabytes and verifies in 17 seconds on a standard CPU, adding roughly 70x to 95x overhead. The only comparable open-source system previously took over 14 hours for 1,000 tokens and produced a 93.6-gigabyte proof. Limitations include a 1,024-token context cap and proofs only when the model picks the single most likely next token.

Anthropic’s new prototype proves which model wrote an AI response without exposing its weights →

Topics: Anthropic

Google unveils Genie 4 for real-time interactive 3D world generation

Google has unveiled Genie 4, an AI system that generates detailed, interactive 3D environments in real time from user input, incorporating physics and soundscapes. Google DeepMind’s system is trained on extensive video data, including from YouTube, to learn how objects move and behave, simulating physics without traditional engines and generating audio aligned with the visuals. Over the past 2.5 years it evolved from basic 2D slideshows to photorealistic interactive worlds. Applications cited include gaming, education through historical reconstructions and biological simulations, filmmaking and robotics. Challenges include high computational costs, limited accessibility mainly through expensive enterprise and researcher subscriptions, and maintaining scene detail over longer simulations. Genie 4 draws on Gemini models, YouTube datasets, proprietary AI chips and robotics research including Waymo.

Google Unveils Genie 4 to Build Real-Time 3D Environments →

Topics: Google DeepMindGoogle

Stanford AI tool finds natural molecule BRP that curbs appetite in animal tests

Stanford Medicine scientists identified a naturally occurring molecule, BRP, that suppressed appetite and promoted fat loss in animal experiments in ways resembling semaglutide, the active ingredient in Ozempic, while potentially avoiding some side effects. BRP appears to act selectively on the hypothalamus rather than on receptors spread across the brain, gut and pancreas. The team found it using an algorithm called Peptide Predictor, which analyzed roughly 20,000 human protein-encoding genes and narrowed the field to 373 prohormones. In lean mice and minipigs injected before feeding, treated animals consumed up to 50 percent less food in the following hour; obese mice given daily injections for 14 days lost an average of 3 grams while controls gained about 3 grams, mostly body fat. The study was published in Nature.

AI helps Stanford scientists discover a “natural Ozempic” with a surprising advantage →

Revolut cuts AI coding costs up to 90% by replacing Claude with self-hosted Z.ai models

Revolut has replaced Anthropic’s Claude with self-hosted, open-weight models from China’s Z.ai, cutting inference costs by as much as 90 percent in under a year. The London-headquartered digital bank had previously tried Google’s Gemini before adopting Claude, but some heavy users spent thousands of dollars per month on tokens without equivalent returns. Revolut Infrastructure Lead David Abitbol said the self-hosted option produced a 70 percent improvement in lead time productivity and a 60 percent improvement in coding time productivity, citing Z.ai’s GLM models for data privacy and auditability. GLM-5.1 brought inference costs to around $0.30 per million input/output tokens, roughly one-tenth of comparable commercial offerings. Revolut now runs its own inference infrastructure, reportedly with 56 Nvidia B200 GPUs.

‘We were highly skeptical about the cost of AI’: Revolut reveals it cut AI coding costs by cutting out Claude entirely →

Topics: GLM