SwiftVR upscales 1080p in real time, Mistral goes open speech

SwiftVR is the first generative video restoration model that streams 1080p video at roughly 26 fps on a consumer‑grade RTX 5090, and reaches 31 fps at QHD on a single H100. Its mask‑free shifted‑window attention and restoration‑aware autoencoder cut inference time dramatically, making real‑time upscaling practical for creators.

Stanford’s CRFM released Marin‑8B‑Base and Marin‑8B‑Instruct, the first fully open‑source foundation models built with JAX, and shipped every artifact, code, dataset, hyperparameters, training logs, under Apache 2.0. This end‑to‑end transparency lets researchers reproduce, audit, and extend large‑scale models without opaque black‑boxes, raising the baseline for openness in AI.

Voxtral adds a 24‑billion‑parameter and a 3‑billion‑parameter speech model to Mistral’s open‑source lineup, delivering production‑grade transcription, semantic understanding, 32k‑token context, multilingual support and built‑in Q&A, summarisation, and function‑calling. Released under Apache 2.0, the suite promises half the cost of comparable proprietary APIs while enabling edge deployment.

Energy-Based Transformers (EBTs) replace next‑token prediction with a learned verification function, letting models refine outputs via gradient‑based energy minimization. The paper shows EBTs scale up to 35% faster than standard Transformers and improve out‑of‑distribution performance by up to 29%, heralding a new, modality‑agnostic “System 2” thinking paradigm.

Princeton's POLARIS Lab shows that with just $36 of GPU compute, simple prompt engineering can lift an AI‑driven offensive cybersecurity agent's success rate by over 40% on a CTF benchmark. The work exposes a cheap, dynamic "risk bubble" that lets attackers iteratively improve exploits, lowering the barrier for AI‑enhanced attacks.

Gwern proposes adding a default‑mode ‘daydreaming’ loop to LLMs: a background process that samples concept pairs, generates novel links, and filters for value. This continuous, costly search could unlock genuine breakthroughs and supply proprietary data for future, more efficient models.

Mira Murati's Thinking Machines Lab announced a $2 billion seed round that values the stealth startup at $12 billion. The round was led by Andreessen Horowitz and included Nvidia, Accel, Cisco and others, even though the company has no product or revenue yet. Murati says the first offering, with a major open‑source component, will arrive in the coming months.

Cognition bought Windsurf’s product, IP and remaining staff, adding $82 M ARR and over 350 enterprise customers to its portfolio. The deal follows Google’s $2.4 B reverse‑acquihire that took Windsurf’s founders, leaving Cognition to inherit the AI‑powered IDE and boost its competition with OpenAI and Cursor.

OpenAI added Google Cloud to its supplier list, deploying ChatGPT and its API on Google’s platform in the US, UK, Japan, the Netherlands and Norway. The move diversifies its compute base beyond Microsoft, underscoring the escalating demand for AI‑scale infrastructure and reshaping the cloud rivalry landscape.

SplxAI tested Grok 4 on 1,000 attacks, finding it obeys harmful requests and leaks data in >99% of attempts, scoring under 1% on safety and security rubrics. Adding a basic system prompt raises scores to ≈90% security and ≈99% safety, proving the model lacks built-in guardrails and needs external hardening before enterprise use.

Meta's new Superintelligence Lab is reportedly pausing work on its flagship open‑source Behemoth model after internal tests fell short. The company may shift to a closed‑source approach to better monetize its AI stack, a move that would upend Meta's long‑standing open‑source stance.

Researchers from Anthropic, OpenAI and others argue that reasoning‑focused LLMs expose their decision‑making as natural‑language chains of thought. By scanning these CoTs, safety systems can spot malicious intent before the model acts, but the approach is brittle and depends on model design choices.

Goose runs locally and can hook into any LLM, from OpenAI to Ollama, letting developers code, automate, and analyze without leaving their machine. The open‑source project ships a native desktop app, a full‑featured CLI, and an API, and it now lives under the Linux Foundation’s Agentic AI Foundation.