Baseten Nears $1.5B Raise at a $13B Valuation — The "Inference Gold Rush" Continues
Baseten, an AI inference infrastructure startup, is reportedly close to finalizing a $1.5 billion funding round at a $13 billion valuation. The raise comes just a few months after the company's last major round, underscoring just how hot the AI inference market has become.
Demand for infrastructure that makes running (inferencing) AI models efficient has surged as models grow larger and enterprise adoption expands. The way inference-focused companies like Baseten have skyrocketed in value over a short period highlights, once again, the strategic importance of the infrastructure layer in the AI industry.
Weibo's VibeThinker-3B Heats Up the Benchmark Controversy
VibeThinker-3B, a small language model released by Chinese social media giant Weibo, has ignited a debate over benchmarks in the AI community. What made waves is that a model with only 3 billion parameters posted scores approaching those of major models on some benchmarks.
In response, skepticism about the reliability of benchmarks and the validity of evaluation methods has been mounting. The phenomenon of small models scoring high on specific metrics has long been pointed out, and the "gap between real-world performance and benchmark scores" is once again in the spotlight. The possibility of benchmark contamination (test data leaking into training data) has also been raised.
ABC-130k — One of the Largest Open-Source Robot Teleoperation Datasets Released
A large-scale robot teleoperation dataset called "ABC-130k" has been released on Hugging Face. It consists of manipulation trajectory data collected on a bimanual robot workstation (YAM), totaling 3,555 hours of recordings across 130,822 episodes. 43,090 of the episodes come with subtask annotations, and the annotations are structured so they can be revised independently.
The data is distributed in the MCAP file format and includes video from a top-mounted camera and two wrist cameras. The scarcity of high-quality real-world data has long been a challenge in robot learning, and the release of an open dataset at this scale should give a major boost to research on general-purpose robot manipulation AI. An accompanying paper plus training and deployment code are also provided.
DiffusionGemma 26B Hits Up to 475 t/s on an RTX 4090
Results from running DiffusionGemma 26B on an RTX 4090 have been reported in Reddit's LocalLLaMA community. Using AWQ-INT4 quantization and a custom vLLM Docker environment, the reporter achieved 475 t/s on the first prompt and generation speeds of 290–700 t/s under normal conditions.
Diffusion-based language models employ a generation approach different from conventional autoregressive models, and theoretical speedups are expected. That said, the reporter also notes several shortcomings: output quality can be inferior to a typical 26B model, the context length is limited (tested at 8k), and it is not well suited to batch processing. Reportedly, "Hayabu struggles with the needle-in-a-haystack test," so information extraction from long contexts still appears to be a weak point.
Ajinomoto Group Cuts Accounting Workload by 76% with an AI Agent
Ajinomoto Financial Solutions, which handles the Ajinomoto Group's finance and accounting operations, has moved a system that lets AI agents autonomously perform expense-report approvals into full production. As a result, the company says it has cut the workload for approval tasks by 76%.
While cautious views remain strong on adopting AI for accounting work — where errors are unacceptable — the Ajinomoto Group has become an early pioneer. With the shortage of accounting talent becoming severe and the scope of the work ever expanding, AI-driven automation looks set to mean more than mere cost reduction. It is a sign that practical agent AI adoption at Japanese companies is steadily reaching the front lines.
Federal Regulators Move to Speed Up Grid Connections for AI Data Centers
The U.S. Federal Energy Regulatory Commission (FERC) has signaled a policy to accelerate grid connections for AI data centers. With data center power consumption becoming a bottleneck amid surging AI compute demand, regulators are rushing to respond.
The expansion of AI infrastructure depends not only on compute resources but also on grid capacity, creating a dynamic in which the speed of regulation can make or break industrial competitiveness. Matching renewable energy supply with AI demand is also likely to emerge as a key challenge.
SubQ — A Sub-Quadratic LLM That Handles Millions of Tokens
SubQ, a sub-quadratic language model, has been making waves on Hacker News. Optimized for long-context inference at the multi-million-token scale, its design stays below the O(n²) computational complexity of conventional Transformer architectures, reportedly slashing inference costs for extremely long contexts.
Long-context processing is in high practical demand but carries enormous computational costs, which is why architecture-level innovation is needed. Efforts like SubQ hold the potential to expand LLM use in practical applications such as analyzing large-scale codebases and summarizing lengthy documents.
---