SAP freezes travel and hiring over AI's soaring cost as embodied world models keep advancing

SAP's decision to halt most travel and hiring over AI's soaring internal cost, paired with reporting that a planned Amazon data center could become the largest climate polluter in the U.S., frames today's clearest strategic story: AI infrastructure spend is now hitting balance sheets and power grids in ways companies can no longer treat as a rounding error. Research activity kept moving forward regardless, with a cluster of embodied world models, FactorJEPA, MASS, and World-to-Wrist, advancing how systems represent dense multi-agent scenes, shared game state, and fine-grained robot manipulation. A parallel cluster of generative media tooling, spanning multimodal tokenizers, multi-shot video editing, object removal, and 3D scene selection, shows the generation stack maturing modality by modality. OpenAI's acquisition of presentation startup NextSlide sits against Meetily's free, open-source meeting-transcription alternative, a small but real signal that some AI product categories are consolidating under labs while others stay open. For SMB-facing AI vendors, the throughline is that infrastructure cost discipline is becoming a board-level topic even as underlying model capability keeps compounding.

11 papers 5 news 4 sources ← Latest

News

4 items

AI infrastructure costs hit balance sheets and the grid

SAP has stopped most corporate travel and hiring as it grapples with AI's soaring internal cost, while a planned Amazon data center in Texas could become the largest single source of climate pollution in the United States, two concrete signs that AI infrastructure spend is now a board-level constraint rather than a background cost.

News Hacker News

Software Giant SAP Stops Most Travel and Hiring Because of AI's Soaring Cost

SAP has curtailed most corporate travel and hiring, citing the rising internal cost of AI as a driver of the belt-tightening.

Why it matters

A major enterprise software vendor treating AI spend as a reason to freeze hiring and travel is a concrete data point on where AI cost pressure actually lands inside large organizations.
If a company selling AI-adjacent enterprise software is cutting costs elsewhere to fund it, smaller AI vendors should expect procurement scrutiny to tighten similarly.

Source →

News TechCrunch AI

Planned Amazon data center could become the biggest climate polluter in the U.S.

A planned Texas data center includes an on-site power plant that could reportedly become the single largest source of climate pollution in the United States.

Why it matters

The scale of on-site power generation being built to feed a single data center illustrates how far compute buildout has outrun grid capacity.
Environmental and permitting scrutiny of hyperscale AI infrastructure is likely to intensify as projects at this scale become public.

infrastructure policy compute

Source →

Presentation and meeting AI splits between lab consolidation and open source

OpenAI acquired presentation startup NextSlide and folded its team into ChatGPT work, while Meetily offers a free, open-source alternative for meeting transcription and summarization, a small pairing that illustrates two different paths AI productivity tooling is taking at once.

News TechCrunch AI

OpenAI acquires presentation startup NextSlide

OpenAI acquired presentation startup NextSlide, with its team now working on ChatGPT.

products market

Source →

News Wired AI

Meetily Lets You Transcribe and Summarize Meetings Without a Subscription—Here's How

Meetily is a free, open-source tool for transcribing and summarizing meetings without a subscription.

tools open-weights products

Source →

Papers

7 items

World models push into dense, multi-agent, and manipulation settings

Three papers advance world modeling beyond clean, single-agent settings: FactorJEPA factorizes predicted futures for crowded Global South urban scenes, MASS gives multiplayer world models a single authoritative shared state instead of per-view latents, and World-to-Wrist conditions robot manipulation on predicted future wrist-camera views.

