Japanese Startup AIdeaLab Releases Anime-Focused Video Generation AI "AnimeGen" Free of Charge
On July 13, AI development company AIdeaLab released "AnimeGen", a video generation AI model specialized in animation. It is licensed under Apache-2.0 and can be used commercially.
It is worth noting that a company with such close ties to Japanese anime culture has developed and open-sourced a specialized model. The approach targets anime-specific art styles and direction that general-purpose models cannot fully express, and its open-source nature makes it approachable for creators. As for the actual generation quality and constraints, we will have to wait for community evaluation.
MiniMax Stock Plunges 80% From Its Peak — JPMorgan Cuts Price Target Yet Again
Shares of Chinese AI model maker MiniMax Group are plunging. The direct trigger was JPMorgan Chase cutting its price target for the second time in less than a week. As of July 13, the stock had fallen as much as 18%, marking new lows for a third consecutive session, and it is now down more than 80% from its March peak.
The decline is reportedly driven by dilution concerns over additional fundraising. It appears to be part of a broader repricing of Chinese AI companies' valuations, and spillover to other China-related AI stocks is worth watching.
GPT-Live's "Dejiru" and "Suyaki" Mishaps — The Strange Slips of a Real-Time Conversational AI
OpenAI's real-time voice AI "GPT-Live" has been praised as "remarkably human-like," but its bizarre verbal slips have become a talking point online.
Reported examples include calling dashi (Japanese soup stock) "dejiru" and describing toast with nothing spread on it as "suyaki" (literally, "baked plain"). While its reasoning capability as an AI is highly regarded, it can make conspicuously unnatural statements in everyday contexts. This new real-time voice interface is highlighting challenges that differ from text-based ones.
Adobe's Stance: "Not a World Where AI Decides Everything"
Adobe has reiterated its position: "We are not aiming for a world where AI decides everything." With generative AI now widely used in production work—centered on the advertising field—and starting to displace some creative and design jobs, the company is posing the question of how to design a collaborative relationship between humans and AI.
Its new initiatives, which explore what "work that remains for humans" looks like in creative fields, offer an important perspective for thinking about work in the AI era. As generative AI continues to spread, how to position the human role is becoming a shared challenge across many industries.
Australian Radio Chart-Topper Under Suspicion of Being AI-Generated
Suspicion has emerged that the song that hit No. 1 on an Australian radio chart may have been generated by AI. Reported by the Guardian, the case suggests that AI-generated music has reached a quality level where it is difficult to distinguish from human work.
How AI-generated content should fit into the music industry involves many unresolved issues—copyright, crediting, and transparency among them—and similar cases are likely to increase going forward.
Flash-MSA: A Sparse Attention Kernel That Accelerates 1M-Token Training (Research Highlight)
Flash-MSA, which is drawing attention in the LocalLLaMA community, is a technique that accelerates training at the 1-million-token scale using a sparse attention kernel. It takes an approach that optimizes GPU compute efficiency and could help reduce the training costs of long-context LLMs.
In a similar research direction, "KV-PRM", published on arXiv, is also attracting attention. It reportedly slashes the inference cost of process reward models (PRMs) via KV cache transfer, achieving FLOPs reductions of up to 5,000x. As an achievement that boosts the practicality of test-time scaling methods, further developments are expected.
---