AI model releases and faster inference keep accelerating

The week in brief

• ● OlmoEarth v1.3: State-of-the-art satellite embeddings Allen Institute for AI
• ● AI Is Getting Good at Math Latent Space
• ● How Cerebras Deploys Trillion-Parameter AI Models Cerebras
• ● Uber brings ChatGPT Work to all employees with OpenAI OpenAI
• ● Grok releases first seven minutes of AI-generated Odyssey TechCrunch AI

Model & Product Releases 4 stories

Allen Institute for AI 3 min read

OlmoEarth v1.3: State-of-the-art satellite embeddings

OlmoEarth v1.3 released a state-of-the-art satellite embedding model with precomputed global annual embeddings at 10m, 64- and 128-dimensions, plus a browser-based Embeddings Explorer. It matters to engineers because the ready-to-use embeddings can be reused cheaply for many tasks without fine-tuning, and the Explorer requires no sign-in or advanced technical knowledge.

Cerebras 3 min read

Cerebras CS-4 delivers up to 30x faster inference

Cerebras published a guide to MoE inference arithmetic and deployment bottlenecks, with CS-4 delivering up to 30x faster inference than GPUs. It matters to engineers because single-GPU deployment is memory-bound, multi-GPU setups are communication-bound, and specialized hardware like Cerebras WSE is compute-bound.

xAI 3 min read

Using Grok Bot to detect and respond to incidents

When a product incident occurs, the support queue can be one of the first places it shows up. Customers start reporting the same issue, sometimes before engineering alerts fire or anyone has declared ...

Latent Space 3 min read

[AINews] Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-Class Open Weights narrows Closed-AI lead to 1 month... and Pro Highspeed API with 2-3× speed

We have Fable at home... or so we are being told by paid shills

Research 4 stories

Latent Space 3 min read

AI Is Getting Good at Math

OpenAI reported that an internal model found a counterexample to Erdős's unit distance conjecture and later published proofs and solutions to the Cycle Double Cover problem, ten more open problems, and Navier Stokes, while Anthropic formalized Fermat's Last Theorem and advanced bounds on Riemann's hypothesis. It matters for engineers because the pattern of which mathematical problems these models solve may help predict and direct where to apply them.

Hugging Face 3 min read

Specializing Vocabulary to Cut SLM Inference Cost

Dharma AI's VocabX project previewed a method that specialized the tokenizer for a production essay-grading model built on Qwen3-4B and cut inference cost by 77% without changing the base model. It matters for engineers because the tokenizer can be a major inference cost lever set before training for domain-specialized small language models.

METR 3 min read

Red-Teaming OpenAI's Agent Monitoring

METR and Redwood Research conducted a three-week high-access red-team pilot of OpenAI's internal agent monitoring system from July 21 to August 11, producing a 61-page internal report that estimated monitoring coverage and whether agents could disable or subvert it. It matters for engineers because it tested whether agentic inference could be monitored or subverted in practice.

Allen Institute for AI 3 min read

LitPivot: Developing research ideas through dynamic contextualization within the literature landscape

LitPivot helps researchers refine early research ideas through literature-supported critiques and revisions as their proposals evolve.

Tooling & Infrastructure 4 stories

Cerebras 3 min read

How Cerebras Deploys Trillion-Parameter AI Models

Cerebras explained how it deploys multi-trillion-parameter frontier models across WSE-3T wafers, using the open-source Kimi K3 with about 3 trillion parameters, at 1,000–2,000+ tokens per second. For engineers building large-model inference, the post details how full models are fit into on-chip memory and mapped across multiple chips.

vLLM 3 min read

Serving DeepSeek-V4.1-Flash at Scale

vLLM showed that with P/D disaggregation and Wide-EP on an NVIDIA GB300 NVL72 rack, it served DeepSeek-V4.1-Flash on SemiAnalysis AgentX at 954K tokens/s per GPU and 196 tokens/s per user at P90, 7.9x its single-node result. It matters because vLLM's decode throughput was within 6% of DeepInfer across batch sizes 8–256, while costing $38.96 per GPU-hour at DeepSeek's off-peak prices.

vLLM 3 min read

vLLM Delivers Day-0 Support for MiniMax-M3.1-Flash

vLLM delivered day-0 support for MiniMax-M3.1-Flash on NVIDIA Blackwell, NVIDIA Rubin, and AMD MI350-series GPUs, with NVFP4 MSA kernels, NVFP4 MegaMoE, DSpark speculative decoding, and day-0 RL. It matters because the post details Blackwell MSA kernels using the CUDA 13.4 QMUL4 instruction for up to 1.7x faster indexer QK decode and 1.3x faster sparse attention decode, plus 20% lower kernel latency in prefill and 7% higher prefill throughput via NVFP4 MegaMoE.

Simon Willison 1 min read

Seam Cutter

Tool: Seam Cutter I built this weird and very specific cropping tool - it lets you create a horizontal or vertical cut across an image, then slide the two sides relative to each other. Effectively it lets you crop out a section in the middle of an image.

Industry & Policy 3 stories

OpenAI 1 min read

Uber brings ChatGPT Work to all employees with OpenAI

OpenAI is rolling out ChatGPT Work to Uber's global workforce of over 30,000 employees, expanding from a pilot of 1,000. Uber will use OpenAI's GPT-6.1 Sol as the default model within ChatGPT Work.

TechCrunch AI 1 min read

AI delivers zero financial payoff to most companies, Apollo says

Apollo's chief economist said AI has delivered zero financial payoff to corporate America outside of tech, with flat profit margins and revenue per worker for the S&P 493. The note cited massive investments and widespread hype but no evidence of improved productivity.

xAI 3 min read

xAI launches SpaceX Super Intelligence Partner Network

xAI announced the SpaceX Super Intelligence Partner Network for companies building with Grok, with over 130 companies in early access. The program includes Transformation Partners and Marketplace Partners for enterprise deployment and plugin integrations.

Also Worth Reading 3 stories

TechCrunch AI 2 min read

Grok releases first seven minutes of AI-generated Odyssey

Grok released the first seven minutes of its AI-generated Odyssey, called Episode 1 of its Odyssey series, three months after Elon Musk announced a full-length, historically accurate movie version of The Odyssey. For engineers building with generative AI, the release offers an early public sample of AI-generated episodic footage, though the excerpt notes the result may not satisfy expectations around artistry or human interest.

Ahead of AI (Sebastian Raschka) 3 min read

Building Reasoning Models from the Ground Up

Sebastian Raschka released Building Reasoning Models from the Ground Up, a full video course with over 11 hours of coding videos on evaluation, inference-time scaling, reinforcement learning, and distillation using Python and PyTorch. For engineers building reasoning models, it provides from-scratch implementations of those methods, which Raschka describes as the best way to learn how they really work.

Interconnects 3 min read

Podcast: Kimi K3.1 Pro, the new cyber reality, and Muse musings