Everyone in AI Is Making Moves Right Now! [AI ROUNDUP]

AI progress is accelerating across text, images, audio, and—most notably—video, with new models pushing speed, realism, and open-source accessibility. Gemini 3.1 Flash Light is positioned as a fast-turnaround option that can generate 2,000 tokens in about five seconds, enabling rapid, in-browser website creation. The demo shows how quickly the model can “regenerate” a simple page from scratch, including interactive edits like generating donation tiers and updating sections on demand. Rumors of a Gemini 3.2 Flash suggest the same direction: faster iteration at lower cost, even if quality remains below top-tier “pro” models.

On the video front, the roundup highlights two contrasting realities: major commercial systems are still constrained, while open-source alternatives are improving fast. Seedance 2.0 (cited as a standout upgrade over OpenAI’s Sora) remains unavailable in the United States due to access restrictions from Caput and Drama, and it launched with heavy censorship—most prominently a ban on realistic faces. Workarounds are circulating, including prompting for hyperrealistic outputs after sketching, but the overall takeaway is that guardrails are tightening creative fidelity. Meanwhile, a brand-new open-source model—described as a single-stream 15 billion transformer that jointly generates audio and video—claims “free” 5-second 1080p clips in about 38 seconds on a single H100 GPU. Quality is described as strong, with realistic faces and fine details, and the model appears geared toward narrative, head-focused scenes and establishing shots rather than extreme body-motion choreography.

The open-source video ecosystem’s current benchmark is LTX 2.3, which can run on consumer hardware; the new model is said to be harder to run at scale, though community members are already discussing distillation to shrink it for less expensive GPUs. A practical test using an input image (a Best Buy worker confronting a “Karen,” with the worker lacking arms) illustrates both promise and limitations: generation can be slow, but outputs keep character movement coherent and avoid some classic failure modes like “growing hands,” even if other issues—such as character acting out another character’s dialogue—still appear.

Image generation also keeps moving toward compositional control and likeness accuracy. Lumalabs’ Uni1 is presented as an “omni” model that can separate layers from a complex composition into backgroundless outputs, effectively extracting multiple elements as distinct images. The workflow angle matters as much as raw generation speed: the layer separation may rely on internal background-removal steps that compress layers into the final result. Photo Labs’ new likeness-focused model emphasizes photorealism and style reference matching, including pet likenesses, but it demands 30 to 50 reference photos per subject to achieve accurate results.

Large language model progress is framed less as a single leap and more as a steady cadence of incremental upgrades and efficiency research. Anthropic’s rumored “Claude Mythos” is described as a very large, researcher-only model with claimed gains in coding, academic reasoning, and cyber security—paired with concerns about misuse risk. Google’s Turbo Quant compression algorithm targets LLM efficiency by reducing key-value cache memory by at least six times and boosting speed up to eight times without accuracy loss, with implementation described as relatively straightforward. Google also launches Gemini 3.1 Flash Live for audio/voice interaction and Laria 3 Pro for longer music tracks (up to three minutes) with more structural control over song sections.

Across the roundup, the common thread is not just better outputs—it’s faster iteration, tighter integration into workflows, and growing emphasis on efficiency and controllability. AI video is “coming into its own,” while LLMs continue to advance through both model updates and infrastructure-level optimizations that can reshape real-world cost and performance.