GPT-5.4 Mini and Nano Released as AI Solves Open Math Problems and Anthropic Ships Dispatch

OpenAI announced GPT-5.4 mini and nano, its latest small models. Mini leaps past GPT-5 mini across coding, reasoning, multimodal understanding, and tool use at 2x the speed, approaching the full 5.4 on benchmarks like SWE-Bench Pro and OSWorld-Verified.

A trillion-parameter mystery model called Hunter Alpha surfaced on OpenRouter without attribution, fueling speculation that DeepSeek is stealth-testing its next generation. The model appeared with a 1M context window and no identifying information, sparking a wave of guesses across the AI community.

DeepMind released a Cognitive Taxonomy decomposing general intelligence into 10 faculties and launched a $200k Kaggle competition to close gaps in metacognition, attention, and social cognition. The framework offers a structured way to measure how close AI systems are to human-level general intelligence across distinct cognitive abilities.

HorizonMath, a benchmark of over 100 predominantly unsolved math problems, found GPT-5.4 Pro has already solved two open problems, proposing novel solutions that improve on best-known published results. The benchmark spans eight mathematical domains and is designed to test whether frontier models can push the boundaries of human mathematical knowledge.

Harmonic released Aristotle Agent, calling it the world's first autonomous mathematician. The agent is live and free to use, capable of independently exploring mathematical problems and constructing proofs.

The DOE is aiming $293 million via the Genesis Mission at using AI to tackle over 20 national challenges in advanced manufacturing, biotech, critical materials, nuclear energy, and quantum science. The program represents one of the largest federal investments in applying AI to scientific grand challenges.

PerturbAI launched with the world's largest in vivo CRISPR atlas, an 8-million-cell, brain-wide dataset capturing real biological circuitry in living tissue. In vivo means in a living organism rather than in a dish, and this atlas was generated in collaboration with NVIDIA and the Allen Institute.

Xaira launched X-Cell, a virtual cell model trained on 25.6 million perturbed cells, predicting gene expression changes across cell types and conditions. The 4.9-billion-parameter model uses a diffusion language architecture and represents a major step toward simulating cellular behavior computationally.

India-based PopVax argues that its generative AI, mRNA platforms, and 10x cheaper Indian R&D costs can make vaccines against neglected diseases like HCV, TB, and Strep A economically viable to develop for the first time. These diseases kill millions annually but have been commercially unattractive for traditional pharma to address.

Anthropic shipped Dispatch in Claude Cowork, a persistent agent running on your computer that you can message from your phone. The feature blurs the line between desktop and mobile AI, letting users kick off tasks remotely and check results later. Observers note Anthropic is building OpenClaw faster than OpenAI.

Jensen Huang called OpenClaw "definitely the next ChatGPT", speaking from the sidelines of GTC with Jim Cramer on Mad Money.

UCSD's Dreamverse now generates 5-second 1080p clips in 4.55 seconds on a single GPU, faster than playback. The system is built on the FastVideo real-time inference stack and represents a leap toward interactive AI video generation.

Korean researchers announced the Seoul World Model, the first city-scale world simulation grounded in a real metropolis via RAG on millions of street-view images. RAG stands for retrieval-augmented generation, where the AI pulls in real data to ground its outputs. The model draws on over 440,000 Seoul street-view images.

Google is expanding "Personal Intelligence" across Search, Gemini, and Chrome, rolling it out to free-tier users in the US.

Microsoft is moving Mustafa Suleyman to first-party AI model development, cutting its dependence on third-party frontier labs. Suleyman will focus on what Microsoft calls "Superintelligence efforts," building world-class models in-house over the next five years.

Nvidia won Beijing's approval to sell H200 chips to China while prepping a Groq version for the same market. The Groq chip variant is expected to be available in May, opening another revenue channel in the Chinese market.

Microsoft is weighing legal action against Amazon and OpenAI over a $50 billion cloud deal that threatens Microsoft's exclusive partnership with OpenAI.

Rural Ohioans want to amend their state constitution to ban data centers above 25 megawatts. A group called Adams County for Responsible Development gathered roughly 1,800 signatures in eight days, reflecting growing local resistance to the energy and land demands of AI infrastructure.

Jensen Huang told GTC that Nvidia is engineering cooling for orbital data centers where convection does not exist. In his words: "In space, there's no convection, there's just radiation."

Swarmer Inc. surged 700% at IPO, the best US debut since Newsmax, on the back of over 100,000 real-world combat drone missions in Ukraine since April 2024. The company's Trident OS and Styx platform power autonomous drone swarms in active conflict.

The Pentagon is planning for AI to train on classified data. Models like Claude are already used in classified settings, but they currently don't learn from that data. Training on classified material would be a significant new step.

Gambling sites now accept bets on which jobs AI will replace. The Action Network launched a tool showing implied odds of AI replacing 756 different jobs, drawing on Anthropic research data, while Kalshi prediction markets have seen over $1.5 million wagered on AI and economy scenarios.

Companies are tracking individual employee token usage, scouting whose AI strategies deserve amplification. Firms like Zapier are measuring how workers use AI at a granular level to identify and scale the most effective approaches.

USPS is auctioning its last-mile delivery for the first time, pushing Amazon to cut volume by roughly two-thirds. Amazon says the USPS "abruptly walked away" from negotiations and introduced the auction concept.

An appeals court temporarily unblocked Perplexity AI's agentic shopping tool on Amazon. The 9th Circuit granted a temporary stay of the lower court injunction, keeping Perplexity's "Comet" shopping agent live for now.

The SEC issued a landmark token taxonomy carving out stablecoins, digital collectibles, and digital commodities as non-securities. The taxonomy creates five categories, with only "digital securities" remaining subject to securities laws.

All five canonical nucleobases have now been found on asteroid Ryugu, strengthening the case that asteroids delivered the ingredients for life to Earth. Nucleobases are the building blocks of DNA and RNA, and the detection of all five — adenine, guanine, cytosine, thymine, and uracil — in samples returned by Japan's Hayabusa2 mission is a landmark finding for the study of life's origins.