Paper Hugging Face

FactorJEPA: Factorizing Monolithic Futures into Layout-Agent-Interaction Channels for Crowded and Chaotic Global South Urban Worlds

FactorJEPA factorizes predicted futures into layout, agent, and interaction channels to model dense, chaotic Global South traffic scenes, releasing a new 1,000-hour, 22-city dataset alongside it.

dataset scale 1,000 hours, 22 citiesranking consistency across backbones rho 0.895-0.978

world-models datasets reinforcement-learning

Source → Arc

Paper Hugging Face

MASS: Multiplayer World Models with Authoritative Shared State

MASS disentangles world dynamics from view rendering with a single learned authoritative shared state, scaling to 1,024 concurrent players over 10,000 recurrent steps in a matched multiplayer benchmark.

concurrent players simulated 1,024

world-models agents

Source → Arc

Paper Hugging Face

World-to-Wrist: Task-Conditioned Future Wrist Modeling for Fine-Grained Robot Manipulation

W2-VLA forecasts future wrist-camera latents conditioned on task and global context to improve fine-grained, contact-sensitive robot manipulation while keeping action generation above 80 Hz.

action-generation rate >80 Hz

robotics vision-language imitation-learning

Source → Arc

Generative media tooling matures modality by modality

A cluster of papers advances practical generative-media infrastructure: KVAE ships a family of audio, image, and video tokenizers matching frontier open-source options, ContextMaster unifies multi-shot video generation and editing under shared context, EffectLearner targets object-induced effects that trip up existing video object removal, and GaussianSelector enables training-free 3D object selection from sparse views.

Paper Hugging Face

KVAE: Family of Tokenizers for Multimodal Generative Models

KVAE releases a family of audio, image, and video tokenizers for text-conditioned generation that match or surpass frontier open-source VAEs from Wan-2.2, HunyuanVideo-1.5, and FLUX.2, with training details published openly.

image compression 8x, 32 channelsaudio sample rate 48 kHz, 50 Hz latent

multimodal audio image-generation video-generation

Source → Arc

Paper Hugging Face

ContextMaster: Interactive Multi-Shot Video Creation via Fixed-Budget Sparse Context Routing

ContextMaster unifies text-to-video generation, reference conditioning, and editing across multiple shots under one model with fixed-budget context routing, reaching 16 FPS on a single GPU.

inference speed 16 FPS, single GPU

video-generation tool-use

Source → Arc

Paper Hugging Face

EffectLearner: World-Aware Object-Effect Reasoning for Real-World Video Object Removal

EffectLearner pairs a VLM-based object-effect reasoner with a video eraser to remove objects along with their induced effects, outperforming baselines on a new EffectWorld benchmark of complex real-world scenes.

video-generation multimodal

Source → Arc

Paper Hugging Face

GaussianSelector: Lightweight Human-Guided Object Selection in 3D Gaussian Splatting with Graph Optimization

GaussianSelector selects complete 3D objects from Gaussian-splat scenes using sparse scribble guidance and graph-cut optimization, matching multi-view SAM methods with far fewer views and less compute.

3d-generation tool-use

Source → Arc

Also today

Paper · Hugging Face Invisible Shortcuts: Why Vision Encoders Know Your Camera — Vision models learn to exploit invisible pixel-level metadata traces from image processing and acquisition, a shortcut that also partly explains why some encoders detect AI-generated images well.
Paper · Hugging Face Interpretable MEG Decoding of Perceived Speech: Cortical Sources and the Stimulus Features That Drive Retrieval — A redesigned MEG-to-speech retrieval model maps its weights to source brain regions and finds silence, sound intensity, vowels, and acoustic onsets drive most of the retrieval signal.
Paper · Hugging Face MameLoshnLM: Yiddish Language Model and Evaluation Benchmark — The first open-source 8B-parameter Yiddish language model, trained on a new high-quality corpus, outperforms general multilingual baselines of similar scale.
Paper · Hugging Face Task-Conditional Flow Matching for Balanced Multilingual Text Embedding Adaptation — TCFM applies flow matching selectively to translation tasks while using task-aligned objectives elsewhere, setting a new state of the art on the Indic Massive Text Embedding Benchmark.
News · Wired AI These AI Barons Are Ready to Give Away Their Fortunes — A new generation of AI-made billionaires is preparing large philanthropic pledges, raising questions about the motives and effects of AI-fortune philanthropy.